INFO 04-11 22:34:15 [__init__.py:216] Automatically detected platform cuda.
2026-04-11 22:34:16,606 | INFO | train_grpo_train | mode=train

Filter:   0%|          | 0/109070 [00:00<?, ? examples/s]
Filter:   1%|          | 1000/109070 [00:00<00:25, 4317.01 examples/s]
Filter:   2%|▏         | 2000/109070 [00:00<00:24, 4400.53 examples/s]
Filter:   3%|▎         | 3000/109070 [00:00<00:23, 4499.08 examples/s]
Filter:   4%|▎         | 4000/109070 [00:00<00:23, 4551.99 examples/s]
Filter:   5%|▍         | 5000/109070 [00:01<00:22, 4550.38 examples/s]
Filter:   6%|▌         | 6000/109070 [00:01<00:22, 4567.18 examples/s]
Filter:   6%|▋         | 7000/109070 [00:01<00:22, 4619.08 examples/s]
Filter:   7%|▋         | 8000/109070 [00:01<00:21, 4626.40 examples/s]
Filter:   8%|▊         | 9000/109070 [00:01<00:21, 4640.74 examples/s]
Filter:   9%|▉         | 10000/109070 [00:02<00:21, 4665.64 examples/s]
Filter:  10%|█         | 11000/109070 [00:02<00:20, 4683.84 examples/s]
Filter:  11%|█         | 12000/109070 [00:02<00:20, 4685.22 examples/s]
Filter:  12%|█▏        | 13000/109070 [00:02<00:20, 4669.21 examples/s]
Filter:  13%|█▎        | 14000/109070 [00:03<00:20, 4644.70 examples/s]
Filter:  14%|█▍        | 15000/109070 [00:03<00:20, 4678.30 examples/s]
Filter:  15%|█▍        | 16000/109070 [00:03<00:19, 4705.64 examples/s]
Filter:  16%|█▌        | 17000/109070 [00:03<00:19, 4680.14 examples/s]
Filter:  17%|█▋        | 18000/109070 [00:03<00:19, 4657.70 examples/s]
Filter:  17%|█▋        | 19000/109070 [00:04<00:19, 4633.46 examples/s]
Filter:  18%|█▊        | 20000/109070 [00:04<00:19, 4631.90 examples/s]
Filter:  19%|█▉        | 21000/109070 [00:04<00:19, 4600.84 examples/s]
Filter:  20%|██        | 22000/109070 [00:04<00:19, 4575.23 examples/s]
Filter:  21%|██        | 23000/109070 [00:04<00:18, 4602.23 examples/s]
Filter:  22%|██▏       | 24000/109070 [00:05<00:18, 4582.62 examples/s]
Filter:  23%|██▎       | 25000/109070 [00:05<00:18, 4599.47 examples/s]
Filter:  24%|██▍       | 26000/109070 [00:05<00:17, 4635.76 examples/s]
Filter:  25%|██▍       | 27000/109070 [00:05<00:18, 4503.11 examples/s]
Filter:  26%|██▌       | 28000/109070 [00:06<00:17, 4563.79 examples/s]
Filter:  27%|██▋       | 29000/109070 [00:06<00:17, 4542.18 examples/s]
Filter:  28%|██▊       | 30000/109070 [00:06<00:18, 4308.56 examples/s]
Filter:  28%|██▊       | 31000/109070 [00:06<00:20, 3754.89 examples/s]
Filter:  29%|██▉       | 32000/109070 [00:07<00:22, 3424.00 examples/s]
Filter:  30%|███       | 33000/109070 [00:07<00:23, 3240.66 examples/s]
Filter:  31%|███       | 34000/109070 [00:07<00:24, 3122.58 examples/s]
Filter:  32%|███▏      | 35000/109070 [00:08<00:24, 3042.72 examples/s]
Filter:  33%|███▎      | 36000/109070 [00:08<00:22, 3263.28 examples/s]
Filter:  34%|███▍      | 37000/109070 [00:08<00:21, 3418.87 examples/s]
Filter:  35%|███▍      | 38000/109070 [00:09<00:21, 3282.14 examples/s]
Filter:  36%|███▌      | 39000/109070 [00:09<00:22, 3142.73 examples/s]
Filter:  37%|███▋      | 40000/109070 [00:09<00:22, 3060.96 examples/s]
Filter:  38%|███▊      | 41000/109070 [00:10<00:22, 3008.71 examples/s]
Filter:  39%|███▊      | 42000/109070 [00:10<00:22, 2969.33 examples/s]
Filter:  39%|███▉      | 43000/109070 [00:10<00:21, 3065.73 examples/s]
Filter:  40%|████      | 44000/109070 [00:11<00:20, 3101.96 examples/s]
Filter:  41%|████▏     | 45000/109070 [00:11<00:21, 3036.52 examples/s]
Filter:  42%|████▏     | 46000/109070 [00:11<00:21, 2988.80 examples/s]
Filter:  43%|████▎     | 47000/109070 [00:12<00:21, 2955.55 examples/s]
Filter:  44%|████▍     | 48000/109070 [00:12<00:20, 2924.22 examples/s]
Filter:  45%|████▍     | 49000/109070 [00:12<00:20, 2919.47 examples/s]
Filter:  46%|████▌     | 50000/109070 [00:13<00:18, 3280.52 examples/s]
Filter:  47%|████▋     | 51000/109070 [00:13<00:17, 3328.05 examples/s]
Filter:  48%|████▊     | 52000/109070 [00:13<00:17, 3178.73 examples/s]
Filter:  49%|████▊     | 53000/109070 [00:14<00:18, 3083.05 examples/s]
Filter:  50%|████▉     | 54000/109070 [00:14<00:18, 3015.34 examples/s]
Filter:  50%|█████     | 55000/109070 [00:14<00:18, 2960.43 examples/s]
Filter:  51%|█████▏    | 56000/109070 [00:15<00:17, 3014.98 examples/s]
Filter:  52%|█████▏    | 57000/109070 [00:15<00:15, 3357.94 examples/s]
Filter:  53%|█████▎    | 58000/109070 [00:15<00:15, 3373.29 examples/s]
Filter:  54%|█████▍    | 59000/109070 [00:15<00:15, 3219.11 examples/s]
Filter:  55%|█████▌    | 60000/109070 [00:16<00:15, 3106.85 examples/s]
Filter:  56%|█████▌    | 61000/109070 [00:16<00:15, 3042.92 examples/s]
Filter:  57%|█████▋    | 62000/109070 [00:16<00:15, 2999.06 examples/s]
Filter:  58%|█████▊    | 63000/109070 [00:17<00:15, 2968.33 examples/s]
Filter:  59%|█████▊    | 64000/109070 [00:17<00:13, 3292.36 examples/s]
Filter:  60%|█████▉    | 65000/109070 [00:17<00:12, 3620.79 examples/s]
Filter:  61%|██████    | 66000/109070 [00:18<00:11, 3860.05 examples/s]
Filter:  61%|██████▏   | 67000/109070 [00:18<00:10, 4011.25 examples/s]
Filter:  62%|██████▏   | 68000/109070 [00:18<00:10, 4071.16 examples/s]
Filter:  63%|██████▎   | 69000/109070 [00:18<00:09, 4121.94 examples/s]
Filter:  64%|██████▍   | 70000/109070 [00:18<00:09, 4227.51 examples/s]
Filter:  65%|██████▌   | 71000/109070 [00:19<00:08, 4367.04 examples/s]
Filter:  66%|██████▌   | 72000/109070 [00:19<00:08, 4429.23 examples/s]
Filter:  67%|██████▋   | 73000/109070 [00:19<00:08, 4483.37 examples/s]
Filter:  68%|██████▊   | 74000/109070 [00:19<00:07, 4591.79 examples/s]
Filter:  69%|██████▉   | 75000/109070 [00:19<00:07, 4554.35 examples/s]
Filter:  70%|██████▉   | 76000/109070 [00:20<00:07, 4404.71 examples/s]
Filter:  71%|███████   | 77000/109070 [00:20<00:07, 4288.60 examples/s]
Filter:  72%|███████▏  | 78000/109070 [00:20<00:07, 4399.14 examples/s]
Filter:  72%|███████▏  | 79000/109070 [00:20<00:06, 4464.79 examples/s]
Filter:  73%|███████▎  | 80000/109070 [00:21<00:06, 4503.77 examples/s]
Filter:  74%|███████▍  | 81000/109070 [00:21<00:06, 4512.31 examples/s]
Filter:  75%|███████▌  | 82000/109070 [00:21<00:05, 4552.51 examples/s]
Filter:  76%|███████▌  | 83000/109070 [00:21<00:05, 4601.04 examples/s]
Filter:  77%|███████▋  | 84000/109070 [00:21<00:05, 4660.76 examples/s]
Filter:  78%|███████▊  | 85000/109070 [00:22<00:05, 4659.09 examples/s]
Filter:  79%|███████▉  | 86000/109070 [00:22<00:04, 4654.82 examples/s]
Filter:  80%|███████▉  | 87000/109070 [00:22<00:04, 4667.98 examples/s]
Filter:  81%|████████  | 88000/109070 [00:22<00:04, 4682.79 examples/s]
Filter:  82%|████████▏ | 89000/109070 [00:23<00:04, 4686.54 examples/s]
Filter:  83%|████████▎ | 90000/109070 [00:23<00:04, 4702.87 examples/s]
Filter:  83%|████████▎ | 91000/109070 [00:23<00:05, 3245.35 examples/s]
Filter:  84%|████████▍ | 92000/109070 [00:24<00:04, 3578.29 examples/s]
Filter:  85%|████████▌ | 93000/109070 [00:24<00:04, 3854.23 examples/s]
Filter:  86%|████████▌ | 94000/109070 [00:24<00:03, 4064.88 examples/s]
Filter:  87%|████████▋ | 95000/109070 [00:24<00:03, 4183.82 examples/s]
Filter:  88%|████████▊ | 96000/109070 [00:24<00:03, 4314.30 examples/s]
Filter:  89%|████████▉ | 97000/109070 [00:25<00:02, 4421.91 examples/s]
Filter:  90%|████████▉ | 98000/109070 [00:25<00:02, 4498.03 examples/s]
Filter:  91%|█████████ | 99000/109070 [00:25<00:02, 4553.69 examples/s]
Filter:  92%|█████████▏| 100000/109070 [00:25<00:01, 4620.93 examples/s]
Filter:  93%|█████████▎| 101000/109070 [00:25<00:01, 4646.10 examples/s]
Filter:  94%|█████████▎| 102000/109070 [00:26<00:01, 4624.67 examples/s]
Filter:  94%|█████████▍| 103000/109070 [00:26<00:01, 3992.77 examples/s]
Filter:  95%|█████████▌| 104000/109070 [00:26<00:01, 3579.24 examples/s]
Filter:  96%|█████████▋| 105000/109070 [00:27<00:01, 3314.52 examples/s]
Filter:  97%|█████████▋| 106000/109070 [00:27<00:00, 3176.85 examples/s]
Filter:  98%|█████████▊| 107000/109070 [00:27<00:00, 3086.87 examples/s]
Filter:  99%|█████████▉| 108000/109070 [00:28<00:00, 3117.29 examples/s]
Filter: 100%|█████████▉| 109000/109070 [00:28<00:00, 3453.12 examples/s]
Filter: 100%|██████████| 109070/109070 [00:28<00:00, 3836.00 examples/s]

Fetching 128 files:   0%|          | 0/128 [00:00<?, ?it/s]
Fetching 128 files: 100%|██████████| 128/128 [00:00<00:00, 3592.89it/s]
`torch_dtype` is deprecated! Use `dtype` instead!

Loading checkpoint shards:   0%|          | 0/3 [00:00<?, ?it/s]
Loading checkpoint shards:  33%|███▎      | 1/3 [00:00<00:01,  1.35it/s]
Loading checkpoint shards:  67%|██████▋   | 2/3 [00:01<00:00,  1.38it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.38it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.38it/s]
The model is already on multiple devices. Skipping the move to device specified in `args`.
INFO 04-11 22:36:28 [utils.py:233] non-default args: {'seed': 0, 'max_model_len': 1536, 'distributed_executor_backend': 'external_launcher', 'gpu_memory_utilization': 0.5, 'max_num_batched_tokens': 4096, 'max_num_seqs': 8, 'disable_log_stats': True, 'model_impl': 'vllm', 'model': 'Navid-AI/Yehia-7B-preview'}
INFO 04-11 22:36:28 [model.py:547] Resolved architecture: LlamaForCausalLM
INFO 04-11 22:36:28 [model.py:1510] Using max model len 1536
INFO 04-11 22:36:28 [parallel.py:380] Using external launcher for distributed inference.
INFO 04-11 22:36:28 [parallel.py:420] Disabling V1 multiprocessing for external launcher.
INFO 04-11 22:36:28 [scheduler.py:205] Chunked prefill is enabled with max_num_batched_tokens=4096.
INFO 04-11 22:36:29 [core.py:77] Initializing a V1 LLM engine (v0.11.0) with config: model='Navid-AI/Yehia-7B-preview', speculative_config=None, tokenizer='Navid-AI/Yehia-7B-preview', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=1536, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Navid-AI/Yehia-7B-preview, enable_prefix_caching=True, chunked_prefill_enabled=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2","vllm.mamba_mixer","vllm.short_conv","vllm.linear_attention","vllm.plamo2_mamba_mixer","vllm.gdn_attention","vllm.sparse_attn_indexer"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"cudagraph_mode":[2,1],"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"use_inductor_graph_partition":false,"pass_config":{},"max_capture_size":16,"local_cache_dir":null}
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
INFO 04-11 22:36:29 [parallel_state.py:1208] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
WARNING 04-11 22:36:30 [topk_topp_sampler.py:66] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
INFO 04-11 22:36:30 [gpu_model_runner.py:2602] Starting to load model Navid-AI/Yehia-7B-preview...
INFO 04-11 22:36:30 [gpu_model_runner.py:2634] Loading model from scratch...
INFO 04-11 22:36:30 [cuda.py:366] Using Flash Attention backend on V1 engine.
INFO 04-11 22:36:30 [weight_utils.py:392] Using model weights format ['*.safetensors']

Loading safetensors checkpoint shards:   0% Completed | 0/3 [00:00<?, ?it/s]

Loading safetensors checkpoint shards:  33% Completed | 1/3 [00:00<00:01,  1.98it/s]

Loading safetensors checkpoint shards:  67% Completed | 2/3 [00:01<00:00,  1.62it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.50it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.56it/s]

INFO 04-11 22:36:33 [default_loader.py:267] Loading weights took 2.14 seconds
INFO 04-11 22:36:33 [gpu_model_runner.py:2653] Model loading took 13.0406 GiB and 2.494626 seconds
INFO 04-11 22:36:39 [backends.py:548] Using cache directory: /root/.cache/vllm/torch_compile_cache/8b012be8b8/rank_0_0/backbone for vLLM's torch.compile
INFO 04-11 22:36:39 [backends.py:559] Dynamo bytecode transform time: 5.36 s
INFO 04-11 22:36:43 [backends.py:164] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.941 s
INFO 04-11 22:36:44 [monitor.py:34] torch.compile takes 5.36 s in total
INFO 04-11 22:36:46 [gpu_worker.py:298] Available KV cache memory: 26.11 GiB
INFO 04-11 22:36:46 [kv_cache_utils.py:1087] GPU KV cache size: 53,456 tokens
INFO 04-11 22:36:46 [kv_cache_utils.py:1091] Maximum concurrency for 1,536 tokens per request: 34.80x

Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):   0%|          | 0/5 [00:00<?, ?it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  40%|████      | 2/5 [00:00<00:00, 14.68it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  80%|████████  | 4/5 [00:00<00:00, 14.92it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 13.58it/s]

Capturing CUDA graphs (decode, FULL):   0%|          | 0/4 [00:00<?, ?it/s]
Capturing CUDA graphs (decode, FULL):  50%|█████     | 2/4 [00:00<00:00, 16.83it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 17.84it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 17.66it/s]
INFO 04-11 22:36:47 [gpu_model_runner.py:3480] Graph capturing finished in 1 secs, took 0.09 GiB
INFO 04-11 22:36:47 [core.py:210] init engine (profile, create kv cache, warmup model) took 14.28 seconds
INFO 04-11 22:36:48 [llm.py:306] Supported_tasks: ('generate',)
2026-04-11 22:36:48,374 | INFO | train_grpo_train | dataset_id=Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1 train_size=24897 eval_size=104
2026-04-11 22:36:48,375 | INFO | train_grpo_train | source_dataset_id=Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits test_size=204 hard_diagnostic_size=3126
2026-04-11 22:36:48,375 | INFO | train_grpo_train | active_rewards=['meter', 'count_adherence', 'arabic_clean', 'repeat_penalty', 'total_composite']
2026-04-11 22:36:48,375 | INFO | train_grpo_train | reward_weights={'meter': 0.0, 'count_adherence': 0.0, 'arabic_clean': 0.0, 'repeat_penalty': 0.0, 'total_composite': 1.0}
2026-04-11 22:36:48,375 | INFO | train_grpo_train | output_repo=Shaer-AI/Shaer-adapters-grpo
2026-04-11 22:36:48,375 | INFO | train_grpo_train | train_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_drop_trio/cap_3000/selected_manifest.csv
2026-04-11 22:36:48,375 | INFO | train_grpo_train | hard_diagnostic_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_drop_trio/hard_diagnostic_cap_256/selected_manifest.csv
2026-04-11 22:36:48,375 | INFO | train_grpo_train | trainable_params=159907840 total_params=7160467456 trainable_ratio=0.022332
2026-04-11 22:36:48,375 | INFO | train_grpo_train | resume_decision={"timestamp_utc": "2026-04-11T22:36:28Z", "resume_mode": "from_path", "requested_spec": "/root/workspace/Shaer/grpo/outputs/train/shaer_grpo_20260411_192107/checkpoint-650", "result": "resume", "reason": "explicit_local_path", "local_resume_path": "/root/workspace/Shaer/grpo/outputs/train/shaer_grpo_20260411_192107/checkpoint-650", "remote_repo": null, "remote_prefix": null, "compatibility": "fingerprint_mismatch", "config_fingerprint": "dddf85cc86110a84f67b563540177d5e59b5c500695e99fe984dfc39dac36c3f"}
2026-04-11 22:36:48,375 | INFO | train_grpo_train | starting trainer.train()
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'pad_token_id': 2}.

  0%|          | 0/3300 [00:00<?, ?it/s]INFO 04-11 22:36:51 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 20%|█▉        | 651/3300 [00:04<00:18, 139.51it/s]2026-04-11 22:36:55,674 | INFO | train_grpo_train | metrics_logged mode=train step=651

                                                   
{'loss': -0.0071, 'grad_norm': 9.100071907043457, 'learning_rate': 8.03030303030303e-06, 'num_tokens': 1409555.0, 'completions/mean_length': 35.125, 'completions/min_length': 33.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9935092329978943, 'rewards/meter/std': 0.000985646271146834, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9935092329978943, 'rewards/total_composite/std': 0.000985646271146834, 'reward': 0.9935092329978943, 'reward_std': 0.0009856420801952481, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04576735198497772, 'sampling/sampling_logp_difference/max': 0.9458228349685669, 'sampling/importance_sampling_ratio/min': 0.3883599042892456, 'sampling/importance_sampling_ratio/mean': 1.011391282081604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17123969458043575, 'clip_ratio/low_mean': 0.028315248200669885, 'clip_ratio/low_min': 0.028315248200669885, 'clip_ratio/high_mean': 0.010521235642954707, 'clip_ratio/high_max': 0.010521235642954707, 'clip_ratio/region_mean': 0.03883648384362459, 'reward_total_mean': 0.9935092329978943, 'reward_meter_mean': 0.9935092329978943, 'reward_meter_std': 0.000985646271146834, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9935092329978943, 'reward_total_composite_std': 0.000985646271146834, 'epoch': 0.03}

 20%|█▉        | 651/3300 [00:04<00:18, 139.51it/s]INFO 04-11 22:36:55 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(
2026-04-11 22:37:05,537 | INFO | train_grpo_train | metrics_logged mode=train step=652

                                                   
{'loss': -0.2894, 'grad_norm': 0.4141400456428528, 'learning_rate': 8.027272727272728e-06, 'num_tokens': 1413749.0, 'completions/mean_length': 374.25, 'completions/min_length': 343.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 354.5714416503906, 'completions/min_terminated_length': 343.0, 'completions/max_terminated_length': 363.0, 'rewards/meter/mean': 0.8732885122299194, 'rewards/meter/std': 0.3528631627559662, 'rewards/count_adherence/mean': 0.824999988079071, 'rewards/count_adherence/std': 0.337003618478775, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.17251461744308472, 'rewards/repeat_penalty/std': 0.33435773849487305, 'rewards/total_composite/mean': 0.04464861750602722, 'rewards/total_composite/std': 0.018085261806845665, 'reward': 0.04464861750602722, 'reward_std': 0.018085261806845665, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004473666660487652, 'sampling/sampling_logp_difference/max': 0.7091238498687744, 'sampling/importance_sampling_ratio/min': 0.4920751452445984, 'sampling/importance_sampling_ratio/mean': 1.0013567209243774, 'sampling/importance_sampling_ratio/max': 1.834702491760254, 'entropy': 0.019337893230840564, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0035446555411908776, 'clip_ratio/high_max': 0.0035446555411908776, 'clip_ratio/region_mean': 0.0035446555411908776, 'reward_total_mean': 0.04464861750602722, 'reward_meter_mean': 0.8732885122299194, 'reward_meter_std': 0.3528631627559662, 'reward_count_adherence_mean': 0.824999988079071, 'reward_count_adherence_std': 0.337003618478775, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.17251461744308472, 'reward_repeat_penalty_std': 0.33435773849487305, 'reward_total_composite_mean': 0.04464861750602722, 'reward_total_composite_std': 0.018085261806845665, 'epoch': 0.03}

 20%|█▉        | 652/3300 [00:14<00:18, 139.51it/s]INFO 04-11 22:37:05 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 652/3300 [00:16<00:18, 139.51it/s]
 20%|█▉        | 653/3300 [00:24<02:10, 20.24it/s] 2026-04-11 22:37:15,021 | INFO | train_grpo_train | metrics_logged mode=train step=653

                                                  
{'loss': 0.1416, 'grad_norm': 1.191830039024353, 'learning_rate': 8.024242424242425e-06, 'num_tokens': 1415617.0, 'completions/mean_length': 133.5, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 79.42857360839844, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9687988758087158, 'rewards/meter/std': 0.06074121594429016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4583333432674408, 'rewards/repeat_penalty/std': 0.24800792336463928, 'rewards/total_composite/mean': 0.43250638246536255, 'rewards/total_composite/std': 0.1944577693939209, 'reward': 0.43250638246536255, 'reward_std': 0.1944577544927597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014217361807823181, 'sampling/sampling_logp_difference/max': 0.9473090171813965, 'sampling/importance_sampling_ratio/min': 0.3877831697463989, 'sampling/importance_sampling_ratio/mean': 1.0012526512145996, 'sampling/importance_sampling_ratio/max': 1.5146371126174927, 'entropy': 0.0790142323821783, 'clip_ratio/low_mean': 0.007873826543800533, 'clip_ratio/low_min': 0.007873826543800533, 'clip_ratio/high_mean': 0.0029761905316263437, 'clip_ratio/high_max': 0.0029761905316263437, 'clip_ratio/region_mean': 0.010850017075426877, 'reward_total_mean': 0.43250638246536255, 'reward_meter_mean': 0.9687988758087158, 'reward_meter_std': 0.06074121594429016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4583333432674408, 'reward_repeat_penalty_std': 0.24800792336463928, 'reward_total_composite_mean': 0.43250638246536255, 'reward_total_composite_std': 0.1944577693939209, 'epoch': 0.03}

 20%|█▉        | 653/3300 [00:24<02:10, 20.24it/s]INFO 04-11 22:37:15 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 654/3300 [00:30<03:02, 14.54it/s]2026-04-11 22:37:21,300 | INFO | train_grpo_train | metrics_logged mode=train step=654

                                                  
{'loss': -0.0232, 'grad_norm': 1.927111268043518, 'learning_rate': 8.021212121212122e-06, 'num_tokens': 1418478.0, 'completions/mean_length': 170.625, 'completions/min_length': 162.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.625, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.9953961968421936, 'rewards/meter/std': 0.0023201238363981247, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.2678571343421936, 'rewards/repeat_penalty/std': 0.17806050181388855, 'rewards/total_composite/mean': 0.2665513753890991, 'rewards/total_composite/std': 0.17725923657417297, 'reward': 0.2665513753890991, 'reward_std': 0.17725922167301178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01804671622812748, 'sampling/sampling_logp_difference/max': 1.4201292991638184, 'sampling/importance_sampling_ratio/min': 0.24168278276920319, 'sampling/importance_sampling_ratio/mean': 1.0009305477142334, 'sampling/importance_sampling_ratio/max': 1.6224781274795532, 'entropy': 0.08471588138490915, 'clip_ratio/low_mean': 0.003801907878369093, 'clip_ratio/low_min': 0.003801907878369093, 'clip_ratio/high_mean': 0.008620842476375401, 'clip_ratio/high_max': 0.008620842476375401, 'clip_ratio/region_mean': 0.012422750354744494, 'reward_total_mean': 0.2665513753890991, 'reward_meter_mean': 0.9953961968421936, 'reward_meter_std': 0.0023201238363981247, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.2678571343421936, 'reward_repeat_penalty_std': 0.17806050181388855, 'reward_total_composite_mean': 0.2665513753890991, 'reward_total_composite_std': 0.17725923657417297, 'epoch': 0.03}

 20%|█▉        | 654/3300 [00:30<03:02, 14.54it/s]INFO 04-11 22:37:21 [block_pool.py:378] Successfully reset prefix cache
2026-04-11 22:37:30,818 | INFO | train_grpo_train | metrics_logged mode=train step=655

                                                  
{'loss': -0.0945, 'grad_norm': 2.2248940467834473, 'learning_rate': 8.018181818181818e-06, 'num_tokens': 1420163.0, 'completions/mean_length': 116.625, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.142860412597656, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7503823637962341, 'rewards/meter/std': 0.44998785853385925, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5, 'rewards/repeat_penalty/std': 0.30860671401023865, 'rewards/total_composite/mean': 0.25364238023757935, 'rewards/total_composite/std': 0.14388985931873322, 'reward': 0.25364238023757935, 'reward_std': 0.1438898742198944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024109482765197754, 'sampling/sampling_logp_difference/max': 1.0410680770874023, 'sampling/importance_sampling_ratio/min': 0.35307735204696655, 'sampling/importance_sampling_ratio/mean': 1.0044450759887695, 'sampling/importance_sampling_ratio/max': 1.6566717624664307, 'entropy': 0.17866009753197432, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.012714299838989973, 'clip_ratio/high_max': 0.012714299838989973, 'clip_ratio/region_mean': 0.01455253513995558, 'reward_total_mean': 0.25364238023757935, 'reward_meter_mean': 0.7503823637962341, 'reward_meter_std': 0.44998785853385925, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5, 'reward_repeat_penalty_std': 0.30860671401023865, 'reward_total_composite_mean': 0.25364238023757935, 'reward_total_composite_std': 0.14388985931873322, 'epoch': 0.03}

 20%|█▉        | 655/3300 [00:39<03:01, 14.54it/s]INFO 04-11 22:37:31 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 655/3300 [00:42<03:01, 14.54it/s]
 20%|█▉        | 656/3300 [00:44<05:43,  7.70it/s]2026-04-11 22:37:35,310 | INFO | train_grpo_train | metrics_logged mode=train step=656

                                                  
{'loss': 0.0094, 'grad_norm': 6.352079391479492, 'learning_rate': 8.015151515151515e-06, 'num_tokens': 1421891.0, 'completions/mean_length': 58.0, 'completions/min_length': 55.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9239930510520935, 'rewards/meter/std': 0.15441425144672394, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.3333333432674408, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.3079977035522461, 'rewards/total_composite/std': 0.051471415907144547, 'reward': 0.3079977035522461, 'reward_std': 0.05147142335772514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022419268265366554, 'sampling/sampling_logp_difference/max': 0.8770105838775635, 'sampling/importance_sampling_ratio/min': 0.41602474451065063, 'sampling/importance_sampling_ratio/mean': 1.0036613941192627, 'sampling/importance_sampling_ratio/max': 1.928532361984253, 'entropy': 0.16807558294385672, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.01278492109850049, 'clip_ratio/high_max': 0.01278492109850049, 'clip_ratio/region_mean': 0.01709526591002941, 'reward_total_mean': 0.3079977035522461, 'reward_meter_mean': 0.9239930510520935, 'reward_meter_std': 0.15441425144672394, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.3333333432674408, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.3079977035522461, 'reward_total_composite_std': 0.051471415907144547, 'epoch': 0.03}

 20%|█▉        | 656/3300 [00:44<05:43,  7.70it/s]INFO 04-11 22:37:35 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 657/3300 [00:50<07:18,  6.02it/s]2026-04-11 22:37:41,214 | INFO | train_grpo_train | metrics_logged mode=train step=657

                                                  
{'loss': 0.0154, 'grad_norm': 2.9878616333007812, 'learning_rate': 8.012121212121214e-06, 'num_tokens': 1424909.0, 'completions/mean_length': 181.25, 'completions/min_length': 175.0, 'completions/max_length': 185.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 181.25, 'completions/min_terminated_length': 175.0, 'completions/max_terminated_length': 185.0, 'rewards/meter/mean': 0.9874005317687988, 'rewards/meter/std': 0.011436098255217075, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.20192307233810425, 'rewards/repeat_penalty/std': 0.2094048410654068, 'rewards/total_composite/mean': 0.1662987768650055, 'rewards/total_composite/std': 0.17312301695346832, 'reward': 0.1662987768650055, 'reward_std': 0.17312301695346832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010927603580057621, 'sampling/sampling_logp_difference/max': 1.7186641693115234, 'sampling/importance_sampling_ratio/min': 0.1793055236339569, 'sampling/importance_sampling_ratio/mean': 1.0016026496887207, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.042974324664101005, 'clip_ratio/low_mean': 0.008925562433432788, 'clip_ratio/low_min': 0.008925562433432788, 'clip_ratio/high_mean': 0.0028169237775728106, 'clip_ratio/high_max': 0.0028169237775728106, 'clip_ratio/region_mean': 0.011742486211005598, 'reward_total_mean': 0.1662987768650055, 'reward_meter_mean': 0.9874005317687988, 'reward_meter_std': 0.011436098255217075, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.20192307233810425, 'reward_repeat_penalty_std': 0.2094048410654068, 'reward_total_composite_mean': 0.1662987768650055, 'reward_total_composite_std': 0.17312301695346832, 'epoch': 0.03}

 20%|█▉        | 657/3300 [00:50<07:18,  6.02it/s]INFO 04-11 22:37:41 [block_pool.py:378] Successfully reset prefix cache
2026-04-11 22:37:51,192 | INFO | train_grpo_train | metrics_logged mode=train step=658

                                                  
{'loss': -0.0762, 'grad_norm': 0.7183297276496887, 'learning_rate': 8.00909090909091e-06, 'num_tokens': 1429785.0, 'completions/mean_length': 428.5, 'completions/min_length': 394.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 416.5714416503906, 'completions/min_terminated_length': 394.0, 'completions/max_terminated_length': 441.0, 'rewards/meter/mean': 0.8735865354537964, 'rewards/meter/std': 0.35298284888267517, 'rewards/count_adherence/mean': 0.6416666507720947, 'rewards/count_adherence/std': 0.26170989871025085, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.17413419485092163, 'rewards/repeat_penalty/std': 0.34882497787475586, 'rewards/total_composite/mean': 0.033445145934820175, 'rewards/total_composite/std': 0.06880706548690796, 'reward': 0.033445145934820175, 'reward_std': 0.06880706548690796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0026275559794157743, 'sampling/sampling_logp_difference/max': 1.0090997219085693, 'sampling/importance_sampling_ratio/min': 0.3645470440387726, 'sampling/importance_sampling_ratio/mean': 1.0009920597076416, 'sampling/importance_sampling_ratio/max': 1.7735869884490967, 'entropy': 0.012874894309788942, 'clip_ratio/low_mean': 0.00030637255986221135, 'clip_ratio/low_min': 0.00030637255986221135, 'clip_ratio/high_mean': 0.0005817760829813778, 'clip_ratio/high_max': 0.0005817760829813778, 'clip_ratio/region_mean': 0.0008881486428435892, 'reward_total_mean': 0.033445145934820175, 'reward_meter_mean': 0.8735865354537964, 'reward_meter_std': 0.35298284888267517, 'reward_count_adherence_mean': 0.6416666507720947, 'reward_count_adherence_std': 0.26170989871025085, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.17413419485092163, 'reward_repeat_penalty_std': 0.34882497787475586, 'reward_total_composite_mean': 0.033445145934820175, 'reward_total_composite_std': 0.06880706548690796, 'epoch': 0.03}

 20%|█▉        | 658/3300 [01:00<07:18,  6.02it/s]INFO 04-11 22:37:51 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 658/3300 [01:02<07:18,  6.02it/s]
 20%|█▉        | 659/3300 [01:09<14:40,  3.00it/s]2026-04-11 22:38:00,609 | INFO | train_grpo_train | metrics_logged mode=train step=659

                                                  
{'loss': -0.1652, 'grad_norm': 1.3609012365341187, 'learning_rate': 8.006060606060607e-06, 'num_tokens': 1431540.0, 'completions/mean_length': 190.375, 'completions/min_length': 81.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 83.16667175292969, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9505882263183594, 'rewards/meter/std': 0.030698692426085472, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.513268232345581, 'rewards/total_composite/std': 0.3378344178199768, 'reward': 0.513268232345581, 'reward_std': 0.3378343880176544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02022649347782135, 'sampling/sampling_logp_difference/max': 1.8950445652008057, 'sampling/importance_sampling_ratio/min': 0.1503116339445114, 'sampling/importance_sampling_ratio/mean': 1.0079931020736694, 'sampling/importance_sampling_ratio/max': 1.7507625818252563, 'entropy': 0.06744233565405011, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.008969587041065097, 'clip_ratio/high_max': 0.008969587041065097, 'clip_ratio/region_mean': 0.008969587041065097, 'reward_total_mean': 0.513268232345581, 'reward_meter_mean': 0.9505882263183594, 'reward_meter_std': 0.030698692426085472, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.513268232345581, 'reward_total_composite_std': 0.3378344178199768, 'epoch': 0.03}

 20%|█▉        | 659/3300 [01:09<14:40,  3.00it/s]INFO 04-11 22:38:00 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 660/3300 [01:14<17:24,  2.53it/s]2026-04-11 22:38:05,980 | INFO | train_grpo_train | metrics_logged mode=train step=660

                                                  
{'loss': 0.0308, 'grad_norm': 2.2579872608184814, 'learning_rate': 8.003030303030304e-06, 'num_tokens': 1433681.0, 'completions/mean_length': 110.625, 'completions/min_length': 104.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.625, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.9927444458007812, 'rewards/meter/std': 0.0012347318697720766, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.23571428656578064, 'rewards/repeat_penalty/std': 0.07284314185380936, 'rewards/total_composite/mean': 0.22219520807266235, 'rewards/total_composite/std': 0.07082200050354004, 'reward': 0.22219520807266235, 'reward_std': 0.07082199305295944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008687403053045273, 'sampling/sampling_logp_difference/max': 1.8446245193481445, 'sampling/importance_sampling_ratio/min': 0.1580846756696701, 'sampling/importance_sampling_ratio/mean': 1.001028299331665, 'sampling/importance_sampling_ratio/max': 1.6747108697891235, 'entropy': 0.020129066659137607, 'clip_ratio/low_mean': 0.006200430449098349, 'clip_ratio/low_min': 0.006200430449098349, 'clip_ratio/high_mean': 0.0036057692486792803, 'clip_ratio/high_max': 0.0036057692486792803, 'clip_ratio/region_mean': 0.009806199697777629, 'reward_total_mean': 0.22219520807266235, 'reward_meter_mean': 0.9927444458007812, 'reward_meter_std': 0.0012347318697720766, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.23571428656578064, 'reward_repeat_penalty_std': 0.07284314185380936, 'reward_total_composite_mean': 0.22219520807266235, 'reward_total_composite_std': 0.07082200050354004, 'epoch': 0.03}

 20%|██        | 660/3300 [01:14<17:24,  2.53it/s]INFO 04-11 22:38:06 [block_pool.py:378] Successfully reset prefix cache
2026-04-11 22:38:15,702 | INFO | train_grpo_train | metrics_logged mode=train step=661

                                                  
{'loss': -0.0515, 'grad_norm': 3.181408405303955, 'learning_rate': 8.000000000000001e-06, 'num_tokens': 1435750.0, 'completions/mean_length': 203.625, 'completions/min_length': 81.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 100.83333587646484, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.43306416273117065, 'rewards/meter/std': 0.44791871309280396, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.6357142925262451, 'rewards/repeat_penalty/std': 0.31916436553001404, 'rewards/total_composite/mean': 0.22007080912590027, 'rewards/total_composite/std': 0.2968771755695343, 'reward': 0.22007080912590027, 'reward_std': 0.2968771457672119, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04375706985592842, 'sampling/sampling_logp_difference/max': 1.372495174407959, 'sampling/importance_sampling_ratio/min': 0.253473699092865, 'sampling/importance_sampling_ratio/mean': 1.0068386793136597, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17050567921251059, 'clip_ratio/low_mean': 0.008919409476220608, 'clip_ratio/low_min': 0.008919409476220608, 'clip_ratio/high_mean': 0.017718179151415825, 'clip_ratio/high_max': 0.017718179151415825, 'clip_ratio/region_mean': 0.026637588627636433, 'reward_total_mean': 0.22007080912590027, 'reward_meter_mean': 0.43306416273117065, 'reward_meter_std': 0.44791871309280396, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.6357142925262451, 'reward_repeat_penalty_std': 0.31916436553001404, 'reward_total_composite_mean': 0.22007080912590027, 'reward_total_composite_std': 0.2968771755695343, 'epoch': 0.03}

 20%|██        | 661/3300 [01:24<17:24,  2.53it/s]INFO 04-11 22:38:15 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 661/3300 [01:26<17:24,  2.53it/s]
 20%|██        | 662/3300 [01:34<31:13,  1.41it/s]2026-04-11 22:38:25,178 | INFO | train_grpo_train | metrics_logged mode=train step=662

                                                  
{'loss': -0.1279, 'grad_norm': 2.6338069438934326, 'learning_rate': 7.996969696969697e-06, 'num_tokens': 1437638.0, 'completions/mean_length': 117.0, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.57143020629883, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9266777634620667, 'rewards/meter/std': 0.1883465051651001, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5833333730697632, 'rewards/repeat_penalty/std': 0.2357022762298584, 'rewards/total_composite/mean': 0.455108642578125, 'rewards/total_composite/std': 0.24599777162075043, 'reward': 0.455108642578125, 'reward_std': 0.24599777162075043, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032812319695949554, 'sampling/sampling_logp_difference/max': 1.3324346542358398, 'sampling/importance_sampling_ratio/min': 0.263834148645401, 'sampling/importance_sampling_ratio/mean': 1.0021228790283203, 'sampling/importance_sampling_ratio/max': 1.474334478378296, 'entropy': 0.17267457023262978, 'clip_ratio/low_mean': 0.004385964944958687, 'clip_ratio/low_min': 0.004385964944958687, 'clip_ratio/high_mean': 0.012073024990968406, 'clip_ratio/high_max': 0.012073024990968406, 'clip_ratio/region_mean': 0.016458989935927093, 'reward_total_mean': 0.455108642578125, 'reward_meter_mean': 0.9266777634620667, 'reward_meter_std': 0.1883465051651001, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5833333730697632, 'reward_repeat_penalty_std': 0.2357022762298584, 'reward_total_composite_mean': 0.455108642578125, 'reward_total_composite_std': 0.24599777162075043, 'epoch': 0.03}

 20%|██        | 662/3300 [01:34<31:13,  1.41it/s]INFO 04-11 22:38:25 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 663/3300 [01:39<35:57,  1.22it/s]2026-04-11 22:38:30,408 | INFO | train_grpo_train | metrics_logged mode=train step=663

                                                  
{'loss': 0.021, 'grad_norm': 4.197593688964844, 'learning_rate': 7.993939393939396e-06, 'num_tokens': 1439950.0, 'completions/mean_length': 108.0, 'completions/min_length': 96.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.0, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9684537053108215, 'rewards/meter/std': 0.027583837509155273, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4226190447807312, 'rewards/repeat_penalty/std': 0.2210753709077835, 'rewards/total_composite/mean': 0.3900730013847351, 'rewards/total_composite/std': 0.19891957938671112, 'reward': 0.3900730013847351, 'reward_std': 0.19891956448554993, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02085985243320465, 'sampling/sampling_logp_difference/max': 1.0295929908752441, 'sampling/importance_sampling_ratio/min': 0.3571523129940033, 'sampling/importance_sampling_ratio/mean': 0.9997091889381409, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11175651382654905, 'clip_ratio/low_mean': 0.00461137923412025, 'clip_ratio/low_min': 0.00461137923412025, 'clip_ratio/high_mean': 0.014703674940392375, 'clip_ratio/high_max': 0.014703674940392375, 'clip_ratio/region_mean': 0.019315054174512625, 'reward_total_mean': 0.3900730013847351, 'reward_meter_mean': 0.9684537053108215, 'reward_meter_std': 0.027583837509155273, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4226190447807312, 'reward_repeat_penalty_std': 0.2210753709077835, 'reward_total_composite_mean': 0.3900730013847351, 'reward_total_composite_std': 0.19891957938671112, 'epoch': 0.03}

 20%|██        | 663/3300 [01:39<35:57,  1.22it/s]INFO 04-11 22:38:30 [block_pool.py:378] Successfully reset prefix cache
2026-04-11 22:38:35,095 | INFO | train_grpo_train | metrics_logged mode=train step=664

                                                  
{'loss': -0.029, 'grad_norm': 7.99793004989624, 'learning_rate': 7.990909090909091e-06, 'num_tokens': 1441763.0, 'completions/mean_length': 70.625, 'completions/min_length': 68.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.625, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9828510880470276, 'rewards/meter/std': 0.025825461372733116, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5833333730697632, 'rewards/repeat_penalty/std': 0.29546841979026794, 'rewards/total_composite/mean': 0.568142294883728, 'rewards/total_composite/std': 0.27455055713653564, 'reward': 0.568142294883728, 'reward_std': 0.27455058693885803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03234035521745682, 'sampling/sampling_logp_difference/max': 1.1230463981628418, 'sampling/importance_sampling_ratio/min': 0.3252873122692108, 'sampling/importance_sampling_ratio/mean': 1.0028775930404663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10323597816750407, 'clip_ratio/low_mean': 0.0055147059028968215, 'clip_ratio/low_min': 0.0055147059028968215, 'clip_ratio/high_mean': 0.023867564275860786, 'clip_ratio/high_max': 0.023867564275860786, 'clip_ratio/region_mean': 0.029382270178757608, 'reward_total_mean': 0.568142294883728, 'reward_meter_mean': 0.9828510880470276, 'reward_meter_std': 0.025825461372733116, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5833333730697632, 'reward_repeat_penalty_std': 0.29546841979026794, 'reward_total_composite_mean': 0.568142294883728, 'reward_total_composite_std': 0.27455055713653564, 'epoch': 0.03}

 20%|██        | 664/3300 [01:44<35:56,  1.22it/s]INFO 04-11 22:38:35 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 665/3300 [01:48<46:38,  1.06s/it]2026-04-11 22:38:39,704 | INFO | train_grpo_train | metrics_logged mode=train step=665

                                                  
{'loss': -0.0097, 'grad_norm': 3.2300848960876465, 'learning_rate': 7.987878787878789e-06, 'num_tokens': 1443616.0, 'completions/mean_length': 80.625, 'completions/min_length': 74.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.625, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9603564143180847, 'rewards/meter/std': 0.023673059418797493, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.7601395845413208, 'rewards/total_composite/std': 0.16597026586532593, 'reward': 0.7601395845413208, 'reward_std': 0.16597026586532593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0124077582731843, 'sampling/sampling_logp_difference/max': 1.0308361053466797, 'sampling/importance_sampling_ratio/min': 0.3567086160182953, 'sampling/importance_sampling_ratio/mean': 1.0014386177062988, 'sampling/importance_sampling_ratio/max': 1.564980387687683, 'entropy': 0.07347342604771256, 'clip_ratio/low_mean': 0.009405238670296967, 'clip_ratio/low_min': 0.009405238670296967, 'clip_ratio/high_mean': 0.0015432098880410194, 'clip_ratio/high_max': 0.0015432098880410194, 'clip_ratio/region_mean': 0.010948448558337986, 'reward_total_mean': 0.7601395845413208, 'reward_meter_mean': 0.9603564143180847, 'reward_meter_std': 0.023673059418797493, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.7601395845413208, 'reward_total_composite_std': 0.16597026586532593, 'epoch': 0.03}

 20%|██        | 665/3300 [01:48<46:38,  1.06s/it]INFO 04-11 22:38:39 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 666/3300 [01:53<54:40,  1.25s/it]2026-04-11 22:38:44,973 | INFO | train_grpo_train | metrics_logged mode=train step=666

                                                  
{'loss': 0.0007, 'grad_norm': 3.160210609436035, 'learning_rate': 7.984848484848486e-06, 'num_tokens': 1445930.0, 'completions/mean_length': 122.25, 'completions/min_length': 120.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.25, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.8794082999229431, 'rewards/meter/std': 0.32782238721847534, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.212053582072258, 'rewards/repeat_penalty/std': 0.2030286192893982, 'rewards/total_composite/mean': 0.12827160954475403, 'rewards/total_composite/std': 0.03277355059981346, 'reward': 0.12827160954475403, 'reward_std': 0.03277355059981346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02072002924978733, 'sampling/sampling_logp_difference/max': 6.419788360595703, 'sampling/importance_sampling_ratio/min': 0.0016290009953081608, 'sampling/importance_sampling_ratio/mean': 0.9980396032333374, 'sampling/importance_sampling_ratio/max': 1.6665557622909546, 'entropy': 0.03877314692363143, 'clip_ratio/low_mean': 0.0010416667209938169, 'clip_ratio/low_min': 0.0010416667209938169, 'clip_ratio/high_mean': 0.006232782383449376, 'clip_ratio/high_max': 0.006232782383449376, 'clip_ratio/region_mean': 0.0072744491044431925, 'reward_total_mean': 0.12827160954475403, 'reward_meter_mean': 0.8794082999229431, 'reward_meter_std': 0.32782238721847534, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.212053582072258, 'reward_repeat_penalty_std': 0.2030286192893982, 'reward_total_composite_mean': 0.12827160954475403, 'reward_total_composite_std': 0.03277355059981346, 'epoch': 0.03}

 20%|██        | 666/3300 [01:53<54:40,  1.25s/it]INFO 04-11 22:38:45 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 667/3300 [01:59<1:04:42,  1.47s/it]2026-04-11 22:38:50,134 | INFO | train_grpo_train | metrics_logged mode=train step=667

                                                    
{'loss': 0.0108, 'grad_norm': 4.775012016296387, 'learning_rate': 7.981818181818183e-06, 'num_tokens': 1448313.0, 'completions/mean_length': 120.875, 'completions/min_length': 116.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.875, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9965612888336182, 'rewards/meter/std': 0.0018690497381612659, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.25, 'rewards/repeat_penalty/std': 0.21257823705673218, 'rewards/total_composite/mean': 0.24911293387413025, 'rewards/total_composite/std': 0.2116239219903946, 'reward': 0.24911293387413025, 'reward_std': 0.2116239219903946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017011119052767754, 'sampling/sampling_logp_difference/max': 1.0500693321228027, 'sampling/importance_sampling_ratio/min': 0.3773258924484253, 'sampling/importance_sampling_ratio/mean': 0.9989076256752014, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07529849279671907, 'clip_ratio/low_mean': 0.0031968391267582774, 'clip_ratio/low_min': 0.0031968391267582774, 'clip_ratio/high_mean': 0.006372183095663786, 'clip_ratio/high_max': 0.006372183095663786, 'clip_ratio/region_mean': 0.009569022222422063, 'reward_total_mean': 0.24911293387413025, 'reward_meter_mean': 0.9965612888336182, 'reward_meter_std': 0.0018690497381612659, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.25, 'reward_repeat_penalty_std': 0.21257823705673218, 'reward_total_composite_mean': 0.24911293387413025, 'reward_total_composite_std': 0.2116239219903946, 'epoch': 0.03}

 20%|██        | 667/3300 [01:59<1:04:42,  1.47s/it]INFO 04-11 22:38:50 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 668/3300 [02:08<1:32:38,  2.11s/it]2026-04-11 22:38:59,869 | INFO | train_grpo_train | metrics_logged mode=train step=668

                                                    
{'loss': -0.1559, 'grad_norm': 2.0396084785461426, 'learning_rate': 7.978787878787879e-06, 'num_tokens': 1451329.0, 'completions/mean_length': 243.0, 'completions/min_length': 195.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 204.57144165039062, 'completions/min_terminated_length': 195.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.9733182787895203, 'rewards/meter/std': 0.02887692302465439, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.27272728085517883, 'rewards/repeat_penalty/std': 0.13744163513183594, 'rewards/total_composite/mean': 0.17570531368255615, 'rewards/total_composite/std': 0.11893084645271301, 'reward': 0.17570531368255615, 'reward_std': 0.11893083900213242, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01857808604836464, 'sampling/sampling_logp_difference/max': 2.0781538486480713, 'sampling/importance_sampling_ratio/min': 0.1251610666513443, 'sampling/importance_sampling_ratio/mean': 0.9996070861816406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05400602100417018, 'clip_ratio/low_mean': 0.007991038146428764, 'clip_ratio/low_min': 0.007991038146428764, 'clip_ratio/high_mean': 0.0030251864809542894, 'clip_ratio/high_max': 0.0030251864809542894, 'clip_ratio/region_mean': 0.011016224627383053, 'reward_total_mean': 0.17570531368255615, 'reward_meter_mean': 0.9733182787895203, 'reward_meter_std': 0.02887692302465439, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.27272728085517883, 'reward_repeat_penalty_std': 0.13744163513183594, 'reward_total_composite_mean': 0.17570531368255615, 'reward_total_composite_std': 0.11893084645271301, 'epoch': 0.03}

 20%|██        | 668/3300 [02:08<1:32:38,  2.11s/it]INFO 04-11 22:39:00 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 669/3300 [02:18<2:04:49,  2.85s/it]2026-04-11 22:39:09,370 | INFO | train_grpo_train | metrics_logged mode=train step=669

                                                    
{'loss': -0.023, 'grad_norm': 2.1599395275115967, 'learning_rate': 7.975757575757576e-06, 'num_tokens': 1453130.0, 'completions/mean_length': 240.125, 'completions/min_length': 74.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 77.0, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.5011062026023865, 'rewards/meter/std': 0.4051038920879364, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4261804223060608, 'rewards/total_composite/std': 0.4667132496833801, 'reward': 0.4261804223060608, 'reward_std': 0.46671321988105774, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.052589599043130875, 'sampling/sampling_logp_difference/max': 1.824247121810913, 'sampling/importance_sampling_ratio/min': 0.16133907437324524, 'sampling/importance_sampling_ratio/mean': 1.0080572366714478, 'sampling/importance_sampling_ratio/max': 1.8760875463485718, 'entropy': 0.16993126086890697, 'clip_ratio/low_mean': 0.009377967799082398, 'clip_ratio/low_min': 0.009377967799082398, 'clip_ratio/high_mean': 0.011646514758467674, 'clip_ratio/high_max': 0.011646514758467674, 'clip_ratio/region_mean': 0.021024482557550073, 'reward_total_mean': 0.4261804223060608, 'reward_meter_mean': 0.5011062026023865, 'reward_meter_std': 0.4051038920879364, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4261804223060608, 'reward_total_composite_std': 0.4667132496833801, 'epoch': 0.03}

 20%|██        | 669/3300 [02:18<2:04:49,  2.85s/it]INFO 04-11 22:39:09 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 670/3300 [02:27<2:41:21,  3.68s/it]2026-04-11 22:39:18,932 | INFO | train_grpo_train | metrics_logged mode=train step=670

                                                    
{'loss': -0.1072, 'grad_norm': 0.766589879989624, 'learning_rate': 7.972727272727273e-06, 'num_tokens': 1454698.0, 'completions/mean_length': 351.0, 'completions/min_length': 73.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 82.66667175292969, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.6174333691596985, 'rewards/meter/std': 0.4055008888244629, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/repeat_penalty/mean': 0.7083333730697632, 'rewards/repeat_penalty/std': 0.11785111576318741, 'rewards/total_composite/mean': 0.249087393283844, 'rewards/total_composite/std': 0.3437734544277191, 'reward': 0.249087393283844, 'reward_std': 0.3437734842300415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03491988405585289, 'sampling/sampling_logp_difference/max': 1.1255141496658325, 'sampling/importance_sampling_ratio/min': 0.4397118091583252, 'sampling/importance_sampling_ratio/mean': 1.004920482635498, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09451978467404842, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.013186233583837748, 'clip_ratio/high_max': 0.013186233583837748, 'clip_ratio/region_mean': 0.013186233583837748, 'reward_total_mean': 0.249087393283844, 'reward_meter_mean': 0.6174333691596985, 'reward_meter_std': 0.4055008888244629, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_repeat_penalty_mean': 0.7083333730697632, 'reward_repeat_penalty_std': 0.11785111576318741, 'reward_total_composite_mean': 0.249087393283844, 'reward_total_composite_std': 0.3437734544277191, 'epoch': 0.03}

 20%|██        | 670/3300 [02:27<2:41:21,  3.68s/it]INFO 04-11 22:39:19 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 671/3300 [02:37<3:21:19,  4.59s/it]2026-04-11 22:39:28,672 | INFO | train_grpo_train | metrics_logged mode=train step=671

                                                    
{'loss': -0.0912, 'grad_norm': 0.7819209098815918, 'learning_rate': 7.96969696969697e-06, 'num_tokens': 1456328.0, 'completions/mean_length': 354.75, 'completions/min_length': 91.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 92.66667175292969, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7872094511985779, 'rewards/meter/std': 0.3365079164505005, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/repeat_penalty/mean': 0.4750000238418579, 'rewards/repeat_penalty/std': 0.2121320366859436, 'rewards/total_composite/mean': 0.19879300892353058, 'rewards/total_composite/std': 0.21251867711544037, 'reward': 0.19879300892353058, 'reward_std': 0.21251867711544037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0315895713865757, 'sampling/sampling_logp_difference/max': 1.163419246673584, 'sampling/importance_sampling_ratio/min': 0.31241610646247864, 'sampling/importance_sampling_ratio/mean': 1.0077892541885376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07205967605113983, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006735671544447541, 'clip_ratio/high_max': 0.006735671544447541, 'clip_ratio/region_mean': 0.006735671544447541, 'reward_total_mean': 0.19879300892353058, 'reward_meter_mean': 0.7872094511985779, 'reward_meter_std': 0.3365079164505005, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_repeat_penalty_mean': 0.4750000238418579, 'reward_repeat_penalty_std': 0.2121320366859436, 'reward_total_composite_mean': 0.19879300892353058, 'reward_total_composite_std': 0.21251867711544037, 'epoch': 0.03}

 20%|██        | 671/3300 [02:37<3:21:19,  4.59s/it]INFO 04-11 22:39:28 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 672/3300 [02:47<4:00:55,  5.50s/it]2026-04-11 22:39:38,378 | INFO | train_grpo_train | metrics_logged mode=train step=672

                                                    
{'loss': -0.1237, 'grad_norm': 0.8036666512489319, 'learning_rate': 7.966666666666668e-06, 'num_tokens': 1458728.0, 'completions/mean_length': 314.0, 'completions/min_length': 160.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 195.1999969482422, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 220.0, 'rewards/meter/mean': 0.7117201685905457, 'rewards/meter/std': 0.3770325481891632, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.1414213627576828, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.5681818127632141, 'rewards/repeat_penalty/std': 0.2368127554655075, 'rewards/total_composite/mean': 0.23602712154388428, 'rewards/total_composite/std': 0.1749935895204544, 'reward': 0.23602712154388428, 'reward_std': 0.1749935895204544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020951304584741592, 'sampling/sampling_logp_difference/max': 1.4866762161254883, 'sampling/importance_sampling_ratio/min': 0.22612299025058746, 'sampling/importance_sampling_ratio/mean': 0.9994598627090454, 'sampling/importance_sampling_ratio/max': 1.857919692993164, 'entropy': 0.06148350611329079, 'clip_ratio/low_mean': 0.0028696630615741014, 'clip_ratio/low_min': 0.0028696630615741014, 'clip_ratio/high_mean': 0.010866477387025952, 'clip_ratio/high_max': 0.010866477387025952, 'clip_ratio/region_mean': 0.013736140448600054, 'reward_total_mean': 0.23602712154388428, 'reward_meter_mean': 0.7117201685905457, 'reward_meter_std': 0.3770325481891632, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.1414213627576828, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.5681818127632141, 'reward_repeat_penalty_std': 0.2368127554655075, 'reward_total_composite_mean': 0.23602712154388428, 'reward_total_composite_std': 0.1749935895204544, 'epoch': 0.03}

 20%|██        | 672/3300 [02:47<4:00:55,  5.50s/it]INFO 04-11 22:39:38 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 673/3300 [02:57<4:40:50,  6.41s/it]2026-04-11 22:39:48,402 | INFO | train_grpo_train | metrics_logged mode=train step=673

                                                    
{'loss': -0.0271, 'grad_norm': 4.668978214263916, 'learning_rate': 7.963636363636365e-06, 'num_tokens': 1460405.0, 'completions/mean_length': 117.625, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.28571701049805, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8629885911941528, 'rewards/meter/std': 0.28332778811454773, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.768899142742157, 'rewards/total_composite/std': 0.3212806284427643, 'reward': 0.768899142742157, 'reward_std': 0.32128065824508667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04510435834527016, 'sampling/sampling_logp_difference/max': 1.6506894826889038, 'sampling/importance_sampling_ratio/min': 0.1919175386428833, 'sampling/importance_sampling_ratio/mean': 1.0077714920043945, 'sampling/importance_sampling_ratio/max': 1.7451122999191284, 'entropy': 0.2579981219023466, 'clip_ratio/low_mean': 0.020833334419876337, 'clip_ratio/low_min': 0.020833334419876337, 'clip_ratio/high_mean': 0.02129602595232427, 'clip_ratio/high_max': 0.02129602595232427, 'clip_ratio/region_mean': 0.04212936037220061, 'reward_total_mean': 0.768899142742157, 'reward_meter_mean': 0.8629885911941528, 'reward_meter_std': 0.28332778811454773, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.768899142742157, 'reward_total_composite_std': 0.3212806284427643, 'epoch': 0.03}

 20%|██        | 673/3300 [02:57<4:40:50,  6.41s/it]INFO 04-11 22:39:48 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 674/3300 [03:07<5:12:31,  7.14s/it]2026-04-11 22:39:58,064 | INFO | train_grpo_train | metrics_logged mode=train step=674

                                                    
{'loss': -0.0956, 'grad_norm': 3.691714286804199, 'learning_rate': 7.96060606060606e-06, 'num_tokens': 1462244.0, 'completions/mean_length': 128.875, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.14286041259766, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.7070336937904358, 'rewards/meter/std': 0.39935943484306335, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.17817415297031403, 'rewards/total_composite/mean': 0.5691444873809814, 'rewards/total_composite/std': 0.343730628490448, 'reward': 0.5691444873809814, 'reward_std': 0.3437305986881256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05939958617091179, 'sampling/sampling_logp_difference/max': 1.2720730304718018, 'sampling/importance_sampling_ratio/min': 0.28025004267692566, 'sampling/importance_sampling_ratio/mean': 1.0029557943344116, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23062355443835258, 'clip_ratio/low_mean': 0.009934040834195912, 'clip_ratio/low_min': 0.009934040834195912, 'clip_ratio/high_mean': 0.04135313397273421, 'clip_ratio/high_max': 0.04135313397273421, 'clip_ratio/region_mean': 0.051287174806930125, 'reward_total_mean': 0.5691444873809814, 'reward_meter_mean': 0.7070336937904358, 'reward_meter_std': 0.39935943484306335, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.17817415297031403, 'reward_total_composite_mean': 0.5691444873809814, 'reward_total_composite_std': 0.343730628490448, 'epoch': 0.03}

 20%|██        | 674/3300 [03:07<5:12:31,  7.14s/it]INFO 04-11 22:39:58 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 675/3300 [03:17<5:42:19,  7.82s/it]2026-04-11 22:40:08,021 | INFO | train_grpo_train | metrics_logged mode=train step=675

                                                    
{'loss': -0.0973, 'grad_norm': 0.7685384154319763, 'learning_rate': 7.957575757575758e-06, 'num_tokens': 1466421.0, 'completions/mean_length': 376.125, 'completions/min_length': 337.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 356.71429443359375, 'completions/min_terminated_length': 337.0, 'completions/max_terminated_length': 367.0, 'rewards/meter/mean': 0.9768602848052979, 'rewards/meter/std': 0.0202656090259552, 'rewards/count_adherence/mean': 0.8624999523162842, 'rewards/count_adherence/std': 0.31139087677001953, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.2991071343421936, 'rewards/repeat_penalty/std': 0.36836329102516174, 'rewards/total_composite/mean': 0.15836204588413239, 'rewards/total_composite/std': 0.21933622658252716, 'reward': 0.15836204588413239, 'reward_std': 0.21933621168136597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005116640590131283, 'sampling/sampling_logp_difference/max': 1.467843770980835, 'sampling/importance_sampling_ratio/min': 0.4014633893966675, 'sampling/importance_sampling_ratio/mean': 1.0009857416152954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0205289286095649, 'clip_ratio/low_mean': 0.0032165506563615054, 'clip_ratio/low_min': 0.0032165506563615054, 'clip_ratio/high_mean': 0.000681198900565505, 'clip_ratio/high_max': 0.000681198900565505, 'clip_ratio/region_mean': 0.0038977495569270104, 'reward_total_mean': 0.15836204588413239, 'reward_meter_mean': 0.9768602848052979, 'reward_meter_std': 0.0202656090259552, 'reward_count_adherence_mean': 0.8624999523162842, 'reward_count_adherence_std': 0.31139087677001953, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.2991071343421936, 'reward_repeat_penalty_std': 0.36836329102516174, 'reward_total_composite_mean': 0.15836204588413239, 'reward_total_composite_std': 0.21933622658252716, 'epoch': 0.03}

 20%|██        | 675/3300 [03:17<5:42:19,  7.82s/it]INFO 04-11 22:40:08 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 676/3300 [03:21<5:05:54,  7.00s/it]2026-04-11 22:40:12,622 | INFO | train_grpo_train | metrics_logged mode=train step=676

                                                    
{'loss': -0.1218, 'grad_norm': 9.489598274230957, 'learning_rate': 7.954545454545455e-06, 'num_tokens': 1468025.0, 'completions/mean_length': 45.5, 'completions/min_length': 41.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9916675090789795, 'rewards/meter/std': 0.0027789692394435406, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9916675090789795, 'rewards/total_composite/std': 0.0027789692394435406, 'reward': 0.9916675090789795, 'reward_std': 0.002778968308120966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040273357182741165, 'sampling/sampling_logp_difference/max': 2.040062665939331, 'sampling/importance_sampling_ratio/min': 0.13002057373523712, 'sampling/importance_sampling_ratio/mean': 0.9964287877082825, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16463111247867346, 'clip_ratio/low_mean': 0.009073751280084252, 'clip_ratio/low_min': 0.009073751280084252, 'clip_ratio/high_mean': 0.01704174862243235, 'clip_ratio/high_max': 0.01704174862243235, 'clip_ratio/region_mean': 0.026115499902516603, 'reward_total_mean': 0.9916675090789795, 'reward_meter_mean': 0.9916675090789795, 'reward_meter_std': 0.0027789692394435406, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9916675090789795, 'reward_total_composite_std': 0.0027789692394435406, 'epoch': 0.03}

 20%|██        | 676/3300 [03:21<5:05:54,  7.00s/it]INFO 04-11 22:40:12 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 677/3300 [03:26<4:37:55,  6.36s/it]2026-04-11 22:40:17,256 | INFO | train_grpo_train | metrics_logged mode=train step=677

                                                    
{'loss': 0.0185, 'grad_norm': 4.125744342803955, 'learning_rate': 7.951515151515152e-06, 'num_tokens': 1469690.0, 'completions/mean_length': 62.125, 'completions/min_length': 58.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9938265085220337, 'rewards/meter/std': 0.0003937912406399846, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6625509858131409, 'rewards/total_composite/std': 0.00026252749375998974, 'reward': 0.6625509858131409, 'reward_std': 0.0002625406195875257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02882186695933342, 'sampling/sampling_logp_difference/max': 5.798013687133789, 'sampling/importance_sampling_ratio/min': 0.0030335744377225637, 'sampling/importance_sampling_ratio/mean': 1.0004545450210571, 'sampling/importance_sampling_ratio/max': 1.5561105012893677, 'entropy': 0.11069364938884974, 'clip_ratio/low_mean': 0.012098872568458319, 'clip_ratio/low_min': 0.012098872568458319, 'clip_ratio/high_mean': 0.0021551724057644606, 'clip_ratio/high_max': 0.0021551724057644606, 'clip_ratio/region_mean': 0.01425404497422278, 'reward_total_mean': 0.6625509858131409, 'reward_meter_mean': 0.9938265085220337, 'reward_meter_std': 0.0003937912406399846, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6625509858131409, 'reward_total_composite_std': 0.00026252749375998974, 'epoch': 0.03}

 21%|██        | 677/3300 [03:26<4:37:55,  6.36s/it]INFO 04-11 22:40:17 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 678/3300 [03:36<5:19:08,  7.30s/it]2026-04-11 22:40:27,014 | INFO | train_grpo_train | metrics_logged mode=train step=678

                                                    
{'loss': -0.15, 'grad_norm': 0.7712501287460327, 'learning_rate': 7.948484848484848e-06, 'num_tokens': 1472788.0, 'completions/mean_length': 336.25, 'completions/min_length': 268.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 277.66668701171875, 'completions/min_terminated_length': 268.0, 'completions/max_terminated_length': 291.0, 'rewards/meter/mean': 0.6280694007873535, 'rewards/meter/std': 0.5047115683555603, 'rewards/count_adherence/mean': 0.8035714626312256, 'rewards/count_adherence/std': 0.15152288973331451, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.41633522510528564, 'rewards/repeat_penalty/std': 0.3482770323753357, 'rewards/total_composite/mean': 0.1397983729839325, 'rewards/total_composite/std': 0.1796947717666626, 'reward': 0.1397983729839325, 'reward_std': 0.1796947568655014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01377029623836279, 'sampling/sampling_logp_difference/max': 5.731827259063721, 'sampling/importance_sampling_ratio/min': 0.003241149475798011, 'sampling/importance_sampling_ratio/mean': 0.9978663921356201, 'sampling/importance_sampling_ratio/max': 1.6066981554031372, 'entropy': 0.03139376197941601, 'clip_ratio/low_mean': 0.004044867469929159, 'clip_ratio/low_min': 0.004044867469929159, 'clip_ratio/high_mean': 0.002251501166028902, 'clip_ratio/high_max': 0.002251501166028902, 'clip_ratio/region_mean': 0.006296368635958061, 'reward_total_mean': 0.1397983729839325, 'reward_meter_mean': 0.6280694007873535, 'reward_meter_std': 0.5047115683555603, 'reward_count_adherence_mean': 0.8035714626312256, 'reward_count_adherence_std': 0.15152288973331451, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.41633522510528564, 'reward_repeat_penalty_std': 0.3482770323753357, 'reward_total_composite_mean': 0.1397983729839325, 'reward_total_composite_std': 0.1796947717666626, 'epoch': 0.03}

 21%|██        | 678/3300 [03:36<5:19:08,  7.30s/it]INFO 04-11 22:40:27 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 679/3300 [03:40<4:42:14,  6.46s/it]2026-04-11 22:40:31,362 | INFO | train_grpo_train | metrics_logged mode=train step=679

                                                    
{'loss': -0.0123, 'grad_norm': 12.517946243286133, 'learning_rate': 7.945454545454547e-06, 'num_tokens': 1474604.0, 'completions/mean_length': 57.0, 'completions/min_length': 53.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.0, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9841916561126709, 'rewards/meter/std': 0.005828971043229103, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9428657293319702, 'rewards/total_composite/std': 0.1139112040400505, 'reward': 0.9428657293319702, 'reward_std': 0.1139112189412117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02025281824171543, 'sampling/sampling_logp_difference/max': 1.1112747192382812, 'sampling/importance_sampling_ratio/min': 0.3291391432285309, 'sampling/importance_sampling_ratio/mean': 0.9998968243598938, 'sampling/importance_sampling_ratio/max': 1.880238652229309, 'entropy': 0.11628487333655357, 'clip_ratio/low_mean': 0.002358490601181984, 'clip_ratio/low_min': 0.002358490601181984, 'clip_ratio/high_mean': 0.01512786210514605, 'clip_ratio/high_max': 0.01512786210514605, 'clip_ratio/region_mean': 0.017486352706328034, 'reward_total_mean': 0.9428657293319702, 'reward_meter_mean': 0.9841916561126709, 'reward_meter_std': 0.005828971043229103, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9428657293319702, 'reward_total_composite_std': 0.1139112040400505, 'epoch': 0.03}

 21%|██        | 679/3300 [03:40<4:42:14,  6.46s/it]INFO 04-11 22:40:31 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 680/3300 [03:45<4:27:05,  6.12s/it]2026-04-11 22:40:36,613 | INFO | train_grpo_train | metrics_logged mode=train step=680

                                                    
{'loss': 0.031, 'grad_norm': 8.33311653137207, 'learning_rate': 7.942424242424242e-06, 'num_tokens': 1476777.0, 'completions/mean_length': 115.625, 'completions/min_length': 101.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.625, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.371232271194458, 'rewards/meter/std': 0.47603076696395874, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.48125001788139343, 'rewards/repeat_penalty/std': 0.13611315190792084, 'rewards/total_composite/mean': 0.22119268774986267, 'rewards/total_composite/std': 0.28687217831611633, 'reward': 0.22119268774986267, 'reward_std': 0.2868722081184387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020960384979844093, 'sampling/sampling_logp_difference/max': 1.9988516569137573, 'sampling/importance_sampling_ratio/min': 0.1354907900094986, 'sampling/importance_sampling_ratio/mean': 1.0001816749572754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08438334474340081, 'clip_ratio/low_mean': 0.014545571291819215, 'clip_ratio/low_min': 0.014545571291819215, 'clip_ratio/high_mean': 0.0033385155256837606, 'clip_ratio/high_max': 0.0033385155256837606, 'clip_ratio/region_mean': 0.017884086817502975, 'reward_total_mean': 0.22119268774986267, 'reward_meter_mean': 0.371232271194458, 'reward_meter_std': 0.47603076696395874, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.48125001788139343, 'reward_repeat_penalty_std': 0.13611315190792084, 'reward_total_composite_mean': 0.22119268774986267, 'reward_total_composite_std': 0.28687217831611633, 'epoch': 0.03}

 21%|██        | 680/3300 [03:45<4:27:05,  6.12s/it]INFO 04-11 22:40:36 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 681/3300 [03:50<4:07:39,  5.67s/it]2026-04-11 22:40:41,212 | INFO | train_grpo_train | metrics_logged mode=train step=681

                                                    
{'loss': 0.0231, 'grad_norm': 3.9222095012664795, 'learning_rate': 7.93939393939394e-06, 'num_tokens': 1478560.0, 'completions/mean_length': 57.875, 'completions/min_length': 57.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9723160266876221, 'rewards/meter/std': 0.027899622917175293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.8934845924377441, 'rewards/total_composite/std': 0.1633896678686142, 'reward': 0.8934845924377441, 'reward_std': 0.163389652967453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023344971239566803, 'sampling/sampling_logp_difference/max': 1.0728743076324463, 'sampling/importance_sampling_ratio/min': 0.342024028301239, 'sampling/importance_sampling_ratio/mean': 1.0034892559051514, 'sampling/importance_sampling_ratio/max': 1.6929512023925781, 'entropy': 0.11835484858602285, 'clip_ratio/low_mean': 0.014689265750348568, 'clip_ratio/low_min': 0.014689265750348568, 'clip_ratio/high_mean': 0.015127861872315407, 'clip_ratio/high_max': 0.015127861872315407, 'clip_ratio/region_mean': 0.029817127622663975, 'reward_total_mean': 0.8934845924377441, 'reward_meter_mean': 0.9723160266876221, 'reward_meter_std': 0.027899622917175293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.8934845924377441, 'reward_total_composite_std': 0.1633896678686142, 'epoch': 0.03}

 21%|██        | 681/3300 [03:50<4:07:39,  5.67s/it]INFO 04-11 22:40:41 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 682/3300 [03:59<4:56:04,  6.79s/it]2026-04-11 22:40:50,664 | INFO | train_grpo_train | metrics_logged mode=train step=682

                                                    
{'loss': -0.0676, 'grad_norm': 1.2039419412612915, 'learning_rate': 7.936363636363637e-06, 'num_tokens': 1480663.0, 'completions/mean_length': 173.875, 'completions/min_length': 109.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 125.5714340209961, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.38672682642936707, 'rewards/meter/std': 0.36102262139320374, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.48750001192092896, 'rewards/repeat_penalty/std': 0.24604006111621857, 'rewards/total_composite/mean': 0.16035160422325134, 'rewards/total_composite/std': 0.24009783565998077, 'reward': 0.16035160422325134, 'reward_std': 0.24009782075881958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013726767152547836, 'sampling/sampling_logp_difference/max': 1.2122516632080078, 'sampling/importance_sampling_ratio/min': 0.29752659797668457, 'sampling/importance_sampling_ratio/mean': 1.0000630617141724, 'sampling/importance_sampling_ratio/max': 1.6329307556152344, 'entropy': 0.049124513287097216, 'clip_ratio/low_mean': 0.005251961061730981, 'clip_ratio/low_min': 0.005251961061730981, 'clip_ratio/high_mean': 0.0029069767333567142, 'clip_ratio/high_max': 0.0029069767333567142, 'clip_ratio/region_mean': 0.008158937795087695, 'reward_total_mean': 0.16035160422325134, 'reward_meter_mean': 0.38672682642936707, 'reward_meter_std': 0.36102262139320374, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.48750001192092896, 'reward_repeat_penalty_std': 0.24604006111621857, 'reward_total_composite_mean': 0.16035160422325134, 'reward_total_composite_std': 0.24009783565998077, 'epoch': 0.03}

 21%|██        | 682/3300 [03:59<4:56:04,  6.79s/it]INFO 04-11 22:40:50 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 683/3300 [04:04<4:25:16,  6.08s/it]2026-04-11 22:40:55,073 | INFO | train_grpo_train | metrics_logged mode=train step=683

                                                    
{'loss': 0.0096, 'grad_norm': 4.130276679992676, 'learning_rate': 7.933333333333334e-06, 'num_tokens': 1482349.0, 'completions/mean_length': 58.75, 'completions/min_length': 57.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9239417314529419, 'rewards/meter/std': 0.05428864806890488, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.8474001884460449, 'rewards/total_composite/std': 0.15459680557250977, 'reward': 0.8474001884460449, 'reward_std': 0.15459680557250977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03419341892004013, 'sampling/sampling_logp_difference/max': 1.8840758800506592, 'sampling/importance_sampling_ratio/min': 0.15196943283081055, 'sampling/importance_sampling_ratio/mean': 0.994130551815033, 'sampling/importance_sampling_ratio/max': 1.6103789806365967, 'entropy': 0.11382754053920507, 'clip_ratio/low_mean': 0.0042372881434857845, 'clip_ratio/low_min': 0.0042372881434857845, 'clip_ratio/high_mean': 0.016894312808290124, 'clip_ratio/high_max': 0.016894312808290124, 'clip_ratio/region_mean': 0.02113160095177591, 'reward_total_mean': 0.8474001884460449, 'reward_meter_mean': 0.9239417314529419, 'reward_meter_std': 0.05428864806890488, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.8474001884460449, 'reward_total_composite_std': 0.15459680557250977, 'epoch': 0.03}

 21%|██        | 683/3300 [04:04<4:25:16,  6.08s/it]INFO 04-11 22:40:55 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 684/3300 [04:08<4:07:19,  5.67s/it]2026-04-11 22:40:59,778 | INFO | train_grpo_train | metrics_logged mode=train step=684

                                                    
{'loss': -0.0064, 'grad_norm': 4.445869445800781, 'learning_rate': 7.930303030303031e-06, 'num_tokens': 1484327.0, 'completions/mean_length': 80.25, 'completions/min_length': 79.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.25, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9331251978874207, 'rewards/meter/std': 0.07388782501220703, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5, 'rewards/repeat_penalty/std': 0.10690450668334961, 'rewards/total_composite/mean': 0.46393126249313354, 'rewards/total_composite/std': 0.09499123692512512, 'reward': 0.46393126249313354, 'reward_std': 0.09499124437570572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01861550658941269, 'sampling/sampling_logp_difference/max': 1.1865754127502441, 'sampling/importance_sampling_ratio/min': 0.3052648901939392, 'sampling/importance_sampling_ratio/mean': 1.0021919012069702, 'sampling/importance_sampling_ratio/max': 1.7777718305587769, 'entropy': 0.0625843945890665, 'clip_ratio/low_mean': 0.007758883642964065, 'clip_ratio/low_min': 0.007758883642964065, 'clip_ratio/high_mean': 0.004629629664123058, 'clip_ratio/high_max': 0.004629629664123058, 'clip_ratio/region_mean': 0.012388513307087123, 'reward_total_mean': 0.46393126249313354, 'reward_meter_mean': 0.9331251978874207, 'reward_meter_std': 0.07388782501220703, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5, 'reward_repeat_penalty_std': 0.10690450668334961, 'reward_total_composite_mean': 0.46393126249313354, 'reward_total_composite_std': 0.09499123692512512, 'epoch': 0.03}

 21%|██        | 684/3300 [04:08<4:07:19,  5.67s/it]INFO 04-11 22:41:00 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 685/3300 [04:14<4:08:53,  5.71s/it]2026-04-11 22:41:05,613 | INFO | train_grpo_train | metrics_logged mode=train step=685

                                                    
{'loss': 0.0202, 'grad_norm': 2.300926685333252, 'learning_rate': 7.927272727272729e-06, 'num_tokens': 1487252.0, 'completions/mean_length': 170.625, 'completions/min_length': 165.0, 'completions/max_length': 178.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.625, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 178.0, 'rewards/meter/mean': 0.8037871718406677, 'rewards/meter/std': 0.32751503586769104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4107142686843872, 'rewards/repeat_penalty/std': 0.050507623702287674, 'rewards/total_composite/mean': 0.34419798851013184, 'rewards/total_composite/std': 0.14113976061344147, 'reward': 0.34419798851013184, 'reward_std': 0.14113974571228027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010442078113555908, 'sampling/sampling_logp_difference/max': 1.643384337425232, 'sampling/importance_sampling_ratio/min': 0.19332465529441833, 'sampling/importance_sampling_ratio/mean': 0.9983497262001038, 'sampling/importance_sampling_ratio/max': 1.4589576721191406, 'entropy': 0.046010758727788925, 'clip_ratio/low_mean': 0.0007022471982054412, 'clip_ratio/low_min': 0.0007022471982054412, 'clip_ratio/high_mean': 0.010335821425542235, 'clip_ratio/high_max': 0.010335821425542235, 'clip_ratio/region_mean': 0.011038068623747677, 'reward_total_mean': 0.34419798851013184, 'reward_meter_mean': 0.8037871718406677, 'reward_meter_std': 0.32751503586769104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4107142686843872, 'reward_repeat_penalty_std': 0.050507623702287674, 'reward_total_composite_mean': 0.34419798851013184, 'reward_total_composite_std': 0.14113976061344147, 'epoch': 0.03}

 21%|██        | 685/3300 [04:14<4:08:53,  5.71s/it]INFO 04-11 22:41:05 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 686/3300 [04:24<4:59:01,  6.86s/it]2026-04-11 22:41:15,153 | INFO | train_grpo_train | metrics_logged mode=train step=686

                                                    
{'loss': -0.0998, 'grad_norm': 1.047049641609192, 'learning_rate': 7.924242424242426e-06, 'num_tokens': 1490154.0, 'completions/mean_length': 251.75, 'completions/min_length': 212.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 214.57144165039062, 'completions/min_terminated_length': 212.0, 'completions/max_terminated_length': 222.0, 'rewards/meter/mean': 0.9798320531845093, 'rewards/meter/std': 0.047762468457221985, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.2828427255153656, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.32499998807907104, 'rewards/repeat_penalty/std': 0.27645719051361084, 'rewards/total_composite/mean': 0.2209167182445526, 'rewards/total_composite/std': 0.04932519793510437, 'reward': 0.2209167182445526, 'reward_std': 0.04932519420981407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011242986656725407, 'sampling/sampling_logp_difference/max': 2.929636001586914, 'sampling/importance_sampling_ratio/min': 0.053416479378938675, 'sampling/importance_sampling_ratio/mean': 0.9986944198608398, 'sampling/importance_sampling_ratio/max': 1.7827941179275513, 'entropy': 0.021744283847510815, 'clip_ratio/low_mean': 0.003458057180978358, 'clip_ratio/low_min': 0.003458057180978358, 'clip_ratio/high_mean': 0.00234195904340595, 'clip_ratio/high_max': 0.00234195904340595, 'clip_ratio/region_mean': 0.005800016224384308, 'reward_total_mean': 0.2209167182445526, 'reward_meter_mean': 0.9798320531845093, 'reward_meter_std': 0.047762468457221985, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.2828427255153656, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.32499998807907104, 'reward_repeat_penalty_std': 0.27645719051361084, 'reward_total_composite_mean': 0.2209167182445526, 'reward_total_composite_std': 0.04932519793510437, 'epoch': 0.03}

 21%|██        | 686/3300 [04:24<4:59:01,  6.86s/it]INFO 04-11 22:41:15 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 687/3300 [04:28<4:28:35,  6.17s/it]2026-04-11 22:41:19,695 | INFO | train_grpo_train | metrics_logged mode=train step=687

                                                    
{'loss': 0.0292, 'grad_norm': 10.655113220214844, 'learning_rate': 7.921212121212122e-06, 'num_tokens': 1491874.0, 'completions/mean_length': 71.0, 'completions/min_length': 67.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8845878839492798, 'rewards/meter/std': 0.2909509837627411, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.8430095314979553, 'rewards/total_composite/std': 0.2961682975292206, 'reward': 0.8430095314979553, 'reward_std': 0.2961682677268982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05931292846798897, 'sampling/sampling_logp_difference/max': 2.937185287475586, 'sampling/importance_sampling_ratio/min': 0.05301474407315254, 'sampling/importance_sampling_ratio/mean': 1.0036027431488037, 'sampling/importance_sampling_ratio/max': 1.9307304620742798, 'entropy': 0.1976525131613016, 'clip_ratio/low_mean': 0.011955027701333165, 'clip_ratio/low_min': 0.011955027701333165, 'clip_ratio/high_mean': 0.020970338257029653, 'clip_ratio/high_max': 0.020970338257029653, 'clip_ratio/region_mean': 0.03292536595836282, 'reward_total_mean': 0.8430095314979553, 'reward_meter_mean': 0.8845878839492798, 'reward_meter_std': 0.2909509837627411, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.8430095314979553, 'reward_total_composite_std': 0.2961682975292206, 'epoch': 0.03}

 21%|██        | 687/3300 [04:28<4:28:35,  6.17s/it]INFO 04-11 22:41:19 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 688/3300 [04:33<4:08:11,  5.70s/it]2026-04-11 22:41:24,297 | INFO | train_grpo_train | metrics_logged mode=train step=688

                                                    
{'loss': -0.0086, 'grad_norm': 1.4717203378677368, 'learning_rate': 7.918181818181819e-06, 'num_tokens': 1493736.0, 'completions/mean_length': 67.75, 'completions/min_length': 66.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9984875321388245, 'rewards/meter/std': 0.0001689638738753274, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6656583547592163, 'rewards/total_composite/std': 0.00011263116175541654, 'reward': 0.6656583547592163, 'reward_std': 0.00011264239583397284, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018908310681581497, 'sampling/sampling_logp_difference/max': 2.604221820831299, 'sampling/importance_sampling_ratio/min': 0.07396066933870316, 'sampling/importance_sampling_ratio/mean': 0.9985640048980713, 'sampling/importance_sampling_ratio/max': 1.6862038373947144, 'entropy': 0.06369170360267162, 'clip_ratio/low_mean': 0.011363636702299118, 'clip_ratio/low_min': 0.011363636702299118, 'clip_ratio/high_mean': 0.010716472752392292, 'clip_ratio/high_max': 0.010716472752392292, 'clip_ratio/region_mean': 0.02208010945469141, 'reward_total_mean': 0.6656583547592163, 'reward_meter_mean': 0.9984875321388245, 'reward_meter_std': 0.0001689638738753274, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6656583547592163, 'reward_total_composite_std': 0.00011263116175541654, 'epoch': 0.03}

 21%|██        | 688/3300 [04:33<4:08:11,  5.70s/it]INFO 04-11 22:41:24 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 689/3300 [04:39<4:14:13,  5.84s/it]2026-04-11 22:41:30,466 | INFO | train_grpo_train | metrics_logged mode=train step=689

                                                    
{'loss': 0.0309, 'grad_norm': 2.0693576335906982, 'learning_rate': 7.915151515151516e-06, 'num_tokens': 1497021.0, 'completions/mean_length': 202.625, 'completions/min_length': 191.0, 'completions/max_length': 212.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 202.625, 'completions/min_terminated_length': 191.0, 'completions/max_terminated_length': 212.0, 'rewards/meter/mean': 0.995306134223938, 'rewards/meter/std': 0.0053658634424209595, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4194444417953491, 'rewards/repeat_penalty/std': 0.13975918292999268, 'rewards/total_composite/mean': 0.4050329029560089, 'rewards/total_composite/std': 0.1355670541524887, 'reward': 0.4050329029560089, 'reward_std': 0.1355670541524887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014271133579313755, 'sampling/sampling_logp_difference/max': 2.4142937660217285, 'sampling/importance_sampling_ratio/min': 0.08943047374486923, 'sampling/importance_sampling_ratio/mean': 0.9986175298690796, 'sampling/importance_sampling_ratio/max': 1.674434781074524, 'entropy': 0.03671248443424702, 'clip_ratio/low_mean': 0.007414351915940642, 'clip_ratio/low_min': 0.007414351915940642, 'clip_ratio/high_mean': 0.0019430051324889064, 'clip_ratio/high_max': 0.0019430051324889064, 'clip_ratio/region_mean': 0.009357357048429549, 'reward_total_mean': 0.4050329029560089, 'reward_meter_mean': 0.995306134223938, 'reward_meter_std': 0.0053658634424209595, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4194444417953491, 'reward_repeat_penalty_std': 0.13975918292999268, 'reward_total_composite_mean': 0.4050329029560089, 'reward_total_composite_std': 0.1355670541524887, 'epoch': 0.03}

 21%|██        | 689/3300 [04:39<4:14:13,  5.84s/it]INFO 04-11 22:41:30 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 690/3300 [04:48<4:49:27,  6.65s/it]2026-04-11 22:41:39,021 | INFO | train_grpo_train | metrics_logged mode=train step=690

                                                    
{'loss': 0.0024, 'grad_norm': 1.1219055652618408, 'learning_rate': 7.912121212121213e-06, 'num_tokens': 1502074.0, 'completions/mean_length': 401.625, 'completions/min_length': 400.0, 'completions/max_length': 402.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 401.625, 'completions/min_terminated_length': 400.0, 'completions/max_terminated_length': 402.0, 'rewards/meter/mean': 0.9981971979141235, 'rewards/meter/std': 0.000627980858553201, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.2503289580345154, 'rewards/repeat_penalty/std': 0.23873279988765717, 'rewards/total_composite/mean': 0.1784874051809311, 'rewards/total_composite/std': 0.17026527225971222, 'reward': 0.1784874051809311, 'reward_std': 0.17026525735855103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005141077097505331, 'sampling/sampling_logp_difference/max': 1.3938775062561035, 'sampling/importance_sampling_ratio/min': 0.24811138212680817, 'sampling/importance_sampling_ratio/mean': 1.0005240440368652, 'sampling/importance_sampling_ratio/max': 1.6704438924789429, 'entropy': 0.027160495053976774, 'clip_ratio/low_mean': 0.0021797263179905713, 'clip_ratio/low_min': 0.0021797263179905713, 'clip_ratio/high_mean': 0.0015578281017951667, 'clip_ratio/high_max': 0.0015578281017951667, 'clip_ratio/region_mean': 0.003737554419785738, 'reward_total_mean': 0.1784874051809311, 'reward_meter_mean': 0.9981971979141235, 'reward_meter_std': 0.000627980858553201, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.2503289580345154, 'reward_repeat_penalty_std': 0.23873279988765717, 'reward_total_composite_mean': 0.1784874051809311, 'reward_total_composite_std': 0.17026527225971222, 'epoch': 0.03}

 21%|██        | 690/3300 [04:48<4:49:27,  6.65s/it]INFO 04-11 22:41:39 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 691/3300 [04:52<4:18:22,  5.94s/it]2026-04-11 22:41:43,353 | INFO | train_grpo_train | metrics_logged mode=train step=691

                                                    
{'loss': -0.1031, 'grad_norm': 11.69098949432373, 'learning_rate': 7.909090909090909e-06, 'num_tokens': 1503707.0, 'completions/mean_length': 38.125, 'completions/min_length': 19.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.4370739161968231, 'rewards/meter/std': 0.2700711488723755, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4370739161968231, 'rewards/total_composite/std': 0.2700711488723755, 'reward': 0.4370739161968231, 'reward_std': 0.2700711488723755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0835486352443695, 'sampling/sampling_logp_difference/max': 1.102844476699829, 'sampling/importance_sampling_ratio/min': 0.3319256007671356, 'sampling/importance_sampling_ratio/mean': 1.0089696645736694, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5944820679724216, 'clip_ratio/low_mean': 0.019354344811290503, 'clip_ratio/low_min': 0.019354344811290503, 'clip_ratio/high_mean': 0.04530784301459789, 'clip_ratio/high_max': 0.04530784301459789, 'clip_ratio/region_mean': 0.0646621878258884, 'reward_total_mean': 0.4370739161968231, 'reward_meter_mean': 0.4370739161968231, 'reward_meter_std': 0.2700711488723755, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4370739161968231, 'reward_total_composite_std': 0.2700711488723755, 'epoch': 0.03}

 21%|██        | 691/3300 [04:52<4:18:22,  5.94s/it]INFO 04-11 22:41:43 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 692/3300 [04:56<4:02:01,  5.57s/it]2026-04-11 22:41:47,992 | INFO | train_grpo_train | metrics_logged mode=train step=692

                                                    
{'loss': 0.0153, 'grad_norm': 5.031876564025879, 'learning_rate': 7.906060606060608e-06, 'num_tokens': 1505712.0, 'completions/mean_length': 86.625, 'completions/min_length': 84.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.625, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9942313432693481, 'rewards/meter/std': 0.0012215422466397285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9942313432693481, 'rewards/total_composite/std': 0.0012215422466397285, 'reward': 0.9942313432693481, 'reward_std': 0.00122154806740582, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005889675114303827, 'sampling/sampling_logp_difference/max': 0.44769424200057983, 'sampling/importance_sampling_ratio/min': 0.7906866669654846, 'sampling/importance_sampling_ratio/mean': 1.003043293952942, 'sampling/importance_sampling_ratio/max': 1.5647001266479492, 'entropy': 0.04205932654440403, 'clip_ratio/low_mean': 0.0014367816038429737, 'clip_ratio/low_min': 0.0014367816038429737, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0014367816038429737, 'reward_total_mean': 0.9942313432693481, 'reward_meter_mean': 0.9942313432693481, 'reward_meter_std': 0.0012215422466397285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9942313432693481, 'reward_total_composite_std': 0.0012215422466397285, 'epoch': 0.03}

 21%|██        | 692/3300 [04:56<4:02:01,  5.57s/it]INFO 04-11 22:41:48 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 693/3300 [05:01<3:45:11,  5.18s/it]2026-04-11 22:41:52,274 | INFO | train_grpo_train | metrics_logged mode=train step=693

                                                    
{'loss': 0.0041, 'grad_norm': 10.364151954650879, 'learning_rate': 7.903030303030303e-06, 'num_tokens': 1507385.0, 'completions/mean_length': 54.125, 'completions/min_length': 52.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9447178244590759, 'rewards/meter/std': 0.019572317600250244, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9447178244590759, 'rewards/total_composite/std': 0.019572317600250244, 'reward': 0.9447178244590759, 'reward_std': 0.01957232505083084, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024356268346309662, 'sampling/sampling_logp_difference/max': 1.9013309478759766, 'sampling/importance_sampling_ratio/min': 0.14936968684196472, 'sampling/importance_sampling_ratio/mean': 1.003846287727356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0895584006793797, 'clip_ratio/low_mean': 0.009437322150915861, 'clip_ratio/low_min': 0.009437322150915861, 'clip_ratio/high_mean': 0.01416083937510848, 'clip_ratio/high_max': 0.01416083937510848, 'clip_ratio/region_mean': 0.02359816152602434, 'reward_total_mean': 0.9447178244590759, 'reward_meter_mean': 0.9447178244590759, 'reward_meter_std': 0.019572317600250244, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9447178244590759, 'reward_total_composite_std': 0.019572317600250244, 'epoch': 0.03}

 21%|██        | 693/3300 [05:01<3:45:11,  5.18s/it]INFO 04-11 22:41:52 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 694/3300 [05:05<3:37:22,  5.00s/it]2026-04-11 22:41:56,868 | INFO | train_grpo_train | metrics_logged mode=train step=694

                                                    
{'loss': 0.0035, 'grad_norm': 6.670380115509033, 'learning_rate': 7.9e-06, 'num_tokens': 1509273.0, 'completions/mean_length': 74.0, 'completions/min_length': 70.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.610145092010498, 'rewards/meter/std': 0.23973627388477325, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.5444035530090332, 'rewards/total_composite/std': 0.24921227991580963, 'reward': 0.5444035530090332, 'reward_std': 0.24921227991580963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030294643715023994, 'sampling/sampling_logp_difference/max': 1.8091429471969604, 'sampling/importance_sampling_ratio/min': 0.16379445791244507, 'sampling/importance_sampling_ratio/mean': 1.0047415494918823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17548873648047447, 'clip_ratio/low_mean': 0.01502489356789738, 'clip_ratio/low_min': 0.01502489356789738, 'clip_ratio/high_mean': 0.02390445303171873, 'clip_ratio/high_max': 0.02390445303171873, 'clip_ratio/region_mean': 0.03892934659961611, 'reward_total_mean': 0.5444035530090332, 'reward_meter_mean': 0.610145092010498, 'reward_meter_std': 0.23973627388477325, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.5444035530090332, 'reward_total_composite_std': 0.24921227991580963, 'epoch': 0.03}

 21%|██        | 694/3300 [05:05<3:37:22,  5.00s/it]INFO 04-11 22:41:57 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 695/3300 [05:10<3:35:18,  4.96s/it]2026-04-11 22:42:01,718 | INFO | train_grpo_train | metrics_logged mode=train step=695

                                                    
{'loss': -0.0124, 'grad_norm': 3.7013967037200928, 'learning_rate': 7.896969696969698e-06, 'num_tokens': 1511199.0, 'completions/mean_length': 80.75, 'completions/min_length': 78.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.75, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9964468479156494, 'rewards/meter/std': 0.0019429969834163785, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.788793683052063, 'rewards/total_composite/std': 0.17156179249286652, 'reward': 0.788793683052063, 'reward_std': 0.17156179249286652, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012971381656825542, 'sampling/sampling_logp_difference/max': 1.7286226749420166, 'sampling/importance_sampling_ratio/min': 0.1775287538766861, 'sampling/importance_sampling_ratio/mean': 1.0024663209915161, 'sampling/importance_sampling_ratio/max': 1.8497363328933716, 'entropy': 0.06276643788442016, 'clip_ratio/low_mean': 0.0062915480230003595, 'clip_ratio/low_min': 0.0062915480230003595, 'clip_ratio/high_mean': 0.004360056365840137, 'clip_ratio/high_max': 0.004360056365840137, 'clip_ratio/region_mean': 0.010651604388840497, 'reward_total_mean': 0.788793683052063, 'reward_meter_mean': 0.9964468479156494, 'reward_meter_std': 0.0019429969834163785, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.788793683052063, 'reward_total_composite_std': 0.17156179249286652, 'epoch': 0.03}

 21%|██        | 695/3300 [05:10<3:35:18,  4.96s/it]INFO 04-11 22:42:01 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 696/3300 [05:15<3:38:42,  5.04s/it]2026-04-11 22:42:06,942 | INFO | train_grpo_train | metrics_logged mode=train step=696

                                                    
{'loss': -0.0063, 'grad_norm': 3.119380235671997, 'learning_rate': 7.893939393939395e-06, 'num_tokens': 1513573.0, 'completions/mean_length': 125.75, 'completions/min_length': 112.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.75, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9228010177612305, 'rewards/meter/std': 0.20124337077140808, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7000000476837158, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.6532999277114868, 'rewards/total_composite/std': 0.18960076570510864, 'reward': 0.6532999277114868, 'reward_std': 0.18960076570510864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016659488901495934, 'sampling/sampling_logp_difference/max': 1.3884687423706055, 'sampling/importance_sampling_ratio/min': 0.2494570016860962, 'sampling/importance_sampling_ratio/mean': 0.9984103441238403, 'sampling/importance_sampling_ratio/max': 1.9363112449645996, 'entropy': 0.07676348416134715, 'clip_ratio/low_mean': 0.0020850637229159474, 'clip_ratio/low_min': 0.0020850637229159474, 'clip_ratio/high_mean': 0.00876707280986011, 'clip_ratio/high_max': 0.00876707280986011, 'clip_ratio/region_mean': 0.010852136532776058, 'reward_total_mean': 0.6532999277114868, 'reward_meter_mean': 0.9228010177612305, 'reward_meter_std': 0.20124337077140808, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7000000476837158, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.6532999277114868, 'reward_total_composite_std': 0.18960076570510864, 'epoch': 0.03}

 21%|██        | 696/3300 [05:15<3:38:42,  5.04s/it]INFO 04-11 22:42:07 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 697/3300 [05:20<3:33:09,  4.91s/it]2026-04-11 22:42:11,562 | INFO | train_grpo_train | metrics_logged mode=train step=697

                                                    
{'loss': 0.0606, 'grad_norm': 4.277256011962891, 'learning_rate': 7.89090909090909e-06, 'num_tokens': 1515360.0, 'completions/mean_length': 59.375, 'completions/min_length': 55.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.375, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9654116630554199, 'rewards/meter/std': 0.034304678440093994, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.8894380331039429, 'rewards/total_composite/std': 0.17405925691127777, 'reward': 0.8894380331039429, 'reward_std': 0.17405925691127777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020010532811284065, 'sampling/sampling_logp_difference/max': 1.0107874870300293, 'sampling/importance_sampling_ratio/min': 0.3639322817325592, 'sampling/importance_sampling_ratio/mean': 0.99860018491745, 'sampling/importance_sampling_ratio/max': 1.4003076553344727, 'entropy': 0.09744885191321373, 'clip_ratio/low_mean': 0.003846153849735856, 'clip_ratio/low_min': 0.003846153849735856, 'clip_ratio/high_mean': 0.01933896285481751, 'clip_ratio/high_max': 0.01933896285481751, 'clip_ratio/region_mean': 0.023185116704553366, 'reward_total_mean': 0.8894380331039429, 'reward_meter_mean': 0.9654116630554199, 'reward_meter_std': 0.034304678440093994, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.8894380331039429, 'reward_total_composite_std': 0.17405925691127777, 'epoch': 0.03}

 21%|██        | 697/3300 [05:20<3:33:09,  4.91s/it]INFO 04-11 22:42:11 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 698/3300 [05:30<4:36:17,  6.37s/it]2026-04-11 22:42:21,335 | INFO | train_grpo_train | metrics_logged mode=train step=698

                                                    
{'loss': -0.0873, 'grad_norm': 2.2967092990875244, 'learning_rate': 7.88787878787879e-06, 'num_tokens': 1519450.0, 'completions/mean_length': 382.25, 'completions/min_length': 321.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 363.71429443359375, 'completions/min_terminated_length': 321.0, 'completions/max_terminated_length': 412.0, 'rewards/meter/mean': 0.6105729937553406, 'rewards/meter/std': 0.4790833294391632, 'rewards/count_adherence/mean': 0.8854166865348816, 'rewards/count_adherence/std': 0.1254950612783432, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.3964124917984009, 'rewards/repeat_penalty/std': 0.25513792037963867, 'rewards/total_composite/mean': 0.18387337028980255, 'rewards/total_composite/std': 0.21844197809696198, 'reward': 0.18387337028980255, 'reward_std': 0.21844197809696198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0396236851811409, 'sampling/sampling_logp_difference/max': 6.693481922149658, 'sampling/importance_sampling_ratio/min': 0.001238961354829371, 'sampling/importance_sampling_ratio/mean': 0.9987672567367554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18515709601342678, 'clip_ratio/low_mean': 0.007161757908761501, 'clip_ratio/low_min': 0.007161757908761501, 'clip_ratio/high_mean': 0.015907755587249994, 'clip_ratio/high_max': 0.015907755587249994, 'clip_ratio/region_mean': 0.023069513496011496, 'reward_total_mean': 0.18387337028980255, 'reward_meter_mean': 0.6105729937553406, 'reward_meter_std': 0.4790833294391632, 'reward_count_adherence_mean': 0.8854166865348816, 'reward_count_adherence_std': 0.1254950612783432, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.3964124917984009, 'reward_repeat_penalty_std': 0.25513792037963867, 'reward_total_composite_mean': 0.18387337028980255, 'reward_total_composite_std': 0.21844197809696198, 'epoch': 0.03}

 21%|██        | 698/3300 [05:30<4:36:17,  6.37s/it]INFO 04-11 22:42:21 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 699/3300 [05:34<4:13:41,  5.85s/it]2026-04-11 22:42:25,978 | INFO | train_grpo_train | metrics_logged mode=train step=699

                                                    
{'loss': 0.0056, 'grad_norm': 8.473615646362305, 'learning_rate': 7.884848484848485e-06, 'num_tokens': 1521331.0, 'completions/mean_length': 67.125, 'completions/min_length': 63.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9927859306335449, 'rewards/meter/std': 0.014541360549628735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.17817415297031403, 'rewards/total_composite/mean': 0.8282062411308289, 'rewards/total_composite/std': 0.18183693289756775, 'reward': 0.8282062411308289, 'reward_std': 0.18183691799640656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03374454379081726, 'sampling/sampling_logp_difference/max': 1.766160249710083, 'sampling/importance_sampling_ratio/min': 0.1795266717672348, 'sampling/importance_sampling_ratio/mean': 0.9997567534446716, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14588068891316652, 'clip_ratio/low_mean': 0.00747219193726778, 'clip_ratio/low_min': 0.00747219193726778, 'clip_ratio/high_mean': 0.02036348171532154, 'clip_ratio/high_max': 0.02036348171532154, 'clip_ratio/region_mean': 0.02783567365258932, 'reward_total_mean': 0.8282062411308289, 'reward_meter_mean': 0.9927859306335449, 'reward_meter_std': 0.014541360549628735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.17817415297031403, 'reward_total_composite_mean': 0.8282062411308289, 'reward_total_composite_std': 0.18183693289756775, 'epoch': 0.03}

 21%|██        | 699/3300 [05:34<4:13:41,  5.85s/it]INFO 04-11 22:42:26 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 700/3300 [05:44<4:56:11,  6.84s/it]2026-04-11 22:42:35,107 | INFO | train_grpo_train | metrics_logged mode=train step=700

                                                    
{'loss': 0.0097, 'grad_norm': 0.8298696279525757, 'learning_rate': 7.881818181818182e-06, 'num_tokens': 1526242.0, 'completions/mean_length': 414.875, 'completions/min_length': 401.0, 'completions/max_length': 444.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 414.875, 'completions/min_terminated_length': 401.0, 'completions/max_terminated_length': 444.0, 'rewards/meter/mean': 0.9960967302322388, 'rewards/meter/std': 0.0013218529056757689, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.029462797567248344, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.19121241569519043, 'rewards/repeat_penalty/std': 0.07494427263736725, 'rewards/total_composite/mean': 0.16019713878631592, 'rewards/total_composite/std': 0.061176449060440063, 'reward': 0.16019713878631592, 'reward_std': 0.061176449060440063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008506695739924908, 'sampling/sampling_logp_difference/max': 1.432920217514038, 'sampling/importance_sampling_ratio/min': 0.238611102104187, 'sampling/importance_sampling_ratio/mean': 1.0012154579162598, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.045434954110533, 'clip_ratio/low_mean': 0.00364298140630126, 'clip_ratio/low_min': 0.00364298140630126, 'clip_ratio/high_mean': 0.006157157360576093, 'clip_ratio/high_max': 0.006157157360576093, 'clip_ratio/region_mean': 0.009800138766877353, 'reward_total_mean': 0.16019713878631592, 'reward_meter_mean': 0.9960967302322388, 'reward_meter_std': 0.0013218529056757689, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.029462797567248344, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.19121241569519043, 'reward_repeat_penalty_std': 0.07494427263736725, 'reward_total_composite_mean': 0.16019713878631592, 'reward_total_composite_std': 0.061176449060440063, 'epoch': 0.03}

 21%|██        | 700/3300 [05:44<4:56:11,  6.84s/it]INFO 04-11 22:42:35 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:39,  3.56s/it]

 23%|██▎       | 3/13 [00:11<00:38,  3.85s/it]

 31%|███       | 4/13 [00:18<00:45,  5.08s/it]

 38%|███▊      | 5/13 [00:25<00:47,  5.88s/it]

 46%|████▌     | 6/13 [00:32<00:42,  6.12s/it]

 54%|█████▍    | 7/13 [00:37<00:35,  5.87s/it]

 62%|██████▏   | 8/13 [00:45<00:31,  6.38s/it]

 69%|██████▉   | 9/13 [00:52<00:26,  6.68s/it]

 77%|███████▋  | 10/13 [00:59<00:20,  6.85s/it]

 85%|████████▍ | 11/13 [01:07<00:13,  6.95s/it]

 92%|█████████▏| 12/13 [01:14<00:07,  7.08s/it]

100%|██████████| 13/13 [01:20<00:00,  6.66s/it]2026-04-11 22:44:02,860 | INFO | train_grpo_train | metrics_logged mode=eval step=700

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 87.7418, 'eval_samples_per_second': 1.185, 'eval_steps_per_second': 0.148, 'eval_num_tokens': 1526242.0, 'eval_completions/mean_length': 249.43269230769232, 'eval_completions/min_length': 68.53846153846153, 'eval_completions/max_length': 471.9230769230769, 'eval_completions/clipped_ratio': 0.11538461538461539, 'eval_completions/mean_terminated_length': 217.6739994929387, 'eval_completions/min_terminated_length': 68.53846153846153, 'eval_completions/max_terminated_length': 415.15384615384613, 'eval_rewards/meter/mean': 0.6321157022164419, 'eval_rewards/meter/std': 0.37711624113413006, 'eval_rewards/count_adherence/mean': 0.8881359283740704, 'eval_rewards/count_adherence/std': 0.16248861929545036, 'eval_rewards/arabic_clean/mean': 0.9038461538461539, 'eval_rewards/arabic_clean/std': 0.2343954168833219, 'eval_rewards/repeat_penalty/mean': 0.5953293947073129, 'eval_rewards/repeat_penalty/std': 0.32899803152451146, 'eval_rewards/total_composite/mean': 0.29438196466519284, 'eval_rewards/total_composite/std': 0.28008361991781455, 'eval_reward': 0.29438196466519284, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.012457216982371531, 'eval_sampling/sampling_logp_difference/max': 0.881988103573139, 'eval_sampling/importance_sampling_ratio/min': 0.4343368663237645, 'eval_sampling/importance_sampling_ratio/mean': 1.0041690973135142, 'eval_sampling/importance_sampling_ratio/max': 1.455959943624643, 'eval_entropy': 0.1376804428604933, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.29438196466519284, 'eval_reward_meter_mean': 0.6321157022164419, 'eval_reward_meter_std': 0.37711624113413006, 'eval_reward_count_adherence_mean': 0.8881359283740704, 'eval_reward_count_adherence_std': 0.16248861929545036, 'eval_reward_arabic_clean_mean': 0.9038461538461539, 'eval_reward_arabic_clean_std': 0.2343954168833219, 'eval_reward_repeat_penalty_mean': 0.5953293947073129, 'eval_reward_repeat_penalty_std': 0.32899803152451146, 'eval_reward_total_composite_mean': 0.29438196466519284, 'eval_reward_total_composite_std': 0.28008361991781455, 'epoch': 0.03}

 21%|██        | 700/3300 [07:11<4:56:11,  6.84s/it]

100%|██████████| 13/13 [01:20<00:00,  6.66s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 21%|██        | 701/3300 [07:26<25:42:36, 35.61s/it]2026-04-11 22:44:17,868 | INFO | train_grpo_train | metrics_logged mode=train step=701

                                                     
{'loss': -0.12, 'grad_norm': 2.5312631130218506, 'learning_rate': 7.87878787878788e-06, 'num_tokens': 1528168.0, 'completions/mean_length': 138.75, 'completions/min_length': 79.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 85.42857360839844, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.646858811378479, 'rewards/meter/std': 0.40135496854782104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6387161016464233, 'rewards/total_composite/std': 0.41526278853416443, 'reward': 0.6387161016464233, 'reward_std': 0.41526278853416443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017966095358133316, 'sampling/sampling_logp_difference/max': 0.638648509979248, 'sampling/importance_sampling_ratio/min': 0.5280055403709412, 'sampling/importance_sampling_ratio/mean': 1.0036530494689941, 'sampling/importance_sampling_ratio/max': 1.6849335432052612, 'entropy': 0.11331802047789097, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.015958538744598627, 'clip_ratio/high_max': 0.015958538744598627, 'clip_ratio/region_mean': 0.015958538744598627, 'reward_total_mean': 0.6387161016464233, 'reward_meter_mean': 0.646858811378479, 'reward_meter_std': 0.40135496854782104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6387161016464233, 'reward_total_composite_std': 0.41526278853416443, 'epoch': 0.03}

 21%|██        | 701/3300 [07:26<25:42:36, 35.61s/it]INFO 04-11 22:44:18 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 702/3300 [07:36<20:05:33, 27.84s/it]2026-04-11 22:44:27,579 | INFO | train_grpo_train | metrics_logged mode=train step=702

                                                     
{'loss': -0.0691, 'grad_norm': 3.0773770809173584, 'learning_rate': 7.875757575757577e-06, 'num_tokens': 1529770.0, 'completions/mean_length': 128.25, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.42857360839844, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.3803873062133789, 'rewards/meter/std': 0.26946189999580383, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.33759811520576477, 'rewards/total_composite/std': 0.30163025856018066, 'reward': 0.33759811520576477, 'reward_std': 0.3016302287578583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04749147966504097, 'sampling/sampling_logp_difference/max': 1.9366450309753418, 'sampling/importance_sampling_ratio/min': 0.14418688416481018, 'sampling/importance_sampling_ratio/mean': 1.011560082435608, 'sampling/importance_sampling_ratio/max': 1.7729606628417969, 'entropy': 0.37723080068826675, 'clip_ratio/low_mean': 0.01317842910066247, 'clip_ratio/low_min': 0.01317842910066247, 'clip_ratio/high_mean': 0.0155344782397151, 'clip_ratio/high_max': 0.0155344782397151, 'clip_ratio/region_mean': 0.02871290734037757, 'reward_total_mean': 0.33759811520576477, 'reward_meter_mean': 0.3803873062133789, 'reward_meter_std': 0.26946189999580383, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.33759811520576477, 'reward_total_composite_std': 0.30163025856018066, 'epoch': 0.03}

 21%|██▏       | 702/3300 [07:36<20:05:33, 27.84s/it]INFO 04-11 22:44:27 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 703/3300 [07:43<15:33:35, 21.57s/it]2026-04-11 22:44:34,512 | INFO | train_grpo_train | metrics_logged mode=train step=703

                                                     
{'loss': -0.0245, 'grad_norm': 1.2964930534362793, 'learning_rate': 7.872727272727273e-06, 'num_tokens': 1533320.0, 'completions/mean_length': 229.75, 'completions/min_length': 206.0, 'completions/max_length': 242.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 229.75, 'completions/min_terminated_length': 206.0, 'completions/max_terminated_length': 242.0, 'rewards/meter/mean': 0.9970081448554993, 'rewards/meter/std': 0.001187506248243153, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.2613636255264282, 'rewards/repeat_penalty/std': 0.197011336684227, 'rewards/total_composite/mean': 0.2607119679450989, 'rewards/total_composite/std': 0.19680176675319672, 'reward': 0.2607119679450989, 'reward_std': 0.1968017816543579, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014087834395468235, 'sampling/sampling_logp_difference/max': 4.6370849609375, 'sampling/importance_sampling_ratio/min': 0.009685891680419445, 'sampling/importance_sampling_ratio/mean': 0.9988393187522888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03166929807048291, 'clip_ratio/low_mean': 0.0028346364269964397, 'clip_ratio/low_min': 0.0028346364269964397, 'clip_ratio/high_mean': 0.004737977171316743, 'clip_ratio/high_max': 0.004737977171316743, 'clip_ratio/region_mean': 0.007572613598313183, 'reward_total_mean': 0.2607119679450989, 'reward_meter_mean': 0.9970081448554993, 'reward_meter_std': 0.001187506248243153, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.2613636255264282, 'reward_repeat_penalty_std': 0.197011336684227, 'reward_total_composite_mean': 0.2607119679450989, 'reward_total_composite_std': 0.19680176675319672, 'epoch': 0.03}

 21%|██▏       | 703/3300 [07:43<15:33:35, 21.57s/it]INFO 04-11 22:44:34 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 704/3300 [07:53<13:02:02, 18.07s/it]2026-04-11 22:44:44,432 | INFO | train_grpo_train | metrics_logged mode=train step=704

                                                     
{'loss': -0.1074, 'grad_norm': 3.015087127685547, 'learning_rate': 7.86969696969697e-06, 'num_tokens': 1536668.0, 'completions/mean_length': 364.5, 'completions/min_length': 229.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 315.3333435058594, 'completions/min_terminated_length': 229.0, 'completions/max_terminated_length': 368.0, 'rewards/meter/mean': 0.20065514743328094, 'rewards/meter/std': 0.3053584098815918, 'rewards/count_adherence/mean': 0.8522727489471436, 'rewards/count_adherence/std': 0.21697448194026947, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.6299689412117004, 'rewards/repeat_penalty/std': 0.28235092759132385, 'rewards/total_composite/mean': 0.0969100296497345, 'rewards/total_composite/std': 0.21232596039772034, 'reward': 0.0969100296497345, 'reward_std': 0.21232594549655914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04316805675625801, 'sampling/sampling_logp_difference/max': 2.4648847579956055, 'sampling/importance_sampling_ratio/min': 0.08501863479614258, 'sampling/importance_sampling_ratio/mean': 0.9969695806503296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2017015889286995, 'clip_ratio/low_mean': 0.01385743310675025, 'clip_ratio/low_min': 0.01385743310675025, 'clip_ratio/high_mean': 0.009915342554450035, 'clip_ratio/high_max': 0.009915342554450035, 'clip_ratio/region_mean': 0.023772775661200285, 'reward_total_mean': 0.0969100296497345, 'reward_meter_mean': 0.20065514743328094, 'reward_meter_std': 0.3053584098815918, 'reward_count_adherence_mean': 0.8522727489471436, 'reward_count_adherence_std': 0.21697448194026947, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.6299689412117004, 'reward_repeat_penalty_std': 0.28235092759132385, 'reward_total_composite_mean': 0.0969100296497345, 'reward_total_composite_std': 0.21232596039772034, 'epoch': 0.03}

 21%|██▏       | 704/3300 [07:53<13:02:02, 18.07s/it]INFO 04-11 22:44:44 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 705/3300 [07:58<10:09:12, 14.09s/it]2026-04-11 22:44:49,210 | INFO | train_grpo_train | metrics_logged mode=train step=705

                                                     
{'loss': -0.0075, 'grad_norm': 2.615363836288452, 'learning_rate': 7.866666666666667e-06, 'num_tokens': 1538402.0, 'completions/mean_length': 58.75, 'completions/min_length': 56.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9921605587005615, 'rewards/meter/std': 0.006017809733748436, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.909184455871582, 'rewards/total_composite/std': 0.15155236423015594, 'reward': 0.909184455871582, 'reward_std': 0.15155236423015594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021948276087641716, 'sampling/sampling_logp_difference/max': 3.973776340484619, 'sampling/importance_sampling_ratio/min': 0.01880229450762272, 'sampling/importance_sampling_ratio/mean': 0.9994084239006042, 'sampling/importance_sampling_ratio/max': 1.5749088525772095, 'entropy': 0.053199955029413104, 'clip_ratio/low_mean': 0.006398809840902686, 'clip_ratio/low_min': 0.006398809840902686, 'clip_ratio/high_mean': 0.006355932215228677, 'clip_ratio/high_max': 0.006355932215228677, 'clip_ratio/region_mean': 0.012754742056131363, 'reward_total_mean': 0.909184455871582, 'reward_meter_mean': 0.9921605587005615, 'reward_meter_std': 0.006017809733748436, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.909184455871582, 'reward_total_composite_std': 0.15155236423015594, 'epoch': 0.03}

 21%|██▏       | 705/3300 [07:58<10:09:12, 14.09s/it]INFO 04-11 22:44:49 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 706/3300 [08:02<8:02:09, 11.15s/it] 2026-04-11 22:44:53,518 | INFO | train_grpo_train | metrics_logged mode=train step=706

                                                    
{'loss': 0.0093, 'grad_norm': 4.515374183654785, 'learning_rate': 7.863636363636364e-06, 'num_tokens': 1539816.0, 'completions/mean_length': 44.75, 'completions/min_length': 43.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9157871603965759, 'rewards/meter/std': 0.05085242539644241, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9157871603965759, 'rewards/total_composite/std': 0.05085242539644241, 'reward': 0.9157871603965759, 'reward_std': 0.05085243284702301, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02820507250726223, 'sampling/sampling_logp_difference/max': 0.7927889823913574, 'sampling/importance_sampling_ratio/min': 0.4525808095932007, 'sampling/importance_sampling_ratio/mean': 0.997020423412323, 'sampling/importance_sampling_ratio/max': 1.2550123929977417, 'entropy': 0.13661748263984919, 'clip_ratio/low_mean': 0.0027777778450399637, 'clip_ratio/low_min': 0.0027777778450399637, 'clip_ratio/high_mean': 0.01414728700183332, 'clip_ratio/high_max': 0.01414728700183332, 'clip_ratio/region_mean': 0.016925064846873283, 'reward_total_mean': 0.9157871603965759, 'reward_meter_mean': 0.9157871603965759, 'reward_meter_std': 0.05085242539644241, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9157871603965759, 'reward_total_composite_std': 0.05085242539644241, 'epoch': 0.03}

 21%|██▏       | 706/3300 [08:02<8:02:09, 11.15s/it]INFO 04-11 22:44:53 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 707/3300 [08:08<6:50:34,  9.50s/it]2026-04-11 22:44:59,164 | INFO | train_grpo_train | metrics_logged mode=train step=707

                                                    
{'loss': -0.0003, 'grad_norm': 2.2196364402770996, 'learning_rate': 7.860606060606062e-06, 'num_tokens': 1541901.0, 'completions/mean_length': 84.625, 'completions/min_length': 83.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.625, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.9956607818603516, 'rewards/meter/std': 0.0007934165187180042, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9956607818603516, 'rewards/total_composite/std': 0.0007934165187180042, 'reward': 0.9956607818603516, 'reward_std': 0.0007934237364679575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026080820709466934, 'sampling/sampling_logp_difference/max': 0.8360247611999512, 'sampling/importance_sampling_ratio/min': 0.4566366374492645, 'sampling/importance_sampling_ratio/mean': 1.0091147422790527, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15222040470689535, 'clip_ratio/low_mean': 0.007405462441965938, 'clip_ratio/low_min': 0.007405462441965938, 'clip_ratio/high_mean': 0.013377037481404841, 'clip_ratio/high_max': 0.013377037481404841, 'clip_ratio/region_mean': 0.02078249992337078, 'reward_total_mean': 0.9956607818603516, 'reward_meter_mean': 0.9956607818603516, 'reward_meter_std': 0.0007934165187180042, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9956607818603516, 'reward_total_composite_std': 0.0007934165187180042, 'epoch': 0.03}

 21%|██▏       | 707/3300 [08:08<6:50:34,  9.50s/it]INFO 04-11 22:44:59 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 708/3300 [08:16<6:36:29,  9.18s/it]2026-04-11 22:45:07,606 | INFO | train_grpo_train | metrics_logged mode=train step=708

                                                    
{'loss': 0.0303, 'grad_norm': 2.581402063369751, 'learning_rate': 7.857575757575759e-06, 'num_tokens': 1546099.0, 'completions/mean_length': 330.75, 'completions/min_length': 284.0, 'completions/max_length': 358.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 330.75, 'completions/min_terminated_length': 284.0, 'completions/max_terminated_length': 358.0, 'rewards/meter/mean': 0.5674466490745544, 'rewards/meter/std': 0.4153376519680023, 'rewards/count_adherence/mean': 0.9027777910232544, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.34049707651138306, 'rewards/repeat_penalty/std': 0.22640277445316315, 'rewards/total_composite/mean': 0.2035118043422699, 'rewards/total_composite/std': 0.23429380357265472, 'reward': 0.2035118043422699, 'reward_std': 0.23429378867149353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023940538987517357, 'sampling/sampling_logp_difference/max': 3.625192880630493, 'sampling/importance_sampling_ratio/min': 0.026643957942724228, 'sampling/importance_sampling_ratio/mean': 1.002025842666626, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10461874585598707, 'clip_ratio/low_mean': 0.012675938894972205, 'clip_ratio/low_min': 0.012675938894972205, 'clip_ratio/high_mean': 0.010587403550744057, 'clip_ratio/high_max': 0.010587403550744057, 'clip_ratio/region_mean': 0.023263342445716262, 'reward_total_mean': 0.2035118043422699, 'reward_meter_mean': 0.5674466490745544, 'reward_meter_std': 0.4153376519680023, 'reward_count_adherence_mean': 0.9027777910232544, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.34049707651138306, 'reward_repeat_penalty_std': 0.22640277445316315, 'reward_total_composite_mean': 0.2035118043422699, 'reward_total_composite_std': 0.23429380357265472, 'epoch': 0.03}

 21%|██▏       | 708/3300 [08:16<6:36:29,  9.18s/it]INFO 04-11 22:45:07 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 709/3300 [08:26<6:50:25,  9.50s/it]2026-04-11 22:45:17,886 | INFO | train_grpo_train | metrics_logged mode=train step=709

                                                    
{'loss': -0.3582, 'grad_norm': 0.8571996688842773, 'learning_rate': 7.854545454545454e-06, 'num_tokens': 1549805.0, 'completions/mean_length': 452.25, 'completions/min_length': 372.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 416.3999938964844, 'completions/min_terminated_length': 372.0, 'completions/max_terminated_length': 472.0, 'rewards/meter/mean': 0.8675932884216309, 'rewards/meter/std': 0.35074320435523987, 'rewards/count_adherence/mean': 0.6590909361839294, 'rewards/count_adherence/std': 0.39101481437683105, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/repeat_penalty/mean': 0.6495236158370972, 'rewards/repeat_penalty/std': 0.30929210782051086, 'rewards/total_composite/mean': 0.25732704997062683, 'rewards/total_composite/std': 0.2531158924102783, 'reward': 0.25732704997062683, 'reward_std': 0.2531158924102783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019949674606323242, 'sampling/sampling_logp_difference/max': 2.8196310997009277, 'sampling/importance_sampling_ratio/min': 0.05962793529033661, 'sampling/importance_sampling_ratio/mean': 1.0035980939865112, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0702343238517642, 'clip_ratio/low_mean': 0.0015743073308840394, 'clip_ratio/low_min': 0.0015743073308840394, 'clip_ratio/high_mean': 0.005005840037483722, 'clip_ratio/high_max': 0.005005840037483722, 'clip_ratio/region_mean': 0.006580147368367761, 'reward_total_mean': 0.25732704997062683, 'reward_meter_mean': 0.8675932884216309, 'reward_meter_std': 0.35074320435523987, 'reward_count_adherence_mean': 0.6590909361839294, 'reward_count_adherence_std': 0.39101481437683105, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_repeat_penalty_mean': 0.6495236158370972, 'reward_repeat_penalty_std': 0.30929210782051086, 'reward_total_composite_mean': 0.25732704997062683, 'reward_total_composite_std': 0.2531158924102783, 'epoch': 0.03}

 21%|██▏       | 709/3300 [08:26<6:50:25,  9.50s/it]INFO 04-11 22:45:18 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 710/3300 [08:37<7:04:48,  9.84s/it]2026-04-11 22:45:28,484 | INFO | train_grpo_train | metrics_logged mode=train step=710

                                                    
{'loss': -0.2247, 'grad_norm': 0.6216875314712524, 'learning_rate': 7.851515151515152e-06, 'num_tokens': 1554988.0, 'completions/mean_length': 474.875, 'completions/min_length': 439.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 469.5714416503906, 'completions/min_terminated_length': 439.0, 'completions/max_terminated_length': 497.0, 'rewards/meter/mean': 0.9897267818450928, 'rewards/meter/std': 0.019744135439395905, 'rewards/count_adherence/mean': 0.7767857313156128, 'rewards/count_adherence/std': 0.20360276103019714, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5224603414535522, 'rewards/repeat_penalty/std': 0.23697951436042786, 'rewards/total_composite/mean': 0.3333708643913269, 'rewards/total_composite/std': 0.18023891746997833, 'reward': 0.3333708643913269, 'reward_std': 0.18023891746997833, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010584630072116852, 'sampling/sampling_logp_difference/max': 2.4926586151123047, 'sampling/importance_sampling_ratio/min': 0.08268983662128448, 'sampling/importance_sampling_ratio/mean': 1.0017552375793457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05670151812955737, 'clip_ratio/low_mean': 0.0021170872496441007, 'clip_ratio/low_min': 0.0021170872496441007, 'clip_ratio/high_mean': 0.004514876694884151, 'clip_ratio/high_max': 0.004514876694884151, 'clip_ratio/region_mean': 0.006631963944528252, 'reward_total_mean': 0.3333708643913269, 'reward_meter_mean': 0.9897267818450928, 'reward_meter_std': 0.019744135439395905, 'reward_count_adherence_mean': 0.7767857313156128, 'reward_count_adherence_std': 0.20360276103019714, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5224603414535522, 'reward_repeat_penalty_std': 0.23697951436042786, 'reward_total_composite_mean': 0.3333708643913269, 'reward_total_composite_std': 0.18023891746997833, 'epoch': 0.03}

 22%|██▏       | 710/3300 [08:37<7:04:48,  9.84s/it]INFO 04-11 22:45:28 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 711/3300 [08:47<7:04:31,  9.84s/it]2026-04-11 22:45:38,315 | INFO | train_grpo_train | metrics_logged mode=train step=711

                                                    
{'loss': -0.0049, 'grad_norm': 0.6046677231788635, 'learning_rate': 7.848484848484849e-06, 'num_tokens': 1560550.0, 'completions/mean_length': 440.25, 'completions/min_length': 436.0, 'completions/max_length': 461.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 440.25, 'completions/min_terminated_length': 436.0, 'completions/max_terminated_length': 461.0, 'rewards/meter/mean': 0.9979845285415649, 'rewards/meter/std': 0.00018585202633403242, 'rewards/count_adherence/mean': 0.7946428060531616, 'rewards/count_adherence/std': 0.025253823027014732, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.0676877498626709, 'rewards/repeat_penalty/std': 0.023803479969501495, 'rewards/total_composite/mean': 0.053851306438446045, 'rewards/total_composite/std': 0.019493911415338516, 'reward': 0.053851306438446045, 'reward_std': 0.019493909552693367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003577793249860406, 'sampling/sampling_logp_difference/max': 1.1886392831802368, 'sampling/importance_sampling_ratio/min': 0.30463549494743347, 'sampling/importance_sampling_ratio/mean': 1.0001322031021118, 'sampling/importance_sampling_ratio/max': 1.460545539855957, 'entropy': 0.011412000167183578, 'clip_ratio/low_mean': 0.0008561643480788916, 'clip_ratio/low_min': 0.0008561643480788916, 'clip_ratio/high_mean': 0.0016890882980078459, 'clip_ratio/high_max': 0.0016890882980078459, 'clip_ratio/region_mean': 0.0025452526460867375, 'reward_total_mean': 0.053851306438446045, 'reward_meter_mean': 0.9979845285415649, 'reward_meter_std': 0.00018585202633403242, 'reward_count_adherence_mean': 0.7946428060531616, 'reward_count_adherence_std': 0.025253823027014732, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.0676877498626709, 'reward_repeat_penalty_std': 0.023803479969501495, 'reward_total_composite_mean': 0.053851306438446045, 'reward_total_composite_std': 0.019493911415338516, 'epoch': 0.03}

 22%|██▏       | 711/3300 [08:47<7:04:31,  9.84s/it]INFO 04-11 22:45:38 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 712/3300 [08:53<6:12:14,  8.63s/it]2026-04-11 22:45:44,126 | INFO | train_grpo_train | metrics_logged mode=train step=712

                                                    
{'loss': 0.0004, 'grad_norm': 4.805282115936279, 'learning_rate': 7.845454545454546e-06, 'num_tokens': 1562410.0, 'completions/mean_length': 69.5, 'completions/min_length': 68.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.8851493000984192, 'rewards/meter/std': 0.1646534502506256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.7055873274803162, 'rewards/total_composite/std': 0.21653573215007782, 'reward': 0.7055873274803162, 'reward_std': 0.216535747051239, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03851144388318062, 'sampling/sampling_logp_difference/max': 1.9700713157653809, 'sampling/importance_sampling_ratio/min': 0.1394468992948532, 'sampling/importance_sampling_ratio/mean': 0.997905969619751, 'sampling/importance_sampling_ratio/max': 1.8839155435562134, 'entropy': 0.1954718241468072, 'clip_ratio/low_mean': 0.016306631732732058, 'clip_ratio/low_min': 0.016306631732732058, 'clip_ratio/high_mean': 0.014235412469133735, 'clip_ratio/high_max': 0.014235412469133735, 'clip_ratio/region_mean': 0.030542044201865792, 'reward_total_mean': 0.7055873274803162, 'reward_meter_mean': 0.8851493000984192, 'reward_meter_std': 0.1646534502506256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.7055873274803162, 'reward_total_composite_std': 0.21653573215007782, 'epoch': 0.03}

 22%|██▏       | 712/3300 [08:53<6:12:14,  8.63s/it]INFO 04-11 22:45:44 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 713/3300 [09:03<6:34:45,  9.16s/it]2026-04-11 22:45:54,507 | INFO | train_grpo_train | metrics_logged mode=train step=713

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.842424242424243e-06, 'num_tokens': 1564234.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9704335927963257, 'rewards/meter/std': 0.07359233498573303, 'rewards/count_adherence/mean': 0.8166667222976685, 'rewards/count_adherence/std': 0.030860668048262596, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5789903998374939, 'rewards/repeat_penalty/std': 0.045423366129398346, 'rewards/total_composite/mean': 0.4566306471824646, 'rewards/total_composite/std': 0.02297402359545231, 'reward': 0.4566306471824646, 'reward_std': 0.02297401800751686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.4566306471824646, 'reward_meter_mean': 0.9704335927963257, 'reward_meter_std': 0.07359233498573303, 'reward_count_adherence_mean': 0.8166667222976685, 'reward_count_adherence_std': 0.030860668048262596, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5789903998374939, 'reward_repeat_penalty_std': 0.045423366129398346, 'reward_total_composite_mean': 0.4566306471824646, 'reward_total_composite_std': 0.02297402359545231, 'epoch': 0.03}

 22%|██▏       | 713/3300 [09:03<6:34:45,  9.16s/it]INFO 04-11 22:45:54 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 714/3300 [09:11<6:20:28,  8.83s/it]2026-04-11 22:46:02,571 | INFO | train_grpo_train | metrics_logged mode=train step=714

                                                    
{'loss': 0.0251, 'grad_norm': 0.7193350791931152, 'learning_rate': 7.83939393939394e-06, 'num_tokens': 1568398.0, 'completions/mean_length': 336.5, 'completions/min_length': 311.0, 'completions/max_length': 345.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 336.5, 'completions/min_terminated_length': 311.0, 'completions/max_terminated_length': 345.0, 'rewards/meter/mean': 0.9977849721908569, 'rewards/meter/std': 0.0006831432110629976, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.19117647409439087, 'rewards/repeat_penalty/std': 0.16262187063694, 'rewards/total_composite/mean': 0.13619501888751984, 'rewards/total_composite/std': 0.1156935766339302, 'reward': 0.13619501888751984, 'reward_std': 0.11569356918334961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003576630027964711, 'sampling/sampling_logp_difference/max': 1.290938377380371, 'sampling/importance_sampling_ratio/min': 0.2750125825405121, 'sampling/importance_sampling_ratio/mean': 0.9997386932373047, 'sampling/importance_sampling_ratio/max': 1.3743659257888794, 'entropy': 0.018028545891866088, 'clip_ratio/low_mean': 0.0018522579048294574, 'clip_ratio/low_min': 0.0018522579048294574, 'clip_ratio/high_mean': 0.0012019231216982007, 'clip_ratio/high_max': 0.0012019231216982007, 'clip_ratio/region_mean': 0.003054181026527658, 'reward_total_mean': 0.13619501888751984, 'reward_meter_mean': 0.9977849721908569, 'reward_meter_std': 0.0006831432110629976, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.19117647409439087, 'reward_repeat_penalty_std': 0.16262187063694, 'reward_total_composite_mean': 0.13619501888751984, 'reward_total_composite_std': 0.1156935766339302, 'epoch': 0.03}

 22%|██▏       | 714/3300 [09:11<6:20:28,  8.83s/it]INFO 04-11 22:46:02 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 715/3300 [09:17<5:36:36,  7.81s/it]2026-04-11 22:46:08,015 | INFO | train_grpo_train | metrics_logged mode=train step=715

                                                    
{'loss': 0.0035, 'grad_norm': 4.87160062789917, 'learning_rate': 7.836363636363638e-06, 'num_tokens': 1570787.0, 'completions/mean_length': 123.625, 'completions/min_length': 115.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.625, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9901050329208374, 'rewards/meter/std': 0.0037112515419721603, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.1608559489250183, 'rewards/total_composite/mean': 0.7246866822242737, 'rewards/total_composite/std': 0.15871340036392212, 'reward': 0.7246866822242737, 'reward_std': 0.15871338546276093, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04571465030312538, 'sampling/sampling_logp_difference/max': 1.6292939186096191, 'sampling/importance_sampling_ratio/min': 0.19606797397136688, 'sampling/importance_sampling_ratio/mean': 0.9998616576194763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23239293694496155, 'clip_ratio/low_mean': 0.023657660058233887, 'clip_ratio/low_min': 0.023657660058233887, 'clip_ratio/high_mean': 0.018382353708148003, 'clip_ratio/high_max': 0.018382353708148003, 'clip_ratio/region_mean': 0.04204001376638189, 'reward_total_mean': 0.7246866822242737, 'reward_meter_mean': 0.9901050329208374, 'reward_meter_std': 0.0037112515419721603, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.1608559489250183, 'reward_total_composite_mean': 0.7246866822242737, 'reward_total_composite_std': 0.15871340036392212, 'epoch': 0.03}

 22%|██▏       | 715/3300 [09:17<5:36:36,  7.81s/it]INFO 04-11 22:46:08 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 716/3300 [09:23<5:20:29,  7.44s/it]2026-04-11 22:46:14,590 | INFO | train_grpo_train | metrics_logged mode=train step=716

                                                    
{'loss': 0.0175, 'grad_norm': 2.7540860176086426, 'learning_rate': 7.833333333333333e-06, 'num_tokens': 1573577.0, 'completions/mean_length': 175.75, 'completions/min_length': 162.0, 'completions/max_length': 215.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 175.75, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 215.0, 'rewards/meter/mean': 0.8235251903533936, 'rewards/meter/std': 0.28830382227897644, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6091269850730896, 'rewards/repeat_penalty/std': 0.19641855359077454, 'rewards/total_composite/mean': 0.4638897478580475, 'rewards/total_composite/std': 0.2275657206773758, 'reward': 0.4638897478580475, 'reward_std': 0.2275657057762146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02091868966817856, 'sampling/sampling_logp_difference/max': 2.359158992767334, 'sampling/importance_sampling_ratio/min': 0.09449966251850128, 'sampling/importance_sampling_ratio/mean': 1.0048706531524658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15365072712302208, 'clip_ratio/low_mean': 0.010836752247996628, 'clip_ratio/low_min': 0.010836752247996628, 'clip_ratio/high_mean': 0.0044064579415135086, 'clip_ratio/high_max': 0.0044064579415135086, 'clip_ratio/region_mean': 0.015243210189510137, 'reward_total_mean': 0.4638897478580475, 'reward_meter_mean': 0.8235251903533936, 'reward_meter_std': 0.28830382227897644, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6091269850730896, 'reward_repeat_penalty_std': 0.19641855359077454, 'reward_total_composite_mean': 0.4638897478580475, 'reward_total_composite_std': 0.2275657206773758, 'epoch': 0.03}

 22%|██▏       | 716/3300 [09:23<5:20:29,  7.44s/it]INFO 04-11 22:46:14 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 717/3300 [09:33<5:52:44,  8.19s/it]2026-04-11 22:46:24,724 | INFO | train_grpo_train | metrics_logged mode=train step=717

                                                    
{'loss': -0.1657, 'grad_norm': 1.3285834789276123, 'learning_rate': 7.83030303030303e-06, 'num_tokens': 1575801.0, 'completions/mean_length': 391.0, 'completions/min_length': 156.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 189.33334350585938, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 210.0, 'rewards/meter/mean': 0.3074490427970886, 'rewards/meter/std': 0.2025083303451538, 'rewards/count_adherence/mean': 0.824999988079071, 'rewards/count_adherence/std': 0.12817399203777313, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/repeat_penalty/mean': 0.8395833373069763, 'rewards/repeat_penalty/std': 0.25571832060813904, 'rewards/total_composite/mean': 0.08441510796546936, 'rewards/total_composite/std': 0.10946666449308395, 'reward': 0.08441510796546936, 'reward_std': 0.10946667194366455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05243195965886116, 'sampling/sampling_logp_difference/max': 8.712602615356445, 'sampling/importance_sampling_ratio/min': 0.0001644995791139081, 'sampling/importance_sampling_ratio/mean': 1.002199649810791, 'sampling/importance_sampling_ratio/max': 1.6081585884094238, 'entropy': 0.13471947237849236, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010442280676215887, 'clip_ratio/high_max': 0.010442280676215887, 'clip_ratio/region_mean': 0.010442280676215887, 'reward_total_mean': 0.08441510796546936, 'reward_meter_mean': 0.3074490427970886, 'reward_meter_std': 0.2025083303451538, 'reward_count_adherence_mean': 0.824999988079071, 'reward_count_adherence_std': 0.12817399203777313, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_repeat_penalty_mean': 0.8395833373069763, 'reward_repeat_penalty_std': 0.25571832060813904, 'reward_total_composite_mean': 0.08441510796546936, 'reward_total_composite_std': 0.10946666449308395, 'epoch': 0.03}

 22%|██▏       | 717/3300 [09:33<5:52:44,  8.19s/it]INFO 04-11 22:46:24 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 718/3300 [09:38<5:10:06,  7.21s/it]2026-04-11 22:46:29,518 | INFO | train_grpo_train | metrics_logged mode=train step=718

                                                    
{'loss': 0.0058, 'grad_norm': 6.637593746185303, 'learning_rate': 7.827272727272728e-06, 'num_tokens': 1577519.0, 'completions/mean_length': 57.75, 'completions/min_length': 57.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9619162082672119, 'rewards/meter/std': 0.06612562388181686, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9619162082672119, 'rewards/total_composite/std': 0.06612562388181686, 'reward': 0.9619162082672119, 'reward_std': 0.06612562388181686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03136257454752922, 'sampling/sampling_logp_difference/max': 1.5830154418945312, 'sampling/importance_sampling_ratio/min': 0.20535492897033691, 'sampling/importance_sampling_ratio/mean': 1.004167914390564, 'sampling/importance_sampling_ratio/max': 1.769487977027893, 'entropy': 0.14472659677267075, 'clip_ratio/low_mean': 0.006465517450124025, 'clip_ratio/low_min': 0.006465517450124025, 'clip_ratio/high_mean': 0.023710741428658366, 'clip_ratio/high_max': 0.023710741428658366, 'clip_ratio/region_mean': 0.03017625887878239, 'reward_total_mean': 0.9619162082672119, 'reward_meter_mean': 0.9619162082672119, 'reward_meter_std': 0.06612562388181686, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9619162082672119, 'reward_total_composite_std': 0.06612562388181686, 'epoch': 0.03}

 22%|██▏       | 718/3300 [09:38<5:10:06,  7.21s/it]INFO 04-11 22:46:29 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 719/3300 [09:48<5:50:49,  8.16s/it]2026-04-11 22:46:39,931 | INFO | train_grpo_train | metrics_logged mode=train step=719

                                                    
{'loss': -0.1972, 'grad_norm': 0.982711136341095, 'learning_rate': 7.824242424242425e-06, 'num_tokens': 1580014.0, 'completions/mean_length': 208.875, 'completions/min_length': 142.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 165.57144165039062, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.8615807294845581, 'rewards/meter/std': 0.1756223738193512, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5022321939468384, 'rewards/repeat_penalty/std': 0.19195686280727386, 'rewards/total_composite/mean': 0.36472243070602417, 'rewards/total_composite/std': 0.1904788464307785, 'reward': 0.36472243070602417, 'reward_std': 0.1904788315296173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011406843550503254, 'sampling/sampling_logp_difference/max': 0.5394062995910645, 'sampling/importance_sampling_ratio/min': 0.5830943584442139, 'sampling/importance_sampling_ratio/mean': 1.0052262544631958, 'sampling/importance_sampling_ratio/max': 1.6336287260055542, 'entropy': 0.07858177460730076, 'clip_ratio/low_mean': 0.0008802816737443209, 'clip_ratio/low_min': 0.0008802816737443209, 'clip_ratio/high_mean': 0.005198180675506592, 'clip_ratio/high_max': 0.005198180675506592, 'clip_ratio/region_mean': 0.006078462349250913, 'reward_total_mean': 0.36472243070602417, 'reward_meter_mean': 0.8615807294845581, 'reward_meter_std': 0.1756223738193512, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5022321939468384, 'reward_repeat_penalty_std': 0.19195686280727386, 'reward_total_composite_mean': 0.36472243070602417, 'reward_total_composite_std': 0.1904788464307785, 'epoch': 0.03}

 22%|██▏       | 719/3300 [09:48<5:50:49,  8.16s/it]INFO 04-11 22:46:40 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 720/3300 [09:56<5:42:27,  7.96s/it]2026-04-11 22:46:47,370 | INFO | train_grpo_train | metrics_logged mode=train step=720

                                                    
{'loss': -0.0478, 'grad_norm': 1.5212147235870361, 'learning_rate': 7.821212121212122e-06, 'num_tokens': 1583419.0, 'completions/mean_length': 228.625, 'completions/min_length': 195.0, 'completions/max_length': 243.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 228.625, 'completions/min_terminated_length': 195.0, 'completions/max_terminated_length': 243.0, 'rewards/meter/mean': 0.9813181161880493, 'rewards/meter/std': 0.009512215852737427, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.527634859085083, 'rewards/repeat_penalty/std': 0.18670302629470825, 'rewards/total_composite/mean': 0.3818710744380951, 'rewards/total_composite/std': 0.13756081461906433, 'reward': 0.3818710744380951, 'reward_std': 0.13756079971790314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014808963052928448, 'sampling/sampling_logp_difference/max': 1.9443225860595703, 'sampling/importance_sampling_ratio/min': 0.14308412373065948, 'sampling/importance_sampling_ratio/mean': 1.0003083944320679, 'sampling/importance_sampling_ratio/max': 1.7837820053100586, 'entropy': 0.08757321583107114, 'clip_ratio/low_mean': 0.0012820513220503926, 'clip_ratio/low_min': 0.0012820513220503926, 'clip_ratio/high_mean': 0.015587894711643457, 'clip_ratio/high_max': 0.015587894711643457, 'clip_ratio/region_mean': 0.01686994603369385, 'reward_total_mean': 0.3818710744380951, 'reward_meter_mean': 0.9813181161880493, 'reward_meter_std': 0.009512215852737427, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.527634859085083, 'reward_repeat_penalty_std': 0.18670302629470825, 'reward_total_composite_mean': 0.3818710744380951, 'reward_total_composite_std': 0.13756081461906433, 'epoch': 0.03}

 22%|██▏       | 720/3300 [09:56<5:42:27,  7.96s/it]INFO 04-11 22:46:47 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 721/3300 [10:01<5:04:53,  7.09s/it]2026-04-11 22:46:52,602 | INFO | train_grpo_train | metrics_logged mode=train step=721

                                                    
{'loss': -0.0114, 'grad_norm': 9.25263786315918, 'learning_rate': 7.81818181818182e-06, 'num_tokens': 1585267.0, 'completions/mean_length': 65.0, 'completions/min_length': 60.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.6475868821144104, 'rewards/meter/std': 0.3421926498413086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.5946333408355713, 'rewards/total_composite/std': 0.34638410806655884, 'reward': 0.5946333408355713, 'reward_std': 0.34638410806655884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1203828752040863, 'sampling/sampling_logp_difference/max': 2.941412925720215, 'sampling/importance_sampling_ratio/min': 0.052791085094213486, 'sampling/importance_sampling_ratio/mean': 0.991131603717804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5916384495794773, 'clip_ratio/low_mean': 0.051155281253159046, 'clip_ratio/low_min': 0.051155281253159046, 'clip_ratio/high_mean': 0.05381742771714926, 'clip_ratio/high_max': 0.05381742771714926, 'clip_ratio/region_mean': 0.1049727089703083, 'reward_total_mean': 0.5946333408355713, 'reward_meter_mean': 0.6475868821144104, 'reward_meter_std': 0.3421926498413086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.5946333408355713, 'reward_total_composite_std': 0.34638410806655884, 'epoch': 0.03}

 22%|██▏       | 721/3300 [10:01<5:04:53,  7.09s/it]INFO 04-11 22:46:52 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 722/3300 [10:06<4:41:00,  6.54s/it]2026-04-11 22:46:57,829 | INFO | train_grpo_train | metrics_logged mode=train step=722

                                                    
{'loss': 0.3273, 'grad_norm': 7.886468410491943, 'learning_rate': 7.815151515151515e-06, 'num_tokens': 1586837.0, 'completions/mean_length': 48.25, 'completions/min_length': 42.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9944906830787659, 'rewards/meter/std': 0.0011895333882421255, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8698122501373291, 'rewards/total_composite/std': 0.35145723819732666, 'reward': 0.8698122501373291, 'reward_std': 0.35145723819732666, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01547156646847725, 'sampling/sampling_logp_difference/max': 0.7411696910858154, 'sampling/importance_sampling_ratio/min': 0.4765561819076538, 'sampling/importance_sampling_ratio/mean': 1.0014286041259766, 'sampling/importance_sampling_ratio/max': 1.3288551568984985, 'entropy': 0.10689277853816748, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.8698122501373291, 'reward_meter_mean': 0.9944906830787659, 'reward_meter_std': 0.0011895333882421255, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8698122501373291, 'reward_total_composite_std': 0.35145723819732666, 'epoch': 0.03}

 22%|██▏       | 722/3300 [10:06<4:41:00,  6.54s/it]INFO 04-11 22:46:58 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 723/3300 [10:17<5:34:29,  7.79s/it]2026-04-11 22:47:08,396 | INFO | train_grpo_train | metrics_logged mode=train step=723

                                                    
{'loss': -0.2014, 'grad_norm': 0.8348656892776489, 'learning_rate': 7.812121212121213e-06, 'num_tokens': 1590899.0, 'completions/mean_length': 485.75, 'completions/min_length': 452.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 470.0, 'completions/min_terminated_length': 452.0, 'completions/max_terminated_length': 494.0, 'rewards/meter/mean': 0.996362566947937, 'rewards/meter/std': 0.0027744658291339874, 'rewards/count_adherence/mean': 0.9464285373687744, 'rewards/count_adherence/std': 0.03306501731276512, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.10470085591077805, 'rewards/repeat_penalty/std': 0.10408195108175278, 'rewards/total_composite/mean': 0.09887628257274628, 'rewards/total_composite/std': 0.09698235988616943, 'reward': 0.09887628257274628, 'reward_std': 0.09698235988616943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007093369495123625, 'sampling/sampling_logp_difference/max': 1.990865707397461, 'sampling/importance_sampling_ratio/min': 0.13657712936401367, 'sampling/importance_sampling_ratio/mean': 1.0003700256347656, 'sampling/importance_sampling_ratio/max': 1.536428451538086, 'entropy': 0.018016068963333964, 'clip_ratio/low_mean': 0.0007961043156683445, 'clip_ratio/low_min': 0.0007961043156683445, 'clip_ratio/high_mean': 0.0007826215587556362, 'clip_ratio/high_max': 0.0007826215587556362, 'clip_ratio/region_mean': 0.0015787258744239807, 'reward_total_mean': 0.09887628257274628, 'reward_meter_mean': 0.996362566947937, 'reward_meter_std': 0.0027744658291339874, 'reward_count_adherence_mean': 0.9464285373687744, 'reward_count_adherence_std': 0.03306501731276512, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.10470085591077805, 'reward_repeat_penalty_std': 0.10408195108175278, 'reward_total_composite_mean': 0.09887628257274628, 'reward_total_composite_std': 0.09698235988616943, 'epoch': 0.03}

 22%|██▏       | 723/3300 [10:17<5:34:29,  7.79s/it]INFO 04-11 22:47:08 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 724/3300 [10:26<5:47:01,  8.08s/it]2026-04-11 22:47:17,111 | INFO | train_grpo_train | metrics_logged mode=train step=724

                                                    
{'loss': 0.0117, 'grad_norm': 1.2350369691848755, 'learning_rate': 7.80909090909091e-06, 'num_tokens': 1595768.0, 'completions/mean_length': 390.625, 'completions/min_length': 379.0, 'completions/max_length': 403.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 390.625, 'completions/min_terminated_length': 379.0, 'completions/max_terminated_length': 403.0, 'rewards/meter/mean': 0.9978216886520386, 'rewards/meter/std': 0.0007184247369877994, 'rewards/count_adherence/mean': 0.734375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.17863407731056213, 'rewards/repeat_penalty/std': 0.16910506784915924, 'rewards/total_composite/mean': 0.1255086362361908, 'rewards/total_composite/std': 0.10828039050102234, 'reward': 0.1255086362361908, 'reward_std': 0.10828038305044174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0057006035931408405, 'sampling/sampling_logp_difference/max': 1.5134329795837402, 'sampling/importance_sampling_ratio/min': 0.22015291452407837, 'sampling/importance_sampling_ratio/mean': 0.9993754625320435, 'sampling/importance_sampling_ratio/max': 1.8996793031692505, 'entropy': 0.028950548847205937, 'clip_ratio/low_mean': 0.0028556776233017445, 'clip_ratio/low_min': 0.0028556776233017445, 'clip_ratio/high_mean': 0.002920488826930523, 'clip_ratio/high_max': 0.002920488826930523, 'clip_ratio/region_mean': 0.005776166450232267, 'reward_total_mean': 0.1255086362361908, 'reward_meter_mean': 0.9978216886520386, 'reward_meter_std': 0.0007184247369877994, 'reward_count_adherence_mean': 0.734375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.17863407731056213, 'reward_repeat_penalty_std': 0.16910506784915924, 'reward_total_composite_mean': 0.1255086362361908, 'reward_total_composite_std': 0.10828039050102234, 'epoch': 0.03}

 22%|██▏       | 724/3300 [10:26<5:47:01,  8.08s/it]INFO 04-11 22:47:17 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 725/3300 [10:35<6:06:22,  8.54s/it]2026-04-11 22:47:26,708 | INFO | train_grpo_train | metrics_logged mode=train step=725

                                                    
{'loss': -0.1213, 'grad_norm': 1.6505030393600464, 'learning_rate': 7.806060606060607e-06, 'num_tokens': 1597686.0, 'completions/mean_length': 136.75, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 83.14286041259766, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5868990421295166, 'rewards/meter/std': 0.38611555099487305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.5530991554260254, 'rewards/total_composite/std': 0.40852445363998413, 'reward': 0.5530991554260254, 'reward_std': 0.4085244834423065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022672384977340698, 'sampling/sampling_logp_difference/max': 0.911320686340332, 'sampling/importance_sampling_ratio/min': 0.4019929766654968, 'sampling/importance_sampling_ratio/mean': 1.0019892454147339, 'sampling/importance_sampling_ratio/max': 1.7016674280166626, 'entropy': 0.09931796230375767, 'clip_ratio/low_mean': 0.004787406767718494, 'clip_ratio/low_min': 0.004787406767718494, 'clip_ratio/high_mean': 0.007217321544885635, 'clip_ratio/high_max': 0.007217321544885635, 'clip_ratio/region_mean': 0.01200472831260413, 'reward_total_mean': 0.5530991554260254, 'reward_meter_mean': 0.5868990421295166, 'reward_meter_std': 0.38611555099487305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.5530991554260254, 'reward_total_composite_std': 0.40852445363998413, 'epoch': 0.03}

 22%|██▏       | 725/3300 [10:35<6:06:22,  8.54s/it]INFO 04-11 22:47:26 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 726/3300 [10:41<5:25:00,  7.58s/it]2026-04-11 22:47:32,040 | INFO | train_grpo_train | metrics_logged mode=train step=726

                                                    
{'loss': -0.0023, 'grad_norm': 1.0429009199142456, 'learning_rate': 7.803030303030303e-06, 'num_tokens': 1600008.0, 'completions/mean_length': 127.25, 'completions/min_length': 126.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.25, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9302235841751099, 'rewards/meter/std': 0.015423719771206379, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5, 'rewards/repeat_penalty/std': 0.1511857956647873, 'rewards/total_composite/mean': 0.46501004695892334, 'rewards/total_composite/std': 0.14159858226776123, 'reward': 0.46501004695892334, 'reward_std': 0.14159856736660004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01078812312334776, 'sampling/sampling_logp_difference/max': 0.9157018661499023, 'sampling/importance_sampling_ratio/min': 0.4002356231212616, 'sampling/importance_sampling_ratio/mean': 1.002754807472229, 'sampling/importance_sampling_ratio/max': 1.5183614492416382, 'entropy': 0.05762754753232002, 'clip_ratio/low_mean': 0.0029605674790218472, 'clip_ratio/low_min': 0.0029605674790218472, 'clip_ratio/high_mean': 0.0009765625, 'clip_ratio/high_max': 0.0009765625, 'clip_ratio/region_mean': 0.003937129979021847, 'reward_total_mean': 0.46501004695892334, 'reward_meter_mean': 0.9302235841751099, 'reward_meter_std': 0.015423719771206379, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5, 'reward_repeat_penalty_std': 0.1511857956647873, 'reward_total_composite_mean': 0.46501004695892334, 'reward_total_composite_std': 0.14159858226776123, 'epoch': 0.03}

 22%|██▏       | 726/3300 [10:41<5:25:00,  7.58s/it]INFO 04-11 22:47:32 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 727/3300 [10:46<4:52:44,  6.83s/it]2026-04-11 22:47:37,119 | INFO | train_grpo_train | metrics_logged mode=train step=727

                                                    
{'loss': 0.0279, 'grad_norm': 7.388904094696045, 'learning_rate': 7.800000000000002e-06, 'num_tokens': 1601923.0, 'completions/mean_length': 99.375, 'completions/min_length': 93.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.375, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.4179952144622803, 'rewards/meter/std': 0.45823782682418823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.3363860249519348, 'rewards/total_composite/std': 0.36587125062942505, 'reward': 0.3363860249519348, 'reward_std': 0.36587125062942505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0847281664609909, 'sampling/sampling_logp_difference/max': 5.028669357299805, 'sampling/importance_sampling_ratio/min': 0.006547517143189907, 'sampling/importance_sampling_ratio/mean': 0.9850558042526245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2684608269482851, 'clip_ratio/low_mean': 0.032587712397798896, 'clip_ratio/low_min': 0.032587712397798896, 'clip_ratio/high_mean': 0.02533797360956669, 'clip_ratio/high_max': 0.02533797360956669, 'clip_ratio/region_mean': 0.057925686007365584, 'reward_total_mean': 0.3363860249519348, 'reward_meter_mean': 0.4179952144622803, 'reward_meter_std': 0.45823782682418823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.3363860249519348, 'reward_total_composite_std': 0.36587125062942505, 'epoch': 0.03}

 22%|██▏       | 727/3300 [10:46<4:52:44,  6.83s/it]INFO 04-11 22:47:37 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 728/3300 [10:55<5:30:43,  7.72s/it]2026-04-11 22:47:46,907 | INFO | train_grpo_train | metrics_logged mode=train step=728

                                                    
{'loss': 0.0076, 'grad_norm': 2.0805506706237793, 'learning_rate': 7.796969696969697e-06, 'num_tokens': 1607332.0, 'completions/mean_length': 459.125, 'completions/min_length': 427.0, 'completions/max_length': 483.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 459.125, 'completions/min_terminated_length': 427.0, 'completions/max_terminated_length': 483.0, 'rewards/meter/mean': 0.28689688444137573, 'rewards/meter/std': 0.3819184899330139, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0431290864944458, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5729086995124817, 'rewards/repeat_penalty/std': 0.018385794013738632, 'rewards/total_composite/mean': 0.12284211814403534, 'rewards/total_composite/std': 0.1601785272359848, 'reward': 0.12284211814403534, 'reward_std': 0.1601785272359848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019973881542682648, 'sampling/sampling_logp_difference/max': 3.37805438041687, 'sampling/importance_sampling_ratio/min': 0.03411376476287842, 'sampling/importance_sampling_ratio/mean': 0.9993253350257874, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06644301256164908, 'clip_ratio/low_mean': 0.008432107453700155, 'clip_ratio/low_min': 0.008432107453700155, 'clip_ratio/high_mean': 0.003258531214669347, 'clip_ratio/high_max': 0.003258531214669347, 'clip_ratio/region_mean': 0.011690638668369502, 'reward_total_mean': 0.12284211814403534, 'reward_meter_mean': 0.28689688444137573, 'reward_meter_std': 0.3819184899330139, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0431290864944458, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5729086995124817, 'reward_repeat_penalty_std': 0.018385794013738632, 'reward_total_composite_mean': 0.12284211814403534, 'reward_total_composite_std': 0.1601785272359848, 'epoch': 0.03}

 22%|██▏       | 728/3300 [10:55<5:30:43,  7.72s/it]INFO 04-11 22:47:47 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 729/3300 [11:05<5:54:08,  8.26s/it]2026-04-11 22:47:56,457 | INFO | train_grpo_train | metrics_logged mode=train step=729

                                                    
{'loss': -0.1369, 'grad_norm': 2.659188747406006, 'learning_rate': 7.793939393939394e-06, 'num_tokens': 1609182.0, 'completions/mean_length': 139.25, 'completions/min_length': 80.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 86.00000762939453, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.700344443321228, 'rewards/meter/std': 0.42837557196617126, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6998236179351807, 'rewards/total_composite/std': 0.4293443560600281, 'reward': 0.6998236179351807, 'reward_std': 0.4293443262577057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022995855659246445, 'sampling/sampling_logp_difference/max': 0.6324782371520996, 'sampling/importance_sampling_ratio/min': 0.5312735438346863, 'sampling/importance_sampling_ratio/mean': 1.0047773122787476, 'sampling/importance_sampling_ratio/max': 1.5019832849502563, 'entropy': 0.1982464650645852, 'clip_ratio/low_mean': 0.0015625000232830644, 'clip_ratio/low_min': 0.0015625000232830644, 'clip_ratio/high_mean': 0.007218867307528853, 'clip_ratio/high_max': 0.007218867307528853, 'clip_ratio/region_mean': 0.008781367330811918, 'reward_total_mean': 0.6998236179351807, 'reward_meter_mean': 0.700344443321228, 'reward_meter_std': 0.42837557196617126, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6998236179351807, 'reward_total_composite_std': 0.4293443560600281, 'epoch': 0.03}

 22%|██▏       | 729/3300 [11:05<5:54:08,  8.26s/it]INFO 04-11 22:47:56 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 730/3300 [11:10<5:10:41,  7.25s/it]2026-04-11 22:48:01,350 | INFO | train_grpo_train | metrics_logged mode=train step=730

                                                    
{'loss': 0.0222, 'grad_norm': 3.446493625640869, 'learning_rate': 7.790909090909092e-06, 'num_tokens': 1611175.0, 'completions/mean_length': 84.125, 'completions/min_length': 78.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.125, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9330159425735474, 'rewards/meter/std': 0.04698435589671135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9330159425735474, 'rewards/total_composite/std': 0.04698435589671135, 'reward': 0.9330159425735474, 'reward_std': 0.04698435962200165, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02119125984609127, 'sampling/sampling_logp_difference/max': 1.4188241958618164, 'sampling/importance_sampling_ratio/min': 0.3314521312713623, 'sampling/importance_sampling_ratio/mean': 1.0041048526763916, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13521220535039902, 'clip_ratio/low_mean': 0.002890269970521331, 'clip_ratio/low_min': 0.002890269970521331, 'clip_ratio/high_mean': 0.010773762594908476, 'clip_ratio/high_max': 0.010773762594908476, 'clip_ratio/region_mean': 0.013664032565429807, 'reward_total_mean': 0.9330159425735474, 'reward_meter_mean': 0.9330159425735474, 'reward_meter_std': 0.04698435589671135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9330159425735474, 'reward_total_composite_std': 0.04698435589671135, 'epoch': 0.03}

 22%|██▏       | 730/3300 [11:10<5:10:41,  7.25s/it]INFO 04-11 22:48:01 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 731/3300 [11:20<5:48:44,  8.15s/it]2026-04-11 22:48:11,574 | INFO | train_grpo_train | metrics_logged mode=train step=731

                                                    
{'loss': -0.2761, 'grad_norm': 3.5135695934295654, 'learning_rate': 7.787878787878789e-06, 'num_tokens': 1613214.0, 'completions/mean_length': 506.875, 'completions/min_length': 471.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 471.0, 'completions/min_terminated_length': 471.0, 'completions/max_terminated_length': 471.0, 'rewards/meter/mean': 0.9958064556121826, 'rewards/meter/std': 0.0007912082364782691, 'rewards/count_adherence/mean': 0.5277777910232544, 'rewards/count_adherence/std': 0.051434461027383804, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.11126373708248138, 'rewards/repeat_penalty/std': 0.07416856288909912, 'rewards/total_composite/mean': 0.055620092898607254, 'rewards/total_composite/std': 0.029501251876354218, 'reward': 0.055620092898607254, 'reward_std': 0.02950124815106392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013655466958880424, 'sampling/sampling_logp_difference/max': 1.3458614349365234, 'sampling/importance_sampling_ratio/min': 0.2603153586387634, 'sampling/importance_sampling_ratio/mean': 0.9985577464103699, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.005113576073199511, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0005307855899445713, 'clip_ratio/high_max': 0.0005307855899445713, 'clip_ratio/region_mean': 0.0005307855899445713, 'reward_total_mean': 0.055620092898607254, 'reward_meter_mean': 0.9958064556121826, 'reward_meter_std': 0.0007912082364782691, 'reward_count_adherence_mean': 0.5277777910232544, 'reward_count_adherence_std': 0.051434461027383804, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.11126373708248138, 'reward_repeat_penalty_std': 0.07416856288909912, 'reward_total_composite_mean': 0.055620092898607254, 'reward_total_composite_std': 0.029501251876354218, 'epoch': 0.03}

 22%|██▏       | 731/3300 [11:20<5:48:44,  8.15s/it]INFO 04-11 22:48:11 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 732/3300 [11:30<6:16:20,  8.79s/it]2026-04-11 22:48:21,879 | INFO | train_grpo_train | metrics_logged mode=train step=732

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.784848484848484e-06, 'num_tokens': 1614982.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9962900280952454, 'rewards/meter/std': 0.001235451316460967, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0345032773911953, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.19743433594703674, 'rewards/repeat_penalty/std': 0.17679214477539062, 'rewards/total_composite/mean': 0.18774111568927765, 'rewards/total_composite/std': 0.16288606822490692, 'reward': 0.18774111568927765, 'reward_std': 0.16288605332374573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.18774111568927765, 'reward_meter_mean': 0.9962900280952454, 'reward_meter_std': 0.001235451316460967, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0345032773911953, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.19743433594703674, 'reward_repeat_penalty_std': 0.17679214477539062, 'reward_total_composite_mean': 0.18774111568927765, 'reward_total_composite_std': 0.16288606822490692, 'epoch': 0.03}

 22%|██▏       | 732/3300 [11:30<6:16:20,  8.79s/it]INFO 04-11 22:48:22 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 733/3300 [11:35<5:24:23,  7.58s/it]2026-04-11 22:48:26,636 | INFO | train_grpo_train | metrics_logged mode=train step=733

                                                    
{'loss': 0.026, 'grad_norm': 5.411409378051758, 'learning_rate': 7.781818181818183e-06, 'num_tokens': 1616806.0, 'completions/mean_length': 72.0, 'completions/min_length': 63.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6948316097259521, 'rewards/meter/std': 0.2794814109802246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.17817415297031403, 'rewards/total_composite/mean': 0.5835638046264648, 'rewards/total_composite/std': 0.28273841738700867, 'reward': 0.5835638046264648, 'reward_std': 0.28273844718933105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.055757418274879456, 'sampling/sampling_logp_difference/max': 3.232696771621704, 'sampling/importance_sampling_ratio/min': 0.03945096582174301, 'sampling/importance_sampling_ratio/mean': 1.0083332061767578, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3269932046532631, 'clip_ratio/low_mean': 0.034895967692136765, 'clip_ratio/low_min': 0.034895967692136765, 'clip_ratio/high_mean': 0.027756539173424244, 'clip_ratio/high_max': 0.027756539173424244, 'clip_ratio/region_mean': 0.06265250686556101, 'reward_total_mean': 0.5835638046264648, 'reward_meter_mean': 0.6948316097259521, 'reward_meter_std': 0.2794814109802246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.17817415297031403, 'reward_total_composite_mean': 0.5835638046264648, 'reward_total_composite_std': 0.28273841738700867, 'epoch': 0.03}

 22%|██▏       | 733/3300 [11:35<5:24:23,  7.58s/it]INFO 04-11 22:48:26 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 734/3300 [11:40<4:52:19,  6.84s/it]2026-04-11 22:48:31,729 | INFO | train_grpo_train | metrics_logged mode=train step=734

                                                    
{'loss': 0.0012, 'grad_norm': 1.1426414251327515, 'learning_rate': 7.778787878787879e-06, 'num_tokens': 1619174.0, 'completions/mean_length': 116.0, 'completions/min_length': 116.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.0, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9807324409484863, 'rewards/meter/std': 0.005540105979889631, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8406277894973755, 'rewards/total_composite/std': 0.004748670384287834, 'reward': 0.8406277894973755, 'reward_std': 0.004748655948787928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004732904955744743, 'sampling/sampling_logp_difference/max': 0.8726806640625, 'sampling/importance_sampling_ratio/min': 0.4178299903869629, 'sampling/importance_sampling_ratio/mean': 1.0012872219085693, 'sampling/importance_sampling_ratio/max': 1.4202854633331299, 'entropy': 0.025226276833564043, 'clip_ratio/low_mean': 0.0021551724057644606, 'clip_ratio/low_min': 0.0021551724057644606, 'clip_ratio/high_mean': 0.0010775862028822303, 'clip_ratio/high_max': 0.0010775862028822303, 'clip_ratio/region_mean': 0.003232758608646691, 'reward_total_mean': 0.8406277894973755, 'reward_meter_mean': 0.9807324409484863, 'reward_meter_std': 0.005540105979889631, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8406277894973755, 'reward_total_composite_std': 0.004748670384287834, 'epoch': 0.03}

 22%|██▏       | 734/3300 [11:40<4:52:19,  6.84s/it]INFO 04-11 22:48:31 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 735/3300 [11:45<4:27:00,  6.25s/it]2026-04-11 22:48:36,599 | INFO | train_grpo_train | metrics_logged mode=train step=735

                                                    
{'loss': 0.0021, 'grad_norm': 6.164267539978027, 'learning_rate': 7.775757575757576e-06, 'num_tokens': 1620938.0, 'completions/mean_length': 77.5, 'completions/min_length': 74.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.8432132005691528, 'rewards/meter/std': 0.11800947040319443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.7734056115150452, 'rewards/total_composite/std': 0.09497449547052383, 'reward': 0.7734056115150452, 'reward_std': 0.09497448056936264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04740333557128906, 'sampling/sampling_logp_difference/max': 1.3389692306518555, 'sampling/importance_sampling_ratio/min': 0.2621157169342041, 'sampling/importance_sampling_ratio/mean': 1.0057860612869263, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2497086301445961, 'clip_ratio/low_mean': 0.026082158088684082, 'clip_ratio/low_min': 0.026082158088684082, 'clip_ratio/high_mean': 0.011470985249616206, 'clip_ratio/high_max': 0.011470985249616206, 'clip_ratio/region_mean': 0.03755314333830029, 'reward_total_mean': 0.7734056115150452, 'reward_meter_mean': 0.8432132005691528, 'reward_meter_std': 0.11800947040319443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.7734056115150452, 'reward_total_composite_std': 0.09497449547052383, 'epoch': 0.03}

 22%|██▏       | 735/3300 [11:45<4:27:00,  6.25s/it]INFO 04-11 22:48:36 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 736/3300 [11:53<4:46:56,  6.71s/it]2026-04-11 22:48:44,408 | INFO | train_grpo_train | metrics_logged mode=train step=736

                                                    
{'loss': 0.46, 'grad_norm': 5.787602424621582, 'learning_rate': 7.772727272727273e-06, 'num_tokens': 1623113.0, 'completions/mean_length': 115.875, 'completions/min_length': 72.0, 'completions/max_length': 353.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 353.0, 'rewards/meter/mean': 0.7237052917480469, 'rewards/meter/std': 0.3117098808288574, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6369646191596985, 'rewards/total_composite/std': 0.40405309200286865, 'reward': 0.6369646191596985, 'reward_std': 0.40405309200286865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08920754492282867, 'sampling/sampling_logp_difference/max': 1.3201618194580078, 'sampling/importance_sampling_ratio/min': 0.2670920789241791, 'sampling/importance_sampling_ratio/mean': 1.0180284976959229, 'sampling/importance_sampling_ratio/max': 1.964148998260498, 'entropy': 1.2863622568547726, 'clip_ratio/low_mean': 0.008429729146882892, 'clip_ratio/low_min': 0.008429729146882892, 'clip_ratio/high_mean': 0.01677176496013999, 'clip_ratio/high_max': 0.01677176496013999, 'clip_ratio/region_mean': 0.02520149410702288, 'reward_total_mean': 0.6369646191596985, 'reward_meter_mean': 0.7237052917480469, 'reward_meter_std': 0.3117098808288574, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6369646191596985, 'reward_total_composite_std': 0.40405309200286865, 'epoch': 0.03}

 22%|██▏       | 736/3300 [11:53<4:46:56,  6.71s/it]INFO 04-11 22:48:44 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 737/3300 [12:03<5:25:07,  7.61s/it]2026-04-11 22:48:54,111 | INFO | train_grpo_train | metrics_logged mode=train step=737

                                                    
{'loss': -0.2132, 'grad_norm': 1.1047923564910889, 'learning_rate': 7.76969696969697e-06, 'num_tokens': 1625144.0, 'completions/mean_length': 294.875, 'completions/min_length': 147.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 164.60000610351562, 'completions/min_terminated_length': 147.0, 'completions/max_terminated_length': 191.0, 'rewards/meter/mean': 0.3803099989891052, 'rewards/meter/std': 0.2973701059818268, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.22160132229328156, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/repeat_penalty/mean': 0.797619104385376, 'rewards/repeat_penalty/std': 0.2185886949300766, 'rewards/total_composite/mean': 0.20652014017105103, 'rewards/total_composite/std': 0.19584397971630096, 'reward': 0.20652014017105103, 'reward_std': 0.19584397971630096, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030263110995292664, 'sampling/sampling_logp_difference/max': 1.4690535068511963, 'sampling/importance_sampling_ratio/min': 0.2301432192325592, 'sampling/importance_sampling_ratio/mean': 1.0015567541122437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12422919739037752, 'clip_ratio/low_mean': 0.0008333333535119891, 'clip_ratio/low_min': 0.0008333333535119891, 'clip_ratio/high_mean': 0.015350477071478963, 'clip_ratio/high_max': 0.015350477071478963, 'clip_ratio/region_mean': 0.016183810424990952, 'reward_total_mean': 0.20652014017105103, 'reward_meter_mean': 0.3803099989891052, 'reward_meter_std': 0.2973701059818268, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.22160132229328156, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_repeat_penalty_mean': 0.797619104385376, 'reward_repeat_penalty_std': 0.2185886949300766, 'reward_total_composite_mean': 0.20652014017105103, 'reward_total_composite_std': 0.19584397971630096, 'epoch': 0.03}

 22%|██▏       | 737/3300 [12:03<5:25:07,  7.61s/it]INFO 04-11 22:48:54 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 738/3300 [12:13<5:58:40,  8.40s/it]2026-04-11 22:49:04,357 | INFO | train_grpo_train | metrics_logged mode=train step=738

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.766666666666666e-06, 'num_tokens': 1626768.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9730854630470276, 'rewards/meter/std': 0.03588658943772316, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.050507619976997375, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.2781907618045807, 'rewards/repeat_penalty/std': 0.2387264519929886, 'rewards/total_composite/mean': 0.2381971776485443, 'rewards/total_composite/std': 0.20977550745010376, 'reward': 0.2381971776485443, 'reward_std': 0.20977550745010376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.2381971776485443, 'reward_meter_mean': 0.9730854630470276, 'reward_meter_std': 0.03588658943772316, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.050507619976997375, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.2781907618045807, 'reward_repeat_penalty_std': 0.2387264519929886, 'reward_total_composite_mean': 0.2381971776485443, 'reward_total_composite_std': 0.20977550745010376, 'epoch': 0.03}

 22%|██▏       | 738/3300 [12:13<5:58:40,  8.40s/it]INFO 04-11 22:49:04 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 739/3300 [12:22<6:14:27,  8.77s/it]2026-04-11 22:49:13,995 | INFO | train_grpo_train | metrics_logged mode=train step=739

                                                    
{'loss': 0.0363, 'grad_norm': 1.4499318599700928, 'learning_rate': 7.763636363636364e-06, 'num_tokens': 1631595.0, 'completions/mean_length': 423.375, 'completions/min_length': 398.0, 'completions/max_length': 483.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 423.375, 'completions/min_terminated_length': 398.0, 'completions/max_terminated_length': 483.0, 'rewards/meter/mean': 0.7534134387969971, 'rewards/meter/std': 0.2928631901741028, 'rewards/count_adherence/mean': 0.3888888955116272, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5623973608016968, 'rewards/repeat_penalty/std': 0.18699996173381805, 'rewards/total_composite/mean': 0.15555939078330994, 'rewards/total_composite/std': 0.07782954722642899, 'reward': 0.15555939078330994, 'reward_std': 0.07782954722642899, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013467294164001942, 'sampling/sampling_logp_difference/max': 3.2660930156707764, 'sampling/importance_sampling_ratio/min': 0.0381552055478096, 'sampling/importance_sampling_ratio/mean': 0.9997313022613525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0637149391695857, 'clip_ratio/low_mean': 0.005539899080758914, 'clip_ratio/low_min': 0.005539899080758914, 'clip_ratio/high_mean': 0.004910050658509135, 'clip_ratio/high_max': 0.004910050658509135, 'clip_ratio/region_mean': 0.01044994973926805, 'reward_total_mean': 0.15555939078330994, 'reward_meter_mean': 0.7534134387969971, 'reward_meter_std': 0.2928631901741028, 'reward_count_adherence_mean': 0.3888888955116272, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5623973608016968, 'reward_repeat_penalty_std': 0.18699996173381805, 'reward_total_composite_mean': 0.15555939078330994, 'reward_total_composite_std': 0.07782954722642899, 'epoch': 0.03}

 22%|██▏       | 739/3300 [12:22<6:14:27,  8.77s/it]INFO 04-11 22:49:14 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 740/3300 [12:28<5:32:19,  7.79s/it]2026-04-11 22:49:19,488 | INFO | train_grpo_train | metrics_logged mode=train step=740

                                                    
{'loss': -0.0085, 'grad_norm': 3.1338510513305664, 'learning_rate': 7.76060606060606e-06, 'num_tokens': 1634096.0, 'completions/mean_length': 133.625, 'completions/min_length': 126.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.625, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.11201652884483337, 'rewards/meter/std': 0.19264455139636993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6964285969734192, 'rewards/repeat_penalty/std': 0.11921756714582443, 'rewards/total_composite/mean': 0.0800618976354599, 'rewards/total_composite/std': 0.13757191598415375, 'reward': 0.0800618976354599, 'reward_std': 0.13757191598415375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02300534024834633, 'sampling/sampling_logp_difference/max': 1.0942306518554688, 'sampling/importance_sampling_ratio/min': 0.33479708433151245, 'sampling/importance_sampling_ratio/mean': 1.005319595336914, 'sampling/importance_sampling_ratio/max': 1.8425383567810059, 'entropy': 0.15879025869071484, 'clip_ratio/low_mean': 0.008456611772999167, 'clip_ratio/low_min': 0.008456611772999167, 'clip_ratio/high_mean': 0.009530608775094151, 'clip_ratio/high_max': 0.009530608775094151, 'clip_ratio/region_mean': 0.01798722054809332, 'reward_total_mean': 0.0800618976354599, 'reward_meter_mean': 0.11201652884483337, 'reward_meter_std': 0.19264455139636993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6964285969734192, 'reward_repeat_penalty_std': 0.11921756714582443, 'reward_total_composite_mean': 0.0800618976354599, 'reward_total_composite_std': 0.13757191598415375, 'epoch': 0.03}

 22%|██▏       | 740/3300 [12:28<5:32:19,  7.79s/it]INFO 04-11 22:49:19 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 741/3300 [12:33<4:54:08,  6.90s/it]2026-04-11 22:49:24,303 | INFO | train_grpo_train | metrics_logged mode=train step=741

                                                    
{'loss': -0.0061, 'grad_norm': 1.2720859050750732, 'learning_rate': 7.757575757575758e-06, 'num_tokens': 1636098.0, 'completions/mean_length': 79.25, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9925205707550049, 'rewards/meter/std': 0.0022588411811739206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.794016420841217, 'rewards/total_composite/std': 0.001807067426852882, 'reward': 0.794016420841217, 'reward_std': 0.0018070697551593184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011411315761506557, 'sampling/sampling_logp_difference/max': 0.8109602928161621, 'sampling/importance_sampling_ratio/min': 0.4444310963153839, 'sampling/importance_sampling_ratio/mean': 1.001147985458374, 'sampling/importance_sampling_ratio/max': 1.5393544435501099, 'entropy': 0.07466331776231527, 'clip_ratio/low_mean': 0.006329114083200693, 'clip_ratio/low_min': 0.006329114083200693, 'clip_ratio/high_mean': 0.007794186705723405, 'clip_ratio/high_max': 0.007794186705723405, 'clip_ratio/region_mean': 0.014123300788924098, 'reward_total_mean': 0.794016420841217, 'reward_meter_mean': 0.9925205707550049, 'reward_meter_std': 0.0022588411811739206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.794016420841217, 'reward_total_composite_std': 0.001807067426852882, 'epoch': 0.03}

 22%|██▏       | 741/3300 [12:33<4:54:08,  6.90s/it]INFO 04-11 22:49:24 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 742/3300 [12:38<4:34:41,  6.44s/it]2026-04-11 22:49:29,687 | INFO | train_grpo_train | metrics_logged mode=train step=742

                                                    
{'loss': 0.012, 'grad_norm': 7.301284313201904, 'learning_rate': 7.754545454545455e-06, 'num_tokens': 1638424.0, 'completions/mean_length': 120.75, 'completions/min_length': 114.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.75, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.7745586633682251, 'rewards/meter/std': 0.25043216347694397, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6964285969734192, 'rewards/repeat_penalty/std': 0.17806050181388855, 'rewards/total_composite/mean': 0.5217785835266113, 'rewards/total_composite/std': 0.1874629557132721, 'reward': 0.5217785835266113, 'reward_std': 0.1874629259109497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04599983990192413, 'sampling/sampling_logp_difference/max': 4.404999732971191, 'sampling/importance_sampling_ratio/min': 0.012216109782457352, 'sampling/importance_sampling_ratio/mean': 0.9962561726570129, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12627906422130764, 'clip_ratio/low_mean': 0.021951976465061307, 'clip_ratio/low_min': 0.021951976465061307, 'clip_ratio/high_mean': 0.018717249389737844, 'clip_ratio/high_max': 0.018717249389737844, 'clip_ratio/region_mean': 0.04066922585479915, 'reward_total_mean': 0.5217785835266113, 'reward_meter_mean': 0.7745586633682251, 'reward_meter_std': 0.25043216347694397, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6964285969734192, 'reward_repeat_penalty_std': 0.17806050181388855, 'reward_total_composite_mean': 0.5217785835266113, 'reward_total_composite_std': 0.1874629557132721, 'epoch': 0.03}

 22%|██▏       | 742/3300 [12:38<4:34:41,  6.44s/it]INFO 04-11 22:49:29 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 743/3300 [12:48<5:23:50,  7.60s/it]2026-04-11 22:49:39,985 | INFO | train_grpo_train | metrics_logged mode=train step=743

                                                    
{'loss': 0.1769, 'grad_norm': 1.0547409057617188, 'learning_rate': 7.751515151515153e-06, 'num_tokens': 1642208.0, 'completions/mean_length': 510.0, 'completions/min_length': 507.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 508.0, 'completions/min_terminated_length': 507.0, 'completions/max_terminated_length': 510.0, 'rewards/meter/mean': 0.45572322607040405, 'rewards/meter/std': 0.1519794911146164, 'rewards/count_adherence/mean': 0.6153846383094788, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5480158925056458, 'rewards/repeat_penalty/std': 0.010451768524944782, 'rewards/total_composite/mean': 0.13946115970611572, 'rewards/total_composite/std': 0.07740650326013565, 'reward': 0.13946115970611572, 'reward_std': 0.07740650326013565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00907099712640047, 'sampling/sampling_logp_difference/max': 1.135289192199707, 'sampling/importance_sampling_ratio/min': 0.32132917642593384, 'sampling/importance_sampling_ratio/mean': 1.0025880336761475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03308352828025818, 'clip_ratio/low_mean': 0.0017248676158487797, 'clip_ratio/low_min': 0.0017248676158487797, 'clip_ratio/high_mean': 0.0012283907853998244, 'clip_ratio/high_max': 0.0012283907853998244, 'clip_ratio/region_mean': 0.002953258401248604, 'reward_total_mean': 0.13946115970611572, 'reward_meter_mean': 0.45572322607040405, 'reward_meter_std': 0.1519794911146164, 'reward_count_adherence_mean': 0.6153846383094788, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5480158925056458, 'reward_repeat_penalty_std': 0.010451768524944782, 'reward_total_composite_mean': 0.13946115970611572, 'reward_total_composite_std': 0.07740650326013565, 'epoch': 0.03}

 23%|██▎       | 743/3300 [12:48<5:23:50,  7.60s/it]INFO 04-11 22:49:40 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 744/3300 [12:53<4:47:00,  6.74s/it]2026-04-11 22:49:44,710 | INFO | train_grpo_train | metrics_logged mode=train step=744

                                                    
{'loss': 0.0283, 'grad_norm': 6.8046875, 'learning_rate': 7.74848484848485e-06, 'num_tokens': 1643996.0, 'completions/mean_length': 64.5, 'completions/min_length': 61.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.5396110415458679, 'rewards/meter/std': 0.20649512112140656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5396110415458679, 'rewards/total_composite/std': 0.20649512112140656, 'reward': 0.5396110415458679, 'reward_std': 0.20649512112140656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06342907249927521, 'sampling/sampling_logp_difference/max': 1.2414631843566895, 'sampling/importance_sampling_ratio/min': 0.28896111249923706, 'sampling/importance_sampling_ratio/mean': 1.0097684860229492, 'sampling/importance_sampling_ratio/max': 1.632293462753296, 'entropy': 0.6070369817316532, 'clip_ratio/low_mean': 0.02111415727995336, 'clip_ratio/low_min': 0.02111415727995336, 'clip_ratio/high_mean': 0.027501578675583005, 'clip_ratio/high_max': 0.027501578675583005, 'clip_ratio/region_mean': 0.048615735955536366, 'reward_total_mean': 0.5396110415458679, 'reward_meter_mean': 0.5396110415458679, 'reward_meter_std': 0.20649512112140656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5396110415458679, 'reward_total_composite_std': 0.20649512112140656, 'epoch': 0.03}

 23%|██▎       | 744/3300 [12:53<4:47:00,  6.74s/it]INFO 04-11 22:49:44 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 745/3300 [12:58<4:20:56,  6.13s/it]2026-04-11 22:49:49,418 | INFO | train_grpo_train | metrics_logged mode=train step=745

                                                    
{'loss': 0.0099, 'grad_norm': 9.705404281616211, 'learning_rate': 7.745454545454545e-06, 'num_tokens': 1645756.0, 'completions/mean_length': 63.0, 'completions/min_length': 62.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9103853106498718, 'rewards/meter/std': 0.22165298461914062, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9103853106498718, 'rewards/total_composite/std': 0.22165298461914062, 'reward': 0.9103853106498718, 'reward_std': 0.22165298461914062, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07189808040857315, 'sampling/sampling_logp_difference/max': 2.013796806335449, 'sampling/importance_sampling_ratio/min': 0.13348090648651123, 'sampling/importance_sampling_ratio/mean': 1.0033738613128662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.25149067025631666, 'clip_ratio/low_mean': 0.008064515888690948, 'clip_ratio/low_min': 0.008064515888690948, 'clip_ratio/high_mean': 0.043645198456943035, 'clip_ratio/high_max': 0.043645198456943035, 'clip_ratio/region_mean': 0.051709714345633984, 'reward_total_mean': 0.9103853106498718, 'reward_meter_mean': 0.9103853106498718, 'reward_meter_std': 0.22165298461914062, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9103853106498718, 'reward_total_composite_std': 0.22165298461914062, 'epoch': 0.03}

 23%|██▎       | 745/3300 [12:58<4:20:56,  6.13s/it]INFO 04-11 22:49:49 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 746/3300 [13:05<4:31:46,  6.38s/it]2026-04-11 22:49:56,400 | INFO | train_grpo_train | metrics_logged mode=train step=746

                                                    
{'loss': 0.0227, 'grad_norm': 2.2934165000915527, 'learning_rate': 7.742424242424244e-06, 'num_tokens': 1649182.0, 'completions/mean_length': 238.25, 'completions/min_length': 205.0, 'completions/max_length': 260.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 238.25, 'completions/min_terminated_length': 205.0, 'completions/max_terminated_length': 260.0, 'rewards/meter/mean': 0.8676279783248901, 'rewards/meter/std': 0.3215867578983307, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.36800700426101685, 'rewards/repeat_penalty/std': 0.24577626585960388, 'rewards/total_composite/mean': 0.2228499948978424, 'rewards/total_composite/std': 0.18318407237529755, 'reward': 0.2228499948978424, 'reward_std': 0.18318407237529755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018860070034861565, 'sampling/sampling_logp_difference/max': 4.0910139083862305, 'sampling/importance_sampling_ratio/min': 0.016722269356250763, 'sampling/importance_sampling_ratio/mean': 1.0014268159866333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06980151077732444, 'clip_ratio/low_mean': 0.003769519622437656, 'clip_ratio/low_min': 0.003769519622437656, 'clip_ratio/high_mean': 0.006430621142499149, 'clip_ratio/high_max': 0.006430621142499149, 'clip_ratio/region_mean': 0.010200140764936805, 'reward_total_mean': 0.2228499948978424, 'reward_meter_mean': 0.8676279783248901, 'reward_meter_std': 0.3215867578983307, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.36800700426101685, 'reward_repeat_penalty_std': 0.24577626585960388, 'reward_total_composite_mean': 0.2228499948978424, 'reward_total_composite_std': 0.18318407237529755, 'epoch': 0.03}

 23%|██▎       | 746/3300 [13:05<4:31:46,  6.38s/it]INFO 04-11 22:49:56 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 747/3300 [13:09<4:06:12,  5.79s/it]2026-04-11 22:50:00,789 | INFO | train_grpo_train | metrics_logged mode=train step=747

                                                    
{'loss': -0.0252, 'grad_norm': 7.876811981201172, 'learning_rate': 7.73939393939394e-06, 'num_tokens': 1650683.0, 'completions/mean_length': 40.625, 'completions/min_length': 38.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9663740396499634, 'rewards/meter/std': 0.011263493448495865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9663740396499634, 'rewards/total_composite/std': 0.011263493448495865, 'reward': 0.9663740396499634, 'reward_std': 0.011263499036431313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03788968175649643, 'sampling/sampling_logp_difference/max': 0.8854336738586426, 'sampling/importance_sampling_ratio/min': 0.41253525018692017, 'sampling/importance_sampling_ratio/mean': 1.0029720067977905, 'sampling/importance_sampling_ratio/max': 1.344473123550415, 'entropy': 0.28702736645936966, 'clip_ratio/low_mean': 0.012351190904155374, 'clip_ratio/low_min': 0.012351190904155374, 'clip_ratio/high_mean': 0.018227352295070887, 'clip_ratio/high_max': 0.018227352295070887, 'clip_ratio/region_mean': 0.03057854319922626, 'reward_total_mean': 0.9663740396499634, 'reward_meter_mean': 0.9663740396499634, 'reward_meter_std': 0.011263493448495865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9663740396499634, 'reward_total_composite_std': 0.011263493448495865, 'epoch': 0.03}

 23%|██▎       | 747/3300 [13:09<4:06:12,  5.79s/it]INFO 04-11 22:50:01 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 748/3300 [13:14<3:54:02,  5.50s/it]2026-04-11 22:50:05,630 | INFO | train_grpo_train | metrics_logged mode=train step=748

                                                    
{'loss': 0.031, 'grad_norm': 4.674757957458496, 'learning_rate': 7.736363636363637e-06, 'num_tokens': 1652564.0, 'completions/mean_length': 82.125, 'completions/min_length': 74.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.125, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.8491703271865845, 'rewards/meter/std': 0.15907908976078033, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8491703271865845, 'rewards/total_composite/std': 0.15907908976078033, 'reward': 0.8491703271865845, 'reward_std': 0.15907907485961914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0295439250767231, 'sampling/sampling_logp_difference/max': 1.300436019897461, 'sampling/importance_sampling_ratio/min': 0.2724129855632782, 'sampling/importance_sampling_ratio/mean': 1.0036686658859253, 'sampling/importance_sampling_ratio/max': 1.843922734260559, 'entropy': 0.22462552785873413, 'clip_ratio/low_mean': 0.002961171790957451, 'clip_ratio/low_min': 0.002961171790957451, 'clip_ratio/high_mean': 0.021474606008268893, 'clip_ratio/high_max': 0.021474606008268893, 'clip_ratio/region_mean': 0.024435777799226344, 'reward_total_mean': 0.8491703271865845, 'reward_meter_mean': 0.8491703271865845, 'reward_meter_std': 0.15907908976078033, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8491703271865845, 'reward_total_composite_std': 0.15907908976078033, 'epoch': 0.03}

 23%|██▎       | 748/3300 [13:14<3:54:02,  5.50s/it]INFO 04-11 22:50:05 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 749/3300 [13:19<3:47:46,  5.36s/it]2026-04-11 22:50:10,654 | INFO | train_grpo_train | metrics_logged mode=train step=749

                                                    
{'loss': 0.0372, 'grad_norm': 6.799373626708984, 'learning_rate': 7.733333333333334e-06, 'num_tokens': 1654625.0, 'completions/mean_length': 93.625, 'completions/min_length': 87.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.625, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9564855694770813, 'rewards/meter/std': 0.10235674679279327, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8075714707374573, 'rewards/total_composite/std': 0.0810769721865654, 'reward': 0.8075714707374573, 'reward_std': 0.0810769572854042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03427453711628914, 'sampling/sampling_logp_difference/max': 2.4858551025390625, 'sampling/importance_sampling_ratio/min': 0.08325432986021042, 'sampling/importance_sampling_ratio/mean': 0.9965777397155762, 'sampling/importance_sampling_ratio/max': 1.903843641281128, 'entropy': 0.12411046819761395, 'clip_ratio/low_mean': 0.026110241888090968, 'clip_ratio/low_min': 0.026110241888090968, 'clip_ratio/high_mean': 0.008620689623057842, 'clip_ratio/high_max': 0.008620689623057842, 'clip_ratio/region_mean': 0.03473093151114881, 'reward_total_mean': 0.8075714707374573, 'reward_meter_mean': 0.9564855694770813, 'reward_meter_std': 0.10235674679279327, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8075714707374573, 'reward_total_composite_std': 0.0810769721865654, 'epoch': 0.03}

 23%|██▎       | 749/3300 [13:19<3:47:46,  5.36s/it]INFO 04-11 22:50:10 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 750/3300 [13:24<3:39:44,  5.17s/it]2026-04-11 22:50:15,387 | INFO | train_grpo_train | metrics_logged mode=train step=750

                                                    
{'loss': 0.031, 'grad_norm': 6.688387870788574, 'learning_rate': 7.730303030303032e-06, 'num_tokens': 1656513.0, 'completions/mean_length': 67.0, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9406866431236267, 'rewards/meter/std': 0.1574798822402954, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.8990964889526367, 'rewards/total_composite/std': 0.18213681876659393, 'reward': 0.8990964889526367, 'reward_std': 0.18213684856891632, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035663776099681854, 'sampling/sampling_logp_difference/max': 0.9646925926208496, 'sampling/importance_sampling_ratio/min': 0.42169255018234253, 'sampling/importance_sampling_ratio/mean': 1.0121128559112549, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21013184823095798, 'clip_ratio/low_mean': 0.00892857147846371, 'clip_ratio/low_min': 0.00892857147846371, 'clip_ratio/high_mean': 0.022817256744019687, 'clip_ratio/high_max': 0.022817256744019687, 'clip_ratio/region_mean': 0.031745828222483397, 'reward_total_mean': 0.8990964889526367, 'reward_meter_mean': 0.9406866431236267, 'reward_meter_std': 0.1574798822402954, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.8990964889526367, 'reward_total_composite_std': 0.18213681876659393, 'epoch': 0.03}

 23%|██▎       | 750/3300 [13:24<3:39:44,  5.17s/it]INFO 04-11 22:50:15 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:40,  3.66s/it]

 23%|██▎       | 3/13 [00:11<00:40,  4.04s/it]

 31%|███       | 4/13 [00:19<00:47,  5.26s/it]

 38%|███▊      | 5/13 [00:26<00:47,  5.99s/it]

 46%|████▌     | 6/13 [00:33<00:44,  6.37s/it]

 54%|█████▍    | 7/13 [00:40<00:39,  6.66s/it]

 62%|██████▏   | 8/13 [00:48<00:34,  6.96s/it]

 69%|██████▉   | 9/13 [00:56<00:28,  7.11s/it]

 77%|███████▋  | 10/13 [01:03<00:21,  7.18s/it]

 85%|████████▍ | 11/13 [01:10<00:14,  7.21s/it]

 92%|█████████▏| 12/13 [01:17<00:07,  7.24s/it]

100%|██████████| 13/13 [01:25<00:00,  7.23s/it]2026-04-11 22:51:48,561 | INFO | train_grpo_train | metrics_logged mode=eval step=750

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 93.1633, 'eval_samples_per_second': 1.116, 'eval_steps_per_second': 0.14, 'eval_num_tokens': 1656513.0, 'eval_completions/mean_length': 284.52884615384613, 'eval_completions/min_length': 61.46153846153846, 'eval_completions/max_length': 496.7692307692308, 'eval_completions/clipped_ratio': 0.23076923076923078, 'eval_completions/mean_terminated_length': 215.78288092980017, 'eval_completions/min_terminated_length': 61.46153846153846, 'eval_completions/max_terminated_length': 393.46153846153845, 'eval_rewards/meter/mean': 0.6660121427132533, 'eval_rewards/meter/std': 0.37975076070198643, 'eval_rewards/count_adherence/mean': 0.7424245018225449, 'eval_rewards/count_adherence/std': 0.2582179422561939, 'eval_rewards/arabic_clean/mean': 0.9711538461538461, 'eval_rewards/arabic_clean/std': 0.08158924258672275, 'eval_rewards/repeat_penalty/mean': 0.638540084545429, 'eval_rewards/repeat_penalty/std': 0.270676647241299, 'eval_rewards/total_composite/mean': 0.33899185634576356, 'eval_rewards/total_composite/std': 0.3308297275350644, 'eval_reward': 0.33899185634576356, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.012646582407447008, 'eval_sampling/sampling_logp_difference/max': 0.8113565261547382, 'eval_sampling/importance_sampling_ratio/min': 0.4615517258644104, 'eval_sampling/importance_sampling_ratio/mean': 1.0036690326837392, 'eval_sampling/importance_sampling_ratio/max': 1.3942875678722675, 'eval_entropy': 0.17351046003974402, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.33899185634576356, 'eval_reward_meter_mean': 0.6660121427132533, 'eval_reward_meter_std': 0.37975076070198643, 'eval_reward_count_adherence_mean': 0.7424245018225449, 'eval_reward_count_adherence_std': 0.2582179422561939, 'eval_reward_arabic_clean_mean': 0.9711538461538461, 'eval_reward_arabic_clean_std': 0.08158924258672275, 'eval_reward_repeat_penalty_mean': 0.638540084545429, 'eval_reward_repeat_penalty_std': 0.270676647241299, 'eval_reward_total_composite_mean': 0.33899185634576356, 'eval_reward_total_composite_std': 0.3308297275350644, 'epoch': 0.03}

 23%|██▎       | 750/3300 [14:57<3:39:44,  5.17s/it]

100%|██████████| 13/13 [01:25<00:00,  7.23s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 23%|██▎       | 751/3300 [15:04<23:48:08, 33.62s/it]2026-04-11 22:51:55,376 | INFO | train_grpo_train | metrics_logged mode=train step=751

                                                     
{'loss': 0.0024, 'grad_norm': 5.31432580947876, 'learning_rate': 7.727272727272727e-06, 'num_tokens': 1658395.0, 'completions/mean_length': 59.25, 'completions/min_length': 59.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9893741607666016, 'rewards/meter/std': 0.007569636683911085, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9893741607666016, 'rewards/total_composite/std': 0.007569636683911085, 'reward': 0.9893741607666016, 'reward_std': 0.007569642271846533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013296845369040966, 'sampling/sampling_logp_difference/max': 0.9290802478790283, 'sampling/importance_sampling_ratio/min': 0.3949167728424072, 'sampling/importance_sampling_ratio/mean': 1.0002750158309937, 'sampling/importance_sampling_ratio/max': 1.868045449256897, 'entropy': 0.052004152443259954, 'clip_ratio/low_mean': 0.006320621585473418, 'clip_ratio/low_min': 0.006320621585473418, 'clip_ratio/high_mean': 0.008403955027461052, 'clip_ratio/high_max': 0.008403955027461052, 'clip_ratio/region_mean': 0.01472457661293447, 'reward_total_mean': 0.9893741607666016, 'reward_meter_mean': 0.9893741607666016, 'reward_meter_std': 0.007569636683911085, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9893741607666016, 'reward_total_composite_std': 0.007569636683911085, 'epoch': 0.03}

 23%|██▎       | 751/3300 [15:04<23:48:08, 33.62s/it]INFO 04-11 22:51:55 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 752/3300 [15:09<17:41:17, 24.99s/it]2026-04-11 22:52:00,240 | INFO | train_grpo_train | metrics_logged mode=train step=752

                                                     
{'loss': 0.0051, 'grad_norm': 4.07772970199585, 'learning_rate': 7.724242424242424e-06, 'num_tokens': 1660117.0, 'completions/mean_length': 59.25, 'completions/min_length': 59.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9949396848678589, 'rewards/meter/std': 0.00028474157443270087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9534660577774048, 'rewards/total_composite/std': 0.11713220924139023, 'reward': 0.9534660577774048, 'reward_std': 0.11713218688964844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009837880730628967, 'sampling/sampling_logp_difference/max': 1.2378616333007812, 'sampling/importance_sampling_ratio/min': 0.2900037169456482, 'sampling/importance_sampling_ratio/mean': 1.0019704103469849, 'sampling/importance_sampling_ratio/max': 1.3358639478683472, 'entropy': 0.03427604655735195, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.006320621585473418, 'clip_ratio/high_max': 0.006320621585473418, 'clip_ratio/region_mean': 0.008403955027461052, 'reward_total_mean': 0.9534660577774048, 'reward_meter_mean': 0.9949396848678589, 'reward_meter_std': 0.00028474157443270087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9534660577774048, 'reward_total_composite_std': 0.11713220924139023, 'epoch': 0.03}

 23%|██▎       | 752/3300 [15:09<17:41:17, 24.99s/it]INFO 04-11 22:52:00 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 753/3300 [15:13<13:19:18, 18.83s/it]2026-04-11 22:52:04,693 | INFO | train_grpo_train | metrics_logged mode=train step=753

                                                     
{'loss': -0.0196, 'grad_norm': 14.107855796813965, 'learning_rate': 7.721212121212122e-06, 'num_tokens': 1661428.0, 'completions/mean_length': 35.875, 'completions/min_length': 34.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9349486827850342, 'rewards/meter/std': 0.1060987114906311, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9349486827850342, 'rewards/total_composite/std': 0.1060987114906311, 'reward': 0.9349486827850342, 'reward_std': 0.1060987114906311, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07551710307598114, 'sampling/sampling_logp_difference/max': 1.5768651962280273, 'sampling/importance_sampling_ratio/min': 0.20662181079387665, 'sampling/importance_sampling_ratio/mean': 1.0009469985961914, 'sampling/importance_sampling_ratio/max': 1.6115726232528687, 'entropy': 0.6193088293075562, 'clip_ratio/low_mean': 0.017439668532460928, 'clip_ratio/low_min': 0.017439668532460928, 'clip_ratio/high_mean': 0.04496192745864391, 'clip_ratio/high_max': 0.04496192745864391, 'clip_ratio/region_mean': 0.06240159599110484, 'reward_total_mean': 0.9349486827850342, 'reward_meter_mean': 0.9349486827850342, 'reward_meter_std': 0.1060987114906311, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9349486827850342, 'reward_total_composite_std': 0.1060987114906311, 'epoch': 0.03}

 23%|██▎       | 753/3300 [15:13<13:19:18, 18.83s/it]INFO 04-11 22:52:04 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 754/3300 [15:18<10:17:33, 14.55s/it]2026-04-11 22:52:09,269 | INFO | train_grpo_train | metrics_logged mode=train step=754

                                                     
{'loss': 0.0091, 'grad_norm': 2.418649673461914, 'learning_rate': 7.718181818181819e-06, 'num_tokens': 1663080.0, 'completions/mean_length': 41.5, 'completions/min_length': 41.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9943797588348389, 'rewards/meter/std': 0.0011224248446524143, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943797588348389, 'rewards/total_composite/std': 0.0011224248446524143, 'reward': 0.9943797588348389, 'reward_std': 0.0011224271729588509, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030370449647307396, 'sampling/sampling_logp_difference/max': 1.4918346405029297, 'sampling/importance_sampling_ratio/min': 0.2249595671892166, 'sampling/importance_sampling_ratio/mean': 0.9925946593284607, 'sampling/importance_sampling_ratio/max': 1.2741270065307617, 'entropy': 0.11915541160851717, 'clip_ratio/low_mean': 0.008928571594879031, 'clip_ratio/low_min': 0.008928571594879031, 'clip_ratio/high_mean': 0.018074912950396538, 'clip_ratio/high_max': 0.018074912950396538, 'clip_ratio/region_mean': 0.02700348454527557, 'reward_total_mean': 0.9943797588348389, 'reward_meter_mean': 0.9943797588348389, 'reward_meter_std': 0.0011224248446524143, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943797588348389, 'reward_total_composite_std': 0.0011224248446524143, 'epoch': 0.03}

 23%|██▎       | 754/3300 [15:18<10:17:33, 14.55s/it]INFO 04-11 22:52:09 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 755/3300 [15:22<8:08:17, 11.51s/it] 2026-04-11 22:52:13,683 | INFO | train_grpo_train | metrics_logged mode=train step=755

                                                    
{'loss': -0.0018, 'grad_norm': 5.144253730773926, 'learning_rate': 7.715151515151516e-06, 'num_tokens': 1664616.0, 'completions/mean_length': 41.0, 'completions/min_length': 41.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9944278001785278, 'rewards/meter/std': 0.0012945194030180573, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9944278001785278, 'rewards/total_composite/std': 0.0012945194030180573, 'reward': 0.9944278001785278, 'reward_std': 0.0012945224298164248, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014835118316113949, 'sampling/sampling_logp_difference/max': 0.617079496383667, 'sampling/importance_sampling_ratio/min': 0.5395178198814392, 'sampling/importance_sampling_ratio/mean': 1.0021907091140747, 'sampling/importance_sampling_ratio/max': 1.2102103233337402, 'entropy': 0.10165042616426945, 'clip_ratio/low_mean': 0.0030487803742289543, 'clip_ratio/low_min': 0.0030487803742289543, 'clip_ratio/high_mean': 0.012195121496915817, 'clip_ratio/high_max': 0.012195121496915817, 'clip_ratio/region_mean': 0.015243901871144772, 'reward_total_mean': 0.9944278001785278, 'reward_meter_mean': 0.9944278001785278, 'reward_meter_std': 0.0012945194030180573, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9944278001785278, 'reward_total_composite_std': 0.0012945194030180573, 'epoch': 0.03}

 23%|██▎       | 755/3300 [15:22<8:08:17, 11.51s/it]INFO 04-11 22:52:13 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 756/3300 [15:28<6:53:20,  9.75s/it]2026-04-11 22:52:19,323 | INFO | train_grpo_train | metrics_logged mode=train step=756

                                                    
{'loss': -0.0099, 'grad_norm': 2.4242031574249268, 'learning_rate': 7.712121212121213e-06, 'num_tokens': 1666878.0, 'completions/mean_length': 113.75, 'completions/min_length': 111.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.75, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.9662827253341675, 'rewards/meter/std': 0.009305375628173351, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6500000357627869, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.6285786032676697, 'rewards/total_composite/std': 0.0941418930888176, 'reward': 0.6285786032676697, 'reward_std': 0.09414192289113998, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011911381967365742, 'sampling/sampling_logp_difference/max': 1.6042141914367676, 'sampling/importance_sampling_ratio/min': 0.20104746520519257, 'sampling/importance_sampling_ratio/mean': 1.0027672052383423, 'sampling/importance_sampling_ratio/max': 1.911029577255249, 'entropy': 0.0643842932768166, 'clip_ratio/low_mean': 0.005512091098353267, 'clip_ratio/low_min': 0.005512091098353267, 'clip_ratio/high_mean': 0.005401917500421405, 'clip_ratio/high_max': 0.005401917500421405, 'clip_ratio/region_mean': 0.010914008598774672, 'reward_total_mean': 0.6285786032676697, 'reward_meter_mean': 0.9662827253341675, 'reward_meter_std': 0.009305375628173351, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6500000357627869, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.6285786032676697, 'reward_total_composite_std': 0.0941418930888176, 'epoch': 0.03}

 23%|██▎       | 756/3300 [15:28<6:53:20,  9.75s/it]INFO 04-11 22:52:19 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 757/3300 [15:35<6:25:39,  9.10s/it]2026-04-11 22:52:26,901 | INFO | train_grpo_train | metrics_logged mode=train step=757

                                                    
{'loss': 0.0259, 'grad_norm': 1.2799259424209595, 'learning_rate': 7.709090909090909e-06, 'num_tokens': 1670368.0, 'completions/mean_length': 243.25, 'completions/min_length': 217.0, 'completions/max_length': 267.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 243.25, 'completions/min_terminated_length': 217.0, 'completions/max_terminated_length': 267.0, 'rewards/meter/mean': 0.9969884157180786, 'rewards/meter/std': 0.000910833477973938, 'rewards/count_adherence/mean': 0.6500000357627869, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4187062978744507, 'rewards/repeat_penalty/std': 0.23771634697914124, 'rewards/total_composite/mean': 0.2822417914867401, 'rewards/total_composite/std': 0.18207497894763947, 'reward': 0.2822417914867401, 'reward_std': 0.18207496404647827, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00974962953478098, 'sampling/sampling_logp_difference/max': 1.2791423797607422, 'sampling/importance_sampling_ratio/min': 0.27827587723731995, 'sampling/importance_sampling_ratio/mean': 1.0003676414489746, 'sampling/importance_sampling_ratio/max': 1.5265671014785767, 'entropy': 0.06725997012108564, 'clip_ratio/low_mean': 0.0010245901066809893, 'clip_ratio/low_min': 0.0010245901066809893, 'clip_ratio/high_mean': 0.007282168400706723, 'clip_ratio/high_max': 0.007282168400706723, 'clip_ratio/region_mean': 0.008306758507387713, 'reward_total_mean': 0.2822417914867401, 'reward_meter_mean': 0.9969884157180786, 'reward_meter_std': 0.000910833477973938, 'reward_count_adherence_mean': 0.6500000357627869, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4187062978744507, 'reward_repeat_penalty_std': 0.23771634697914124, 'reward_total_composite_mean': 0.2822417914867401, 'reward_total_composite_std': 0.18207497894763947, 'epoch': 0.03}

 23%|██▎       | 757/3300 [15:35<6:25:39,  9.10s/it]INFO 04-11 22:52:27 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 758/3300 [15:46<6:41:46,  9.48s/it]2026-04-11 22:52:37,283 | INFO | train_grpo_train | metrics_logged mode=train step=758

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.706060606060606e-06, 'num_tokens': 1672112.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.8747541308403015, 'rewards/meter/std': 0.1777074635028839, 'rewards/count_adherence/mean': 0.15625, 'rewards/count_adherence/std': 0.11080066114664078, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.42480412125587463, 'rewards/repeat_penalty/std': 0.2929826080799103, 'rewards/total_composite/mean': 0.059182293713092804, 'rewards/total_composite/std': 0.06593845039606094, 'reward': 0.059182293713092804, 'reward_std': 0.06593845039606094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.059182293713092804, 'reward_meter_mean': 0.8747541308403015, 'reward_meter_std': 0.1777074635028839, 'reward_count_adherence_mean': 0.15625, 'reward_count_adherence_std': 0.11080066114664078, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.42480412125587463, 'reward_repeat_penalty_std': 0.2929826080799103, 'reward_total_composite_mean': 0.059182293713092804, 'reward_total_composite_std': 0.06593845039606094, 'epoch': 0.03}

 23%|██▎       | 758/3300 [15:46<6:41:46,  9.48s/it]INFO 04-11 22:52:37 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 759/3300 [15:51<5:42:55,  8.10s/it]2026-04-11 22:52:42,145 | INFO | train_grpo_train | metrics_logged mode=train step=759

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.703030303030304e-06, 'num_tokens': 1674144.0, 'completions/mean_length': 87.0, 'completions/min_length': 87.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.0, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9949710369110107, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6000000238418579, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5969825983047485, 'rewards/total_composite/std': 0.0, 'reward': 0.5969825983047485, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.001654994674026966, 'sampling/sampling_logp_difference/max': 0.1641908884048462, 'sampling/importance_sampling_ratio/min': 0.848580002784729, 'sampling/importance_sampling_ratio/mean': 1.0006695985794067, 'sampling/importance_sampling_ratio/max': 1.0613410472869873, 'entropy': 0.014668526826426387, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.5969825983047485, 'reward_meter_mean': 0.9949710369110107, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6000000238418579, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5969825983047485, 'reward_total_composite_std': 0.0, 'epoch': 0.03}

 23%|██▎       | 759/3300 [15:51<5:42:55,  8.10s/it]INFO 04-11 22:52:42 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 760/3300 [15:55<4:58:01,  7.04s/it]2026-04-11 22:52:46,717 | INFO | train_grpo_train | metrics_logged mode=train step=760

                                                    
{'loss': -0.0037, 'grad_norm': 4.493424415588379, 'learning_rate': 7.7e-06, 'num_tokens': 1675945.0, 'completions/mean_length': 54.125, 'completions/min_length': 54.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9517734050750732, 'rewards/meter/std': 0.0011382178636267781, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7083333730697632, 'rewards/repeat_penalty/std': 0.11785111576318741, 'rewards/total_composite/mean': 0.6740554571151733, 'rewards/total_composite/std': 0.11107677966356277, 'reward': 0.6740554571151733, 'reward_std': 0.11107677221298218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005725136026740074, 'sampling/sampling_logp_difference/max': 0.4777810573577881, 'sampling/importance_sampling_ratio/min': 0.7561957836151123, 'sampling/importance_sampling_ratio/mean': 1.0033941268920898, 'sampling/importance_sampling_ratio/max': 1.612492322921753, 'entropy': 0.042751661501824856, 'clip_ratio/low_mean': 0.004629629664123058, 'clip_ratio/low_min': 0.004629629664123058, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004629629664123058, 'reward_total_mean': 0.6740554571151733, 'reward_meter_mean': 0.9517734050750732, 'reward_meter_std': 0.0011382178636267781, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7083333730697632, 'reward_repeat_penalty_std': 0.11785111576318741, 'reward_total_composite_mean': 0.6740554571151733, 'reward_total_composite_std': 0.11107677966356277, 'epoch': 0.03}

 23%|██▎       | 760/3300 [15:55<4:58:01,  7.04s/it]INFO 04-11 22:52:46 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 761/3300 [16:00<4:26:48,  6.31s/it]2026-04-11 22:52:51,308 | INFO | train_grpo_train | metrics_logged mode=train step=761

                                                    
{'loss': -0.0023, 'grad_norm': 0.8903390765190125, 'learning_rate': 7.696969696969696e-06, 'num_tokens': 1677747.0, 'completions/mean_length': 58.25, 'completions/min_length': 58.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9949068427085876, 'rewards/meter/std': 0.0007968654972501099, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7083333730697632, 'rewards/repeat_penalty/std': 0.11785111576318741, 'rewards/total_composite/mean': 0.7047345042228699, 'rewards/total_composite/std': 0.1173337996006012, 'reward': 0.7047345042228699, 'reward_std': 0.1173337996006012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011640225537121296, 'sampling/sampling_logp_difference/max': 1.0540056228637695, 'sampling/importance_sampling_ratio/min': 0.3485388159751892, 'sampling/importance_sampling_ratio/mean': 1.0014004707336426, 'sampling/importance_sampling_ratio/max': 1.6007080078125, 'entropy': 0.05586709058843553, 'clip_ratio/low_mean': 0.006428988883271813, 'clip_ratio/low_min': 0.006428988883271813, 'clip_ratio/high_mean': 0.0021186440717428923, 'clip_ratio/high_max': 0.0021186440717428923, 'clip_ratio/region_mean': 0.008547632955014706, 'reward_total_mean': 0.7047345042228699, 'reward_meter_mean': 0.9949068427085876, 'reward_meter_std': 0.0007968654972501099, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7083333730697632, 'reward_repeat_penalty_std': 0.11785111576318741, 'reward_total_composite_mean': 0.7047345042228699, 'reward_total_composite_std': 0.1173337996006012, 'epoch': 0.03}

 23%|██▎       | 761/3300 [16:00<4:26:48,  6.31s/it]INFO 04-11 22:52:51 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 762/3300 [16:05<4:08:59,  5.89s/it]2026-04-11 22:52:56,217 | INFO | train_grpo_train | metrics_logged mode=train step=762

                                                    
{'loss': 0.024, 'grad_norm': 5.6463236808776855, 'learning_rate': 7.693939393939395e-06, 'num_tokens': 1679613.0, 'completions/mean_length': 78.25, 'completions/min_length': 75.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.25, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.8964411020278931, 'rewards/meter/std': 0.12688924372196198, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6750000715255737, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.6016985774040222, 'rewards/total_composite/std': 0.1109333410859108, 'reward': 0.6016985774040222, 'reward_std': 0.11093335598707199, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02804330736398697, 'sampling/sampling_logp_difference/max': 0.9042620658874512, 'sampling/importance_sampling_ratio/min': 0.4048405587673187, 'sampling/importance_sampling_ratio/mean': 1.0003811120986938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11404913989827037, 'clip_ratio/low_mean': 0.012610982405021787, 'clip_ratio/low_min': 0.012610982405021787, 'clip_ratio/high_mean': 0.006666666595265269, 'clip_ratio/high_max': 0.006666666595265269, 'clip_ratio/region_mean': 0.019277649000287056, 'reward_total_mean': 0.6016985774040222, 'reward_meter_mean': 0.8964411020278931, 'reward_meter_std': 0.12688924372196198, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6750000715255737, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.6016985774040222, 'reward_total_composite_std': 0.1109333410859108, 'epoch': 0.03}

 23%|██▎       | 762/3300 [16:05<4:08:59,  5.89s/it]INFO 04-11 22:52:56 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 763/3300 [16:15<5:05:43,  7.23s/it]2026-04-11 22:53:06,584 | INFO | train_grpo_train | metrics_logged mode=train step=763

                                                    
{'loss': 0.0481, 'grad_norm': 1.2447540760040283, 'learning_rate': 7.690909090909091e-06, 'num_tokens': 1684742.0, 'completions/mean_length': 438.125, 'completions/min_length': 391.0, 'completions/max_length': 498.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 438.125, 'completions/min_terminated_length': 391.0, 'completions/max_terminated_length': 498.0, 'rewards/meter/mean': 0.9878873825073242, 'rewards/meter/std': 0.019713442772626877, 'rewards/count_adherence/mean': 0.2142857164144516, 'rewards/count_adherence/std': 0.07636035978794098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5576170682907104, 'rewards/repeat_penalty/std': 0.0946657732129097, 'rewards/total_composite/mean': 0.12146620452404022, 'rewards/total_composite/std': 0.05504889413714409, 'reward': 0.12146620452404022, 'reward_std': 0.05504889413714409, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010950884781777859, 'sampling/sampling_logp_difference/max': 1.4895976781845093, 'sampling/importance_sampling_ratio/min': 0.22546334564685822, 'sampling/importance_sampling_ratio/mean': 1.0031057596206665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07848789915442467, 'clip_ratio/low_mean': 0.005670643062330782, 'clip_ratio/low_min': 0.005670643062330782, 'clip_ratio/high_mean': 0.005744358117226511, 'clip_ratio/high_max': 0.005744358117226511, 'clip_ratio/region_mean': 0.011415001179557294, 'reward_total_mean': 0.12146620452404022, 'reward_meter_mean': 0.9878873825073242, 'reward_meter_std': 0.019713442772626877, 'reward_count_adherence_mean': 0.2142857164144516, 'reward_count_adherence_std': 0.07636035978794098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5576170682907104, 'reward_repeat_penalty_std': 0.0946657732129097, 'reward_total_composite_mean': 0.12146620452404022, 'reward_total_composite_std': 0.05504889413714409, 'epoch': 0.03}

 23%|██▎       | 763/3300 [16:15<5:05:43,  7.23s/it]INFO 04-11 22:53:06 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 764/3300 [16:20<4:34:45,  6.50s/it]2026-04-11 22:53:11,381 | INFO | train_grpo_train | metrics_logged mode=train step=764

                                                    
{'loss': 0.0189, 'grad_norm': 5.640087127685547, 'learning_rate': 7.687878787878788e-06, 'num_tokens': 1686539.0, 'completions/mean_length': 63.625, 'completions/min_length': 60.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.625, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.874701976776123, 'rewards/meter/std': 0.18508003652095795, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.7850605845451355, 'rewards/total_composite/std': 0.2699912190437317, 'reward': 0.7850605845451355, 'reward_std': 0.2699912190437317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032904475927352905, 'sampling/sampling_logp_difference/max': 1.8050861358642578, 'sampling/importance_sampling_ratio/min': 0.16446030139923096, 'sampling/importance_sampling_ratio/mean': 0.9924401044845581, 'sampling/importance_sampling_ratio/max': 1.438651442527771, 'entropy': 0.1180117940530181, 'clip_ratio/low_mean': 0.005871212342754006, 'clip_ratio/low_min': 0.005871212342754006, 'clip_ratio/high_mean': 0.02714023506268859, 'clip_ratio/high_max': 0.02714023506268859, 'clip_ratio/region_mean': 0.033011447405442595, 'reward_total_mean': 0.7850605845451355, 'reward_meter_mean': 0.874701976776123, 'reward_meter_std': 0.18508003652095795, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.7850605845451355, 'reward_total_composite_std': 0.2699912190437317, 'epoch': 0.03}

 23%|██▎       | 764/3300 [16:20<4:34:45,  6.50s/it]INFO 04-11 22:53:11 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 765/3300 [16:25<4:14:25,  6.02s/it]2026-04-11 22:53:16,287 | INFO | train_grpo_train | metrics_logged mode=train step=765

                                                    
{'loss': 0.0008, 'grad_norm': 4.968373775482178, 'learning_rate': 7.684848484848485e-06, 'num_tokens': 1688548.0, 'completions/mean_length': 87.125, 'completions/min_length': 87.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.125, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9950021505355835, 'rewards/meter/std': 8.806584810372442e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6500000357627869, 'rewards/repeat_penalty/std': 0.1414213478565216, 'rewards/total_composite/mean': 0.6467622518539429, 'rewards/total_composite/std': 0.14079822599887848, 'reward': 0.6467622518539429, 'reward_std': 0.14079821109771729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004592791199684143, 'sampling/sampling_logp_difference/max': 1.7399308681488037, 'sampling/importance_sampling_ratio/min': 0.1755325347185135, 'sampling/importance_sampling_ratio/mean': 0.9999979138374329, 'sampling/importance_sampling_ratio/max': 1.767101764678955, 'entropy': 0.008648848335724324, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0028409091755747795, 'clip_ratio/high_max': 0.0028409091755747795, 'clip_ratio/region_mean': 0.0028409091755747795, 'reward_total_mean': 0.6467622518539429, 'reward_meter_mean': 0.9950021505355835, 'reward_meter_std': 8.806584810372442e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6500000357627869, 'reward_repeat_penalty_std': 0.1414213478565216, 'reward_total_composite_mean': 0.6467622518539429, 'reward_total_composite_std': 0.14079822599887848, 'epoch': 0.03}

 23%|██▎       | 765/3300 [16:25<4:14:25,  6.02s/it]INFO 04-11 22:53:16 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 766/3300 [16:30<3:58:46,  5.65s/it]2026-04-11 22:53:21,082 | INFO | train_grpo_train | metrics_logged mode=train step=766

                                                    
{'loss': 0.0014, 'grad_norm': 10.79392147064209, 'learning_rate': 7.681818181818183e-06, 'num_tokens': 1690236.0, 'completions/mean_length': 52.0, 'completions/min_length': 51.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8247180581092834, 'rewards/meter/std': 0.3005145490169525, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8247180581092834, 'rewards/total_composite/std': 0.3005145490169525, 'reward': 0.8247180581092834, 'reward_std': 0.3005145490169525, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04232970252633095, 'sampling/sampling_logp_difference/max': 1.6089832782745361, 'sampling/importance_sampling_ratio/min': 0.2000909447669983, 'sampling/importance_sampling_ratio/mean': 0.997559666633606, 'sampling/importance_sampling_ratio/max': 1.6643987894058228, 'entropy': 0.19545143470168114, 'clip_ratio/low_mean': 0.004807692486792803, 'clip_ratio/low_min': 0.004807692486792803, 'clip_ratio/high_mean': 0.03126057400368154, 'clip_ratio/high_max': 0.03126057400368154, 'clip_ratio/region_mean': 0.03606826649047434, 'reward_total_mean': 0.8247180581092834, 'reward_meter_mean': 0.8247180581092834, 'reward_meter_std': 0.3005145490169525, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8247180581092834, 'reward_total_composite_std': 0.3005145490169525, 'epoch': 0.03}

 23%|██▎       | 766/3300 [16:30<3:58:46,  5.65s/it]INFO 04-11 22:53:21 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 767/3300 [16:35<3:49:34,  5.44s/it]2026-04-11 22:53:26,017 | INFO | train_grpo_train | metrics_logged mode=train step=767

                                                    
{'loss': 0.0103, 'grad_norm': 12.172296524047852, 'learning_rate': 7.678787878787878e-06, 'num_tokens': 1692065.0, 'completions/mean_length': 72.625, 'completions/min_length': 67.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.3632194995880127, 'rewards/meter/std': 0.32341283559799194, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.36088046431541443, 'rewards/total_composite/std': 0.3263150453567505, 'reward': 0.36088046431541443, 'reward_std': 0.3263150453567505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07868395745754242, 'sampling/sampling_logp_difference/max': 5.154880046844482, 'sampling/importance_sampling_ratio/min': 0.005771172232925892, 'sampling/importance_sampling_ratio/mean': 0.9972518086433411, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46072521805763245, 'clip_ratio/low_mean': 0.03220835281535983, 'clip_ratio/low_min': 0.03220835281535983, 'clip_ratio/high_mean': 0.026226354064419866, 'clip_ratio/high_max': 0.026226354064419866, 'clip_ratio/region_mean': 0.058434706879779696, 'reward_total_mean': 0.36088046431541443, 'reward_meter_mean': 0.3632194995880127, 'reward_meter_std': 0.32341283559799194, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.36088046431541443, 'reward_total_composite_std': 0.3263150453567505, 'epoch': 0.03}

 23%|██▎       | 767/3300 [16:35<3:49:34,  5.44s/it]INFO 04-11 22:53:26 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 768/3300 [16:40<3:45:50,  5.35s/it]2026-04-11 22:53:31,167 | INFO | train_grpo_train | metrics_logged mode=train step=768

                                                    
{'loss': -0.0095, 'grad_norm': 2.0709922313690186, 'learning_rate': 7.675757575757577e-06, 'num_tokens': 1694344.0, 'completions/mean_length': 107.875, 'completions/min_length': 104.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.875, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.7045435905456543, 'rewards/meter/std': 0.13543730974197388, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5636348724365234, 'rewards/total_composite/std': 0.10834985971450806, 'reward': 0.5636348724365234, 'reward_std': 0.10834983736276627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010335095226764679, 'sampling/sampling_logp_difference/max': 0.9596023559570312, 'sampling/importance_sampling_ratio/min': 0.38304516673088074, 'sampling/importance_sampling_ratio/mean': 1.0031026601791382, 'sampling/importance_sampling_ratio/max': 1.4763460159301758, 'entropy': 0.09003173373639584, 'clip_ratio/low_mean': 0.0059091257862746716, 'clip_ratio/low_min': 0.0059091257862746716, 'clip_ratio/high_mean': 0.0022935778833925724, 'clip_ratio/high_max': 0.0022935778833925724, 'clip_ratio/region_mean': 0.008202703669667244, 'reward_total_mean': 0.5636348724365234, 'reward_meter_mean': 0.7045435905456543, 'reward_meter_std': 0.13543730974197388, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5636348724365234, 'reward_total_composite_std': 0.10834985971450806, 'epoch': 0.03}

 23%|██▎       | 768/3300 [16:40<3:45:50,  5.35s/it]INFO 04-11 22:53:31 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 769/3300 [16:46<3:53:02,  5.52s/it]2026-04-11 22:53:37,094 | INFO | train_grpo_train | metrics_logged mode=train step=769

                                                    
{'loss': 0.0714, 'grad_norm': 7.707916736602783, 'learning_rate': 7.672727272727273e-06, 'num_tokens': 1696973.0, 'completions/mean_length': 143.625, 'completions/min_length': 126.0, 'completions/max_length': 172.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.625, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 172.0, 'rewards/meter/mean': 0.7441333532333374, 'rewards/meter/std': 0.4554755687713623, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6174242496490479, 'rewards/repeat_penalty/std': 0.1036364957690239, 'rewards/total_composite/mean': 0.4105571210384369, 'rewards/total_composite/std': 0.2803230583667755, 'reward': 0.4105571210384369, 'reward_std': 0.2803230583667755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03509281575679779, 'sampling/sampling_logp_difference/max': 4.862173557281494, 'sampling/importance_sampling_ratio/min': 0.0077336556278169155, 'sampling/importance_sampling_ratio/mean': 0.9966462850570679, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0793102509342134, 'clip_ratio/low_mean': 0.009834110038354993, 'clip_ratio/low_min': 0.009834110038354993, 'clip_ratio/high_mean': 0.0102224723668769, 'clip_ratio/high_max': 0.0102224723668769, 'clip_ratio/region_mean': 0.020056582405231893, 'reward_total_mean': 0.4105571210384369, 'reward_meter_mean': 0.7441333532333374, 'reward_meter_std': 0.4554755687713623, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6174242496490479, 'reward_repeat_penalty_std': 0.1036364957690239, 'reward_total_composite_mean': 0.4105571210384369, 'reward_total_composite_std': 0.2803230583667755, 'epoch': 0.03}

 23%|██▎       | 769/3300 [16:46<3:53:02,  5.52s/it]INFO 04-11 22:53:37 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 770/3300 [16:56<4:53:56,  6.97s/it]2026-04-11 22:53:47,440 | INFO | train_grpo_train | metrics_logged mode=train step=770

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.66969696969697e-06, 'num_tokens': 1698965.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.8939934968948364, 'rewards/meter/std': 0.13721102476119995, 'rewards/count_adherence/mean': 0.7828947305679321, 'rewards/count_adherence/std': 0.03373000770807266, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5402884483337402, 'rewards/repeat_penalty/std': 0.2157072126865387, 'rewards/total_composite/mean': 0.3744324743747711, 'rewards/total_composite/std': 0.17441345751285553, 'reward': 0.3744324743747711, 'reward_std': 0.17441345751285553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.3744324743747711, 'reward_meter_mean': 0.8939934968948364, 'reward_meter_std': 0.13721102476119995, 'reward_count_adherence_mean': 0.7828947305679321, 'reward_count_adherence_std': 0.03373000770807266, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5402884483337402, 'reward_repeat_penalty_std': 0.2157072126865387, 'reward_total_composite_mean': 0.3744324743747711, 'reward_total_composite_std': 0.17441345751285553, 'epoch': 0.03}

 23%|██▎       | 770/3300 [16:56<4:53:56,  6.97s/it]INFO 04-11 22:53:47 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 771/3300 [17:03<4:54:42,  6.99s/it]2026-04-11 22:53:54,481 | INFO | train_grpo_train | metrics_logged mode=train step=771

                                                    
{'loss': 0.0526, 'grad_norm': 4.9908342361450195, 'learning_rate': 7.666666666666667e-06, 'num_tokens': 1702409.0, 'completions/mean_length': 230.5, 'completions/min_length': 180.0, 'completions/max_length': 263.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 230.5, 'completions/min_terminated_length': 180.0, 'completions/max_terminated_length': 263.0, 'rewards/meter/mean': 0.7852883338928223, 'rewards/meter/std': 0.3382474482059479, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5688130855560303, 'rewards/repeat_penalty/std': 0.20688475668430328, 'rewards/total_composite/mean': 0.3491661250591278, 'rewards/total_composite/std': 0.22600221633911133, 'reward': 0.3491661250591278, 'reward_std': 0.22600221633911133, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018919790163636208, 'sampling/sampling_logp_difference/max': 1.5345475673675537, 'sampling/importance_sampling_ratio/min': 0.21555320918560028, 'sampling/importance_sampling_ratio/mean': 1.0030508041381836, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13555688876658678, 'clip_ratio/low_mean': 0.004294316866435111, 'clip_ratio/low_min': 0.004294316866435111, 'clip_ratio/high_mean': 0.010813763190526515, 'clip_ratio/high_max': 0.010813763190526515, 'clip_ratio/region_mean': 0.015108080056961626, 'reward_total_mean': 0.3491661250591278, 'reward_meter_mean': 0.7852883338928223, 'reward_meter_std': 0.3382474482059479, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5688130855560303, 'reward_repeat_penalty_std': 0.20688475668430328, 'reward_total_composite_mean': 0.3491661250591278, 'reward_total_composite_std': 0.22600221633911133, 'epoch': 0.03}

 23%|██▎       | 771/3300 [17:03<4:54:42,  6.99s/it]INFO 04-11 22:53:54 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 772/3300 [17:09<4:38:29,  6.61s/it]2026-04-11 22:54:00,203 | INFO | train_grpo_train | metrics_logged mode=train step=772

                                                    
{'loss': -0.1232, 'grad_norm': 5.677700042724609, 'learning_rate': 7.663636363636364e-06, 'num_tokens': 1704516.0, 'completions/mean_length': 99.375, 'completions/min_length': 90.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.375, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9852504730224609, 'rewards/meter/std': 0.008375532925128937, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7785714864730835, 'rewards/repeat_penalty/std': 0.039677999913692474, 'rewards/total_composite/mean': 0.6195021271705627, 'rewards/total_composite/std': 0.05528084933757782, 'reward': 0.6195021271705627, 'reward_std': 0.05528085306286812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020249057561159134, 'sampling/sampling_logp_difference/max': 1.5644612312316895, 'sampling/importance_sampling_ratio/min': 0.2092006951570511, 'sampling/importance_sampling_ratio/mean': 1.0006357431411743, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08717024885118008, 'clip_ratio/low_mean': 0.015005706925876439, 'clip_ratio/low_min': 0.015005706925876439, 'clip_ratio/high_mean': 0.006097560748457909, 'clip_ratio/high_max': 0.006097560748457909, 'clip_ratio/region_mean': 0.021103267674334347, 'reward_total_mean': 0.6195021271705627, 'reward_meter_mean': 0.9852504730224609, 'reward_meter_std': 0.008375532925128937, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7785714864730835, 'reward_repeat_penalty_std': 0.039677999913692474, 'reward_total_composite_mean': 0.6195021271705627, 'reward_total_composite_std': 0.05528084933757782, 'epoch': 0.03}

 23%|██▎       | 772/3300 [17:09<4:38:29,  6.61s/it]INFO 04-11 22:54:00 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 773/3300 [17:18<5:18:47,  7.57s/it]2026-04-11 22:54:10,006 | INFO | train_grpo_train | metrics_logged mode=train step=773

                                                    
{'loss': 0.0245, 'grad_norm': 5.519010066986084, 'learning_rate': 7.660606060606062e-06, 'num_tokens': 1709201.0, 'completions/mean_length': 409.625, 'completions/min_length': 346.0, 'completions/max_length': 460.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 409.625, 'completions/min_terminated_length': 346.0, 'completions/max_terminated_length': 460.0, 'rewards/meter/mean': 0.9907213449478149, 'rewards/meter/std': 0.006963523104786873, 'rewards/count_adherence/mean': 0.3571428656578064, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.45148104429244995, 'rewards/repeat_penalty/std': 0.23929926753044128, 'rewards/total_composite/mean': 0.17074038088321686, 'rewards/total_composite/std': 0.10653632879257202, 'reward': 0.17074038088321686, 'reward_std': 0.10653632134199142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02135242149233818, 'sampling/sampling_logp_difference/max': 5.103124141693115, 'sampling/importance_sampling_ratio/min': 0.006077729165554047, 'sampling/importance_sampling_ratio/mean': 0.9987974166870117, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07645870675332844, 'clip_ratio/low_mean': 0.003349777136463672, 'clip_ratio/low_min': 0.003349777136463672, 'clip_ratio/high_mean': 0.008041649358347058, 'clip_ratio/high_max': 0.008041649358347058, 'clip_ratio/region_mean': 0.01139142649481073, 'reward_total_mean': 0.17074038088321686, 'reward_meter_mean': 0.9907213449478149, 'reward_meter_std': 0.006963523104786873, 'reward_count_adherence_mean': 0.3571428656578064, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.45148104429244995, 'reward_repeat_penalty_std': 0.23929926753044128, 'reward_total_composite_mean': 0.17074038088321686, 'reward_total_composite_std': 0.10653632879257202, 'epoch': 0.03}

 23%|██▎       | 773/3300 [17:19<5:18:47,  7.57s/it]INFO 04-11 22:54:10 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 774/3300 [17:23<4:43:23,  6.73s/it]2026-04-11 22:54:14,783 | INFO | train_grpo_train | metrics_logged mode=train step=774

                                                    
{'loss': -0.0134, 'grad_norm': 6.2788872718811035, 'learning_rate': 7.657575757575757e-06, 'num_tokens': 1711058.0, 'completions/mean_length': 69.125, 'completions/min_length': 65.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.7258920669555664, 'rewards/meter/std': 0.40674278140068054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7258920669555664, 'rewards/total_composite/std': 0.40674278140068054, 'reward': 0.7258920669555664, 'reward_std': 0.40674278140068054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05154965817928314, 'sampling/sampling_logp_difference/max': 2.2386436462402344, 'sampling/importance_sampling_ratio/min': 0.1066029891371727, 'sampling/importance_sampling_ratio/mean': 1.0082036256790161, 'sampling/importance_sampling_ratio/max': 1.882003903388977, 'entropy': 0.37627044692635536, 'clip_ratio/low_mean': 0.005654420121572912, 'clip_ratio/low_min': 0.005654420121572912, 'clip_ratio/high_mean': 0.022313665016554296, 'clip_ratio/high_max': 0.022313665016554296, 'clip_ratio/region_mean': 0.027968085138127208, 'reward_total_mean': 0.7258920669555664, 'reward_meter_mean': 0.7258920669555664, 'reward_meter_std': 0.40674278140068054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7258920669555664, 'reward_total_composite_std': 0.40674278140068054, 'epoch': 0.03}

 23%|██▎       | 774/3300 [17:23<4:43:23,  6.73s/it]INFO 04-11 22:54:15 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 775/3300 [17:29<4:34:19,  6.52s/it]2026-04-11 22:54:20,807 | INFO | train_grpo_train | metrics_logged mode=train step=775

                                                    
{'loss': 0.0636, 'grad_norm': 2.807114601135254, 'learning_rate': 7.654545454545456e-06, 'num_tokens': 1713733.0, 'completions/mean_length': 168.375, 'completions/min_length': 147.0, 'completions/max_length': 188.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.375, 'completions/min_terminated_length': 147.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.9743660688400269, 'rewards/meter/std': 0.04519447684288025, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6060605645179749, 'rewards/repeat_penalty/std': 0.13551926612854004, 'rewards/total_composite/mean': 0.5430476665496826, 'rewards/total_composite/std': 0.16461656987667084, 'reward': 0.5430476665496826, 'reward_std': 0.16461656987667084, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026703810319304466, 'sampling/sampling_logp_difference/max': 12.626859664916992, 'sampling/importance_sampling_ratio/min': 3.2826496862981003e-06, 'sampling/importance_sampling_ratio/mean': 0.9989210367202759, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06733021000400186, 'clip_ratio/low_mean': 0.00418766331858933, 'clip_ratio/low_min': 0.00418766331858933, 'clip_ratio/high_mean': 0.007139897206798196, 'clip_ratio/high_max': 0.007139897206798196, 'clip_ratio/region_mean': 0.011327560525387526, 'reward_total_mean': 0.5430476665496826, 'reward_meter_mean': 0.9743660688400269, 'reward_meter_std': 0.04519447684288025, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6060605645179749, 'reward_repeat_penalty_std': 0.13551926612854004, 'reward_total_composite_mean': 0.5430476665496826, 'reward_total_composite_std': 0.16461656987667084, 'epoch': 0.03}

 23%|██▎       | 775/3300 [17:29<4:34:19,  6.52s/it]INFO 04-11 22:54:21 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 776/3300 [17:34<4:09:03,  5.92s/it]2026-04-11 22:54:25,331 | INFO | train_grpo_train | metrics_logged mode=train step=776

                                                    
{'loss': 0.0012, 'grad_norm': 7.4227614402771, 'learning_rate': 7.651515151515152e-06, 'num_tokens': 1715409.0, 'completions/mean_length': 54.5, 'completions/min_length': 53.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.5, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9904471635818481, 'rewards/meter/std': 0.002613567281514406, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9904471635818481, 'rewards/total_composite/std': 0.002613567281514406, 'reward': 0.9904471635818481, 'reward_std': 0.002613575430586934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024185476824641228, 'sampling/sampling_logp_difference/max': 1.0824594497680664, 'sampling/importance_sampling_ratio/min': 0.3387613296508789, 'sampling/importance_sampling_ratio/mean': 0.9967302083969116, 'sampling/importance_sampling_ratio/max': 1.4630638360977173, 'entropy': 0.09682174911722541, 'clip_ratio/low_mean': 0.009263938991352916, 'clip_ratio/low_min': 0.009263938991352916, 'clip_ratio/high_mean': 0.009050324792042375, 'clip_ratio/high_max': 0.009050324792042375, 'clip_ratio/region_mean': 0.01831426378339529, 'reward_total_mean': 0.9904471635818481, 'reward_meter_mean': 0.9904471635818481, 'reward_meter_std': 0.002613567281514406, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9904471635818481, 'reward_total_composite_std': 0.002613567281514406, 'epoch': 0.03}

 24%|██▎       | 776/3300 [17:34<4:09:03,  5.92s/it]INFO 04-11 22:54:25 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 777/3300 [17:38<3:48:58,  5.45s/it]2026-04-11 22:54:29,668 | INFO | train_grpo_train | metrics_logged mode=train step=777

                                                    
{'loss': 0.0046, 'grad_norm': 8.215063095092773, 'learning_rate': 7.648484848484849e-06, 'num_tokens': 1716925.0, 'completions/mean_length': 36.5, 'completions/min_length': 36.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.978127121925354, 'rewards/meter/std': 0.0063257296569645405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.978127121925354, 'rewards/total_composite/std': 0.0063257296569645405, 'reward': 0.978127121925354, 'reward_std': 0.006325736176222563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03158137574791908, 'sampling/sampling_logp_difference/max': 1.4631445407867432, 'sampling/importance_sampling_ratio/min': 0.23150713741779327, 'sampling/importance_sampling_ratio/mean': 1.0036948919296265, 'sampling/importance_sampling_ratio/max': 1.6868826150894165, 'entropy': 0.11897748988121748, 'clip_ratio/low_mean': 0.013795045437291265, 'clip_ratio/low_min': 0.013795045437291265, 'clip_ratio/high_mean': 0.013795045204460621, 'clip_ratio/high_max': 0.013795045204460621, 'clip_ratio/region_mean': 0.027590090641751885, 'reward_total_mean': 0.978127121925354, 'reward_meter_mean': 0.978127121925354, 'reward_meter_std': 0.0063257296569645405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.978127121925354, 'reward_total_composite_std': 0.0063257296569645405, 'epoch': 0.03}

 24%|██▎       | 777/3300 [17:38<3:48:58,  5.45s/it]INFO 04-11 22:54:29 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 778/3300 [17:43<3:39:56,  5.23s/it]2026-04-11 22:54:34,404 | INFO | train_grpo_train | metrics_logged mode=train step=778

                                                    
{'loss': 0.014, 'grad_norm': 5.06115198135376, 'learning_rate': 7.645454545454546e-06, 'num_tokens': 1718951.0, 'completions/mean_length': 72.25, 'completions/min_length': 68.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9873093366622925, 'rewards/meter/std': 0.01036460418254137, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9873093366622925, 'rewards/total_composite/std': 0.01036460418254137, 'reward': 0.9873093366622925, 'reward_std': 0.010364595800638199, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03534568101167679, 'sampling/sampling_logp_difference/max': 0.8904938697814941, 'sampling/importance_sampling_ratio/min': 0.4104529917240143, 'sampling/importance_sampling_ratio/mean': 1.0149582624435425, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.25925541296601295, 'clip_ratio/low_mean': 0.02250340231694281, 'clip_ratio/low_min': 0.02250340231694281, 'clip_ratio/high_mean': 0.020928236190229654, 'clip_ratio/high_max': 0.020928236190229654, 'clip_ratio/region_mean': 0.043431638507172465, 'reward_total_mean': 0.9873093366622925, 'reward_meter_mean': 0.9873093366622925, 'reward_meter_std': 0.01036460418254137, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9873093366622925, 'reward_total_composite_std': 0.01036460418254137, 'epoch': 0.03}

 24%|██▎       | 778/3300 [17:43<3:39:56,  5.23s/it]INFO 04-11 22:54:34 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 779/3300 [17:54<4:53:37,  6.99s/it]2026-04-11 22:54:45,494 | INFO | train_grpo_train | metrics_logged mode=train step=779

                                                    
{'loss': -0.1238, 'grad_norm': 1.4244983196258545, 'learning_rate': 7.642424242424244e-06, 'num_tokens': 1720483.0, 'completions/mean_length': 182.5, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 72.66667175292969, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.6190794110298157, 'rewards/meter/std': 0.3674232065677643, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5616785287857056, 'rewards/total_composite/std': 0.43257132172584534, 'reward': 0.5616785287857056, 'reward_std': 0.43257129192352295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07225891947746277, 'sampling/sampling_logp_difference/max': 1.2201738357543945, 'sampling/importance_sampling_ratio/min': 0.2951788604259491, 'sampling/importance_sampling_ratio/mean': 1.0154547691345215, 'sampling/importance_sampling_ratio/max': 1.9660524129867554, 'entropy': 0.38310878723859787, 'clip_ratio/low_mean': 0.008561643771827221, 'clip_ratio/low_min': 0.008561643771827221, 'clip_ratio/high_mean': 0.030813875840976834, 'clip_ratio/high_max': 0.030813875840976834, 'clip_ratio/region_mean': 0.039375519612804055, 'reward_total_mean': 0.5616785287857056, 'reward_meter_mean': 0.6190794110298157, 'reward_meter_std': 0.3674232065677643, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5616785287857056, 'reward_total_composite_std': 0.43257132172584534, 'epoch': 0.03}

 24%|██▎       | 779/3300 [17:54<4:53:37,  6.99s/it]INFO 04-11 22:54:45 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 780/3300 [18:05<5:38:55,  8.07s/it]2026-04-11 22:54:56,084 | INFO | train_grpo_train | metrics_logged mode=train step=780

                                                    
{'loss': -0.0498, 'grad_norm': 3.119354009628296, 'learning_rate': 7.639393939393939e-06, 'num_tokens': 1722862.0, 'completions/mean_length': 174.375, 'completions/min_length': 111.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 126.14286041259766, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.6168656349182129, 'rewards/meter/std': 0.4164867103099823, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.7492559552192688, 'rewards/repeat_penalty/std': 0.14768067002296448, 'rewards/total_composite/mean': 0.4575069546699524, 'rewards/total_composite/std': 0.3478102684020996, 'reward': 0.4575069546699524, 'reward_std': 0.3478102684020996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03274521231651306, 'sampling/sampling_logp_difference/max': 1.078242540359497, 'sampling/importance_sampling_ratio/min': 0.34019285440444946, 'sampling/importance_sampling_ratio/mean': 1.0024807453155518, 'sampling/importance_sampling_ratio/max': 1.7341415882110596, 'entropy': 0.24183030799031258, 'clip_ratio/low_mean': 0.004807692486792803, 'clip_ratio/low_min': 0.004807692486792803, 'clip_ratio/high_mean': 0.029140884289518, 'clip_ratio/high_max': 0.029140884289518, 'clip_ratio/region_mean': 0.0339485767763108, 'reward_total_mean': 0.4575069546699524, 'reward_meter_mean': 0.6168656349182129, 'reward_meter_std': 0.4164867103099823, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.7492559552192688, 'reward_repeat_penalty_std': 0.14768067002296448, 'reward_total_composite_mean': 0.4575069546699524, 'reward_total_composite_std': 0.3478102684020996, 'epoch': 0.03}

 24%|██▎       | 780/3300 [18:05<5:38:55,  8.07s/it]INFO 04-11 22:54:56 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 781/3300 [18:10<5:00:26,  7.16s/it]2026-04-11 22:55:01,108 | INFO | train_grpo_train | metrics_logged mode=train step=781

                                                    
{'loss': -0.0116, 'grad_norm': 12.43602180480957, 'learning_rate': 7.636363636363638e-06, 'num_tokens': 1724537.0, 'completions/mean_length': 47.375, 'completions/min_length': 44.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5619980096817017, 'rewards/meter/std': 0.3257886469364166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.4856577217578888, 'rewards/total_composite/std': 0.3285104036331177, 'reward': 0.4856577217578888, 'reward_std': 0.3285104036331177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09212920814752579, 'sampling/sampling_logp_difference/max': 1.6957062482833862, 'sampling/importance_sampling_ratio/min': 0.18346960842609406, 'sampling/importance_sampling_ratio/mean': 0.9970747232437134, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9118200056254864, 'clip_ratio/low_mean': 0.04398810095153749, 'clip_ratio/low_min': 0.04398810095153749, 'clip_ratio/high_mean': 0.016590908635407686, 'clip_ratio/high_max': 0.016590908635407686, 'clip_ratio/region_mean': 0.060579009586945176, 'reward_total_mean': 0.4856577217578888, 'reward_meter_mean': 0.5619980096817017, 'reward_meter_std': 0.3257886469364166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.4856577217578888, 'reward_total_composite_std': 0.3285104036331177, 'epoch': 0.03}

 24%|██▎       | 781/3300 [18:10<5:00:26,  7.16s/it]INFO 04-11 22:55:01 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 782/3300 [18:19<5:32:40,  7.93s/it]2026-04-11 22:55:10,832 | INFO | train_grpo_train | metrics_logged mode=train step=782

                                                    
{'loss': -0.1573, 'grad_norm': 1.6532626152038574, 'learning_rate': 7.633333333333334e-06, 'num_tokens': 1726623.0, 'completions/mean_length': 166.75, 'completions/min_length': 111.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 117.42857360839844, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.6895759105682373, 'rewards/meter/std': 0.26801931858062744, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.824999988079071, 'rewards/repeat_penalty/std': 0.19820624589920044, 'rewards/total_composite/mean': 0.5303109288215637, 'rewards/total_composite/std': 0.28132033348083496, 'reward': 0.5303109288215637, 'reward_std': 0.28132033348083496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03518048673868179, 'sampling/sampling_logp_difference/max': 1.1620054244995117, 'sampling/importance_sampling_ratio/min': 0.31285813450813293, 'sampling/importance_sampling_ratio/mean': 1.006204605102539, 'sampling/importance_sampling_ratio/max': 1.9478799104690552, 'entropy': 0.2663711039349437, 'clip_ratio/low_mean': 0.008460594224743545, 'clip_ratio/low_min': 0.008460594224743545, 'clip_ratio/high_mean': 0.013982121949084103, 'clip_ratio/high_max': 0.013982121949084103, 'clip_ratio/region_mean': 0.022442716173827648, 'reward_total_mean': 0.5303109288215637, 'reward_meter_mean': 0.6895759105682373, 'reward_meter_std': 0.26801931858062744, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.824999988079071, 'reward_repeat_penalty_std': 0.19820624589920044, 'reward_total_composite_mean': 0.5303109288215637, 'reward_total_composite_std': 0.28132033348083496, 'epoch': 0.03}

 24%|██▎       | 782/3300 [18:19<5:32:40,  7.93s/it]INFO 04-11 22:55:11 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 783/3300 [18:24<4:54:59,  7.03s/it]2026-04-11 22:55:15,776 | INFO | train_grpo_train | metrics_logged mode=train step=783

                                                    
{'loss': 0.0077, 'grad_norm': 4.994919776916504, 'learning_rate': 7.630303030303031e-06, 'num_tokens': 1728611.0, 'completions/mean_length': 79.5, 'completions/min_length': 76.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.6204426288604736, 'rewards/meter/std': 0.3357177972793579, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6204426288604736, 'rewards/total_composite/std': 0.3357177972793579, 'reward': 0.6204426288604736, 'reward_std': 0.3357177972793579, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04477819800376892, 'sampling/sampling_logp_difference/max': 1.0638208389282227, 'sampling/importance_sampling_ratio/min': 0.34513458609580994, 'sampling/importance_sampling_ratio/mean': 1.011016845703125, 'sampling/importance_sampling_ratio/max': 1.7397865056991577, 'entropy': 0.3720816671848297, 'clip_ratio/low_mean': 0.017455301131121814, 'clip_ratio/low_min': 0.017455301131121814, 'clip_ratio/high_mean': 0.013986280770041049, 'clip_ratio/high_max': 0.013986280770041049, 'clip_ratio/region_mean': 0.03144158190116286, 'reward_total_mean': 0.6204426288604736, 'reward_meter_mean': 0.6204426288604736, 'reward_meter_std': 0.3357177972793579, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6204426288604736, 'reward_total_composite_std': 0.3357177972793579, 'epoch': 0.03}

 24%|██▎       | 783/3300 [18:24<4:54:59,  7.03s/it]INFO 04-11 22:55:16 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 784/3300 [18:29<4:24:51,  6.32s/it]2026-04-11 22:55:20,422 | INFO | train_grpo_train | metrics_logged mode=train step=784

                                                    
{'loss': 0.0052, 'grad_norm': 7.485208988189697, 'learning_rate': 7.627272727272727e-06, 'num_tokens': 1730350.0, 'completions/mean_length': 61.375, 'completions/min_length': 56.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5055776834487915, 'rewards/meter/std': 0.32300832867622375, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.17817415297031403, 'rewards/total_composite/mean': 0.453016996383667, 'rewards/total_composite/std': 0.33306846022605896, 'reward': 0.453016996383667, 'reward_std': 0.33306846022605896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047304026782512665, 'sampling/sampling_logp_difference/max': 1.7305419445037842, 'sampling/importance_sampling_ratio/min': 0.17718835175037384, 'sampling/importance_sampling_ratio/mean': 0.9948754906654358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2184693105518818, 'clip_ratio/low_mean': 0.018428327050060034, 'clip_ratio/low_min': 0.018428327050060034, 'clip_ratio/high_mean': 0.012099922401830554, 'clip_ratio/high_max': 0.012099922401830554, 'clip_ratio/region_mean': 0.030528249451890588, 'reward_total_mean': 0.453016996383667, 'reward_meter_mean': 0.5055776834487915, 'reward_meter_std': 0.32300832867622375, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.17817415297031403, 'reward_total_composite_mean': 0.453016996383667, 'reward_total_composite_std': 0.33306846022605896, 'epoch': 0.03}

 24%|██▍       | 784/3300 [18:29<4:24:51,  6.32s/it]INFO 04-11 22:55:20 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 785/3300 [18:34<4:12:09,  6.02s/it]2026-04-11 22:55:25,740 | INFO | train_grpo_train | metrics_logged mode=train step=785

                                                    
{'loss': 0.0036, 'grad_norm': 6.059953212738037, 'learning_rate': 7.6242424242424254e-06, 'num_tokens': 1732524.0, 'completions/mean_length': 104.75, 'completions/min_length': 99.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.75, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.49319469928741455, 'rewards/meter/std': 0.2711309790611267, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.46458256244659424, 'rewards/total_composite/std': 0.2836271822452545, 'reward': 0.46458256244659424, 'reward_std': 0.2836271822452545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042036473751068115, 'sampling/sampling_logp_difference/max': 1.2848865985870361, 'sampling/importance_sampling_ratio/min': 0.27668195962905884, 'sampling/importance_sampling_ratio/mean': 1.0086406469345093, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31474705785512924, 'clip_ratio/low_mean': 0.010869022691622376, 'clip_ratio/low_min': 0.010869022691622376, 'clip_ratio/high_mean': 0.01767290150746703, 'clip_ratio/high_max': 0.01767290150746703, 'clip_ratio/region_mean': 0.028541924199089408, 'reward_total_mean': 0.46458256244659424, 'reward_meter_mean': 0.49319469928741455, 'reward_meter_std': 0.2711309790611267, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.46458256244659424, 'reward_total_composite_std': 0.2836271822452545, 'epoch': 0.03}

 24%|██▍       | 785/3300 [18:34<4:12:09,  6.02s/it]INFO 04-11 22:55:25 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 786/3300 [18:43<4:46:23,  6.84s/it]2026-04-11 22:55:34,483 | INFO | train_grpo_train | metrics_logged mode=train step=786

                                                    
{'loss': -0.0567, 'grad_norm': 2.024092435836792, 'learning_rate': 7.621212121212122e-06, 'num_tokens': 1737190.0, 'completions/mean_length': 368.25, 'completions/min_length': 322.0, 'completions/max_length': 406.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 368.25, 'completions/min_terminated_length': 322.0, 'completions/max_terminated_length': 406.0, 'rewards/meter/mean': 0.7014279365539551, 'rewards/meter/std': 0.43751856684684753, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5806276798248291, 'rewards/repeat_penalty/std': 0.06668182462453842, 'rewards/total_composite/mean': 0.3579794764518738, 'rewards/total_composite/std': 0.22696760296821594, 'reward': 0.3579794764518738, 'reward_std': 0.22696760296821594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012620110996067524, 'sampling/sampling_logp_difference/max': 1.491776466369629, 'sampling/importance_sampling_ratio/min': 0.2249726504087448, 'sampling/importance_sampling_ratio/mean': 1.0009483098983765, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07080868305638433, 'clip_ratio/low_mean': 0.004607091657817364, 'clip_ratio/low_min': 0.004607091657817364, 'clip_ratio/high_mean': 0.0042421949619892985, 'clip_ratio/high_max': 0.0042421949619892985, 'clip_ratio/region_mean': 0.008849286619806662, 'reward_total_mean': 0.3579794764518738, 'reward_meter_mean': 0.7014279365539551, 'reward_meter_std': 0.43751856684684753, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5806276798248291, 'reward_repeat_penalty_std': 0.06668182462453842, 'reward_total_composite_mean': 0.3579794764518738, 'reward_total_composite_std': 0.22696760296821594, 'epoch': 0.03}

 24%|██▍       | 786/3300 [18:43<4:46:23,  6.84s/it]INFO 04-11 22:55:34 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 787/3300 [18:49<4:33:32,  6.53s/it]2026-04-11 22:55:40,304 | INFO | train_grpo_train | metrics_logged mode=train step=787

                                                    
{'loss': 0.0123, 'grad_norm': 3.6839067935943604, 'learning_rate': 7.618181818181819e-06, 'num_tokens': 1739852.0, 'completions/mean_length': 153.75, 'completions/min_length': 135.0, 'completions/max_length': 162.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 153.75, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.7929247617721558, 'rewards/meter/std': 0.10911522805690765, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6818181872367859, 'rewards/repeat_penalty/std': 0.06872082501649857, 'rewards/total_composite/mean': 0.435272216796875, 'rewards/total_composite/std': 0.09194310009479523, 'reward': 0.435272216796875, 'reward_std': 0.09194309264421463, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028897186741232872, 'sampling/sampling_logp_difference/max': 1.9267759323120117, 'sampling/importance_sampling_ratio/min': 0.14561693370342255, 'sampling/importance_sampling_ratio/mean': 0.9994563460350037, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11316746287047863, 'clip_ratio/low_mean': 0.01558087719604373, 'clip_ratio/low_min': 0.01558087719604373, 'clip_ratio/high_mean': 0.007280809339135885, 'clip_ratio/high_max': 0.007280809339135885, 'clip_ratio/region_mean': 0.022861686535179615, 'reward_total_mean': 0.435272216796875, 'reward_meter_mean': 0.7929247617721558, 'reward_meter_std': 0.10911522805690765, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6818181872367859, 'reward_repeat_penalty_std': 0.06872082501649857, 'reward_total_composite_mean': 0.435272216796875, 'reward_total_composite_std': 0.09194310009479523, 'epoch': 0.03}

 24%|██▍       | 787/3300 [18:49<4:33:32,  6.53s/it]INFO 04-11 22:55:40 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 788/3300 [18:53<4:09:35,  5.96s/it]2026-04-11 22:55:44,938 | INFO | train_grpo_train | metrics_logged mode=train step=788

                                                    
{'loss': -0.0051, 'grad_norm': 7.856208324432373, 'learning_rate': 7.6151515151515155e-06, 'num_tokens': 1741588.0, 'completions/mean_length': 58.0, 'completions/min_length': 57.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6469426155090332, 'rewards/meter/std': 0.24868278205394745, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6469426155090332, 'rewards/total_composite/std': 0.24868278205394745, 'reward': 0.6469426155090332, 'reward_std': 0.24868276715278625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022744039073586464, 'sampling/sampling_logp_difference/max': 0.6672461032867432, 'sampling/importance_sampling_ratio/min': 0.5131197571754456, 'sampling/importance_sampling_ratio/mean': 1.0034902095794678, 'sampling/importance_sampling_ratio/max': 1.8282591104507446, 'entropy': 0.16343743726611137, 'clip_ratio/low_mean': 0.012860541231930256, 'clip_ratio/low_min': 0.012860541231930256, 'clip_ratio/high_mean': 0.01074189692735672, 'clip_ratio/high_max': 0.01074189692735672, 'clip_ratio/region_mean': 0.023602438159286976, 'reward_total_mean': 0.6469426155090332, 'reward_meter_mean': 0.6469426155090332, 'reward_meter_std': 0.24868278205394745, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6469426155090332, 'reward_total_composite_std': 0.24868278205394745, 'epoch': 0.03}

 24%|██▍       | 788/3300 [18:53<4:09:35,  5.96s/it]INFO 04-11 22:55:45 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 789/3300 [18:58<3:48:37,  5.46s/it]2026-04-11 22:55:49,240 | INFO | train_grpo_train | metrics_logged mode=train step=789

                                                    
{'loss': 0.0131, 'grad_norm': 13.922452926635742, 'learning_rate': 7.612121212121213e-06, 'num_tokens': 1743054.0, 'completions/mean_length': 31.25, 'completions/min_length': 30.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9813005924224854, 'rewards/meter/std': 0.016189413145184517, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9813005924224854, 'rewards/total_composite/std': 0.016189413145184517, 'reward': 0.9813005924224854, 'reward_std': 0.016189415007829666, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033597491681575775, 'sampling/sampling_logp_difference/max': 1.0770900249481201, 'sampling/importance_sampling_ratio/min': 0.3405851721763611, 'sampling/importance_sampling_ratio/mean': 0.9976152181625366, 'sampling/importance_sampling_ratio/max': 1.8372166156768799, 'entropy': 0.17746192403137684, 'clip_ratio/low_mean': 0.008072916883975267, 'clip_ratio/low_min': 0.008072916883975267, 'clip_ratio/high_mean': 0.008072916883975267, 'clip_ratio/high_max': 0.008072916883975267, 'clip_ratio/region_mean': 0.016145833767950535, 'reward_total_mean': 0.9813005924224854, 'reward_meter_mean': 0.9813005924224854, 'reward_meter_std': 0.016189413145184517, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9813005924224854, 'reward_total_composite_std': 0.016189413145184517, 'epoch': 0.03}

 24%|██▍       | 789/3300 [18:58<3:48:37,  5.46s/it]INFO 04-11 22:55:49 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 790/3300 [19:03<3:43:20,  5.34s/it]2026-04-11 22:55:54,287 | INFO | train_grpo_train | metrics_logged mode=train step=790

                                                    
{'loss': -0.0039, 'grad_norm': 6.977390766143799, 'learning_rate': 7.609090909090909e-06, 'num_tokens': 1745187.0, 'completions/mean_length': 89.625, 'completions/min_length': 85.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.625, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.8486406207084656, 'rewards/meter/std': 0.2562169134616852, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6789125204086304, 'rewards/total_composite/std': 0.20497353374958038, 'reward': 0.6789125204086304, 'reward_std': 0.20497353374958038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021279064938426018, 'sampling/sampling_logp_difference/max': 0.814723014831543, 'sampling/importance_sampling_ratio/min': 0.44276192784309387, 'sampling/importance_sampling_ratio/mean': 1.0061084032058716, 'sampling/importance_sampling_ratio/max': 1.8353419303894043, 'entropy': 0.11805877834558487, 'clip_ratio/low_mean': 0.009868005756288767, 'clip_ratio/low_min': 0.009868005756288767, 'clip_ratio/high_mean': 0.019598963437601924, 'clip_ratio/high_max': 0.019598963437601924, 'clip_ratio/region_mean': 0.02946696919389069, 'reward_total_mean': 0.6789125204086304, 'reward_meter_mean': 0.8486406207084656, 'reward_meter_std': 0.2562169134616852, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6789125204086304, 'reward_total_composite_std': 0.20497353374958038, 'epoch': 0.03}

 24%|██▍       | 790/3300 [19:03<3:43:20,  5.34s/it]INFO 04-11 22:55:54 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 791/3300 [19:10<4:03:38,  5.83s/it]2026-04-11 22:56:01,250 | INFO | train_grpo_train | metrics_logged mode=train step=791

                                                    
{'loss': -0.0395, 'grad_norm': 1.8626375198364258, 'learning_rate': 7.606060606060606e-06, 'num_tokens': 1748773.0, 'completions/mean_length': 247.25, 'completions/min_length': 228.0, 'completions/max_length': 258.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 247.25, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 258.0, 'rewards/meter/mean': 0.9959322214126587, 'rewards/meter/std': 0.004204806871712208, 'rewards/count_adherence/mean': 0.910714328289032, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.49047619104385376, 'rewards/repeat_penalty/std': 0.17105023562908173, 'rewards/total_composite/mean': 0.4338645935058594, 'rewards/total_composite/std': 0.12500374019145966, 'reward': 0.4338645935058594, 'reward_std': 0.12500372529029846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014479896053671837, 'sampling/sampling_logp_difference/max': 2.833479881286621, 'sampling/importance_sampling_ratio/min': 0.0588078573346138, 'sampling/importance_sampling_ratio/mean': 1.0004253387451172, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.042171002831310034, 'clip_ratio/low_mean': 0.0032260402804240584, 'clip_ratio/low_min': 0.0032260402804240584, 'clip_ratio/high_mean': 0.00779381615575403, 'clip_ratio/high_max': 0.00779381615575403, 'clip_ratio/region_mean': 0.011019856436178088, 'reward_total_mean': 0.4338645935058594, 'reward_meter_mean': 0.9959322214126587, 'reward_meter_std': 0.004204806871712208, 'reward_count_adherence_mean': 0.910714328289032, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.49047619104385376, 'reward_repeat_penalty_std': 0.17105023562908173, 'reward_total_composite_mean': 0.4338645935058594, 'reward_total_composite_std': 0.12500374019145966, 'epoch': 0.03}

 24%|██▍       | 791/3300 [19:10<4:03:38,  5.83s/it]INFO 04-11 22:56:01 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 792/3300 [19:15<4:02:39,  5.81s/it]2026-04-11 22:56:07,007 | INFO | train_grpo_train | metrics_logged mode=train step=792

                                                    
{'loss': 0.023, 'grad_norm': 3.4738333225250244, 'learning_rate': 7.603030303030303e-06, 'num_tokens': 1751485.0, 'completions/mean_length': 135.0, 'completions/min_length': 124.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.0, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.991753101348877, 'rewards/meter/std': 0.002629074966534972, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6611686944961548, 'rewards/total_composite/std': 0.0017527303425595164, 'reward': 0.6611686944961548, 'reward_std': 0.0017527244053781033, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019971484318375587, 'sampling/sampling_logp_difference/max': 2.0009312629699707, 'sampling/importance_sampling_ratio/min': 0.13520930707454681, 'sampling/importance_sampling_ratio/mean': 1.0068343877792358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0634374669753015, 'clip_ratio/low_mean': 0.010955466306768358, 'clip_ratio/low_min': 0.010955466306768358, 'clip_ratio/high_mean': 0.00658258656039834, 'clip_ratio/high_max': 0.00658258656039834, 'clip_ratio/region_mean': 0.017538052867166698, 'reward_total_mean': 0.6611686944961548, 'reward_meter_mean': 0.991753101348877, 'reward_meter_std': 0.002629074966534972, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6611686944961548, 'reward_total_composite_std': 0.0017527303425595164, 'epoch': 0.03}

 24%|██▍       | 792/3300 [19:16<4:02:39,  5.81s/it]INFO 04-11 22:56:07 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 793/3300 [19:20<3:48:47,  5.48s/it]2026-04-11 22:56:11,714 | INFO | train_grpo_train | metrics_logged mode=train step=793

                                                    
{'loss': 0.0066, 'grad_norm': 5.889811038970947, 'learning_rate': 7.600000000000001e-06, 'num_tokens': 1753253.0, 'completions/mean_length': 56.0, 'completions/min_length': 51.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9871419668197632, 'rewards/meter/std': 0.005153529345989227, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.9050641655921936, 'rewards/total_composite/std': 0.15341757237911224, 'reward': 0.9050641655921936, 'reward_std': 0.15341757237911224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054076679050922394, 'sampling/sampling_logp_difference/max': 2.979457139968872, 'sampling/importance_sampling_ratio/min': 0.05082041397690773, 'sampling/importance_sampling_ratio/mean': 0.9988625645637512, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15502143744379282, 'clip_ratio/low_mean': 0.006658692145720124, 'clip_ratio/low_min': 0.006658692145720124, 'clip_ratio/high_mean': 0.0335467669647187, 'clip_ratio/high_max': 0.0335467669647187, 'clip_ratio/region_mean': 0.040205459110438824, 'reward_total_mean': 0.9050641655921936, 'reward_meter_mean': 0.9871419668197632, 'reward_meter_std': 0.005153529345989227, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.9050641655921936, 'reward_total_composite_std': 0.15341757237911224, 'epoch': 0.03}

 24%|██▍       | 793/3300 [19:20<3:48:47,  5.48s/it]INFO 04-11 22:56:11 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 794/3300 [19:25<3:40:43,  5.28s/it]2026-04-11 22:56:16,556 | INFO | train_grpo_train | metrics_logged mode=train step=794

                                                    
{'loss': 0.002, 'grad_norm': 6.565411567687988, 'learning_rate': 7.596969696969697e-06, 'num_tokens': 1755042.0, 'completions/mean_length': 71.625, 'completions/min_length': 69.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9123142957687378, 'rewards/meter/std': 0.09913700073957443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9123142957687378, 'rewards/total_composite/std': 0.09913700073957443, 'reward': 0.9123142957687378, 'reward_std': 0.09913701564073563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034922052174806595, 'sampling/sampling_logp_difference/max': 1.3885717391967773, 'sampling/importance_sampling_ratio/min': 0.249431312084198, 'sampling/importance_sampling_ratio/mean': 1.0043187141418457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15924948174506426, 'clip_ratio/low_mean': 0.019052310031838715, 'clip_ratio/low_min': 0.019052310031838715, 'clip_ratio/high_mean': 0.007044379832223058, 'clip_ratio/high_max': 0.007044379832223058, 'clip_ratio/region_mean': 0.026096689864061773, 'reward_total_mean': 0.9123142957687378, 'reward_meter_mean': 0.9123142957687378, 'reward_meter_std': 0.09913700073957443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9123142957687378, 'reward_total_composite_std': 0.09913700073957443, 'epoch': 0.03}

 24%|██▍       | 794/3300 [19:25<3:40:43,  5.28s/it]INFO 04-11 22:56:16 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 795/3300 [19:30<3:34:04,  5.13s/it]2026-04-11 22:56:21,313 | INFO | train_grpo_train | metrics_logged mode=train step=795

                                                    
{'loss': 0.0046, 'grad_norm': 2.6046133041381836, 'learning_rate': 7.593939393939395e-06, 'num_tokens': 1756866.0, 'completions/mean_length': 73.0, 'completions/min_length': 73.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.0, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.613559365272522, 'rewards/meter/std': 0.046057041734457016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.613559365272522, 'rewards/total_composite/std': 0.046057041734457016, 'reward': 0.613559365272522, 'reward_std': 0.04605703800916672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01630816049873829, 'sampling/sampling_logp_difference/max': 0.9375072717666626, 'sampling/importance_sampling_ratio/min': 0.5393000841140747, 'sampling/importance_sampling_ratio/mean': 1.0046606063842773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07319271843880415, 'clip_ratio/low_mean': 0.008561643771827221, 'clip_ratio/low_min': 0.008561643771827221, 'clip_ratio/high_mean': 0.0034246575087308884, 'clip_ratio/high_max': 0.0034246575087308884, 'clip_ratio/region_mean': 0.01198630128055811, 'reward_total_mean': 0.613559365272522, 'reward_meter_mean': 0.613559365272522, 'reward_meter_std': 0.046057041734457016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.613559365272522, 'reward_total_composite_std': 0.046057041734457016, 'epoch': 0.03}

 24%|██▍       | 795/3300 [19:30<3:34:04,  5.13s/it]INFO 04-11 22:56:21 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 796/3300 [19:37<4:02:08,  5.80s/it]2026-04-11 22:56:28,688 | INFO | train_grpo_train | metrics_logged mode=train step=796

                                                    
{'loss': 0.0182, 'grad_norm': 1.5290881395339966, 'learning_rate': 7.590909090909091e-06, 'num_tokens': 1760371.0, 'completions/mean_length': 250.125, 'completions/min_length': 248.0, 'completions/max_length': 264.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 250.125, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 264.0, 'rewards/meter/mean': 0.6284428834915161, 'rewards/meter/std': 0.1288241744041443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6098901033401489, 'rewards/repeat_penalty/std': 0.015540807507932186, 'rewards/total_composite/mean': 0.3849101662635803, 'rewards/total_composite/std': 0.08409524708986282, 'reward': 0.3849101662635803, 'reward_std': 0.08409524708986282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006375753786414862, 'sampling/sampling_logp_difference/max': 1.0706424713134766, 'sampling/importance_sampling_ratio/min': 0.3427882194519043, 'sampling/importance_sampling_ratio/mean': 1.0002148151397705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.023278776556253433, 'clip_ratio/low_mean': 0.0004734848625957966, 'clip_ratio/low_min': 0.0004734848625957966, 'clip_ratio/high_mean': 0.002516112755984068, 'clip_ratio/high_max': 0.002516112755984068, 'clip_ratio/region_mean': 0.0029895976185798645, 'reward_total_mean': 0.3849101662635803, 'reward_meter_mean': 0.6284428834915161, 'reward_meter_std': 0.1288241744041443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6098901033401489, 'reward_repeat_penalty_std': 0.015540807507932186, 'reward_total_composite_mean': 0.3849101662635803, 'reward_total_composite_std': 0.08409524708986282, 'epoch': 0.03}

 24%|██▍       | 796/3300 [19:37<4:02:08,  5.80s/it]INFO 04-11 22:56:28 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 797/3300 [19:42<3:51:53,  5.56s/it]2026-04-11 22:56:33,682 | INFO | train_grpo_train | metrics_logged mode=train step=797

                                                    
{'loss': -0.0121, 'grad_norm': 8.541836738586426, 'learning_rate': 7.587878787878788e-06, 'num_tokens': 1762455.0, 'completions/mean_length': 77.5, 'completions/min_length': 72.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9952297210693359, 'rewards/meter/std': 0.002311403863132, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.891382098197937, 'rewards/total_composite/std': 0.1964196413755417, 'reward': 0.891382098197937, 'reward_std': 0.1964196413755417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034574542194604874, 'sampling/sampling_logp_difference/max': 3.3448140621185303, 'sampling/importance_sampling_ratio/min': 0.03526677191257477, 'sampling/importance_sampling_ratio/mean': 0.9954510927200317, 'sampling/importance_sampling_ratio/max': 1.482370376586914, 'entropy': 0.14471599273383617, 'clip_ratio/low_mean': 0.003246753243729472, 'clip_ratio/low_min': 0.003246753243729472, 'clip_ratio/high_mean': 0.009473072132095695, 'clip_ratio/high_max': 0.009473072132095695, 'clip_ratio/region_mean': 0.012719825375825167, 'reward_total_mean': 0.891382098197937, 'reward_meter_mean': 0.9952297210693359, 'reward_meter_std': 0.002311403863132, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.891382098197937, 'reward_total_composite_std': 0.1964196413755417, 'epoch': 0.03}

 24%|██▍       | 797/3300 [19:42<3:51:53,  5.56s/it]INFO 04-11 22:56:33 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 798/3300 [19:47<3:41:52,  5.32s/it]2026-04-11 22:56:38,447 | INFO | train_grpo_train | metrics_logged mode=train step=798

                                                    
{'loss': 0.0167, 'grad_norm': 6.7140374183654785, 'learning_rate': 7.584848484848486e-06, 'num_tokens': 1764291.0, 'completions/mean_length': 66.5, 'completions/min_length': 64.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6813795566558838, 'rewards/meter/std': 0.04148668423295021, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6813795566558838, 'rewards/total_composite/std': 0.04148668423295021, 'reward': 0.6813795566558838, 'reward_std': 0.04148669168353081, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023791270330548286, 'sampling/sampling_logp_difference/max': 0.9566974639892578, 'sampling/importance_sampling_ratio/min': 0.5285674333572388, 'sampling/importance_sampling_ratio/mean': 1.010377049446106, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1323620891198516, 'clip_ratio/low_mean': 0.020607191254384816, 'clip_ratio/low_min': 0.020607191254384816, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.024513441254384816, 'reward_total_mean': 0.6813795566558838, 'reward_meter_mean': 0.6813795566558838, 'reward_meter_std': 0.04148668423295021, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6813795566558838, 'reward_total_composite_std': 0.04148668423295021, 'epoch': 0.03}

 24%|██▍       | 798/3300 [19:47<3:41:52,  5.32s/it]INFO 04-11 22:56:38 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 799/3300 [19:54<3:59:53,  5.76s/it]2026-04-11 22:56:45,214 | INFO | train_grpo_train | metrics_logged mode=train step=799

                                                    
{'loss': 0.0044, 'grad_norm': 1.2270196676254272, 'learning_rate': 7.581818181818183e-06, 'num_tokens': 1767790.0, 'completions/mean_length': 239.375, 'completions/min_length': 237.0, 'completions/max_length': 240.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 239.375, 'completions/min_terminated_length': 237.0, 'completions/max_terminated_length': 240.0, 'rewards/meter/mean': 0.7462765574455261, 'rewards/meter/std': 0.09833642095327377, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6000000238418579, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.38379937410354614, 'rewards/total_composite/std': 0.05057300627231598, 'reward': 0.38379937410354614, 'reward_std': 0.050573013722896576, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007426128257066011, 'sampling/sampling_logp_difference/max': 2.6020090579986572, 'sampling/importance_sampling_ratio/min': 0.07412450760602951, 'sampling/importance_sampling_ratio/mean': 0.9996190071105957, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.030161422211676836, 'clip_ratio/low_mean': 0.003645833523478359, 'clip_ratio/low_min': 0.003645833523478359, 'clip_ratio/high_mean': 0.002109704539179802, 'clip_ratio/high_max': 0.002109704539179802, 'clip_ratio/region_mean': 0.005755538062658161, 'reward_total_mean': 0.38379937410354614, 'reward_meter_mean': 0.7462765574455261, 'reward_meter_std': 0.09833642095327377, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6000000238418579, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.38379937410354614, 'reward_total_composite_std': 0.05057300627231598, 'epoch': 0.03}

 24%|██▍       | 799/3300 [19:54<3:59:53,  5.76s/it]INFO 04-11 22:56:45 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 800/3300 [19:59<3:55:49,  5.66s/it]2026-04-11 22:56:50,658 | INFO | train_grpo_train | metrics_logged mode=train step=800

                                                    
{'loss': 0.0432, 'grad_norm': 4.79632568359375, 'learning_rate': 7.57878787878788e-06, 'num_tokens': 1770025.0, 'completions/mean_length': 119.375, 'completions/min_length': 110.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.375, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.5170607566833496, 'rewards/meter/std': 0.48065322637557983, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.4120614528656006, 'rewards/total_composite/std': 0.386055052280426, 'reward': 0.4120614528656006, 'reward_std': 0.386055052280426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04103344678878784, 'sampling/sampling_logp_difference/max': 1.1491317749023438, 'sampling/importance_sampling_ratio/min': 0.3169117867946625, 'sampling/importance_sampling_ratio/mean': 0.9977750182151794, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23464529775083065, 'clip_ratio/low_mean': 0.024260351667180657, 'clip_ratio/low_min': 0.024260351667180657, 'clip_ratio/high_mean': 0.0143592240056023, 'clip_ratio/high_max': 0.0143592240056023, 'clip_ratio/region_mean': 0.03861957567278296, 'reward_total_mean': 0.4120614528656006, 'reward_meter_mean': 0.5170607566833496, 'reward_meter_std': 0.48065322637557983, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.4120614528656006, 'reward_total_composite_std': 0.386055052280426, 'epoch': 0.03}

 24%|██▍       | 800/3300 [19:59<3:55:49,  5.66s/it]INFO 04-11 22:56:50 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:39,  3.57s/it]

 23%|██▎       | 3/13 [00:11<00:37,  3.75s/it]

 31%|███       | 4/13 [00:18<00:44,  4.92s/it]

 38%|███▊      | 5/13 [00:24<00:44,  5.60s/it]

 46%|████▌     | 6/13 [00:32<00:42,  6.11s/it]

 54%|█████▍    | 7/13 [00:37<00:36,  6.05s/it]

 62%|██████▏   | 8/13 [00:45<00:32,  6.48s/it]

 69%|██████▉   | 9/13 [00:52<00:27,  6.77s/it]

 77%|███████▋  | 10/13 [00:57<00:18,  6.26s/it]

 85%|████████▍ | 11/13 [01:04<00:12,  6.26s/it]

 92%|█████████▏| 12/13 [01:11<00:06,  6.47s/it]

100%|██████████| 13/13 [01:18<00:00,  6.68s/it]2026-04-11 22:58:16,586 | INFO | train_grpo_train | metrics_logged mode=eval step=800

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 85.9157, 'eval_samples_per_second': 1.21, 'eval_steps_per_second': 0.151, 'eval_num_tokens': 1770025.0, 'eval_completions/mean_length': 234.29807692307693, 'eval_completions/min_length': 63.84615384615385, 'eval_completions/max_length': 461.9230769230769, 'eval_completions/clipped_ratio': 0.057692307692307696, 'eval_completions/mean_terminated_length': 216.9835216815655, 'eval_completions/min_terminated_length': 63.84615384615385, 'eval_completions/max_terminated_length': 414.46153846153845, 'eval_rewards/meter/mean': 0.6419616112342248, 'eval_rewards/meter/std': 0.3754527878302794, 'eval_rewards/count_adherence/mean': 0.9552615697567279, 'eval_rewards/count_adherence/std': 0.08013723160211857, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.6556147245260385, 'eval_rewards/repeat_penalty/std': 0.24277657327743676, 'eval_rewards/total_composite/mean': 0.3991968219096844, 'eval_rewards/total_composite/std': 0.3098094039238416, 'eval_reward': 0.3991968219096844, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0054089168552309275, 'eval_sampling/sampling_logp_difference/max': 0.7075341939926147, 'eval_sampling/importance_sampling_ratio/min': 0.5230464408030877, 'eval_sampling/importance_sampling_ratio/mean': 1.0009051194557776, 'eval_sampling/importance_sampling_ratio/max': 1.2961730773632343, 'eval_entropy': 0.047992275741237864, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.3991968219096844, 'eval_reward_meter_mean': 0.6419616112342248, 'eval_reward_meter_std': 0.3754527878302794, 'eval_reward_count_adherence_mean': 0.9552615697567279, 'eval_reward_count_adherence_std': 0.08013723160211857, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.6556147245260385, 'eval_reward_repeat_penalty_std': 0.24277657327743676, 'eval_reward_total_composite_mean': 0.3991968219096844, 'eval_reward_total_composite_std': 0.3098094039238416, 'epoch': 0.03}

 24%|██▍       | 800/3300 [21:25<3:55:49,  5.66s/it]

100%|██████████| 13/13 [01:18<00:00,  6.68s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 24%|██▍       | 801/3300 [21:32<22:03:41, 31.78s/it]2026-04-11 22:58:23,384 | INFO | train_grpo_train | metrics_logged mode=train step=801

                                                     
{'loss': 0.0018, 'grad_norm': 4.925754070281982, 'learning_rate': 7.5757575757575764e-06, 'num_tokens': 1771810.0, 'completions/mean_length': 67.125, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7430859804153442, 'rewards/meter/std': 0.04014641046524048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7430859804153442, 'rewards/total_composite/std': 0.04014641046524048, 'reward': 0.7430859804153442, 'reward_std': 0.040146395564079285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02094501443207264, 'sampling/sampling_logp_difference/max': 1.098017692565918, 'sampling/importance_sampling_ratio/min': 0.33353158831596375, 'sampling/importance_sampling_ratio/mean': 1.0070921182632446, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08063888642936945, 'clip_ratio/low_mean': 0.018601843621581793, 'clip_ratio/low_min': 0.018601843621581793, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.018601843621581793, 'reward_total_mean': 0.7430859804153442, 'reward_meter_mean': 0.7430859804153442, 'reward_meter_std': 0.04014641046524048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7430859804153442, 'reward_total_composite_std': 0.04014641046524048, 'epoch': 0.03}

 24%|██▍       | 801/3300 [21:32<22:03:41, 31.78s/it]INFO 04-11 22:58:23 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 802/3300 [21:37<16:24:40, 23.65s/it]2026-04-11 22:58:28,067 | INFO | train_grpo_train | metrics_logged mode=train step=802

                                                     
{'loss': 0.0172, 'grad_norm': 7.339356422424316, 'learning_rate': 7.572727272727274e-06, 'num_tokens': 1773359.0, 'completions/mean_length': 37.625, 'completions/min_length': 36.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.13143374025821686, 'rewards/meter/std': 0.11004206538200378, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.13143374025821686, 'rewards/total_composite/std': 0.11004206538200378, 'reward': 0.13143374025821686, 'reward_std': 0.11004206538200378, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03102065436542034, 'sampling/sampling_logp_difference/max': 1.9875779151916504, 'sampling/importance_sampling_ratio/min': 0.13702690601348877, 'sampling/importance_sampling_ratio/mean': 0.9923664331436157, 'sampling/importance_sampling_ratio/max': 1.671715497970581, 'entropy': 0.1240494973026216, 'clip_ratio/low_mean': 0.009703947464004159, 'clip_ratio/low_min': 0.009703947464004159, 'clip_ratio/high_mean': 0.00995732587762177, 'clip_ratio/high_max': 0.00995732587762177, 'clip_ratio/region_mean': 0.01966127334162593, 'reward_total_mean': 0.13143374025821686, 'reward_meter_mean': 0.13143374025821686, 'reward_meter_std': 0.11004206538200378, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.13143374025821686, 'reward_total_composite_std': 0.11004206538200378, 'epoch': 0.03}

 24%|██▍       | 802/3300 [21:37<16:24:40, 23.65s/it]INFO 04-11 22:58:28 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 803/3300 [21:41<12:29:57, 18.02s/it]2026-04-11 22:58:32,953 | INFO | train_grpo_train | metrics_logged mode=train step=803

                                                     
{'loss': 0.0049, 'grad_norm': 4.46676778793335, 'learning_rate': 7.56969696969697e-06, 'num_tokens': 1775325.0, 'completions/mean_length': 70.75, 'completions/min_length': 70.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.75, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9594069123268127, 'rewards/meter/std': 0.05366376414895058, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9594069123268127, 'rewards/total_composite/std': 0.05366376414895058, 'reward': 0.9594069123268127, 'reward_std': 0.05366375669836998, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03665493428707123, 'sampling/sampling_logp_difference/max': 1.9053394794464111, 'sampling/importance_sampling_ratio/min': 0.14877213537693024, 'sampling/importance_sampling_ratio/mean': 0.9985708594322205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1510819010436535, 'clip_ratio/low_mean': 0.005281690042465925, 'clip_ratio/low_min': 0.005281690042465925, 'clip_ratio/high_mean': 0.026509752846322954, 'clip_ratio/high_max': 0.026509752846322954, 'clip_ratio/region_mean': 0.03179144288878888, 'reward_total_mean': 0.9594069123268127, 'reward_meter_mean': 0.9594069123268127, 'reward_meter_std': 0.05366376414895058, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9594069123268127, 'reward_total_composite_std': 0.05366376414895058, 'epoch': 0.03}

 24%|██▍       | 803/3300 [21:41<12:29:57, 18.02s/it]INFO 04-11 22:58:33 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 804/3300 [21:53<11:02:56, 15.94s/it]2026-04-11 22:58:44,019 | INFO | train_grpo_train | metrics_logged mode=train step=804

                                                     
{'loss': 0.0506, 'grad_norm': 0.5559709072113037, 'learning_rate': 7.566666666666667e-06, 'num_tokens': 1780007.0, 'completions/mean_length': 505.25, 'completions/min_length': 493.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 503.0, 'completions/min_terminated_length': 493.0, 'completions/max_terminated_length': 508.0, 'rewards/meter/mean': 0.9569629430770874, 'rewards/meter/std': 0.102637879550457, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.0235702246427536, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4376780688762665, 'rewards/repeat_penalty/std': 0.22492921352386475, 'rewards/total_composite/mean': 0.38479870557785034, 'rewards/total_composite/std': 0.20677605271339417, 'reward': 0.38479870557785034, 'reward_std': 0.20677603781223297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006570629775524139, 'sampling/sampling_logp_difference/max': 2.7531256675720215, 'sampling/importance_sampling_ratio/min': 0.06372835487127304, 'sampling/importance_sampling_ratio/mean': 1.0004603862762451, 'sampling/importance_sampling_ratio/max': 1.9659916162490845, 'entropy': 0.014583299867808819, 'clip_ratio/low_mean': 0.0004926113178953528, 'clip_ratio/low_min': 0.0004926113178953528, 'clip_ratio/high_mean': 0.003243225917685777, 'clip_ratio/high_max': 0.003243225917685777, 'clip_ratio/region_mean': 0.00373583723558113, 'reward_total_mean': 0.38479870557785034, 'reward_meter_mean': 0.9569629430770874, 'reward_meter_std': 0.102637879550457, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.0235702246427536, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4376780688762665, 'reward_repeat_penalty_std': 0.22492921352386475, 'reward_total_composite_mean': 0.38479870557785034, 'reward_total_composite_std': 0.20677605271339417, 'epoch': 0.03}

 24%|██▍       | 804/3300 [21:53<11:02:56, 15.94s/it]INFO 04-11 22:58:44 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 805/3300 [21:57<8:43:21, 12.59s/it] 2026-04-11 22:58:48,788 | INFO | train_grpo_train | metrics_logged mode=train step=805

                                                    
{'loss': 0.0024, 'grad_norm': 4.009762287139893, 'learning_rate': 7.563636363636364e-06, 'num_tokens': 1781817.0, 'completions/mean_length': 67.25, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7574520111083984, 'rewards/meter/std': 0.03009505569934845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7574520111083984, 'rewards/total_composite/std': 0.03009505569934845, 'reward': 0.7574520111083984, 'reward_std': 0.0300950538367033, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020206134766340256, 'sampling/sampling_logp_difference/max': 1.227433443069458, 'sampling/importance_sampling_ratio/min': 0.29304373264312744, 'sampling/importance_sampling_ratio/mean': 0.9979568719863892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05657489877194166, 'clip_ratio/low_mean': 0.011085874866694212, 'clip_ratio/low_min': 0.011085874866694212, 'clip_ratio/high_mean': 0.00932835799176246, 'clip_ratio/high_max': 0.00932835799176246, 'clip_ratio/region_mean': 0.02041423285845667, 'reward_total_mean': 0.7574520111083984, 'reward_meter_mean': 0.7574520111083984, 'reward_meter_std': 0.03009505569934845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7574520111083984, 'reward_total_composite_std': 0.03009505569934845, 'epoch': 0.03}

 24%|██▍       | 805/3300 [21:57<8:43:21, 12.59s/it]INFO 04-11 22:58:49 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 806/3300 [22:02<7:04:06, 10.20s/it]2026-04-11 22:58:53,432 | INFO | train_grpo_train | metrics_logged mode=train step=806

                                                    
{'loss': 0.0021, 'grad_norm': 3.6810977458953857, 'learning_rate': 7.560606060606062e-06, 'num_tokens': 1783594.0, 'completions/mean_length': 58.125, 'completions/min_length': 57.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.125, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.990125298500061, 'rewards/meter/std': 0.0014512698398903012, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.17817415297031403, 'rewards/total_composite/mean': 0.8252000212669373, 'rewards/total_composite/std': 0.17693018913269043, 'reward': 0.8252000212669373, 'reward_std': 0.17693018913269043, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023440338671207428, 'sampling/sampling_logp_difference/max': 1.1792564392089844, 'sampling/importance_sampling_ratio/min': 0.30750730633735657, 'sampling/importance_sampling_ratio/mean': 0.9979320168495178, 'sampling/importance_sampling_ratio/max': 1.3977869749069214, 'entropy': 0.11172830406576395, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.01062974869273603, 'clip_ratio/high_max': 0.01062974869273603, 'clip_ratio/region_mean': 0.01494009350426495, 'reward_total_mean': 0.8252000212669373, 'reward_meter_mean': 0.990125298500061, 'reward_meter_std': 0.0014512698398903012, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.17817415297031403, 'reward_total_composite_mean': 0.8252000212669373, 'reward_total_composite_std': 0.17693018913269043, 'epoch': 0.03}

 24%|██▍       | 806/3300 [22:02<7:04:06, 10.20s/it]INFO 04-11 22:58:53 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 807/3300 [22:07<5:57:48,  8.61s/it]2026-04-11 22:58:58,329 | INFO | train_grpo_train | metrics_logged mode=train step=807

                                                    
{'loss': -0.0084, 'grad_norm': 5.640214920043945, 'learning_rate': 7.557575757575758e-06, 'num_tokens': 1785577.0, 'completions/mean_length': 75.875, 'completions/min_length': 72.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7918260097503662, 'rewards/meter/std': 0.3541868031024933, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7303093671798706, 'rewards/total_composite/std': 0.35869720578193665, 'reward': 0.7303093671798706, 'reward_std': 0.35869717597961426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02571277692914009, 'sampling/sampling_logp_difference/max': 1.1890771389007568, 'sampling/importance_sampling_ratio/min': 0.4486599266529083, 'sampling/importance_sampling_ratio/mean': 1.0031059980392456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12689008563756943, 'clip_ratio/low_mean': 0.00854225957300514, 'clip_ratio/low_min': 0.00854225957300514, 'clip_ratio/high_mean': 0.008098726975731552, 'clip_ratio/high_max': 0.008098726975731552, 'clip_ratio/region_mean': 0.01664098654873669, 'reward_total_mean': 0.7303093671798706, 'reward_meter_mean': 0.7918260097503662, 'reward_meter_std': 0.3541868031024933, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7303093671798706, 'reward_total_composite_std': 0.35869720578193665, 'epoch': 0.03}

 24%|██▍       | 807/3300 [22:07<5:57:48,  8.61s/it]INFO 04-11 22:58:58 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 808/3300 [22:11<5:03:34,  7.31s/it]2026-04-11 22:59:02,599 | INFO | train_grpo_train | metrics_logged mode=train step=808

                                                    
{'loss': -0.0018, 'grad_norm': 14.179214477539062, 'learning_rate': 7.5545454545454555e-06, 'num_tokens': 1787137.0, 'completions/mean_length': 38.0, 'completions/min_length': 38.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8494601845741272, 'rewards/meter/std': 0.3245563507080078, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8494601845741272, 'rewards/total_composite/std': 0.3245563507080078, 'reward': 0.8494601845741272, 'reward_std': 0.3245563209056854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026692412793636322, 'sampling/sampling_logp_difference/max': 2.183236598968506, 'sampling/importance_sampling_ratio/min': 0.11267625540494919, 'sampling/importance_sampling_ratio/mean': 0.9977914690971375, 'sampling/importance_sampling_ratio/max': 1.5366783142089844, 'entropy': 0.09128655772656202, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.016447368543595076, 'clip_ratio/high_max': 0.016447368543595076, 'clip_ratio/region_mean': 0.016447368543595076, 'reward_total_mean': 0.8494601845741272, 'reward_meter_mean': 0.8494601845741272, 'reward_meter_std': 0.3245563507080078, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8494601845741272, 'reward_total_composite_std': 0.3245563507080078, 'epoch': 0.03}

 24%|██▍       | 808/3300 [22:11<5:03:34,  7.31s/it]INFO 04-11 22:59:02 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 809/3300 [22:18<4:55:03,  7.11s/it]2026-04-11 22:59:09,234 | INFO | train_grpo_train | metrics_logged mode=train step=809

                                                    
{'loss': -0.0031, 'grad_norm': 1.1023591756820679, 'learning_rate': 7.551515151515152e-06, 'num_tokens': 1790539.0, 'completions/mean_length': 218.25, 'completions/min_length': 206.0, 'completions/max_length': 230.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 218.25, 'completions/min_terminated_length': 206.0, 'completions/max_terminated_length': 230.0, 'rewards/meter/mean': 0.9922986030578613, 'rewards/meter/std': 0.004149852320551872, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5681818723678589, 'rewards/repeat_penalty/std': 0.1735115498304367, 'rewards/total_composite/mean': 0.5638895034790039, 'rewards/total_composite/std': 0.17219732701778412, 'reward': 0.5638895034790039, 'reward_std': 0.17219732701778412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011151120997965336, 'sampling/sampling_logp_difference/max': 1.5731306076049805, 'sampling/importance_sampling_ratio/min': 0.20739488303661346, 'sampling/importance_sampling_ratio/mean': 1.0026495456695557, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03606141824275255, 'clip_ratio/low_mean': 0.005849295761436224, 'clip_ratio/low_min': 0.005849295761436224, 'clip_ratio/high_mean': 0.003968499368056655, 'clip_ratio/high_max': 0.003968499368056655, 'clip_ratio/region_mean': 0.009817795129492879, 'reward_total_mean': 0.5638895034790039, 'reward_meter_mean': 0.9922986030578613, 'reward_meter_std': 0.004149852320551872, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5681818723678589, 'reward_repeat_penalty_std': 0.1735115498304367, 'reward_total_composite_mean': 0.5638895034790039, 'reward_total_composite_std': 0.17219732701778412, 'epoch': 0.03}

 25%|██▍       | 809/3300 [22:18<4:55:03,  7.11s/it]INFO 04-11 22:59:09 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 810/3300 [22:23<4:27:10,  6.44s/it]2026-04-11 22:59:14,111 | INFO | train_grpo_train | metrics_logged mode=train step=810

                                                    
{'loss': 0.0017, 'grad_norm': 5.80548095703125, 'learning_rate': 7.548484848484849e-06, 'num_tokens': 1792366.0, 'completions/mean_length': 71.375, 'completions/min_length': 70.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9742876887321472, 'rewards/meter/std': 0.029914017766714096, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9332020878791809, 'rewards/total_composite/std': 0.11529982835054398, 'reward': 0.9332020878791809, 'reward_std': 0.11529984325170517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0373886339366436, 'sampling/sampling_logp_difference/max': 1.9366556406021118, 'sampling/importance_sampling_ratio/min': 0.14418534934520721, 'sampling/importance_sampling_ratio/mean': 1.0060287714004517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15044390503317118, 'clip_ratio/low_mean': 0.005306840990670025, 'clip_ratio/low_min': 0.005306840990670025, 'clip_ratio/high_mean': 0.03143186215311289, 'clip_ratio/high_max': 0.03143186215311289, 'clip_ratio/region_mean': 0.036738703143782914, 'reward_total_mean': 0.9332020878791809, 'reward_meter_mean': 0.9742876887321472, 'reward_meter_std': 0.029914017766714096, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9332020878791809, 'reward_total_composite_std': 0.11529982835054398, 'epoch': 0.03}

 25%|██▍       | 810/3300 [22:23<4:27:10,  6.44s/it]INFO 04-11 22:59:14 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 811/3300 [22:29<4:30:53,  6.53s/it]2026-04-11 22:59:20,864 | INFO | train_grpo_train | metrics_logged mode=train step=811

                                                    
{'loss': 0.0192, 'grad_norm': 2.172598123550415, 'learning_rate': 7.545454545454546e-06, 'num_tokens': 1795876.0, 'completions/mean_length': 203.75, 'completions/min_length': 196.0, 'completions/max_length': 229.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 203.75, 'completions/min_terminated_length': 196.0, 'completions/max_terminated_length': 229.0, 'rewards/meter/mean': 0.710443377494812, 'rewards/meter/std': 0.2666179835796356, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4431818127632141, 'rewards/repeat_penalty/std': 0.22498852014541626, 'rewards/total_composite/mean': 0.3340369462966919, 'rewards/total_composite/std': 0.2214316725730896, 'reward': 0.3340369462966919, 'reward_std': 0.2214316576719284, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021875398233532906, 'sampling/sampling_logp_difference/max': 4.384680271148682, 'sampling/importance_sampling_ratio/min': 0.012466873973608017, 'sampling/importance_sampling_ratio/mean': 1.0012760162353516, 'sampling/importance_sampling_ratio/max': 1.9448559284210205, 'entropy': 0.1404771413654089, 'clip_ratio/low_mean': 0.006675369921140373, 'clip_ratio/low_min': 0.006675369921140373, 'clip_ratio/high_mean': 0.011296228156425059, 'clip_ratio/high_max': 0.011296228156425059, 'clip_ratio/region_mean': 0.01797159807756543, 'reward_total_mean': 0.3340369462966919, 'reward_meter_mean': 0.710443377494812, 'reward_meter_std': 0.2666179835796356, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4431818127632141, 'reward_repeat_penalty_std': 0.22498852014541626, 'reward_total_composite_mean': 0.3340369462966919, 'reward_total_composite_std': 0.2214316725730896, 'epoch': 0.03}

 25%|██▍       | 811/3300 [22:29<4:30:53,  6.53s/it]INFO 04-11 22:59:21 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 812/3300 [22:34<4:12:26,  6.09s/it]2026-04-11 22:59:25,914 | INFO | train_grpo_train | metrics_logged mode=train step=812

                                                    
{'loss': 0.0139, 'grad_norm': 8.286386489868164, 'learning_rate': 7.542424242424244e-06, 'num_tokens': 1797566.0, 'completions/mean_length': 65.25, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9978810548782349, 'rewards/meter/std': 0.0005103643052279949, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9563552141189575, 'rewards/total_composite/std': 0.11796265840530396, 'reward': 0.9563552141189575, 'reward_std': 0.11796264350414276, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007206229493021965, 'sampling/sampling_logp_difference/max': 0.5880947113037109, 'sampling/importance_sampling_ratio/min': 0.5553844571113586, 'sampling/importance_sampling_ratio/mean': 1.002901315689087, 'sampling/importance_sampling_ratio/max': 1.3499990701675415, 'entropy': 0.036765412194654346, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005769230774603784, 'clip_ratio/high_max': 0.005769230774603784, 'clip_ratio/region_mean': 0.005769230774603784, 'reward_total_mean': 0.9563552141189575, 'reward_meter_mean': 0.9978810548782349, 'reward_meter_std': 0.0005103643052279949, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9563552141189575, 'reward_total_composite_std': 0.11796265840530396, 'epoch': 0.03}

 25%|██▍       | 812/3300 [22:34<4:12:26,  6.09s/it]INFO 04-11 22:59:26 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 813/3300 [22:43<4:39:17,  6.74s/it]2026-04-11 22:59:34,169 | INFO | train_grpo_train | metrics_logged mode=train step=813

                                                    
{'loss': 0.0079, 'grad_norm': 1.0310794115066528, 'learning_rate': 7.53939393939394e-06, 'num_tokens': 1801761.0, 'completions/mean_length': 332.375, 'completions/min_length': 302.0, 'completions/max_length': 353.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 332.375, 'completions/min_terminated_length': 302.0, 'completions/max_terminated_length': 353.0, 'rewards/meter/mean': 0.9938352704048157, 'rewards/meter/std': 0.008550022728741169, 'rewards/count_adherence/mean': 0.890625, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5850183963775635, 'rewards/repeat_penalty/std': 0.009098809212446213, 'rewards/total_composite/mean': 0.517345666885376, 'rewards/total_composite/std': 0.012124452739953995, 'reward': 0.517345666885376, 'reward_std': 0.012124458327889442, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0052232746966183186, 'sampling/sampling_logp_difference/max': 0.7512289881706238, 'sampling/importance_sampling_ratio/min': 0.47178637981414795, 'sampling/importance_sampling_ratio/mean': 1.0009092092514038, 'sampling/importance_sampling_ratio/max': 1.6740483045578003, 'entropy': 0.02348946128040552, 'clip_ratio/low_mean': 0.005218350415816531, 'clip_ratio/low_min': 0.005218350415816531, 'clip_ratio/high_mean': 0.0019113150192424655, 'clip_ratio/high_max': 0.0019113150192424655, 'clip_ratio/region_mean': 0.007129665435058996, 'reward_total_mean': 0.517345666885376, 'reward_meter_mean': 0.9938352704048157, 'reward_meter_std': 0.008550022728741169, 'reward_count_adherence_mean': 0.890625, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5850183963775635, 'reward_repeat_penalty_std': 0.009098809212446213, 'reward_total_composite_mean': 0.517345666885376, 'reward_total_composite_std': 0.012124452739953995, 'epoch': 0.03}

 25%|██▍       | 813/3300 [22:43<4:39:17,  6.74s/it]INFO 04-11 22:59:34 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 814/3300 [22:49<4:29:08,  6.50s/it]2026-04-11 22:59:40,099 | INFO | train_grpo_train | metrics_logged mode=train step=814

                                                    
{'loss': -0.001, 'grad_norm': 13.524036407470703, 'learning_rate': 7.536363636363637e-06, 'num_tokens': 1804516.0, 'completions/mean_length': 159.375, 'completions/min_length': 158.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.375, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9979825019836426, 'rewards/meter/std': 0.0006693408940918744, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6653216481208801, 'rewards/total_composite/std': 0.00044622019049711525, 'reward': 0.6653216481208801, 'reward_std': 0.00044622053974308074, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005735776387155056, 'sampling/sampling_logp_difference/max': 2.232571601867676, 'sampling/importance_sampling_ratio/min': 0.10725226998329163, 'sampling/importance_sampling_ratio/mean': 0.9996488094329834, 'sampling/importance_sampling_ratio/max': 1.493585467338562, 'entropy': 0.03128327080048621, 'clip_ratio/low_mean': 0.00237341778120026, 'clip_ratio/low_min': 0.00237341778120026, 'clip_ratio/high_mean': 0.00237341778120026, 'clip_ratio/high_max': 0.00237341778120026, 'clip_ratio/region_mean': 0.00474683556240052, 'reward_total_mean': 0.6653216481208801, 'reward_meter_mean': 0.9979825019836426, 'reward_meter_std': 0.0006693408940918744, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6653216481208801, 'reward_total_composite_std': 0.00044622019049711525, 'epoch': 0.03}

 25%|██▍       | 814/3300 [22:49<4:29:08,  6.50s/it]INFO 04-11 22:59:40 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 815/3300 [22:53<4:07:16,  5.97s/it]2026-04-11 22:59:44,842 | INFO | train_grpo_train | metrics_logged mode=train step=815

                                                    
{'loss': 0.0458, 'grad_norm': 8.952130317687988, 'learning_rate': 7.533333333333334e-06, 'num_tokens': 1806317.0, 'completions/mean_length': 62.125, 'completions/min_length': 54.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9308251142501831, 'rewards/meter/std': 0.15261007845401764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.8894620537757874, 'rewards/total_composite/std': 0.17642506957054138, 'reward': 0.8894620537757874, 'reward_std': 0.1764250546693802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024382825940847397, 'sampling/sampling_logp_difference/max': 1.424285888671875, 'sampling/importance_sampling_ratio/min': 0.2406802922487259, 'sampling/importance_sampling_ratio/mean': 0.9998781085014343, 'sampling/importance_sampling_ratio/max': 1.5671416521072388, 'entropy': 0.10703426506370306, 'clip_ratio/low_mean': 0.011092530796304345, 'clip_ratio/low_min': 0.011092530796304345, 'clip_ratio/high_mean': 0.016365812392905354, 'clip_ratio/high_max': 0.016365812392905354, 'clip_ratio/region_mean': 0.0274583431892097, 'reward_total_mean': 0.8894620537757874, 'reward_meter_mean': 0.9308251142501831, 'reward_meter_std': 0.15261007845401764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.8894620537757874, 'reward_total_composite_std': 0.17642506957054138, 'epoch': 0.03}

 25%|██▍       | 815/3300 [22:53<4:07:16,  5.97s/it]INFO 04-11 22:59:45 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 816/3300 [22:58<3:47:45,  5.50s/it]2026-04-11 22:59:49,253 | INFO | train_grpo_train | metrics_logged mode=train step=816

                                                    
{'loss': -0.0066, 'grad_norm': 8.190361022949219, 'learning_rate': 7.530303030303031e-06, 'num_tokens': 1807821.0, 'completions/mean_length': 39.0, 'completions/min_length': 38.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7446458339691162, 'rewards/meter/std': 0.30041757225990295, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7446458339691162, 'rewards/total_composite/std': 0.30041757225990295, 'reward': 0.7446458339691162, 'reward_std': 0.30041757225990295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059971921145915985, 'sampling/sampling_logp_difference/max': 2.114863395690918, 'sampling/importance_sampling_ratio/min': 0.12064976990222931, 'sampling/importance_sampling_ratio/mean': 0.9998385310173035, 'sampling/importance_sampling_ratio/max': 1.8297884464263916, 'entropy': 0.16819044947624207, 'clip_ratio/low_mean': 0.023026316426694393, 'clip_ratio/low_min': 0.023026316426694393, 'clip_ratio/high_mean': 0.028449730249121785, 'clip_ratio/high_max': 0.028449730249121785, 'clip_ratio/region_mean': 0.05147604667581618, 'reward_total_mean': 0.7446458339691162, 'reward_meter_mean': 0.7446458339691162, 'reward_meter_std': 0.30041757225990295, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7446458339691162, 'reward_total_composite_std': 0.30041757225990295, 'epoch': 0.03}

 25%|██▍       | 816/3300 [22:58<3:47:45,  5.50s/it]INFO 04-11 22:59:49 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 817/3300 [23:03<3:44:21,  5.42s/it]2026-04-11 22:59:54,485 | INFO | train_grpo_train | metrics_logged mode=train step=817

                                                    
{'loss': 0.0413, 'grad_norm': 3.7302210330963135, 'learning_rate': 7.5272727272727274e-06, 'num_tokens': 1810070.0, 'completions/mean_length': 107.125, 'completions/min_length': 102.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.125, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.3294169008731842, 'rewards/meter/std': 0.28731057047843933, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7750000357627869, 'rewards/repeat_penalty/std': 0.12817399203777313, 'rewards/total_composite/mean': 0.2755756676197052, 'rewards/total_composite/std': 0.2410556823015213, 'reward': 0.2755756676197052, 'reward_std': 0.2410556823015213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01802036352455616, 'sampling/sampling_logp_difference/max': 1.4399476051330566, 'sampling/importance_sampling_ratio/min': 0.2369401901960373, 'sampling/importance_sampling_ratio/mean': 1.0044699907302856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05588802928104997, 'clip_ratio/low_mean': 0.01120058260858059, 'clip_ratio/low_min': 0.01120058260858059, 'clip_ratio/high_mean': 0.013432800536975265, 'clip_ratio/high_max': 0.013432800536975265, 'clip_ratio/region_mean': 0.024633383145555854, 'reward_total_mean': 0.2755756676197052, 'reward_meter_mean': 0.3294169008731842, 'reward_meter_std': 0.28731057047843933, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7750000357627869, 'reward_repeat_penalty_std': 0.12817399203777313, 'reward_total_composite_mean': 0.2755756676197052, 'reward_total_composite_std': 0.2410556823015213, 'epoch': 0.03}

 25%|██▍       | 817/3300 [23:03<3:44:21,  5.42s/it]INFO 04-11 22:59:54 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 818/3300 [23:12<4:24:46,  6.40s/it]2026-04-11 23:00:03,172 | INFO | train_grpo_train | metrics_logged mode=train step=818

                                                    
{'loss': 0.0108, 'grad_norm': 1.0928229093551636, 'learning_rate': 7.524242424242425e-06, 'num_tokens': 1814452.0, 'completions/mean_length': 357.75, 'completions/min_length': 346.0, 'completions/max_length': 384.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 357.75, 'completions/min_terminated_length': 346.0, 'completions/max_terminated_length': 384.0, 'rewards/meter/mean': 0.9965507984161377, 'rewards/meter/std': 0.0037117046304047108, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.05750546231865883, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5523655414581299, 'rewards/repeat_penalty/std': 0.10810358822345734, 'rewards/total_composite/mean': 0.5283541083335876, 'rewards/total_composite/std': 0.11482103914022446, 'reward': 0.5283541083335876, 'reward_std': 0.11482104659080505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00884958729147911, 'sampling/sampling_logp_difference/max': 1.062759518623352, 'sampling/importance_sampling_ratio/min': 0.3455010652542114, 'sampling/importance_sampling_ratio/mean': 1.0001471042633057, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03675817488692701, 'clip_ratio/low_mean': 0.003796361561398953, 'clip_ratio/low_min': 0.003796361561398953, 'clip_ratio/high_mean': 0.006020042230375111, 'clip_ratio/high_max': 0.006020042230375111, 'clip_ratio/region_mean': 0.009816403791774064, 'reward_total_mean': 0.5283541083335876, 'reward_meter_mean': 0.9965507984161377, 'reward_meter_std': 0.0037117046304047108, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.05750546231865883, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5523655414581299, 'reward_repeat_penalty_std': 0.10810358822345734, 'reward_total_composite_mean': 0.5283541083335876, 'reward_total_composite_std': 0.11482103914022446, 'epoch': 0.03}

 25%|██▍       | 818/3300 [23:12<4:24:46,  6.40s/it]INFO 04-11 23:00:03 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 819/3300 [23:17<4:06:50,  5.97s/it]2026-04-11 23:00:08,137 | INFO | train_grpo_train | metrics_logged mode=train step=819

                                                    
{'loss': 0.0279, 'grad_norm': 6.570438861846924, 'learning_rate': 7.521212121212121e-06, 'num_tokens': 1816193.0, 'completions/mean_length': 61.625, 'completions/min_length': 59.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9746619462966919, 'rewards/meter/std': 0.020627282559871674, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9746619462966919, 'rewards/total_composite/std': 0.020627282559871674, 'reward': 0.9746619462966919, 'reward_std': 0.020627308636903763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02879003807902336, 'sampling/sampling_logp_difference/max': 1.3429758548736572, 'sampling/importance_sampling_ratio/min': 0.26106762886047363, 'sampling/importance_sampling_ratio/mean': 0.9997745156288147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10824673250317574, 'clip_ratio/low_mean': 0.013770792167633772, 'clip_ratio/low_min': 0.013770792167633772, 'clip_ratio/high_mean': 0.022682114504277706, 'clip_ratio/high_max': 0.022682114504277706, 'clip_ratio/region_mean': 0.03645290667191148, 'reward_total_mean': 0.9746619462966919, 'reward_meter_mean': 0.9746619462966919, 'reward_meter_std': 0.020627282559871674, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9746619462966919, 'reward_total_composite_std': 0.020627282559871674, 'epoch': 0.03}

 25%|██▍       | 819/3300 [23:17<4:06:50,  5.97s/it]INFO 04-11 23:00:08 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 820/3300 [23:23<4:07:04,  5.98s/it]2026-04-11 23:00:14,131 | INFO | train_grpo_train | metrics_logged mode=train step=820

                                                    
{'loss': 0.0185, 'grad_norm': 2.3457443714141846, 'learning_rate': 7.518181818181819e-06, 'num_tokens': 1818812.0, 'completions/mean_length': 149.375, 'completions/min_length': 143.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 149.375, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.7805307507514954, 'rewards/meter/std': 0.1253480166196823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5714285373687744, 'rewards/repeat_penalty/std': 0.17074695229530334, 'rewards/total_composite/mean': 0.4300089478492737, 'rewards/total_composite/std': 0.0887664407491684, 'reward': 0.4300089478492737, 'reward_std': 0.0887664407491684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015637392178177834, 'sampling/sampling_logp_difference/max': 1.2027087211608887, 'sampling/importance_sampling_ratio/min': 0.3003794550895691, 'sampling/importance_sampling_ratio/mean': 1.0010299682617188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06724543264135718, 'clip_ratio/low_mean': 0.007311595429200679, 'clip_ratio/low_min': 0.007311595429200679, 'clip_ratio/high_mean': 0.008506174897775054, 'clip_ratio/high_max': 0.008506174897775054, 'clip_ratio/region_mean': 0.015817770326975733, 'reward_total_mean': 0.4300089478492737, 'reward_meter_mean': 0.7805307507514954, 'reward_meter_std': 0.1253480166196823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5714285373687744, 'reward_repeat_penalty_std': 0.17074695229530334, 'reward_total_composite_mean': 0.4300089478492737, 'reward_total_composite_std': 0.0887664407491684, 'epoch': 0.03}

 25%|██▍       | 820/3300 [23:23<4:07:04,  5.98s/it]INFO 04-11 23:00:14 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 821/3300 [23:32<4:51:07,  7.05s/it]2026-04-11 23:00:23,670 | INFO | train_grpo_train | metrics_logged mode=train step=821

                                                    
{'loss': -0.1471, 'grad_norm': 0.9305291175842285, 'learning_rate': 7.515151515151516e-06, 'num_tokens': 1820662.0, 'completions/mean_length': 190.25, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 83.0, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.5668963193893433, 'rewards/meter/std': 0.39650967717170715, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5668963193893433, 'rewards/total_composite/std': 0.39650967717170715, 'reward': 0.5668963193893433, 'reward_std': 0.39650964736938477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019788384437561035, 'sampling/sampling_logp_difference/max': 0.7466448545455933, 'sampling/importance_sampling_ratio/min': 0.6215139627456665, 'sampling/importance_sampling_ratio/mean': 1.0027402639389038, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07134009897708893, 'clip_ratio/low_mean': 0.007159017724916339, 'clip_ratio/low_min': 0.007159017724916339, 'clip_ratio/high_mean': 0.01558885129634291, 'clip_ratio/high_max': 0.01558885129634291, 'clip_ratio/region_mean': 0.022747869021259248, 'reward_total_mean': 0.5668963193893433, 'reward_meter_mean': 0.5668963193893433, 'reward_meter_std': 0.39650967717170715, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5668963193893433, 'reward_total_composite_std': 0.39650967717170715, 'epoch': 0.03}

 25%|██▍       | 821/3300 [23:32<4:51:07,  7.05s/it]INFO 04-11 23:00:23 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 822/3300 [23:37<4:24:28,  6.40s/it]2026-04-11 23:00:28,575 | INFO | train_grpo_train | metrics_logged mode=train step=822

                                                    
{'loss': 0.0105, 'grad_norm': 3.704124689102173, 'learning_rate': 7.512121212121213e-06, 'num_tokens': 1822669.0, 'completions/mean_length': 77.875, 'completions/min_length': 75.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.8737984299659729, 'rewards/meter/std': 0.14763900637626648, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8737984299659729, 'rewards/total_composite/std': 0.14763900637626648, 'reward': 0.8737984299659729, 'reward_std': 0.14763899147510529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018405567854642868, 'sampling/sampling_logp_difference/max': 1.2704942226409912, 'sampling/importance_sampling_ratio/min': 0.28069284558296204, 'sampling/importance_sampling_ratio/mean': 1.0058021545410156, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09962678607553244, 'clip_ratio/low_mean': 0.006382113788276911, 'clip_ratio/low_min': 0.006382113788276911, 'clip_ratio/high_mean': 0.009740259731188416, 'clip_ratio/high_max': 0.009740259731188416, 'clip_ratio/region_mean': 0.016122373519465327, 'reward_total_mean': 0.8737984299659729, 'reward_meter_mean': 0.8737984299659729, 'reward_meter_std': 0.14763900637626648, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8737984299659729, 'reward_total_composite_std': 0.14763900637626648, 'epoch': 0.03}

 25%|██▍       | 822/3300 [23:37<4:24:28,  6.40s/it]INFO 04-11 23:00:28 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 823/3300 [23:44<4:37:08,  6.71s/it]2026-04-11 23:00:36,010 | INFO | train_grpo_train | metrics_logged mode=train step=823

                                                    
{'loss': 0.0055, 'grad_norm': 2.836785078048706, 'learning_rate': 7.509090909090909e-06, 'num_tokens': 1826360.0, 'completions/mean_length': 284.375, 'completions/min_length': 272.0, 'completions/max_length': 297.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 284.375, 'completions/min_terminated_length': 272.0, 'completions/max_terminated_length': 297.0, 'rewards/meter/mean': 0.9983997344970703, 'rewards/meter/std': 0.0007759786094538867, 'rewards/count_adherence/mean': 0.9821428656578064, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5952796936035156, 'rewards/repeat_penalty/std': 0.08133196830749512, 'rewards/total_composite/mean': 0.5862284898757935, 'rewards/total_composite/std': 0.09859947860240936, 'reward': 0.5862284898757935, 'reward_std': 0.09859946370124817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015353661961853504, 'sampling/sampling_logp_difference/max': 2.6265687942504883, 'sampling/importance_sampling_ratio/min': 0.07232620567083359, 'sampling/importance_sampling_ratio/mean': 1.0005059242248535, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03474967950023711, 'clip_ratio/low_mean': 0.002257478976389393, 'clip_ratio/low_min': 0.002257478976389393, 'clip_ratio/high_mean': 0.0065078792395070195, 'clip_ratio/high_max': 0.0065078792395070195, 'clip_ratio/region_mean': 0.008765358215896413, 'reward_total_mean': 0.5862284898757935, 'reward_meter_mean': 0.9983997344970703, 'reward_meter_std': 0.0007759786094538867, 'reward_count_adherence_mean': 0.9821428656578064, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5952796936035156, 'reward_repeat_penalty_std': 0.08133196830749512, 'reward_total_composite_mean': 0.5862284898757935, 'reward_total_composite_std': 0.09859947860240936, 'epoch': 0.03}

 25%|██▍       | 823/3300 [23:45<4:37:08,  6.71s/it]INFO 04-11 23:00:36 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 824/3300 [23:49<4:13:30,  6.14s/it]2026-04-11 23:00:40,825 | INFO | train_grpo_train | metrics_logged mode=train step=824

                                                    
{'loss': 0.0168, 'grad_norm': 4.631432056427002, 'learning_rate': 7.5060606060606065e-06, 'num_tokens': 1828232.0, 'completions/mean_length': 71.0, 'completions/min_length': 70.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7449524998664856, 'rewards/meter/std': 0.04608132317662239, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7449524998664856, 'rewards/total_composite/std': 0.04608132317662239, 'reward': 0.7449524998664856, 'reward_std': 0.04608132690191269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01667422614991665, 'sampling/sampling_logp_difference/max': 2.1282317638397217, 'sampling/importance_sampling_ratio/min': 0.1190476045012474, 'sampling/importance_sampling_ratio/mean': 0.9986546039581299, 'sampling/importance_sampling_ratio/max': 1.426641583442688, 'entropy': 0.04505776287987828, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.0070436508394777775, 'clip_ratio/high_max': 0.0070436508394777775, 'clip_ratio/region_mean': 0.008779761963523924, 'reward_total_mean': 0.7449524998664856, 'reward_meter_mean': 0.7449524998664856, 'reward_meter_std': 0.04608132317662239, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7449524998664856, 'reward_total_composite_std': 0.04608132317662239, 'epoch': 0.03}

 25%|██▍       | 824/3300 [23:49<4:13:30,  6.14s/it]INFO 04-11 23:00:41 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 825/3300 [23:54<3:56:51,  5.74s/it]2026-04-11 23:00:45,630 | INFO | train_grpo_train | metrics_logged mode=train step=825

                                                    
{'loss': 0.0481, 'grad_norm': 4.750690937042236, 'learning_rate': 7.503030303030303e-06, 'num_tokens': 1830123.0, 'completions/mean_length': 68.375, 'completions/min_length': 64.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.4837471544742584, 'rewards/meter/std': 0.22915734350681305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4837471544742584, 'rewards/total_composite/std': 0.22915734350681305, 'reward': 0.4837471544742584, 'reward_std': 0.22915734350681305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031915146857500076, 'sampling/sampling_logp_difference/max': 1.8651485443115234, 'sampling/importance_sampling_ratio/min': 0.15487320721149445, 'sampling/importance_sampling_ratio/mean': 0.9999799728393555, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10288840066641569, 'clip_ratio/low_mean': 0.024154700804501772, 'clip_ratio/low_min': 0.024154700804501772, 'clip_ratio/high_mean': 0.028171515092253685, 'clip_ratio/high_max': 0.028171515092253685, 'clip_ratio/region_mean': 0.05232621589675546, 'reward_total_mean': 0.4837471544742584, 'reward_meter_mean': 0.4837471544742584, 'reward_meter_std': 0.22915734350681305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4837471544742584, 'reward_total_composite_std': 0.22915734350681305, 'epoch': 0.03}

 25%|██▌       | 825/3300 [23:54<3:56:51,  5.74s/it]INFO 04-11 23:00:45 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 826/3300 [24:03<4:37:26,  6.73s/it]2026-04-11 23:00:54,663 | INFO | train_grpo_train | metrics_logged mode=train step=826

                                                    
{'loss': 0.0159, 'grad_norm': 1.1773563623428345, 'learning_rate': 7.500000000000001e-06, 'num_tokens': 1835244.0, 'completions/mean_length': 400.125, 'completions/min_length': 375.0, 'completions/max_length': 421.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 400.125, 'completions/min_terminated_length': 375.0, 'completions/max_terminated_length': 421.0, 'rewards/meter/mean': 0.7846977710723877, 'rewards/meter/std': 0.39648592472076416, 'rewards/count_adherence/mean': 0.9545454978942871, 'rewards/count_adherence/std': 0.0485929399728775, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5507364273071289, 'rewards/repeat_penalty/std': 0.050370436161756516, 'rewards/total_composite/mean': 0.3971686363220215, 'rewards/total_composite/std': 0.1935662180185318, 'reward': 0.3971686363220215, 'reward_std': 0.1935662031173706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007981406524777412, 'sampling/sampling_logp_difference/max': 1.5345345735549927, 'sampling/importance_sampling_ratio/min': 0.25283992290496826, 'sampling/importance_sampling_ratio/mean': 1.0009510517120361, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0334128841641359, 'clip_ratio/low_mean': 0.0033315176842734218, 'clip_ratio/low_min': 0.0033315176842734218, 'clip_ratio/high_mean': 0.005249909590929747, 'clip_ratio/high_max': 0.005249909590929747, 'clip_ratio/region_mean': 0.008581427275203168, 'reward_total_mean': 0.3971686363220215, 'reward_meter_mean': 0.7846977710723877, 'reward_meter_std': 0.39648592472076416, 'reward_count_adherence_mean': 0.9545454978942871, 'reward_count_adherence_std': 0.0485929399728775, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5507364273071289, 'reward_repeat_penalty_std': 0.050370436161756516, 'reward_total_composite_mean': 0.3971686363220215, 'reward_total_composite_std': 0.1935662180185318, 'epoch': 0.03}

 25%|██▌       | 826/3300 [24:03<4:37:26,  6.73s/it]INFO 04-11 23:00:54 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 827/3300 [24:08<4:12:15,  6.12s/it]2026-04-11 23:00:59,364 | INFO | train_grpo_train | metrics_logged mode=train step=827

                                                    
{'loss': -0.0012, 'grad_norm': 3.7937092781066895, 'learning_rate': 7.496969696969698e-06, 'num_tokens': 1837278.0, 'completions/mean_length': 68.25, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.15423895418643951, 'rewards/meter/std': 0.07883358746767044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.15423895418643951, 'rewards/total_composite/std': 0.07883358746767044, 'reward': 0.15423895418643951, 'reward_std': 0.07883358746767044, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016422174870967865, 'sampling/sampling_logp_difference/max': 0.6457552909851074, 'sampling/importance_sampling_ratio/min': 0.5242664217948914, 'sampling/importance_sampling_ratio/mean': 0.9993041157722473, 'sampling/importance_sampling_ratio/max': 1.73274564743042, 'entropy': 0.05680654477328062, 'clip_ratio/low_mean': 0.009191176504828036, 'clip_ratio/low_min': 0.009191176504828036, 'clip_ratio/high_mean': 0.005434782709926367, 'clip_ratio/high_max': 0.005434782709926367, 'clip_ratio/region_mean': 0.014625959214754403, 'reward_total_mean': 0.15423895418643951, 'reward_meter_mean': 0.15423895418643951, 'reward_meter_std': 0.07883358746767044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.15423895418643951, 'reward_total_composite_std': 0.07883358746767044, 'epoch': 0.03}

 25%|██▌       | 827/3300 [24:08<4:12:15,  6.12s/it]INFO 04-11 23:00:59 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 828/3300 [24:13<3:54:39,  5.70s/it]2026-04-11 23:01:04,067 | INFO | train_grpo_train | metrics_logged mode=train step=828

                                                    
{'loss': 0.0145, 'grad_norm': 6.0185723304748535, 'learning_rate': 7.493939393939395e-06, 'num_tokens': 1839058.0, 'completions/mean_length': 63.5, 'completions/min_length': 61.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.5, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9804123640060425, 'rewards/meter/std': 0.014910156838595867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9804123640060425, 'rewards/total_composite/std': 0.014910156838595867, 'reward': 0.9804123640060425, 'reward_std': 0.014910157769918442, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028718305751681328, 'sampling/sampling_logp_difference/max': 1.3754212856292725, 'sampling/importance_sampling_ratio/min': 0.25273311138153076, 'sampling/importance_sampling_ratio/mean': 1.0006060600280762, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11623660661280155, 'clip_ratio/low_mean': 0.0038470644503831863, 'clip_ratio/low_min': 0.0038470644503831863, 'clip_ratio/high_mean': 0.01206992007791996, 'clip_ratio/high_max': 0.01206992007791996, 'clip_ratio/region_mean': 0.015916984528303146, 'reward_total_mean': 0.9804123640060425, 'reward_meter_mean': 0.9804123640060425, 'reward_meter_std': 0.014910156838595867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9804123640060425, 'reward_total_composite_std': 0.014910156838595867, 'epoch': 0.03}

 25%|██▌       | 828/3300 [24:13<3:54:39,  5.70s/it]INFO 04-11 23:01:04 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 829/3300 [24:19<3:58:39,  5.79s/it]2026-04-11 23:01:10,094 | INFO | train_grpo_train | metrics_logged mode=train step=829

                                                    
{'loss': 0.0432, 'grad_norm': 1.7107356786727905, 'learning_rate': 7.490909090909092e-06, 'num_tokens': 1841757.0, 'completions/mean_length': 173.375, 'completions/min_length': 151.0, 'completions/max_length': 186.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.375, 'completions/min_terminated_length': 151.0, 'completions/max_terminated_length': 186.0, 'rewards/meter/mean': 0.9961064457893372, 'rewards/meter/std': 0.0029631657525897026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.625, 'rewards/repeat_penalty/std': 0.16967642307281494, 'rewards/total_composite/mean': 0.6229597926139832, 'rewards/total_composite/std': 0.17044374346733093, 'reward': 0.6229597926139832, 'reward_std': 0.17044374346733093, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017439143732190132, 'sampling/sampling_logp_difference/max': 1.495316505432129, 'sampling/importance_sampling_ratio/min': 0.22417764365673065, 'sampling/importance_sampling_ratio/mean': 0.9979692101478577, 'sampling/importance_sampling_ratio/max': 1.7753796577453613, 'entropy': 0.05448292032815516, 'clip_ratio/low_mean': 0.002016129030380398, 'clip_ratio/low_min': 0.002016129030380398, 'clip_ratio/high_mean': 0.016328177880495787, 'clip_ratio/high_max': 0.016328177880495787, 'clip_ratio/region_mean': 0.018344306910876185, 'reward_total_mean': 0.6229597926139832, 'reward_meter_mean': 0.9961064457893372, 'reward_meter_std': 0.0029631657525897026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.625, 'reward_repeat_penalty_std': 0.16967642307281494, 'reward_total_composite_mean': 0.6229597926139832, 'reward_total_composite_std': 0.17044374346733093, 'epoch': 0.03}

 25%|██▌       | 829/3300 [24:19<3:58:39,  5.79s/it]INFO 04-11 23:01:10 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 830/3300 [24:24<3:56:24,  5.74s/it]2026-04-11 23:01:15,715 | INFO | train_grpo_train | metrics_logged mode=train step=830

                                                    
{'loss': 0.034, 'grad_norm': 12.176047325134277, 'learning_rate': 7.487878787878788e-06, 'num_tokens': 1843710.0, 'completions/mean_length': 74.125, 'completions/min_length': 68.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9026094079017639, 'rewards/meter/std': 0.14784321188926697, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9026094079017639, 'rewards/total_composite/std': 0.14784321188926697, 'reward': 0.9026094079017639, 'reward_std': 0.14784321188926697, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.060764994472265244, 'sampling/sampling_logp_difference/max': 5.8967156410217285, 'sampling/importance_sampling_ratio/min': 0.0027484570164233446, 'sampling/importance_sampling_ratio/mean': 0.9942511916160583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12799056991934776, 'clip_ratio/low_mean': 0.007117270142771304, 'clip_ratio/low_min': 0.007117270142771304, 'clip_ratio/high_mean': 0.019965628627687693, 'clip_ratio/high_max': 0.019965628627687693, 'clip_ratio/region_mean': 0.027082898770458996, 'reward_total_mean': 0.9026094079017639, 'reward_meter_mean': 0.9026094079017639, 'reward_meter_std': 0.14784321188926697, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9026094079017639, 'reward_total_composite_std': 0.14784321188926697, 'epoch': 0.03}

 25%|██▌       | 830/3300 [24:24<3:56:24,  5.74s/it]INFO 04-11 23:01:15 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 831/3300 [24:30<3:57:54,  5.78s/it]2026-04-11 23:01:21,586 | INFO | train_grpo_train | metrics_logged mode=train step=831

                                                    
{'loss': -0.003, 'grad_norm': 4.106361389160156, 'learning_rate': 7.484848484848486e-06, 'num_tokens': 1845975.0, 'completions/mean_length': 108.125, 'completions/min_length': 105.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.125, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.3786888122558594, 'rewards/meter/std': 0.35290807485580444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.3143823742866516, 'rewards/total_composite/std': 0.282669335603714, 'reward': 0.3143823742866516, 'reward_std': 0.282669335603714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034197013825178146, 'sampling/sampling_logp_difference/max': 2.057035207748413, 'sampling/importance_sampling_ratio/min': 0.12783241271972656, 'sampling/importance_sampling_ratio/mean': 0.9974236488342285, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12897953018546104, 'clip_ratio/low_mean': 0.016522707068361342, 'clip_ratio/low_min': 0.016522707068361342, 'clip_ratio/high_mean': 0.01371849060524255, 'clip_ratio/high_max': 0.01371849060524255, 'clip_ratio/region_mean': 0.030241197673603892, 'reward_total_mean': 0.3143823742866516, 'reward_meter_mean': 0.3786888122558594, 'reward_meter_std': 0.35290807485580444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.3143823742866516, 'reward_total_composite_std': 0.282669335603714, 'epoch': 0.03}

 25%|██▌       | 831/3300 [24:30<3:57:54,  5.78s/it]INFO 04-11 23:01:21 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 832/3300 [24:41<4:55:13,  7.18s/it]2026-04-11 23:01:32,020 | INFO | train_grpo_train | metrics_logged mode=train step=832

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.481818181818182e-06, 'num_tokens': 1847831.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.6360248923301697, 'rewards/meter/std': 0.25572946667671204, 'rewards/count_adherence/mean': 0.8602941036224365, 'rewards/count_adherence/std': 0.0304440688341856, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5477695465087891, 'rewards/repeat_penalty/std': 0.011733362451195717, 'rewards/total_composite/mean': 0.3008078336715698, 'rewards/total_composite/std': 0.12469936162233353, 'reward': 0.3008078336715698, 'reward_std': 0.12469936162233353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.3008078336715698, 'reward_meter_mean': 0.6360248923301697, 'reward_meter_std': 0.25572946667671204, 'reward_count_adherence_mean': 0.8602941036224365, 'reward_count_adherence_std': 0.0304440688341856, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5477695465087891, 'reward_repeat_penalty_std': 0.011733362451195717, 'reward_total_composite_mean': 0.3008078336715698, 'reward_total_composite_std': 0.12469936162233353, 'epoch': 0.03}

 25%|██▌       | 832/3300 [24:41<4:55:13,  7.18s/it]INFO 04-11 23:01:32 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 833/3300 [24:45<4:19:06,  6.30s/it]2026-04-11 23:01:36,282 | INFO | train_grpo_train | metrics_logged mode=train step=833

                                                    
{'loss': 0.0048, 'grad_norm': 8.433792114257812, 'learning_rate': 7.47878787878788e-06, 'num_tokens': 1849335.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9924166798591614, 'rewards/meter/std': 0.00017309709801338613, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924166798591614, 'rewards/total_composite/std': 0.00017309709801338613, 'reward': 0.9924166798591614, 'reward_std': 0.00017310312250629067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008245354518294334, 'sampling/sampling_logp_difference/max': 0.6460732221603394, 'sampling/importance_sampling_ratio/min': 0.5240997672080994, 'sampling/importance_sampling_ratio/mean': 1.0005563497543335, 'sampling/importance_sampling_ratio/max': 1.1176583766937256, 'entropy': 0.04187649488449097, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9924166798591614, 'reward_meter_mean': 0.9924166798591614, 'reward_meter_std': 0.00017309709801338613, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924166798591614, 'reward_total_composite_std': 0.00017309709801338613, 'epoch': 0.03}

 25%|██▌       | 833/3300 [24:45<4:19:06,  6.30s/it]INFO 04-11 23:01:36 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 834/3300 [24:50<4:00:18,  5.85s/it]2026-04-11 23:01:41,068 | INFO | train_grpo_train | metrics_logged mode=train step=834

                                                    
{'loss': 0.0342, 'grad_norm': 7.6343560218811035, 'learning_rate': 7.4757575757575765e-06, 'num_tokens': 1851111.0, 'completions/mean_length': 68.0, 'completions/min_length': 62.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.40049153566360474, 'rewards/meter/std': 0.3537753224372864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.40049153566360474, 'rewards/total_composite/std': 0.3537753224372864, 'reward': 0.40049153566360474, 'reward_std': 0.3537753224372864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0617922842502594, 'sampling/sampling_logp_difference/max': 3.229440689086914, 'sampling/importance_sampling_ratio/min': 0.03957962989807129, 'sampling/importance_sampling_ratio/mean': 1.0015736818313599, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33059297781437635, 'clip_ratio/low_mean': 0.035882155993022025, 'clip_ratio/low_min': 0.035882155993022025, 'clip_ratio/high_mean': 0.009836265817284584, 'clip_ratio/high_max': 0.009836265817284584, 'clip_ratio/region_mean': 0.04571842181030661, 'reward_total_mean': 0.40049153566360474, 'reward_meter_mean': 0.40049153566360474, 'reward_meter_std': 0.3537753224372864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.40049153566360474, 'reward_total_composite_std': 0.3537753224372864, 'epoch': 0.03}

 25%|██▌       | 834/3300 [24:50<4:00:18,  5.85s/it]INFO 04-11 23:01:41 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 835/3300 [24:55<3:59:21,  5.83s/it]2026-04-11 23:01:46,843 | INFO | train_grpo_train | metrics_logged mode=train step=835

                                                    
{'loss': 0.0177, 'grad_norm': 1.6731441020965576, 'learning_rate': 7.472727272727274e-06, 'num_tokens': 1853656.0, 'completions/mean_length': 150.125, 'completions/min_length': 149.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.125, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.9238545894622803, 'rewards/meter/std': 0.2039637267589569, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6159030199050903, 'rewards/total_composite/std': 0.13597580790519714, 'reward': 0.6159030199050903, 'reward_std': 0.13597580790519714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00628118310123682, 'sampling/sampling_logp_difference/max': 1.1030478477478027, 'sampling/importance_sampling_ratio/min': 0.3318580687046051, 'sampling/importance_sampling_ratio/mean': 0.9989848136901855, 'sampling/importance_sampling_ratio/max': 1.6532901525497437, 'entropy': 0.03071731375530362, 'clip_ratio/low_mean': 0.0007961783558130264, 'clip_ratio/low_min': 0.0007961783558130264, 'clip_ratio/high_mean': 0.004172259592451155, 'clip_ratio/high_max': 0.004172259592451155, 'clip_ratio/region_mean': 0.004968437948264182, 'reward_total_mean': 0.6159030199050903, 'reward_meter_mean': 0.9238545894622803, 'reward_meter_std': 0.2039637267589569, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6159030199050903, 'reward_total_composite_std': 0.13597580790519714, 'epoch': 0.03}

 25%|██▌       | 835/3300 [24:55<3:59:21,  5.83s/it]INFO 04-11 23:01:47 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 836/3300 [25:01<3:57:58,  5.79s/it]2026-04-11 23:01:52,569 | INFO | train_grpo_train | metrics_logged mode=train step=836

                                                    
{'loss': -0.0007, 'grad_norm': 2.8023297786712646, 'learning_rate': 7.46969696969697e-06, 'num_tokens': 1856194.0, 'completions/mean_length': 142.25, 'completions/min_length': 139.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.25, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.8496776819229126, 'rewards/meter/std': 0.3326415419578552, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8214285373687744, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.7131145596504211, 'rewards/total_composite/std': 0.2920505106449127, 'reward': 0.7131145596504211, 'reward_std': 0.29205048084259033, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012105772271752357, 'sampling/sampling_logp_difference/max': 1.3084237575531006, 'sampling/importance_sampling_ratio/min': 0.2702457010746002, 'sampling/importance_sampling_ratio/mean': 0.9984580278396606, 'sampling/importance_sampling_ratio/max': 1.5496021509170532, 'entropy': 0.044805840123444796, 'clip_ratio/low_mean': 0.001773406460415572, 'clip_ratio/low_min': 0.001773406460415572, 'clip_ratio/high_mean': 0.007069471699651331, 'clip_ratio/high_max': 0.007069471699651331, 'clip_ratio/region_mean': 0.008842878160066903, 'reward_total_mean': 0.7131145596504211, 'reward_meter_mean': 0.8496776819229126, 'reward_meter_std': 0.3326415419578552, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8214285373687744, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.7131145596504211, 'reward_total_composite_std': 0.2920505106449127, 'epoch': 0.03}

 25%|██▌       | 836/3300 [25:01<3:57:58,  5.79s/it]INFO 04-11 23:01:52 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 837/3300 [25:05<3:38:07,  5.31s/it]2026-04-11 23:01:56,759 | INFO | train_grpo_train | metrics_logged mode=train step=837

                                                    
{'loss': 0.0053, 'grad_norm': 11.809463500976562, 'learning_rate': 7.4666666666666675e-06, 'num_tokens': 1857618.0, 'completions/mean_length': 33.0, 'completions/min_length': 33.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9952504634857178, 'rewards/meter/std': 0.0004852505517192185, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952504634857178, 'rewards/total_composite/std': 0.0004852505517192185, 'reward': 0.9952504634857178, 'reward_std': 0.0004852571291849017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017783869057893753, 'sampling/sampling_logp_difference/max': 0.7500922679901123, 'sampling/importance_sampling_ratio/min': 0.4723230004310608, 'sampling/importance_sampling_ratio/mean': 1.0051133632659912, 'sampling/importance_sampling_ratio/max': 1.5762113332748413, 'entropy': 0.0944258663803339, 'clip_ratio/low_mean': 0.01515151560306549, 'clip_ratio/low_min': 0.01515151560306549, 'clip_ratio/high_mean': 0.011363636702299118, 'clip_ratio/high_max': 0.011363636702299118, 'clip_ratio/region_mean': 0.02651515230536461, 'reward_total_mean': 0.9952504634857178, 'reward_meter_mean': 0.9952504634857178, 'reward_meter_std': 0.0004852505517192185, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952504634857178, 'reward_total_composite_std': 0.0004852505517192185, 'epoch': 0.03}

 25%|██▌       | 837/3300 [25:05<3:38:07,  5.31s/it]INFO 04-11 23:01:57 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 838/3300 [25:13<4:10:03,  6.09s/it]2026-04-11 23:02:04,669 | INFO | train_grpo_train | metrics_logged mode=train step=838

                                                    
{'loss': 0.0061, 'grad_norm': 2.168991804122925, 'learning_rate': 7.463636363636364e-06, 'num_tokens': 1861717.0, 'completions/mean_length': 295.375, 'completions/min_length': 264.0, 'completions/max_length': 331.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 295.375, 'completions/min_terminated_length': 264.0, 'completions/max_terminated_length': 331.0, 'rewards/meter/mean': 0.869155764579773, 'rewards/meter/std': 0.35062167048454285, 'rewards/count_adherence/mean': 0.828125, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5626935362815857, 'rewards/repeat_penalty/std': 0.04896574467420578, 'rewards/total_composite/mean': 0.4079222083091736, 'rewards/total_composite/std': 0.1708296537399292, 'reward': 0.4079222083091736, 'reward_std': 0.170829638838768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009027719497680664, 'sampling/sampling_logp_difference/max': 1.8500938415527344, 'sampling/importance_sampling_ratio/min': 0.15722240507602692, 'sampling/importance_sampling_ratio/mean': 0.9985241889953613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.02827131818048656, 'clip_ratio/low_mean': 0.0017137863032985479, 'clip_ratio/low_min': 0.0017137863032985479, 'clip_ratio/high_mean': 0.004315092111937702, 'clip_ratio/high_max': 0.004315092111937702, 'clip_ratio/region_mean': 0.0060288784152362496, 'reward_total_mean': 0.4079222083091736, 'reward_meter_mean': 0.869155764579773, 'reward_meter_std': 0.35062167048454285, 'reward_count_adherence_mean': 0.828125, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5626935362815857, 'reward_repeat_penalty_std': 0.04896574467420578, 'reward_total_composite_mean': 0.4079222083091736, 'reward_total_composite_std': 0.1708296537399292, 'epoch': 0.03}

 25%|██▌       | 838/3300 [25:13<4:10:03,  6.09s/it]INFO 04-11 23:02:04 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 839/3300 [25:18<4:00:22,  5.86s/it]2026-04-11 23:02:09,984 | INFO | train_grpo_train | metrics_logged mode=train step=839

                                                    
{'loss': 0.0131, 'grad_norm': 2.164283275604248, 'learning_rate': 7.460606060606061e-06, 'num_tokens': 1863934.0, 'completions/mean_length': 112.125, 'completions/min_length': 109.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.125, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.9943764209747314, 'rewards/meter/std': 0.0010588886216282845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.795501172542572, 'rewards/total_composite/std': 0.0008471080800518394, 'reward': 0.795501172542572, 'reward_std': 0.0008471118635497987, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008892908692359924, 'sampling/sampling_logp_difference/max': 0.7849429249763489, 'sampling/importance_sampling_ratio/min': 0.4561457335948944, 'sampling/importance_sampling_ratio/mean': 1.0025136470794678, 'sampling/importance_sampling_ratio/max': 1.7390761375427246, 'entropy': 0.047560357954353094, 'clip_ratio/low_mean': 0.006605345057323575, 'clip_ratio/low_min': 0.006605345057323575, 'clip_ratio/high_mean': 0.0022935778833925724, 'clip_ratio/high_max': 0.0022935778833925724, 'clip_ratio/region_mean': 0.008898922940716147, 'reward_total_mean': 0.795501172542572, 'reward_meter_mean': 0.9943764209747314, 'reward_meter_std': 0.0010588886216282845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.795501172542572, 'reward_total_composite_std': 0.0008471080800518394, 'epoch': 0.03}

 25%|██▌       | 839/3300 [25:18<4:00:22,  5.86s/it]INFO 04-11 23:02:10 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 840/3300 [25:23<3:46:58,  5.54s/it]2026-04-11 23:02:14,773 | INFO | train_grpo_train | metrics_logged mode=train step=840

                                                    
{'loss': 0.036, 'grad_norm': 5.7492289543151855, 'learning_rate': 7.4575757575757575e-06, 'num_tokens': 1865375.0, 'completions/mean_length': 37.125, 'completions/min_length': 36.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7395036816596985, 'rewards/meter/std': 0.2953052818775177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7395036816596985, 'rewards/total_composite/std': 0.2953052818775177, 'reward': 0.7395036816596985, 'reward_std': 0.2953052818775177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01924203522503376, 'sampling/sampling_logp_difference/max': 0.40451645851135254, 'sampling/importance_sampling_ratio/min': 0.6779606938362122, 'sampling/importance_sampling_ratio/mean': 1.0085185766220093, 'sampling/importance_sampling_ratio/max': 1.4985777139663696, 'entropy': 0.11148730758577585, 'clip_ratio/low_mean': 0.006410256493836641, 'clip_ratio/low_min': 0.006410256493836641, 'clip_ratio/high_mean': 0.010228979168459773, 'clip_ratio/high_max': 0.010228979168459773, 'clip_ratio/region_mean': 0.016639235662296414, 'reward_total_mean': 0.7395036816596985, 'reward_meter_mean': 0.7395036816596985, 'reward_meter_std': 0.2953052818775177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7395036816596985, 'reward_total_composite_std': 0.2953052818775177, 'epoch': 0.03}

 25%|██▌       | 840/3300 [25:23<3:46:58,  5.54s/it]INFO 04-11 23:02:15 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 841/3300 [25:28<3:38:14,  5.33s/it]2026-04-11 23:02:19,615 | INFO | train_grpo_train | metrics_logged mode=train step=841

                                                    
{'loss': -0.0141, 'grad_norm': 3.028710126876831, 'learning_rate': 7.454545454545456e-06, 'num_tokens': 1867245.0, 'completions/mean_length': 64.75, 'completions/min_length': 64.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9528952240943909, 'rewards/meter/std': 0.01815449446439743, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9528952240943909, 'rewards/total_composite/std': 0.01815449446439743, 'reward': 0.9528952240943909, 'reward_std': 0.018154479563236237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01176523882895708, 'sampling/sampling_logp_difference/max': 0.8644721508026123, 'sampling/importance_sampling_ratio/min': 0.5180370807647705, 'sampling/importance_sampling_ratio/mean': 1.0040351152420044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.045040544122457504, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.005771921598352492, 'reward_total_mean': 0.9528952240943909, 'reward_meter_mean': 0.9528952240943909, 'reward_meter_std': 0.01815449446439743, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9528952240943909, 'reward_total_composite_std': 0.01815449446439743, 'epoch': 0.03}

 25%|██▌       | 841/3300 [25:28<3:38:14,  5.33s/it]INFO 04-11 23:02:19 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 842/3300 [25:32<3:26:59,  5.05s/it]2026-04-11 23:02:24,014 | INFO | train_grpo_train | metrics_logged mode=train step=842

                                                    
{'loss': 0.0297, 'grad_norm': 9.17272663116455, 'learning_rate': 7.451515151515152e-06, 'num_tokens': 1868749.0, 'completions/mean_length': 37.0, 'completions/min_length': 36.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.700322687625885, 'rewards/meter/std': 0.3828321695327759, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.700322687625885, 'rewards/total_composite/std': 0.3828321695327759, 'reward': 0.700322687625885, 'reward_std': 0.3828321397304535, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011855069547891617, 'sampling/sampling_logp_difference/max': 0.964850664138794, 'sampling/importance_sampling_ratio/min': 0.6239942312240601, 'sampling/importance_sampling_ratio/mean': 1.0074673891067505, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04681507125496864, 'clip_ratio/low_mean': 0.009699730202555656, 'clip_ratio/low_min': 0.009699730202555656, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.009699730202555656, 'reward_total_mean': 0.700322687625885, 'reward_meter_mean': 0.700322687625885, 'reward_meter_std': 0.3828321695327759, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.700322687625885, 'reward_total_composite_std': 0.3828321695327759, 'epoch': 0.03}

 26%|██▌       | 842/3300 [25:33<3:26:59,  5.05s/it]INFO 04-11 23:02:24 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 843/3300 [25:42<4:23:09,  6.43s/it]2026-04-11 23:02:33,646 | INFO | train_grpo_train | metrics_logged mode=train step=843

                                                    
{'loss': 0.0211, 'grad_norm': 2.23907470703125, 'learning_rate': 7.448484848484849e-06, 'num_tokens': 1874562.0, 'completions/mean_length': 452.625, 'completions/min_length': 426.0, 'completions/max_length': 461.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 452.625, 'completions/min_terminated_length': 426.0, 'completions/max_terminated_length': 461.0, 'rewards/meter/mean': 0.9962563514709473, 'rewards/meter/std': 0.001195572316646576, 'rewards/count_adherence/mean': 0.9270833730697632, 'rewards/count_adherence/std': 0.029462777078151703, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5606521368026733, 'rewards/repeat_penalty/std': 0.0018446200992912054, 'rewards/total_composite/mean': 0.5178769826889038, 'rewards/total_composite/std': 0.01842796988785267, 'reward': 0.5178769826889038, 'reward_std': 0.018427973613142967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004932690877467394, 'sampling/sampling_logp_difference/max': 4.1077117919921875, 'sampling/importance_sampling_ratio/min': 0.01644536294043064, 'sampling/importance_sampling_ratio/mean': 0.999667763710022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.01467959355795756, 'clip_ratio/low_mean': 0.0011018531513400376, 'clip_ratio/low_min': 0.0011018531513400376, 'clip_ratio/high_mean': 0.0002934272342827171, 'clip_ratio/high_max': 0.0002934272342827171, 'clip_ratio/region_mean': 0.0013952803856227547, 'reward_total_mean': 0.5178769826889038, 'reward_meter_mean': 0.9962563514709473, 'reward_meter_std': 0.001195572316646576, 'reward_count_adherence_mean': 0.9270833730697632, 'reward_count_adherence_std': 0.029462777078151703, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5606521368026733, 'reward_repeat_penalty_std': 0.0018446200992912054, 'reward_total_composite_mean': 0.5178769826889038, 'reward_total_composite_std': 0.01842796988785267, 'epoch': 0.03}

 26%|██▌       | 843/3300 [25:42<4:23:09,  6.43s/it]INFO 04-11 23:02:33 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 844/3300 [25:48<4:15:14,  6.24s/it]2026-04-11 23:02:39,436 | INFO | train_grpo_train | metrics_logged mode=train step=844

                                                    
{'loss': -0.0116, 'grad_norm': 2.640986919403076, 'learning_rate': 7.445454545454546e-06, 'num_tokens': 1877374.0, 'completions/mean_length': 149.5, 'completions/min_length': 144.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 149.5, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.5467606782913208, 'rewards/meter/std': 0.4782540798187256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.3916763961315155, 'rewards/total_composite/std': 0.34031400084495544, 'reward': 0.3916763961315155, 'reward_std': 0.34031397104263306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019425617530941963, 'sampling/sampling_logp_difference/max': 3.396358013153076, 'sampling/importance_sampling_ratio/min': 0.03349503502249718, 'sampling/importance_sampling_ratio/mean': 1.0008183717727661, 'sampling/importance_sampling_ratio/max': 1.7641745805740356, 'entropy': 0.09837253391742706, 'clip_ratio/low_mean': 0.011086393264122307, 'clip_ratio/low_min': 0.011086393264122307, 'clip_ratio/high_mean': 0.002457805967424065, 'clip_ratio/high_max': 0.002457805967424065, 'clip_ratio/region_mean': 0.013544199231546372, 'reward_total_mean': 0.3916763961315155, 'reward_meter_mean': 0.5467606782913208, 'reward_meter_std': 0.4782540798187256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.3916763961315155, 'reward_total_composite_std': 0.34031400084495544, 'epoch': 0.03}

 26%|██▌       | 844/3300 [25:48<4:15:14,  6.24s/it]INFO 04-11 23:02:39 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 845/3300 [25:53<4:03:29,  5.95s/it]2026-04-11 23:02:44,722 | INFO | train_grpo_train | metrics_logged mode=train step=845

                                                    
{'loss': 0.0112, 'grad_norm': 6.265551567077637, 'learning_rate': 7.442424242424243e-06, 'num_tokens': 1879205.0, 'completions/mean_length': 71.875, 'completions/min_length': 68.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.522850513458252, 'rewards/meter/std': 0.4167693555355072, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.5038976669311523, 'rewards/total_composite/std': 0.42369261384010315, 'reward': 0.5038976669311523, 'reward_std': 0.42369258403778076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03976872190833092, 'sampling/sampling_logp_difference/max': 1.2724390029907227, 'sampling/importance_sampling_ratio/min': 0.2801474928855896, 'sampling/importance_sampling_ratio/mean': 1.0023622512817383, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16639000456780195, 'clip_ratio/low_mean': 0.01767799479421228, 'clip_ratio/low_min': 0.01767799479421228, 'clip_ratio/high_mean': 0.017289764247834682, 'clip_ratio/high_max': 0.017289764247834682, 'clip_ratio/region_mean': 0.034967759042046964, 'reward_total_mean': 0.5038976669311523, 'reward_meter_mean': 0.522850513458252, 'reward_meter_std': 0.4167693555355072, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.5038976669311523, 'reward_total_composite_std': 0.42369261384010315, 'epoch': 0.03}

 26%|██▌       | 845/3300 [25:53<4:03:29,  5.95s/it]INFO 04-11 23:02:44 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 846/3300 [26:01<4:30:27,  6.61s/it]2026-04-11 23:02:52,881 | INFO | train_grpo_train | metrics_logged mode=train step=846

                                                    
{'loss': -0.0053, 'grad_norm': 4.094456672668457, 'learning_rate': 7.439393939393939e-06, 'num_tokens': 1883732.0, 'completions/mean_length': 321.875, 'completions/min_length': 309.0, 'completions/max_length': 336.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 321.875, 'completions/min_terminated_length': 309.0, 'completions/max_terminated_length': 336.0, 'rewards/meter/mean': 0.6540272831916809, 'rewards/meter/std': 0.39356982707977295, 'rewards/count_adherence/mean': 0.9861111044883728, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4677932560443878, 'rewards/repeat_penalty/std': 0.22017233073711395, 'rewards/total_composite/mean': 0.3497573733329773, 'rewards/total_composite/std': 0.2734247148036957, 'reward': 0.3497573733329773, 'reward_std': 0.2734247148036957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007702399045228958, 'sampling/sampling_logp_difference/max': 1.841371774673462, 'sampling/importance_sampling_ratio/min': 0.15859971940517426, 'sampling/importance_sampling_ratio/mean': 0.9985901713371277, 'sampling/importance_sampling_ratio/max': 1.768311619758606, 'entropy': 0.03188518015667796, 'clip_ratio/low_mean': 0.0011160714784637094, 'clip_ratio/low_min': 0.0011160714784637094, 'clip_ratio/high_mean': 0.0058011687360703945, 'clip_ratio/high_max': 0.0058011687360703945, 'clip_ratio/region_mean': 0.006917240214534104, 'reward_total_mean': 0.3497573733329773, 'reward_meter_mean': 0.6540272831916809, 'reward_meter_std': 0.39356982707977295, 'reward_count_adherence_mean': 0.9861111044883728, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4677932560443878, 'reward_repeat_penalty_std': 0.22017233073711395, 'reward_total_composite_mean': 0.3497573733329773, 'reward_total_composite_std': 0.2734247148036957, 'epoch': 0.03}

 26%|██▌       | 846/3300 [26:01<4:30:27,  6.61s/it]INFO 04-11 23:02:53 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 847/3300 [26:06<4:09:25,  6.10s/it]2026-04-11 23:02:57,786 | INFO | train_grpo_train | metrics_logged mode=train step=847

                                                    
{'loss': 0.045, 'grad_norm': 6.007223606109619, 'learning_rate': 7.4363636363636375e-06, 'num_tokens': 1885713.0, 'completions/mean_length': 72.625, 'completions/min_length': 68.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.26388847827911377, 'rewards/meter/std': 0.3744986653327942, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.2634860873222351, 'rewards/total_composite/std': 0.3748124837875366, 'reward': 0.2634860873222351, 'reward_std': 0.3748124837875366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04990806803107262, 'sampling/sampling_logp_difference/max': 1.3789244890213013, 'sampling/importance_sampling_ratio/min': 0.2518492639064789, 'sampling/importance_sampling_ratio/mean': 1.007526159286499, 'sampling/importance_sampling_ratio/max': 1.518433690071106, 'entropy': 0.2849195022135973, 'clip_ratio/low_mean': 0.02969917980954051, 'clip_ratio/low_min': 0.02969917980954051, 'clip_ratio/high_mean': 0.019930581096559763, 'clip_ratio/high_max': 0.019930581096559763, 'clip_ratio/region_mean': 0.04962976090610027, 'reward_total_mean': 0.2634860873222351, 'reward_meter_mean': 0.26388847827911377, 'reward_meter_std': 0.3744986653327942, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.2634860873222351, 'reward_total_composite_std': 0.3748124837875366, 'epoch': 0.03}

 26%|██▌       | 847/3300 [26:06<4:09:25,  6.10s/it]INFO 04-11 23:02:58 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 848/3300 [26:11<3:58:18,  5.83s/it]2026-04-11 23:03:02,988 | INFO | train_grpo_train | metrics_logged mode=train step=848

                                                    
{'loss': 0.0001, 'grad_norm': 2.973163366317749, 'learning_rate': 7.433333333333334e-06, 'num_tokens': 1888010.0, 'completions/mean_length': 106.125, 'completions/min_length': 100.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.125, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.6358721256256104, 'rewards/meter/std': 0.32246294617652893, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.5016140341758728, 'rewards/total_composite/std': 0.2674255073070526, 'reward': 0.5016140341758728, 'reward_std': 0.2674255073070526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022296659648418427, 'sampling/sampling_logp_difference/max': 1.8602898120880127, 'sampling/importance_sampling_ratio/min': 0.1556275188922882, 'sampling/importance_sampling_ratio/mean': 0.9971781969070435, 'sampling/importance_sampling_ratio/max': 1.6591869592666626, 'entropy': 0.08717937674373388, 'clip_ratio/low_mean': 0.005942982388660312, 'clip_ratio/low_min': 0.005942982388660312, 'clip_ratio/high_mean': 0.013045326224528253, 'clip_ratio/high_max': 0.013045326224528253, 'clip_ratio/region_mean': 0.018988308613188565, 'reward_total_mean': 0.5016140341758728, 'reward_meter_mean': 0.6358721256256104, 'reward_meter_std': 0.32246294617652893, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.5016140341758728, 'reward_total_composite_std': 0.2674255073070526, 'epoch': 0.03}

 26%|██▌       | 848/3300 [26:11<3:58:18,  5.83s/it]INFO 04-11 23:03:03 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 849/3300 [26:19<4:16:34,  6.28s/it]2026-04-11 23:03:10,319 | INFO | train_grpo_train | metrics_logged mode=train step=849

                                                    
{'loss': 0.001, 'grad_norm': 0.23050236701965332, 'learning_rate': 7.430303030303031e-06, 'num_tokens': 1892024.0, 'completions/mean_length': 283.75, 'completions/min_length': 283.0, 'completions/max_length': 284.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 283.75, 'completions/min_terminated_length': 283.0, 'completions/max_terminated_length': 284.0, 'rewards/meter/mean': 0.9965326189994812, 'rewards/meter/std': 0.00021112659305799752, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6000000238418579, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5979195833206177, 'rewards/total_composite/std': 0.00012666928523685783, 'reward': 0.5979195833206177, 'reward_std': 0.000126663115224801, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0016167466528713703, 'sampling/sampling_logp_difference/max': 0.7609295845031738, 'sampling/importance_sampling_ratio/min': 0.4672318696975708, 'sampling/importance_sampling_ratio/mean': 1.0001165866851807, 'sampling/importance_sampling_ratio/max': 1.2486354112625122, 'entropy': 0.010985135799273849, 'clip_ratio/low_mean': 0.00044014083687216043, 'clip_ratio/low_min': 0.00044014083687216043, 'clip_ratio/high_mean': 0.00044169611646793783, 'clip_ratio/high_max': 0.00044169611646793783, 'clip_ratio/region_mean': 0.0008818369533400983, 'reward_total_mean': 0.5979195833206177, 'reward_meter_mean': 0.9965326189994812, 'reward_meter_std': 0.00021112659305799752, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6000000238418579, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5979195833206177, 'reward_total_composite_std': 0.00012666928523685783, 'epoch': 0.03}

 26%|██▌       | 849/3300 [26:19<4:16:34,  6.28s/it]INFO 04-11 23:03:10 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 850/3300 [26:24<4:00:12,  5.88s/it]2026-04-11 23:03:15,274 | INFO | train_grpo_train | metrics_logged mode=train step=850

                                                    
{'loss': -0.0004, 'grad_norm': 7.60121488571167, 'learning_rate': 7.4272727272727275e-06, 'num_tokens': 1893564.0, 'completions/mean_length': 35.5, 'completions/min_length': 33.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9917958974838257, 'rewards/meter/std': 0.008829712867736816, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8676612377166748, 'rewards/total_composite/std': 0.35069888830184937, 'reward': 0.8676612377166748, 'reward_std': 0.350698858499527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04963725060224533, 'sampling/sampling_logp_difference/max': 3.3897039890289307, 'sampling/importance_sampling_ratio/min': 0.03371865674853325, 'sampling/importance_sampling_ratio/mean': 1.0052940845489502, 'sampling/importance_sampling_ratio/max': 1.9962736368179321, 'entropy': 0.15574337635189295, 'clip_ratio/low_mean': 0.0071428571827709675, 'clip_ratio/low_min': 0.0071428571827709675, 'clip_ratio/high_mean': 0.031166881788522005, 'clip_ratio/high_max': 0.031166881788522005, 'clip_ratio/region_mean': 0.03830973897129297, 'reward_total_mean': 0.8676612377166748, 'reward_meter_mean': 0.9917958974838257, 'reward_meter_std': 0.008829712867736816, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8676612377166748, 'reward_total_composite_std': 0.35069888830184937, 'epoch': 0.03}

 26%|██▌       | 850/3300 [26:24<4:00:12,  5.88s/it]INFO 04-11 23:03:15 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:35,  3.19s/it]

 23%|██▎       | 3/13 [00:09<00:33,  3.31s/it]

 31%|███       | 4/13 [00:15<00:37,  4.15s/it]

 38%|███▊      | 5/13 [00:20<00:36,  4.61s/it]

 46%|████▌     | 6/13 [00:27<00:38,  5.43s/it]

 54%|█████▍    | 7/13 [00:33<00:32,  5.42s/it]

 62%|██████▏   | 8/13 [00:39<00:28,  5.77s/it]

 69%|██████▉   | 9/13 [00:46<00:24,  6.18s/it]

 77%|███████▋  | 10/13 [00:51<00:17,  5.68s/it]

 85%|████████▍ | 11/13 [00:56<00:11,  5.53s/it]

 92%|█████████▏| 12/13 [01:02<00:05,  5.66s/it]

100%|██████████| 13/13 [01:08<00:00,  5.74s/it]2026-04-11 23:04:30,996 | INFO | train_grpo_train | metrics_logged mode=eval step=850

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 75.7065, 'eval_samples_per_second': 1.374, 'eval_steps_per_second': 0.172, 'eval_num_tokens': 1893564.0, 'eval_completions/mean_length': 209.85576923076923, 'eval_completions/min_length': 62.0, 'eval_completions/max_length': 402.6923076923077, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 206.50412104679987, 'eval_completions/min_terminated_length': 62.0, 'eval_completions/max_terminated_length': 389.9230769230769, 'eval_rewards/meter/mean': 0.6430152883896461, 'eval_rewards/meter/std': 0.3854568119232471, 'eval_rewards/count_adherence/mean': 0.9535174920008733, 'eval_rewards/count_adherence/std': 0.07147554422800358, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.6578034116671636, 'eval_rewards/repeat_penalty/std': 0.21721924497531012, 'eval_rewards/total_composite/mean': 0.4121822485556969, 'eval_rewards/total_composite/std': 0.3158151931487597, 'eval_reward': 0.4121822485556969, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.00518767936871602, 'eval_sampling/sampling_logp_difference/max': 0.7618373265633216, 'eval_sampling/importance_sampling_ratio/min': 0.4886489510536194, 'eval_sampling/importance_sampling_ratio/mean': 1.001105854144463, 'eval_sampling/importance_sampling_ratio/max': 1.308758864035973, 'eval_entropy': 0.046050483074325785, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.4121822485556969, 'eval_reward_meter_mean': 0.6430152883896461, 'eval_reward_meter_std': 0.3854568119232471, 'eval_reward_count_adherence_mean': 0.9535174920008733, 'eval_reward_count_adherence_std': 0.07147554422800358, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.6578034116671636, 'eval_reward_repeat_penalty_std': 0.21721924497531012, 'eval_reward_total_composite_mean': 0.4121822485556969, 'eval_reward_total_composite_std': 0.3158151931487597, 'epoch': 0.03}

 26%|██▌       | 850/3300 [27:39<4:00:12,  5.88s/it]

100%|██████████| 13/13 [01:08<00:00,  5.74s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 26%|██▌       | 851/3300 [27:48<19:58:34, 29.36s/it]2026-04-11 23:04:39,430 | INFO | train_grpo_train | metrics_logged mode=train step=851

                                                     
{'loss': 0.0004, 'grad_norm': 2.013568162918091, 'learning_rate': 7.424242424242425e-06, 'num_tokens': 1896556.0, 'completions/mean_length': 186.0, 'completions/min_length': 178.0, 'completions/max_length': 199.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.0, 'completions/min_terminated_length': 178.0, 'completions/max_terminated_length': 199.0, 'rewards/meter/mean': 0.989942193031311, 'rewards/meter/std': 0.007896743714809418, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5795454382896423, 'rewards/repeat_penalty/std': 0.16070608794689178, 'rewards/total_composite/mean': 0.5742474794387817, 'rewards/total_composite/std': 0.16005170345306396, 'reward': 0.5742474794387817, 'reward_std': 0.16005171835422516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011838030070066452, 'sampling/sampling_logp_difference/max': 1.5690075159072876, 'sampling/importance_sampling_ratio/min': 0.3101670742034912, 'sampling/importance_sampling_ratio/mean': 1.0001335144042969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.033908116864040494, 'clip_ratio/low_mean': 0.0013513513840734959, 'clip_ratio/low_min': 0.0013513513840734959, 'clip_ratio/high_mean': 0.00752471067244187, 'clip_ratio/high_max': 0.00752471067244187, 'clip_ratio/region_mean': 0.008876062056515366, 'reward_total_mean': 0.5742474794387817, 'reward_meter_mean': 0.989942193031311, 'reward_meter_std': 0.007896743714809418, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5795454382896423, 'reward_repeat_penalty_std': 0.16070608794689178, 'reward_total_composite_mean': 0.5742474794387817, 'reward_total_composite_std': 0.16005170345306396, 'epoch': 0.03}

 26%|██▌       | 851/3300 [27:48<19:58:34, 29.36s/it]INFO 04-11 23:04:39 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 852/3300 [27:53<15:01:08, 22.09s/it]2026-04-11 23:04:44,534 | INFO | train_grpo_train | metrics_logged mode=train step=852

                                                     
{'loss': -0.0116, 'grad_norm': 6.178599834442139, 'learning_rate': 7.421212121212121e-06, 'num_tokens': 1898384.0, 'completions/mean_length': 72.5, 'completions/min_length': 68.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6383301019668579, 'rewards/meter/std': 0.34624141454696655, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.6031950116157532, 'rewards/total_composite/std': 0.3757425546646118, 'reward': 0.6031950116157532, 'reward_std': 0.3757425546646118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03135836869478226, 'sampling/sampling_logp_difference/max': 2.203850746154785, 'sampling/importance_sampling_ratio/min': 0.11037729680538177, 'sampling/importance_sampling_ratio/mean': 0.9961123466491699, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10962006729096174, 'clip_ratio/low_mean': 0.006956097553484142, 'clip_ratio/low_min': 0.006956097553484142, 'clip_ratio/high_mean': 0.006779896444641054, 'clip_ratio/high_max': 0.006779896444641054, 'clip_ratio/region_mean': 0.013735993998125196, 'reward_total_mean': 0.6031950116157532, 'reward_meter_mean': 0.6383301019668579, 'reward_meter_std': 0.34624141454696655, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.6031950116157532, 'reward_total_composite_std': 0.3757425546646118, 'epoch': 0.03}

 26%|██▌       | 852/3300 [27:53<15:01:08, 22.09s/it]INFO 04-11 23:04:44 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 853/3300 [28:02<12:18:50, 18.12s/it]2026-04-11 23:04:53,386 | INFO | train_grpo_train | metrics_logged mode=train step=853

                                                     
{'loss': 0.0349, 'grad_norm': 1.2332594394683838, 'learning_rate': 7.4181818181818185e-06, 'num_tokens': 1903150.0, 'completions/mean_length': 373.75, 'completions/min_length': 351.0, 'completions/max_length': 405.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 373.75, 'completions/min_terminated_length': 351.0, 'completions/max_terminated_length': 405.0, 'rewards/meter/mean': 0.6268174648284912, 'rewards/meter/std': 0.39189305901527405, 'rewards/count_adherence/mean': 0.7692307829856873, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5203947424888611, 'rewards/repeat_penalty/std': 0.10010629892349243, 'rewards/total_composite/mean': 0.2684391736984253, 'rewards/total_composite/std': 0.18215535581111908, 'reward': 0.2684391736984253, 'reward_std': 0.18215534090995789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00892745703458786, 'sampling/sampling_logp_difference/max': 1.788743495941162, 'sampling/importance_sampling_ratio/min': 0.1743081957101822, 'sampling/importance_sampling_ratio/mean': 1.0005637407302856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03790131723508239, 'clip_ratio/low_mean': 0.002888624498154968, 'clip_ratio/low_min': 0.002888624498154968, 'clip_ratio/high_mean': 0.0017242564936168492, 'clip_ratio/high_max': 0.0017242564936168492, 'clip_ratio/region_mean': 0.004612880991771817, 'reward_total_mean': 0.2684391736984253, 'reward_meter_mean': 0.6268174648284912, 'reward_meter_std': 0.39189305901527405, 'reward_count_adherence_mean': 0.7692307829856873, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5203947424888611, 'reward_repeat_penalty_std': 0.10010629892349243, 'reward_total_composite_mean': 0.2684391736984253, 'reward_total_composite_std': 0.18215535581111908, 'epoch': 0.03}

 26%|██▌       | 853/3300 [28:02<12:18:50, 18.12s/it]INFO 04-11 23:04:53 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 854/3300 [28:07<9:45:28, 14.36s/it] 2026-04-11 23:04:58,988 | INFO | train_grpo_train | metrics_logged mode=train step=854

                                                    
{'loss': 0.023, 'grad_norm': 3.387570858001709, 'learning_rate': 7.415151515151515e-06, 'num_tokens': 1905516.0, 'completions/mean_length': 125.75, 'completions/min_length': 118.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.75, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9104515314102173, 'rewards/meter/std': 0.12032636255025864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6607142686843872, 'rewards/repeat_penalty/std': 0.15152288973331451, 'rewards/total_composite/mean': 0.5973337292671204, 'rewards/total_composite/std': 0.1518002152442932, 'reward': 0.5973337292671204, 'reward_std': 0.1518002152442932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014981008134782314, 'sampling/sampling_logp_difference/max': 2.355228900909424, 'sampling/importance_sampling_ratio/min': 0.09487178921699524, 'sampling/importance_sampling_ratio/mean': 0.9986640214920044, 'sampling/importance_sampling_ratio/max': 1.784404993057251, 'entropy': 0.06485871458426118, 'clip_ratio/low_mean': 0.0039122195448726416, 'clip_ratio/low_min': 0.0039122195448726416, 'clip_ratio/high_mean': 0.008279926958493888, 'clip_ratio/high_max': 0.008279926958493888, 'clip_ratio/region_mean': 0.01219214650336653, 'reward_total_mean': 0.5973337292671204, 'reward_meter_mean': 0.9104515314102173, 'reward_meter_std': 0.12032636255025864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6607142686843872, 'reward_repeat_penalty_std': 0.15152288973331451, 'reward_total_composite_mean': 0.5973337292671204, 'reward_total_composite_std': 0.1518002152442932, 'epoch': 0.03}

 26%|██▌       | 854/3300 [28:07<9:45:28, 14.36s/it]INFO 04-11 23:04:59 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 855/3300 [28:12<7:47:17, 11.47s/it]2026-04-11 23:05:03,700 | INFO | train_grpo_train | metrics_logged mode=train step=855

                                                    
{'loss': -0.006, 'grad_norm': 5.122466564178467, 'learning_rate': 7.412121212121213e-06, 'num_tokens': 1907251.0, 'completions/mean_length': 64.875, 'completions/min_length': 63.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9911971092224121, 'rewards/meter/std': 0.0026730033569037914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9497148990631104, 'rewards/total_composite/std': 0.11557892709970474, 'reward': 0.9497148990631104, 'reward_std': 0.11557891964912415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021537287160754204, 'sampling/sampling_logp_difference/max': 1.6212968826293945, 'sampling/importance_sampling_ratio/min': 0.32319414615631104, 'sampling/importance_sampling_ratio/mean': 0.9993307590484619, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06852800818160176, 'clip_ratio/low_mean': 0.003968254197388887, 'clip_ratio/low_min': 0.003968254197388887, 'clip_ratio/high_mean': 0.01156850962433964, 'clip_ratio/high_max': 0.01156850962433964, 'clip_ratio/region_mean': 0.015536763821728528, 'reward_total_mean': 0.9497148990631104, 'reward_meter_mean': 0.9911971092224121, 'reward_meter_std': 0.0026730033569037914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9497148990631104, 'reward_total_composite_std': 0.11557892709970474, 'epoch': 0.03}

 26%|██▌       | 855/3300 [28:12<7:47:17, 11.47s/it]INFO 04-11 23:05:03 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 856/3300 [28:17<6:23:06,  9.41s/it]2026-04-11 23:05:08,294 | INFO | train_grpo_train | metrics_logged mode=train step=856

                                                    
{'loss': 0.0002, 'grad_norm': 6.793356418609619, 'learning_rate': 7.40909090909091e-06, 'num_tokens': 1909129.0, 'completions/mean_length': 64.75, 'completions/min_length': 63.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9881373643875122, 'rewards/meter/std': 0.009240454062819481, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9881373643875122, 'rewards/total_composite/std': 0.009240454062819481, 'reward': 0.9881373643875122, 'reward_std': 0.00924046989530325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017327219247817993, 'sampling/sampling_logp_difference/max': 1.138974666595459, 'sampling/importance_sampling_ratio/min': 0.3201471269130707, 'sampling/importance_sampling_ratio/mean': 0.9964679479598999, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07705914694815874, 'clip_ratio/low_mean': 0.001923076924867928, 'clip_ratio/low_min': 0.001923076924867928, 'clip_ratio/high_mean': 0.011599511839449406, 'clip_ratio/high_max': 0.011599511839449406, 'clip_ratio/region_mean': 0.013522588764317334, 'reward_total_mean': 0.9881373643875122, 'reward_meter_mean': 0.9881373643875122, 'reward_meter_std': 0.009240454062819481, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9881373643875122, 'reward_total_composite_std': 0.009240454062819481, 'epoch': 0.03}

 26%|██▌       | 856/3300 [28:17<6:23:06,  9.41s/it]INFO 04-11 23:05:08 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 857/3300 [28:22<5:27:22,  8.04s/it]2026-04-11 23:05:13,157 | INFO | train_grpo_train | metrics_logged mode=train step=857

                                                    
{'loss': 0.0072, 'grad_norm': 3.6093900203704834, 'learning_rate': 7.406060606060607e-06, 'num_tokens': 1910935.0, 'completions/mean_length': 75.75, 'completions/min_length': 68.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.991020679473877, 'rewards/meter/std': 0.00623312359675765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9496715664863586, 'rewards/total_composite/std': 0.11656945198774338, 'reward': 0.9496715664863586, 'reward_std': 0.11656944453716278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041222382336854935, 'sampling/sampling_logp_difference/max': 3.455476760864258, 'sampling/importance_sampling_ratio/min': 0.03157224878668785, 'sampling/importance_sampling_ratio/mean': 0.9909502863883972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07773020630702376, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.018700583139434457, 'clip_ratio/high_max': 0.018700583139434457, 'clip_ratio/region_mean': 0.018700583139434457, 'reward_total_mean': 0.9496715664863586, 'reward_meter_mean': 0.991020679473877, 'reward_meter_std': 0.00623312359675765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9496715664863586, 'reward_total_composite_std': 0.11656945198774338, 'epoch': 0.03}

 26%|██▌       | 857/3300 [28:22<5:27:22,  8.04s/it]INFO 04-11 23:05:13 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 858/3300 [28:29<5:16:03,  7.77s/it]2026-04-11 23:05:20,276 | INFO | train_grpo_train | metrics_logged mode=train step=858

                                                    
{'loss': 0.0285, 'grad_norm': 1.946842908859253, 'learning_rate': 7.403030303030304e-06, 'num_tokens': 1914420.0, 'completions/mean_length': 254.625, 'completions/min_length': 233.0, 'completions/max_length': 268.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 254.625, 'completions/min_terminated_length': 233.0, 'completions/max_terminated_length': 268.0, 'rewards/meter/mean': 0.6067160367965698, 'rewards/meter/std': 0.3461027145385742, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5166666507720947, 'rewards/repeat_penalty/std': 0.17728106677532196, 'rewards/total_composite/mean': 0.3126515746116638, 'rewards/total_composite/std': 0.2052178531885147, 'reward': 0.3126515746116638, 'reward_std': 0.2052178531885147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020115505903959274, 'sampling/sampling_logp_difference/max': 1.3195867538452148, 'sampling/importance_sampling_ratio/min': 0.26724570989608765, 'sampling/importance_sampling_ratio/mean': 1.0007023811340332, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1408333508297801, 'clip_ratio/low_mean': 0.009092815220355988, 'clip_ratio/low_min': 0.009092815220355988, 'clip_ratio/high_mean': 0.006283877417445183, 'clip_ratio/high_max': 0.006283877417445183, 'clip_ratio/region_mean': 0.01537669263780117, 'reward_total_mean': 0.3126515746116638, 'reward_meter_mean': 0.6067160367965698, 'reward_meter_std': 0.3461027145385742, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5166666507720947, 'reward_repeat_penalty_std': 0.17728106677532196, 'reward_total_composite_mean': 0.3126515746116638, 'reward_total_composite_std': 0.2052178531885147, 'epoch': 0.03}

 26%|██▌       | 858/3300 [28:29<5:16:03,  7.77s/it]INFO 04-11 23:05:20 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 859/3300 [28:34<4:49:03,  7.11s/it]2026-04-11 23:05:25,840 | INFO | train_grpo_train | metrics_logged mode=train step=859

                                                    
{'loss': 0.0073, 'grad_norm': 1.447547435760498, 'learning_rate': 7.4e-06, 'num_tokens': 1916992.0, 'completions/mean_length': 136.5, 'completions/min_length': 129.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.5, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9829227924346924, 'rewards/meter/std': 0.003464686218649149, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5535714626312256, 'rewards/repeat_penalty/std': 0.23458294570446014, 'rewards/total_composite/mean': 0.5436913371086121, 'rewards/total_composite/std': 0.22969432175159454, 'reward': 0.5436913371086121, 'reward_std': 0.22969430685043335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012742106802761555, 'sampling/sampling_logp_difference/max': 1.1751036643981934, 'sampling/importance_sampling_ratio/min': 0.30878695845603943, 'sampling/importance_sampling_ratio/mean': 1.0010943412780762, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.048957847990095615, 'clip_ratio/low_mean': 0.003665413591079414, 'clip_ratio/low_min': 0.003665413591079414, 'clip_ratio/high_mean': 0.008425170031841844, 'clip_ratio/high_max': 0.008425170031841844, 'clip_ratio/region_mean': 0.012090583622921258, 'reward_total_mean': 0.5436913371086121, 'reward_meter_mean': 0.9829227924346924, 'reward_meter_std': 0.003464686218649149, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5535714626312256, 'reward_repeat_penalty_std': 0.23458294570446014, 'reward_total_composite_mean': 0.5436913371086121, 'reward_total_composite_std': 0.22969432175159454, 'epoch': 0.03}

 26%|██▌       | 859/3300 [28:34<4:49:03,  7.11s/it]INFO 04-11 23:05:26 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 860/3300 [28:40<4:31:00,  6.66s/it]2026-04-11 23:05:31,474 | INFO | train_grpo_train | metrics_logged mode=train step=860

                                                    
{'loss': 0.0014, 'grad_norm': 1.1458154916763306, 'learning_rate': 7.396969696969698e-06, 'num_tokens': 1919482.0, 'completions/mean_length': 128.25, 'completions/min_length': 122.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.25, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9833224415779114, 'rewards/meter/std': 0.010524172335863113, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6428571939468384, 'rewards/repeat_penalty/std': 0.15272071957588196, 'rewards/total_composite/mean': 0.6316357851028442, 'rewards/total_composite/std': 0.14875362813472748, 'reward': 0.6316357851028442, 'reward_std': 0.14875362813472748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011860202066600323, 'sampling/sampling_logp_difference/max': 1.2248728275299072, 'sampling/importance_sampling_ratio/min': 0.29379504919052124, 'sampling/importance_sampling_ratio/mean': 1.0009667873382568, 'sampling/importance_sampling_ratio/max': 1.9968863725662231, 'entropy': 0.03678457555361092, 'clip_ratio/low_mean': 0.002871762844733894, 'clip_ratio/low_min': 0.002871762844733894, 'clip_ratio/high_mean': 0.005814186413772404, 'clip_ratio/high_max': 0.005814186413772404, 'clip_ratio/region_mean': 0.008685949258506298, 'reward_total_mean': 0.6316357851028442, 'reward_meter_mean': 0.9833224415779114, 'reward_meter_std': 0.010524172335863113, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6428571939468384, 'reward_repeat_penalty_std': 0.15272071957588196, 'reward_total_composite_mean': 0.6316357851028442, 'reward_total_composite_std': 0.14875362813472748, 'epoch': 0.03}

 26%|██▌       | 860/3300 [28:40<4:31:00,  6.66s/it]INFO 04-11 23:05:31 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 861/3300 [28:45<4:16:21,  6.31s/it]2026-04-11 23:05:36,945 | INFO | train_grpo_train | metrics_logged mode=train step=861

                                                    
{'loss': -0.0084, 'grad_norm': 3.195672035217285, 'learning_rate': 7.393939393939395e-06, 'num_tokens': 1921968.0, 'completions/mean_length': 130.75, 'completions/min_length': 123.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.75, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9875681400299072, 'rewards/meter/std': 0.003578658914193511, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.660714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.6525848507881165, 'rewards/total_composite/std': 0.07386582344770432, 'reward': 0.6525848507881165, 'reward_std': 0.07386582344770432, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009233459830284119, 'sampling/sampling_logp_difference/max': 1.330092430114746, 'sampling/importance_sampling_ratio/min': 0.26445281505584717, 'sampling/importance_sampling_ratio/mean': 0.999982476234436, 'sampling/importance_sampling_ratio/max': 1.8728368282318115, 'entropy': 0.034579032799229026, 'clip_ratio/low_mean': 0.002939337049610913, 'clip_ratio/low_min': 0.002939337049610913, 'clip_ratio/high_mean': 0.007541382568888366, 'clip_ratio/high_max': 0.007541382568888366, 'clip_ratio/region_mean': 0.010480719618499279, 'reward_total_mean': 0.6525848507881165, 'reward_meter_mean': 0.9875681400299072, 'reward_meter_std': 0.003578658914193511, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.660714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.6525848507881165, 'reward_total_composite_std': 0.07386582344770432, 'epoch': 0.03}

 26%|██▌       | 861/3300 [28:45<4:16:21,  6.31s/it]INFO 04-11 23:05:37 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 862/3300 [28:51<4:07:21,  6.09s/it]2026-04-11 23:05:42,523 | INFO | train_grpo_train | metrics_logged mode=train step=862

                                                    
{'loss': -0.026, 'grad_norm': 4.800323963165283, 'learning_rate': 7.390909090909092e-06, 'num_tokens': 1924096.0, 'completions/mean_length': 137.0, 'completions/min_length': 123.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.0, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.8122666478157043, 'rewards/meter/std': 0.3459046185016632, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7142857313156128, 'rewards/repeat_penalty/std': 0.1322600096464157, 'rewards/total_composite/mean': 0.5928546190261841, 'rewards/total_composite/std': 0.2950608432292938, 'reward': 0.5928546190261841, 'reward_std': 0.29506081342697144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024774322286248207, 'sampling/sampling_logp_difference/max': 1.3954243659973145, 'sampling/importance_sampling_ratio/min': 0.2477279007434845, 'sampling/importance_sampling_ratio/mean': 1.0024539232254028, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0997177753597498, 'clip_ratio/low_mean': 0.0039816161734052, 'clip_ratio/low_min': 0.0039816161734052, 'clip_ratio/high_mean': 0.010707754408940673, 'clip_ratio/high_max': 0.010707754408940673, 'clip_ratio/region_mean': 0.014689370582345873, 'reward_total_mean': 0.5928546190261841, 'reward_meter_mean': 0.8122666478157043, 'reward_meter_std': 0.3459046185016632, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7142857313156128, 'reward_repeat_penalty_std': 0.1322600096464157, 'reward_total_composite_mean': 0.5928546190261841, 'reward_total_composite_std': 0.2950608432292938, 'epoch': 0.03}

 26%|██▌       | 862/3300 [28:51<4:07:21,  6.09s/it]INFO 04-11 23:05:42 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 863/3300 [28:57<4:02:48,  5.98s/it]2026-04-11 23:05:48,245 | INFO | train_grpo_train | metrics_logged mode=train step=863

                                                    
{'loss': 0.0474, 'grad_norm': 3.5590970516204834, 'learning_rate': 7.3878787878787885e-06, 'num_tokens': 1926759.0, 'completions/mean_length': 147.875, 'completions/min_length': 131.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.875, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9950444102287292, 'rewards/meter/std': 0.0004635074583347887, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6904761791229248, 'rewards/repeat_penalty/std': 0.06734350323677063, 'rewards/total_composite/mean': 0.6657319068908691, 'rewards/total_composite/std': 0.006685574073344469, 'reward': 0.6657319068908691, 'reward_std': 0.00668557733297348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01787300407886505, 'sampling/sampling_logp_difference/max': 2.7439072132110596, 'sampling/importance_sampling_ratio/min': 0.06431855261325836, 'sampling/importance_sampling_ratio/mean': 0.9978508353233337, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.036845392314717174, 'clip_ratio/low_mean': 0.009212537202984095, 'clip_ratio/low_min': 0.009212537202984095, 'clip_ratio/high_mean': 0.0028625954873859882, 'clip_ratio/high_max': 0.0028625954873859882, 'clip_ratio/region_mean': 0.012075132690370083, 'reward_total_mean': 0.6657319068908691, 'reward_meter_mean': 0.9950444102287292, 'reward_meter_std': 0.0004635074583347887, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6904761791229248, 'reward_repeat_penalty_std': 0.06734350323677063, 'reward_total_composite_mean': 0.6657319068908691, 'reward_total_composite_std': 0.006685574073344469, 'epoch': 0.03}

 26%|██▌       | 863/3300 [28:57<4:02:48,  5.98s/it]INFO 04-11 23:05:48 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 864/3300 [29:02<3:48:40,  5.63s/it]2026-04-11 23:05:53,074 | INFO | train_grpo_train | metrics_logged mode=train step=864

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.384848484848486e-06, 'num_tokens': 1928559.0, 'completions/mean_length': 76.0, 'completions/min_length': 76.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.0, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9935055375099182, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9935055375099182, 'rewards/total_composite/std': 0.0, 'reward': 0.9935055375099182, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0010724789462983608, 'sampling/sampling_logp_difference/max': 0.06501199305057526, 'sampling/importance_sampling_ratio/min': 0.9370561838150024, 'sampling/importance_sampling_ratio/mean': 1.0007295608520508, 'sampling/importance_sampling_ratio/max': 1.0238041877746582, 'entropy': 0.010270603699609637, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9935055375099182, 'reward_meter_mean': 0.9935055375099182, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9935055375099182, 'reward_total_composite_std': 0.0, 'epoch': 0.03}

 26%|██▌       | 864/3300 [29:02<3:48:40,  5.63s/it]INFO 04-11 23:05:53 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 865/3300 [29:07<3:40:55,  5.44s/it]2026-04-11 23:05:58,076 | INFO | train_grpo_train | metrics_logged mode=train step=865

                                                    
{'loss': -0.002, 'grad_norm': 3.38623309135437, 'learning_rate': 7.381818181818182e-06, 'num_tokens': 1930412.0, 'completions/mean_length': 74.625, 'completions/min_length': 74.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.625, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.995256781578064, 'rewards/meter/std': 0.0008067663875408471, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.995256781578064, 'rewards/total_composite/std': 0.0008067663875408471, 'reward': 0.995256781578064, 'reward_std': 0.0008067605085670948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016164077445864677, 'sampling/sampling_logp_difference/max': 1.9147498607635498, 'sampling/importance_sampling_ratio/min': 0.14737869799137115, 'sampling/importance_sampling_ratio/mean': 0.9971731305122375, 'sampling/importance_sampling_ratio/max': 1.6714001893997192, 'entropy': 0.07608733791857958, 'clip_ratio/low_mean': 0.0016891892300918698, 'clip_ratio/low_min': 0.0016891892300918698, 'clip_ratio/high_mean': 0.00829179841093719, 'clip_ratio/high_max': 0.00829179841093719, 'clip_ratio/region_mean': 0.00998098764102906, 'reward_total_mean': 0.995256781578064, 'reward_meter_mean': 0.995256781578064, 'reward_meter_std': 0.0008067663875408471, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.995256781578064, 'reward_total_composite_std': 0.0008067663875408471, 'epoch': 0.03}

 26%|██▌       | 865/3300 [29:07<3:40:55,  5.44s/it]INFO 04-11 23:05:58 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 866/3300 [29:12<3:38:01,  5.37s/it]2026-04-11 23:06:03,287 | INFO | train_grpo_train | metrics_logged mode=train step=866

                                                    
{'loss': 0.1348, 'grad_norm': 4.458425045013428, 'learning_rate': 7.378787878787879e-06, 'num_tokens': 1932315.0, 'completions/mean_length': 79.875, 'completions/min_length': 72.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9195418953895569, 'rewards/meter/std': 0.21018952131271362, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8895903825759888, 'rewards/total_composite/std': 0.2949047088623047, 'reward': 0.8895903825759888, 'reward_std': 0.2949047088623047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011505667120218277, 'sampling/sampling_logp_difference/max': 1.7326593399047852, 'sampling/importance_sampling_ratio/min': 0.17681357264518738, 'sampling/importance_sampling_ratio/mean': 1.0000438690185547, 'sampling/importance_sampling_ratio/max': 1.7624181509017944, 'entropy': 0.029443061910569668, 'clip_ratio/low_mean': 0.0022522523067891598, 'clip_ratio/low_min': 0.0022522523067891598, 'clip_ratio/high_mean': 0.004934210563078523, 'clip_ratio/high_max': 0.004934210563078523, 'clip_ratio/region_mean': 0.0071864628698676825, 'reward_total_mean': 0.8895903825759888, 'reward_meter_mean': 0.9195418953895569, 'reward_meter_std': 0.21018952131271362, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8895903825759888, 'reward_total_composite_std': 0.2949047088623047, 'epoch': 0.03}

 26%|██▌       | 866/3300 [29:12<3:38:01,  5.37s/it]INFO 04-11 23:06:03 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 867/3300 [29:17<3:36:12,  5.33s/it]2026-04-11 23:06:08,522 | INFO | train_grpo_train | metrics_logged mode=train step=867

                                                    
{'loss': 0.0392, 'grad_norm': 2.170355796813965, 'learning_rate': 7.375757575757576e-06, 'num_tokens': 1934484.0, 'completions/mean_length': 98.125, 'completions/min_length': 92.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.125, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.8766216039657593, 'rewards/meter/std': 0.1744304597377777, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7012972831726074, 'rewards/total_composite/std': 0.13954438269138336, 'reward': 0.7012972831726074, 'reward_std': 0.13954436779022217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016189197078347206, 'sampling/sampling_logp_difference/max': 2.60568904876709, 'sampling/importance_sampling_ratio/min': 0.07385222613811493, 'sampling/importance_sampling_ratio/mean': 1.000335454940796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05557233979925513, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.010424387757666409, 'clip_ratio/high_max': 0.010424387757666409, 'clip_ratio/region_mean': 0.012739202589727938, 'reward_total_mean': 0.7012972831726074, 'reward_meter_mean': 0.8766216039657593, 'reward_meter_std': 0.1744304597377777, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7012972831726074, 'reward_total_composite_std': 0.13954438269138336, 'epoch': 0.03}

 26%|██▋       | 867/3300 [29:17<3:36:12,  5.33s/it]INFO 04-11 23:06:08 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 868/3300 [29:22<3:35:06,  5.31s/it]2026-04-11 23:06:13,770 | INFO | train_grpo_train | metrics_logged mode=train step=868

                                                    
{'loss': -0.0072, 'grad_norm': 1.8654179573059082, 'learning_rate': 7.372727272727274e-06, 'num_tokens': 1936683.0, 'completions/mean_length': 110.875, 'completions/min_length': 107.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.875, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.8900142908096313, 'rewards/meter/std': 0.2951919138431549, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.712011456489563, 'rewards/total_composite/std': 0.23615355789661407, 'reward': 0.712011456489563, 'reward_std': 0.23615355789661407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004938235506415367, 'sampling/sampling_logp_difference/max': 1.0195584297180176, 'sampling/importance_sampling_ratio/min': 0.36075422167778015, 'sampling/importance_sampling_ratio/mean': 1.0000348091125488, 'sampling/importance_sampling_ratio/max': 1.3221184015274048, 'entropy': 0.028168250457383692, 'clip_ratio/low_mean': 0.004629629664123058, 'clip_ratio/low_min': 0.004629629664123058, 'clip_ratio/high_mean': 0.002336448524147272, 'clip_ratio/high_max': 0.002336448524147272, 'clip_ratio/region_mean': 0.00696607818827033, 'reward_total_mean': 0.712011456489563, 'reward_meter_mean': 0.8900142908096313, 'reward_meter_std': 0.2951919138431549, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.712011456489563, 'reward_total_composite_std': 0.23615355789661407, 'epoch': 0.03}

 26%|██▋       | 868/3300 [29:22<3:35:06,  5.31s/it]INFO 04-11 23:06:14 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 869/3300 [29:28<3:35:26,  5.32s/it]2026-04-11 23:06:19,110 | INFO | train_grpo_train | metrics_logged mode=train step=869

                                                    
{'loss': -0.019, 'grad_norm': 3.4994313716888428, 'learning_rate': 7.36969696969697e-06, 'num_tokens': 1938765.0, 'completions/mean_length': 107.25, 'completions/min_length': 101.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9882686138153076, 'rewards/meter/std': 0.002476355992257595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8894296884536743, 'rewards/total_composite/std': 0.10556221753358841, 'reward': 0.8894296884536743, 'reward_std': 0.10556221753358841, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02293246239423752, 'sampling/sampling_logp_difference/max': 1.518688678741455, 'sampling/importance_sampling_ratio/min': 0.21899887919425964, 'sampling/importance_sampling_ratio/mean': 1.004834771156311, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12705577816814184, 'clip_ratio/low_mean': 0.008417388889938593, 'clip_ratio/low_min': 0.008417388889938593, 'clip_ratio/high_mean': 0.010045865317806602, 'clip_ratio/high_max': 0.010045865317806602, 'clip_ratio/region_mean': 0.018463254207745194, 'reward_total_mean': 0.8894296884536743, 'reward_meter_mean': 0.9882686138153076, 'reward_meter_std': 0.002476355992257595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8894296884536743, 'reward_total_composite_std': 0.10556221753358841, 'epoch': 0.03}

 26%|██▋       | 869/3300 [29:28<3:35:26,  5.32s/it]INFO 04-11 23:06:19 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 870/3300 [29:34<3:44:41,  5.55s/it]2026-04-11 23:06:25,197 | INFO | train_grpo_train | metrics_logged mode=train step=870

                                                    
{'loss': 0.0088, 'grad_norm': 3.510559320449829, 'learning_rate': 7.3666666666666676e-06, 'num_tokens': 1941320.0, 'completions/mean_length': 138.375, 'completions/min_length': 133.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.375, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9983220100402832, 'rewards/meter/std': 0.0005191374220885336, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8214285373687744, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.6560331583023071, 'rewards/total_composite/std': 0.05272570997476578, 'reward': 0.6560331583023071, 'reward_std': 0.052725713700056076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014895547181367874, 'sampling/sampling_logp_difference/max': 1.4788875579833984, 'sampling/importance_sampling_ratio/min': 0.22789107263088226, 'sampling/importance_sampling_ratio/mean': 0.9968522191047668, 'sampling/importance_sampling_ratio/max': 1.6303558349609375, 'entropy': 0.05070830974727869, 'clip_ratio/low_mean': 0.0017879948718473315, 'clip_ratio/low_min': 0.0017879948718473315, 'clip_ratio/high_mean': 0.015459735121112317, 'clip_ratio/high_max': 0.015459735121112317, 'clip_ratio/region_mean': 0.01724772999295965, 'reward_total_mean': 0.6560331583023071, 'reward_meter_mean': 0.9983220100402832, 'reward_meter_std': 0.0005191374220885336, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8214285373687744, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.6560331583023071, 'reward_total_composite_std': 0.05272570997476578, 'epoch': 0.03}

 26%|██▋       | 870/3300 [29:34<3:44:41,  5.55s/it]INFO 04-11 23:06:25 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 871/3300 [29:40<3:53:02,  5.76s/it]2026-04-11 23:06:31,440 | INFO | train_grpo_train | metrics_logged mode=train step=871

                                                    
{'loss': -0.0038, 'grad_norm': 1.7928707599639893, 'learning_rate': 7.363636363636364e-06, 'num_tokens': 1944242.0, 'completions/mean_length': 190.25, 'completions/min_length': 175.0, 'completions/max_length': 201.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 190.25, 'completions/min_terminated_length': 175.0, 'completions/max_terminated_length': 201.0, 'rewards/meter/mean': 0.8820419907569885, 'rewards/meter/std': 0.2552332878112793, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7638888955116272, 'rewards/repeat_penalty/std': 0.0927247703075409, 'rewards/total_composite/mean': 0.5602418184280396, 'rewards/total_composite/std': 0.1772141009569168, 'reward': 0.5602418184280396, 'reward_std': 0.17721408605575562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015695709735155106, 'sampling/sampling_logp_difference/max': 1.124608039855957, 'sampling/importance_sampling_ratio/min': 0.324779748916626, 'sampling/importance_sampling_ratio/mean': 1.0031383037567139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11951877502724528, 'clip_ratio/low_mean': 0.0053914261516183615, 'clip_ratio/low_min': 0.0053914261516183615, 'clip_ratio/high_mean': 0.009079249284695834, 'clip_ratio/high_max': 0.009079249284695834, 'clip_ratio/region_mean': 0.014470675436314195, 'reward_total_mean': 0.5602418184280396, 'reward_meter_mean': 0.8820419907569885, 'reward_meter_std': 0.2552332878112793, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7638888955116272, 'reward_repeat_penalty_std': 0.0927247703075409, 'reward_total_composite_mean': 0.5602418184280396, 'reward_total_composite_std': 0.1772141009569168, 'epoch': 0.03}

 26%|██▋       | 871/3300 [29:40<3:53:02,  5.76s/it]INFO 04-11 23:06:31 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 872/3300 [29:46<3:53:23,  5.77s/it]2026-04-11 23:06:37,233 | INFO | train_grpo_train | metrics_logged mode=train step=872

                                                    
{'loss': -0.0316, 'grad_norm': 3.245631694793701, 'learning_rate': 7.360606060606061e-06, 'num_tokens': 1946920.0, 'completions/mean_length': 144.75, 'completions/min_length': 129.0, 'completions/max_length': 162.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 144.75, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.9848458170890808, 'rewards/meter/std': 0.03508731722831726, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7250000238418579, 'rewards/repeat_penalty/std': 0.030304575338959694, 'rewards/total_composite/mean': 0.6906325221061707, 'rewards/total_composite/std': 0.06134637072682381, 'reward': 0.6906325221061707, 'reward_std': 0.061346374452114105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010910775512456894, 'sampling/sampling_logp_difference/max': 1.9047117233276367, 'sampling/importance_sampling_ratio/min': 0.14886555075645447, 'sampling/importance_sampling_ratio/mean': 1.0042208433151245, 'sampling/importance_sampling_ratio/max': 1.8771061897277832, 'entropy': 0.05419332440942526, 'clip_ratio/low_mean': 0.0038759689778089523, 'clip_ratio/low_min': 0.0038759689778089523, 'clip_ratio/high_mean': 0.005582162644714117, 'clip_ratio/high_max': 0.005582162644714117, 'clip_ratio/region_mean': 0.00945813162252307, 'reward_total_mean': 0.6906325221061707, 'reward_meter_mean': 0.9848458170890808, 'reward_meter_std': 0.03508731722831726, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7250000238418579, 'reward_repeat_penalty_std': 0.030304575338959694, 'reward_total_composite_mean': 0.6906325221061707, 'reward_total_composite_std': 0.06134637072682381, 'epoch': 0.04}

 26%|██▋       | 872/3300 [29:46<3:53:23,  5.77s/it]INFO 04-11 23:06:37 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 873/3300 [29:50<3:40:27,  5.45s/it]2026-04-11 23:06:41,943 | INFO | train_grpo_train | metrics_logged mode=train step=873

                                                    
{'loss': 0.0209, 'grad_norm': 5.7755126953125, 'learning_rate': 7.357575757575758e-06, 'num_tokens': 1948679.0, 'completions/mean_length': 64.875, 'completions/min_length': 63.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9520989656448364, 'rewards/meter/std': 0.055099111050367355, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9520989656448364, 'rewards/total_composite/std': 0.055099111050367355, 'reward': 0.9520989656448364, 'reward_std': 0.05509910732507706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032992783933877945, 'sampling/sampling_logp_difference/max': 1.5907907485961914, 'sampling/importance_sampling_ratio/min': 0.2037644237279892, 'sampling/importance_sampling_ratio/mean': 1.0100115537643433, 'sampling/importance_sampling_ratio/max': 1.7566462755203247, 'entropy': 0.17148890905082226, 'clip_ratio/low_mean': 0.013034759555011988, 'clip_ratio/low_min': 0.013034759555011988, 'clip_ratio/high_mean': 0.027383167180232704, 'clip_ratio/high_max': 0.027383167180232704, 'clip_ratio/region_mean': 0.04041792673524469, 'reward_total_mean': 0.9520989656448364, 'reward_meter_mean': 0.9520989656448364, 'reward_meter_std': 0.055099111050367355, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9520989656448364, 'reward_total_composite_std': 0.055099111050367355, 'epoch': 0.04}

 26%|██▋       | 873/3300 [29:50<3:40:27,  5.45s/it]INFO 04-11 23:06:42 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 874/3300 [29:56<3:40:23,  5.45s/it]2026-04-11 23:06:47,396 | INFO | train_grpo_train | metrics_logged mode=train step=874

                                                    
{'loss': 0.0149, 'grad_norm': 1.7160921096801758, 'learning_rate': 7.354545454545456e-06, 'num_tokens': 1950913.0, 'completions/mean_length': 107.25, 'completions/min_length': 105.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.9955088496208191, 'rewards/meter/std': 0.0011983213480561972, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7964070439338684, 'rewards/total_composite/std': 0.0009586615487933159, 'reward': 0.7964070439338684, 'reward_std': 0.0009586560190655291, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008256379514932632, 'sampling/sampling_logp_difference/max': 0.7744824886322021, 'sampling/importance_sampling_ratio/min': 0.46094223856925964, 'sampling/importance_sampling_ratio/mean': 0.9997270703315735, 'sampling/importance_sampling_ratio/max': 1.3036043643951416, 'entropy': 0.04265120159834623, 'clip_ratio/low_mean': 0.0022123893722891808, 'clip_ratio/low_min': 0.0022123893722891808, 'clip_ratio/high_mean': 0.007142857299186289, 'clip_ratio/high_max': 0.007142857299186289, 'clip_ratio/region_mean': 0.00935524667147547, 'reward_total_mean': 0.7964070439338684, 'reward_meter_mean': 0.9955088496208191, 'reward_meter_std': 0.0011983213480561972, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7964070439338684, 'reward_total_composite_std': 0.0009586615487933159, 'epoch': 0.04}

 26%|██▋       | 874/3300 [29:56<3:40:23,  5.45s/it]INFO 04-11 23:06:47 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 875/3300 [30:01<3:37:27,  5.38s/it]2026-04-11 23:06:52,612 | INFO | train_grpo_train | metrics_logged mode=train step=875

                                                    
{'loss': -0.0005, 'grad_norm': 2.395378351211548, 'learning_rate': 7.351515151515151e-06, 'num_tokens': 1953074.0, 'completions/mean_length': 94.125, 'completions/min_length': 87.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.125, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.992639422416687, 'rewards/meter/std': 0.006307392846792936, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.7692322134971619, 'rewards/total_composite/std': 0.06972821056842804, 'reward': 0.7692322134971619, 'reward_std': 0.06972822546958923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02180575206875801, 'sampling/sampling_logp_difference/max': 2.036391258239746, 'sampling/importance_sampling_ratio/min': 0.13049879670143127, 'sampling/importance_sampling_ratio/mean': 0.9954232573509216, 'sampling/importance_sampling_ratio/max': 1.5037858486175537, 'entropy': 0.07790858251973987, 'clip_ratio/low_mean': 0.0026881720405071974, 'clip_ratio/low_min': 0.0026881720405071974, 'clip_ratio/high_mean': 0.023883325862698257, 'clip_ratio/high_max': 0.023883325862698257, 'clip_ratio/region_mean': 0.026571497903205454, 'reward_total_mean': 0.7692322134971619, 'reward_meter_mean': 0.992639422416687, 'reward_meter_std': 0.006307392846792936, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.7692322134971619, 'reward_total_composite_std': 0.06972821056842804, 'epoch': 0.04}

 27%|██▋       | 875/3300 [30:01<3:37:27,  5.38s/it]INFO 04-11 23:06:52 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 876/3300 [30:08<3:57:10,  5.87s/it]2026-04-11 23:06:59,628 | INFO | train_grpo_train | metrics_logged mode=train step=876

                                                    
{'loss': -0.0419, 'grad_norm': 2.0017435550689697, 'learning_rate': 7.348484848484849e-06, 'num_tokens': 1956524.0, 'completions/mean_length': 217.25, 'completions/min_length': 206.0, 'completions/max_length': 236.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 217.25, 'completions/min_terminated_length': 206.0, 'completions/max_terminated_length': 236.0, 'rewards/meter/mean': 0.9868931770324707, 'rewards/meter/std': 0.006763860583305359, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6858974695205688, 'rewards/repeat_penalty/std': 0.011869488283991814, 'rewards/total_composite/mean': 0.6128624677658081, 'rewards/total_composite/std': 0.02833697944879532, 'reward': 0.6128624677658081, 'reward_std': 0.02833697572350502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012169601395726204, 'sampling/sampling_logp_difference/max': 1.237346887588501, 'sampling/importance_sampling_ratio/min': 0.29015299677848816, 'sampling/importance_sampling_ratio/mean': 1.0016402006149292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.052224091021344066, 'clip_ratio/low_mean': 0.007715584768448025, 'clip_ratio/low_min': 0.007715584768448025, 'clip_ratio/high_mean': 0.0031779659911990166, 'clip_ratio/high_max': 0.0031779659911990166, 'clip_ratio/region_mean': 0.010893550759647042, 'reward_total_mean': 0.6128624677658081, 'reward_meter_mean': 0.9868931770324707, 'reward_meter_std': 0.006763860583305359, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6858974695205688, 'reward_repeat_penalty_std': 0.011869488283991814, 'reward_total_composite_mean': 0.6128624677658081, 'reward_total_composite_std': 0.02833697944879532, 'epoch': 0.04}

 27%|██▋       | 876/3300 [30:08<3:57:10,  5.87s/it]INFO 04-11 23:06:59 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 877/3300 [30:13<3:47:13,  5.63s/it]2026-04-11 23:07:04,683 | INFO | train_grpo_train | metrics_logged mode=train step=877

                                                    
{'loss': -0.0079, 'grad_norm': 6.4066290855407715, 'learning_rate': 7.345454545454546e-06, 'num_tokens': 1958416.0, 'completions/mean_length': 79.5, 'completions/min_length': 76.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9876222610473633, 'rewards/meter/std': 0.013733073137700558, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9876222610473633, 'rewards/total_composite/std': 0.013733073137700558, 'reward': 0.9876222610473633, 'reward_std': 0.01373306754976511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02791396901011467, 'sampling/sampling_logp_difference/max': 1.3935667276382446, 'sampling/importance_sampling_ratio/min': 0.24818849563598633, 'sampling/importance_sampling_ratio/mean': 1.0038245916366577, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1384354429319501, 'clip_ratio/low_mean': 0.003247300977818668, 'clip_ratio/low_min': 0.003247300977818668, 'clip_ratio/high_mean': 0.015677204821258783, 'clip_ratio/high_max': 0.015677204821258783, 'clip_ratio/region_mean': 0.01892450579907745, 'reward_total_mean': 0.9876222610473633, 'reward_meter_mean': 0.9876222610473633, 'reward_meter_std': 0.013733073137700558, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9876222610473633, 'reward_total_composite_std': 0.013733073137700558, 'epoch': 0.04}

 27%|██▋       | 877/3300 [30:13<3:47:13,  5.63s/it]INFO 04-11 23:07:04 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 878/3300 [30:18<3:36:14,  5.36s/it]2026-04-11 23:07:09,412 | INFO | train_grpo_train | metrics_logged mode=train step=878

                                                    
{'loss': -0.0205, 'grad_norm': 3.125847816467285, 'learning_rate': 7.342424242424243e-06, 'num_tokens': 1960243.0, 'completions/mean_length': 69.375, 'completions/min_length': 64.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9980708360671997, 'rewards/meter/std': 0.00046432489762082696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980708360671997, 'rewards/total_composite/std': 0.00046432489762082696, 'reward': 0.9980708360671997, 'reward_std': 0.00046432187082245946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021711641922593117, 'sampling/sampling_logp_difference/max': 0.9192183017730713, 'sampling/importance_sampling_ratio/min': 0.39883068203926086, 'sampling/importance_sampling_ratio/mean': 0.999176561832428, 'sampling/importance_sampling_ratio/max': 1.6681042909622192, 'entropy': 0.09143292810767889, 'clip_ratio/low_mean': 0.016726079280488193, 'clip_ratio/low_min': 0.016726079280488193, 'clip_ratio/high_mean': 0.014086603187024593, 'clip_ratio/high_max': 0.014086603187024593, 'clip_ratio/region_mean': 0.030812682467512786, 'reward_total_mean': 0.9980708360671997, 'reward_meter_mean': 0.9980708360671997, 'reward_meter_std': 0.00046432489762082696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980708360671997, 'reward_total_composite_std': 0.00046432489762082696, 'epoch': 0.04}

 27%|██▋       | 878/3300 [30:18<3:36:14,  5.36s/it]INFO 04-11 23:07:09 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 879/3300 [30:22<3:27:03,  5.13s/it]2026-04-11 23:07:14,017 | INFO | train_grpo_train | metrics_logged mode=train step=879

                                                    
{'loss': -0.0466, 'grad_norm': 4.135697841644287, 'learning_rate': 7.3393939393939395e-06, 'num_tokens': 1962012.0, 'completions/mean_length': 65.125, 'completions/min_length': 58.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7186502814292908, 'rewards/meter/std': 0.37441036105155945, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.7184880971908569, 'rewards/total_composite/std': 0.3747643232345581, 'reward': 0.7184880971908569, 'reward_std': 0.37476426362991333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025105303153395653, 'sampling/sampling_logp_difference/max': 1.4957494735717773, 'sampling/importance_sampling_ratio/min': 0.22408060729503632, 'sampling/importance_sampling_ratio/mean': 0.9979657530784607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.072424097917974, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.01865671609994024, 'clip_ratio/high_max': 0.01865671609994024, 'clip_ratio/region_mean': 0.01865671609994024, 'reward_total_mean': 0.7184880971908569, 'reward_meter_mean': 0.7186502814292908, 'reward_meter_std': 0.37441036105155945, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.7184880971908569, 'reward_total_composite_std': 0.3747643232345581, 'epoch': 0.04}

 27%|██▋       | 879/3300 [30:23<3:27:03,  5.13s/it]INFO 04-11 23:07:14 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 880/3300 [30:28<3:25:55,  5.11s/it]2026-04-11 23:07:19,062 | INFO | train_grpo_train | metrics_logged mode=train step=880

                                                    
{'loss': -0.0202, 'grad_norm': 4.517680644989014, 'learning_rate': 7.336363636363637e-06, 'num_tokens': 1964149.0, 'completions/mean_length': 94.125, 'completions/min_length': 90.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.125, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.3598088026046753, 'rewards/meter/std': 0.3960496783256531, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.28257015347480774, 'rewards/total_composite/std': 0.31944534182548523, 'reward': 0.28257015347480774, 'reward_std': 0.31944531202316284, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015417142771184444, 'sampling/sampling_logp_difference/max': 1.0558514595031738, 'sampling/importance_sampling_ratio/min': 0.3478960692882538, 'sampling/importance_sampling_ratio/mean': 0.9979188442230225, 'sampling/importance_sampling_ratio/max': 1.8587523698806763, 'entropy': 0.04804056999273598, 'clip_ratio/low_mean': 0.008119725622236729, 'clip_ratio/low_min': 0.008119725622236729, 'clip_ratio/high_mean': 0.008080808212980628, 'clip_ratio/high_max': 0.008080808212980628, 'clip_ratio/region_mean': 0.016200533835217357, 'reward_total_mean': 0.28257015347480774, 'reward_meter_mean': 0.3598088026046753, 'reward_meter_std': 0.3960496783256531, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.28257015347480774, 'reward_total_composite_std': 0.31944534182548523, 'epoch': 0.04}

 27%|██▋       | 880/3300 [30:28<3:25:55,  5.11s/it]INFO 04-11 23:07:19 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 881/3300 [30:36<4:00:48,  5.97s/it]2026-04-11 23:07:27,069 | INFO | train_grpo_train | metrics_logged mode=train step=881

                                                    
{'loss': -0.0185, 'grad_norm': 1.2314221858978271, 'learning_rate': 7.333333333333333e-06, 'num_tokens': 1968226.0, 'completions/mean_length': 292.625, 'completions/min_length': 277.0, 'completions/max_length': 332.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 292.625, 'completions/min_terminated_length': 277.0, 'completions/max_terminated_length': 332.0, 'rewards/meter/mean': 0.9959769248962402, 'rewards/meter/std': 0.0035888429265469313, 'rewards/count_adherence/mean': 0.7386363744735718, 'rewards/count_adherence/std': 0.03214120864868164, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5985294580459595, 'rewards/repeat_penalty/std': 0.004159451462328434, 'rewards/total_composite/mean': 0.4401511549949646, 'rewards/total_composite/std': 0.014130757190287113, 'reward': 0.4401511549949646, 'reward_std': 0.014130760915577412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007294100243598223, 'sampling/sampling_logp_difference/max': 0.9965987205505371, 'sampling/importance_sampling_ratio/min': 0.3691328465938568, 'sampling/importance_sampling_ratio/mean': 1.0001860857009888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04085008194670081, 'clip_ratio/low_mean': 0.003381519520189613, 'clip_ratio/low_min': 0.003381519520189613, 'clip_ratio/high_mean': 0.002022653818130493, 'clip_ratio/high_max': 0.002022653818130493, 'clip_ratio/region_mean': 0.005404173338320106, 'reward_total_mean': 0.4401511549949646, 'reward_meter_mean': 0.9959769248962402, 'reward_meter_std': 0.0035888429265469313, 'reward_count_adherence_mean': 0.7386363744735718, 'reward_count_adherence_std': 0.03214120864868164, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5985294580459595, 'reward_repeat_penalty_std': 0.004159451462328434, 'reward_total_composite_mean': 0.4401511549949646, 'reward_total_composite_std': 0.014130757190287113, 'epoch': 0.04}

 27%|██▋       | 881/3300 [30:36<4:00:48,  5.97s/it]INFO 04-11 23:07:27 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 882/3300 [30:42<4:02:32,  6.02s/it]2026-04-11 23:07:33,182 | INFO | train_grpo_train | metrics_logged mode=train step=882

                                                    
{'loss': -0.002, 'grad_norm': 3.3222203254699707, 'learning_rate': 7.330303030303031e-06, 'num_tokens': 1970988.0, 'completions/mean_length': 144.25, 'completions/min_length': 120.0, 'completions/max_length': 166.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 144.25, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.7575975060462952, 'rewards/meter/std': 0.32098710536956787, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.783730149269104, 'rewards/repeat_penalty/std': 0.05935349687933922, 'rewards/total_composite/mean': 0.5672966837882996, 'rewards/total_composite/std': 0.25382858514785767, 'reward': 0.5672966837882996, 'reward_std': 0.2538285553455353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013289229944348335, 'sampling/sampling_logp_difference/max': 0.987497091293335, 'sampling/importance_sampling_ratio/min': 0.37250787019729614, 'sampling/importance_sampling_ratio/mean': 0.9993391633033752, 'sampling/importance_sampling_ratio/max': 1.9269946813583374, 'entropy': 0.05786414351314306, 'clip_ratio/low_mean': 0.0054536922834813595, 'clip_ratio/low_min': 0.0054536922834813595, 'clip_ratio/high_mean': 0.0039639262249693274, 'clip_ratio/high_max': 0.0039639262249693274, 'clip_ratio/region_mean': 0.009417618508450687, 'reward_total_mean': 0.5672966837882996, 'reward_meter_mean': 0.7575975060462952, 'reward_meter_std': 0.32098710536956787, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.783730149269104, 'reward_repeat_penalty_std': 0.05935349687933922, 'reward_total_composite_mean': 0.5672966837882996, 'reward_total_composite_std': 0.25382858514785767, 'epoch': 0.04}

 27%|██▋       | 882/3300 [30:42<4:02:32,  6.02s/it]INFO 04-11 23:07:33 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 883/3300 [30:46<3:47:20,  5.64s/it]2026-04-11 23:07:37,961 | INFO | train_grpo_train | metrics_logged mode=train step=883

                                                    
{'loss': 0.0339, 'grad_norm': 23.914064407348633, 'learning_rate': 7.3272727272727285e-06, 'num_tokens': 1972789.0, 'completions/mean_length': 61.125, 'completions/min_length': 55.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9858500361442566, 'rewards/meter/std': 0.013770055025815964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9858500361442566, 'rewards/total_composite/std': 0.013770055025815964, 'reward': 0.9858500361442566, 'reward_std': 0.013770047575235367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058846138417720795, 'sampling/sampling_logp_difference/max': 3.47511887550354, 'sampling/importance_sampling_ratio/min': 0.030958153307437897, 'sampling/importance_sampling_ratio/mean': 1.0076637268066406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1264942679554224, 'clip_ratio/low_mean': 0.00797146384138614, 'clip_ratio/low_min': 0.00797146384138614, 'clip_ratio/high_mean': 0.02063214615918696, 'clip_ratio/high_max': 0.02063214615918696, 'clip_ratio/region_mean': 0.0286036100005731, 'reward_total_mean': 0.9858500361442566, 'reward_meter_mean': 0.9858500361442566, 'reward_meter_std': 0.013770055025815964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9858500361442566, 'reward_total_composite_std': 0.013770055025815964, 'epoch': 0.04}

 27%|██▋       | 883/3300 [30:46<3:47:20,  5.64s/it]INFO 04-11 23:07:38 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 884/3300 [30:53<3:56:27,  5.87s/it]2026-04-11 23:07:44,365 | INFO | train_grpo_train | metrics_logged mode=train step=884

                                                    
{'loss': 0.0017, 'grad_norm': 2.2200710773468018, 'learning_rate': 7.324242424242425e-06, 'num_tokens': 1975890.0, 'completions/mean_length': 178.625, 'completions/min_length': 172.0, 'completions/max_length': 187.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.625, 'completions/min_terminated_length': 172.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.6974109411239624, 'rewards/meter/std': 0.3949660360813141, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6805555820465088, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.39888328313827515, 'rewards/total_composite/std': 0.23110367357730865, 'reward': 0.39888328313827515, 'reward_std': 0.23110367357730865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015560545958578587, 'sampling/sampling_logp_difference/max': 2.096597194671631, 'sampling/importance_sampling_ratio/min': 0.12287382781505585, 'sampling/importance_sampling_ratio/mean': 0.9986303448677063, 'sampling/importance_sampling_ratio/max': 1.8987165689468384, 'entropy': 0.05528515903279185, 'clip_ratio/low_mean': 0.003500614082440734, 'clip_ratio/low_min': 0.003500614082440734, 'clip_ratio/high_mean': 0.008322211273480207, 'clip_ratio/high_max': 0.008322211273480207, 'clip_ratio/region_mean': 0.01182282535592094, 'reward_total_mean': 0.39888328313827515, 'reward_meter_mean': 0.6974109411239624, 'reward_meter_std': 0.3949660360813141, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6805555820465088, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.39888328313827515, 'reward_total_composite_std': 0.23110367357730865, 'epoch': 0.04}

 27%|██▋       | 884/3300 [30:53<3:56:27,  5.87s/it]INFO 04-11 23:07:44 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 885/3300 [30:58<3:42:47,  5.54s/it]2026-04-11 23:07:49,107 | INFO | train_grpo_train | metrics_logged mode=train step=885

                                                    
{'loss': 0.0102, 'grad_norm': 8.853222846984863, 'learning_rate': 7.321212121212122e-06, 'num_tokens': 1977749.0, 'completions/mean_length': 70.375, 'completions/min_length': 65.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.778793215751648, 'rewards/meter/std': 0.4060738682746887, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.778793215751648, 'rewards/total_composite/std': 0.4060738682746887, 'reward': 0.778793215751648, 'reward_std': 0.4060738682746887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028910191729664803, 'sampling/sampling_logp_difference/max': 2.095900535583496, 'sampling/importance_sampling_ratio/min': 0.12295946478843689, 'sampling/importance_sampling_ratio/mean': 0.9921159744262695, 'sampling/importance_sampling_ratio/max': 1.356745958328247, 'entropy': 0.10360199213027954, 'clip_ratio/low_mean': 0.007144315168261528, 'clip_ratio/low_min': 0.007144315168261528, 'clip_ratio/high_mean': 0.025362646207213402, 'clip_ratio/high_max': 0.025362646207213402, 'clip_ratio/region_mean': 0.03250696137547493, 'reward_total_mean': 0.778793215751648, 'reward_meter_mean': 0.778793215751648, 'reward_meter_std': 0.4060738682746887, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.778793215751648, 'reward_total_composite_std': 0.4060738682746887, 'epoch': 0.04}

 27%|██▋       | 885/3300 [30:58<3:42:47,  5.54s/it]INFO 04-11 23:07:49 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 886/3300 [31:05<4:08:22,  6.17s/it]2026-04-11 23:07:56,775 | INFO | train_grpo_train | metrics_logged mode=train step=886

                                                    
{'loss': 0.0196, 'grad_norm': 2.6481733322143555, 'learning_rate': 7.3181818181818186e-06, 'num_tokens': 1982015.0, 'completions/mean_length': 294.25, 'completions/min_length': 291.0, 'completions/max_length': 308.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 294.25, 'completions/min_terminated_length': 291.0, 'completions/max_terminated_length': 308.0, 'rewards/meter/mean': 0.8633161783218384, 'rewards/meter/std': 0.3173806965351105, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5424836874008179, 'rewards/repeat_penalty/std': 0.1294051706790924, 'rewards/total_composite/mean': 0.45384329557418823, 'rewards/total_composite/std': 0.1770940124988556, 'reward': 0.45384329557418823, 'reward_std': 0.1770940124988556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003619949799031019, 'sampling/sampling_logp_difference/max': 0.655219316482544, 'sampling/importance_sampling_ratio/min': 0.537617027759552, 'sampling/importance_sampling_ratio/mean': 1.0008361339569092, 'sampling/importance_sampling_ratio/max': 1.9255647659301758, 'entropy': 0.01867874281015247, 'clip_ratio/low_mean': 0.0012175324372947216, 'clip_ratio/low_min': 0.0012175324372947216, 'clip_ratio/high_mean': 0.002134879759978503, 'clip_ratio/high_max': 0.002134879759978503, 'clip_ratio/region_mean': 0.0033524121972732246, 'reward_total_mean': 0.45384329557418823, 'reward_meter_mean': 0.8633161783218384, 'reward_meter_std': 0.3173806965351105, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5424836874008179, 'reward_repeat_penalty_std': 0.1294051706790924, 'reward_total_composite_mean': 0.45384329557418823, 'reward_total_composite_std': 0.1770940124988556, 'epoch': 0.04}

 27%|██▋       | 886/3300 [31:05<4:08:22,  6.17s/it]INFO 04-11 23:07:57 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 887/3300 [31:10<3:56:22,  5.88s/it]2026-04-11 23:08:01,963 | INFO | train_grpo_train | metrics_logged mode=train step=887

                                                    
{'loss': -0.0139, 'grad_norm': 2.381399393081665, 'learning_rate': 7.315151515151516e-06, 'num_tokens': 1984181.0, 'completions/mean_length': 105.75, 'completions/min_length': 102.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.75, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9935408234596252, 'rewards/meter/std': 0.003044598735868931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8444880843162537, 'rewards/total_composite/std': 0.09176833927631378, 'reward': 0.8444880843162537, 'reward_std': 0.09176833927631378, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010191258043050766, 'sampling/sampling_logp_difference/max': 0.7071137428283691, 'sampling/importance_sampling_ratio/min': 0.4930652678012848, 'sampling/importance_sampling_ratio/mean': 1.0018950700759888, 'sampling/importance_sampling_ratio/max': 1.6447700262069702, 'entropy': 0.06286561442539096, 'clip_ratio/low_mean': 0.005965773249045014, 'clip_ratio/low_min': 0.005965773249045014, 'clip_ratio/high_mean': 0.0046083927154541016, 'clip_ratio/high_max': 0.0046083927154541016, 'clip_ratio/region_mean': 0.010574165964499116, 'reward_total_mean': 0.8444880843162537, 'reward_meter_mean': 0.9935408234596252, 'reward_meter_std': 0.003044598735868931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8444880843162537, 'reward_total_composite_std': 0.09176833927631378, 'epoch': 0.04}

 27%|██▋       | 887/3300 [31:10<3:56:22,  5.88s/it]INFO 04-11 23:08:02 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 888/3300 [31:15<3:41:21,  5.51s/it]2026-04-11 23:08:06,598 | INFO | train_grpo_train | metrics_logged mode=train step=888

                                                    
{'loss': 0.0134, 'grad_norm': 5.54829740524292, 'learning_rate': 7.312121212121212e-06, 'num_tokens': 1986101.0, 'completions/mean_length': 62.0, 'completions/min_length': 60.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9698032140731812, 'rewards/meter/std': 0.021262098103761673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9698032140731812, 'rewards/total_composite/std': 0.021262098103761673, 'reward': 0.9698032140731812, 'reward_std': 0.021262086927890778, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019981540739536285, 'sampling/sampling_logp_difference/max': 0.9673044681549072, 'sampling/importance_sampling_ratio/min': 0.38010627031326294, 'sampling/importance_sampling_ratio/mean': 1.0010225772857666, 'sampling/importance_sampling_ratio/max': 1.7581480741500854, 'entropy': 0.08477870561182499, 'clip_ratio/low_mean': 0.003969253972172737, 'clip_ratio/low_min': 0.003969253972172737, 'clip_ratio/high_mean': 0.006115591386333108, 'clip_ratio/high_max': 0.006115591386333108, 'clip_ratio/region_mean': 0.010084845358505845, 'reward_total_mean': 0.9698032140731812, 'reward_meter_mean': 0.9698032140731812, 'reward_meter_std': 0.021262098103761673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9698032140731812, 'reward_total_composite_std': 0.021262098103761673, 'epoch': 0.04}

 27%|██▋       | 888/3300 [31:15<3:41:21,  5.51s/it]INFO 04-11 23:08:06 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 889/3300 [31:20<3:33:14,  5.31s/it]2026-04-11 23:08:11,444 | INFO | train_grpo_train | metrics_logged mode=train step=889

                                                    
{'loss': 0.0078, 'grad_norm': 3.3663084506988525, 'learning_rate': 7.30909090909091e-06, 'num_tokens': 1987943.0, 'completions/mean_length': 72.25, 'completions/min_length': 69.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9931055307388306, 'rewards/meter/std': 0.0025689646136015654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9931055307388306, 'rewards/total_composite/std': 0.0025689646136015654, 'reward': 0.9931055307388306, 'reward_std': 0.002568964147940278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021964317187666893, 'sampling/sampling_logp_difference/max': 1.340165615081787, 'sampling/importance_sampling_ratio/min': 0.2618023157119751, 'sampling/importance_sampling_ratio/mean': 1.0010665655136108, 'sampling/importance_sampling_ratio/max': 1.9120845794677734, 'entropy': 0.09350762516260147, 'clip_ratio/low_mean': 0.010253431391902268, 'clip_ratio/low_min': 0.010253431391902268, 'clip_ratio/high_mean': 0.005093896761536598, 'clip_ratio/high_max': 0.005093896761536598, 'clip_ratio/region_mean': 0.015347328153438866, 'reward_total_mean': 0.9931055307388306, 'reward_meter_mean': 0.9931055307388306, 'reward_meter_std': 0.0025689646136015654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9931055307388306, 'reward_total_composite_std': 0.0025689646136015654, 'epoch': 0.04}

 27%|██▋       | 889/3300 [31:20<3:33:14,  5.31s/it]INFO 04-11 23:08:11 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 890/3300 [31:26<3:42:24,  5.54s/it]2026-04-11 23:08:17,513 | INFO | train_grpo_train | metrics_logged mode=train step=890

                                                    
{'loss': -0.005, 'grad_norm': 2.6016640663146973, 'learning_rate': 7.306060606060607e-06, 'num_tokens': 1990330.0, 'completions/mean_length': 145.375, 'completions/min_length': 134.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 145.375, 'completions/min_terminated_length': 134.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9830442667007446, 'rewards/meter/std': 0.01036095805466175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.7374788522720337, 'rewards/total_composite/std': 0.0681464672088623, 'reward': 0.7374788522720337, 'reward_std': 0.06814645975828171, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021481409668922424, 'sampling/sampling_logp_difference/max': 1.0175132751464844, 'sampling/importance_sampling_ratio/min': 0.361492782831192, 'sampling/importance_sampling_ratio/mean': 1.0078076124191284, 'sampling/importance_sampling_ratio/max': 1.7661346197128296, 'entropy': 0.146129728294909, 'clip_ratio/low_mean': 0.014987432223279029, 'clip_ratio/low_min': 0.014987432223279029, 'clip_ratio/high_mean': 0.002556844614446163, 'clip_ratio/high_max': 0.002556844614446163, 'clip_ratio/region_mean': 0.017544276837725192, 'reward_total_mean': 0.7374788522720337, 'reward_meter_mean': 0.9830442667007446, 'reward_meter_std': 0.01036095805466175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.7374788522720337, 'reward_total_composite_std': 0.0681464672088623, 'epoch': 0.04}

 27%|██▋       | 890/3300 [31:26<3:42:24,  5.54s/it]INFO 04-11 23:08:17 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 891/3300 [31:31<3:33:53,  5.33s/it]2026-04-11 23:08:22,354 | INFO | train_grpo_train | metrics_logged mode=train step=891

                                                    
{'loss': 0.0093, 'grad_norm': 2.81315279006958, 'learning_rate': 7.303030303030304e-06, 'num_tokens': 1992094.0, 'completions/mean_length': 73.5, 'completions/min_length': 71.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.8690831661224365, 'rewards/meter/std': 0.3090604245662689, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8690831661224365, 'rewards/total_composite/std': 0.3090604245662689, 'reward': 0.8690831661224365, 'reward_std': 0.30906039476394653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03247151896357536, 'sampling/sampling_logp_difference/max': 1.3969697952270508, 'sampling/importance_sampling_ratio/min': 0.24734534323215485, 'sampling/importance_sampling_ratio/mean': 1.0030601024627686, 'sampling/importance_sampling_ratio/max': 1.6693646907806396, 'entropy': 0.19349019043147564, 'clip_ratio/low_mean': 0.0016666667070239782, 'clip_ratio/low_min': 0.0016666667070239782, 'clip_ratio/high_mean': 0.018530045170336962, 'clip_ratio/high_max': 0.018530045170336962, 'clip_ratio/region_mean': 0.02019671187736094, 'reward_total_mean': 0.8690831661224365, 'reward_meter_mean': 0.8690831661224365, 'reward_meter_std': 0.3090604245662689, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8690831661224365, 'reward_total_composite_std': 0.3090604245662689, 'epoch': 0.04}

 27%|██▋       | 891/3300 [31:31<3:33:53,  5.33s/it]INFO 04-11 23:08:22 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 892/3300 [31:36<3:32:43,  5.30s/it]2026-04-11 23:08:27,590 | INFO | train_grpo_train | metrics_logged mode=train step=892

                                                    
{'loss': 0.1432, 'grad_norm': 4.598532199859619, 'learning_rate': 7.3e-06, 'num_tokens': 1994275.0, 'completions/mean_length': 99.625, 'completions/min_length': 72.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.994989275932312, 'rewards/meter/std': 0.0016277554677799344, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.12817399203777313, 'rewards/total_composite/mean': 0.5349372625350952, 'rewards/total_composite/std': 0.2865643799304962, 'reward': 0.5349372625350952, 'reward_std': 0.2865643799304962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02673969604074955, 'sampling/sampling_logp_difference/max': 1.9094719886779785, 'sampling/importance_sampling_ratio/min': 0.14815859496593475, 'sampling/importance_sampling_ratio/mean': 1.0079816579818726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1309506855905056, 'clip_ratio/low_mean': 0.013621801743283868, 'clip_ratio/low_min': 0.013621801743283868, 'clip_ratio/high_mean': 0.003402777831070125, 'clip_ratio/high_max': 0.003402777831070125, 'clip_ratio/region_mean': 0.017024579574353993, 'reward_total_mean': 0.5349372625350952, 'reward_meter_mean': 0.994989275932312, 'reward_meter_std': 0.0016277554677799344, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.12817399203777313, 'reward_total_composite_mean': 0.5349372625350952, 'reward_total_composite_std': 0.2865643799304962, 'epoch': 0.04}

 27%|██▋       | 892/3300 [31:36<3:32:43,  5.30s/it]INFO 04-11 23:08:27 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 893/3300 [31:42<3:40:08,  5.49s/it]2026-04-11 23:08:33,512 | INFO | train_grpo_train | metrics_logged mode=train step=893

                                                    
{'loss': -0.0066, 'grad_norm': 3.3807826042175293, 'learning_rate': 7.296969696969698e-06, 'num_tokens': 1996973.0, 'completions/mean_length': 146.25, 'completions/min_length': 143.0, 'completions/max_length': 151.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 146.25, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.024410676211118698, 'rewards/meter/std': 0.046276357024908066, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.017994364723563194, 'rewards/total_composite/std': 0.03303234279155731, 'reward': 0.017994364723563194, 'reward_std': 0.033032339066267014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023801125586032867, 'sampling/sampling_logp_difference/max': 1.7393649816513062, 'sampling/importance_sampling_ratio/min': 0.17563189566135406, 'sampling/importance_sampling_ratio/mean': 1.000627875328064, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13604794908314943, 'clip_ratio/low_mean': 0.006828531681094319, 'clip_ratio/low_min': 0.006828531681094319, 'clip_ratio/high_mean': 0.009196062572300434, 'clip_ratio/high_max': 0.009196062572300434, 'clip_ratio/region_mean': 0.016024594253394753, 'reward_total_mean': 0.017994364723563194, 'reward_meter_mean': 0.024410676211118698, 'reward_meter_std': 0.046276357024908066, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.017994364723563194, 'reward_total_composite_std': 0.03303234279155731, 'epoch': 0.04}

 27%|██▋       | 893/3300 [31:42<3:40:08,  5.49s/it]INFO 04-11 23:08:33 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 894/3300 [31:47<3:29:48,  5.23s/it]2026-04-11 23:08:38,152 | INFO | train_grpo_train | metrics_logged mode=train step=894

                                                    
{'loss': 0.0158, 'grad_norm': 3.678330659866333, 'learning_rate': 7.293939393939394e-06, 'num_tokens': 1998826.0, 'completions/mean_length': 64.625, 'completions/min_length': 63.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.957672119140625, 'rewards/meter/std': 0.010208838619291782, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.957672119140625, 'rewards/total_composite/std': 0.010208838619291782, 'reward': 0.957672119140625, 'reward_std': 0.01020884234458208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00946664996445179, 'sampling/sampling_logp_difference/max': 1.2841720581054688, 'sampling/importance_sampling_ratio/min': 0.2768797278404236, 'sampling/importance_sampling_ratio/mean': 0.9999814033508301, 'sampling/importance_sampling_ratio/max': 1.4367311000823975, 'entropy': 0.04335960140451789, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005890377098694444, 'clip_ratio/high_max': 0.005890377098694444, 'clip_ratio/region_mean': 0.005890377098694444, 'reward_total_mean': 0.957672119140625, 'reward_meter_mean': 0.957672119140625, 'reward_meter_std': 0.010208838619291782, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.957672119140625, 'reward_total_composite_std': 0.010208838619291782, 'epoch': 0.04}

 27%|██▋       | 894/3300 [31:47<3:29:48,  5.23s/it]INFO 04-11 23:08:38 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 895/3300 [31:52<3:29:27,  5.23s/it]2026-04-11 23:08:43,368 | INFO | train_grpo_train | metrics_logged mode=train step=895

                                                    
{'loss': -0.0048, 'grad_norm': 3.20367169380188, 'learning_rate': 7.290909090909092e-06, 'num_tokens': 2000706.0, 'completions/mean_length': 67.0, 'completions/min_length': 66.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.6329219341278076, 'rewards/meter/std': 0.4643419682979584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6329219341278076, 'rewards/total_composite/std': 0.4643419682979584, 'reward': 0.6329219341278076, 'reward_std': 0.4643419682979584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015903156250715256, 'sampling/sampling_logp_difference/max': 0.6511527895927429, 'sampling/importance_sampling_ratio/min': 0.5214443206787109, 'sampling/importance_sampling_ratio/mean': 1.006116509437561, 'sampling/importance_sampling_ratio/max': 1.8657705783843994, 'entropy': 0.07575948582962155, 'clip_ratio/low_mean': 0.011278833262622356, 'clip_ratio/low_min': 0.011278833262622356, 'clip_ratio/high_mean': 0.012951546465046704, 'clip_ratio/high_max': 0.012951546465046704, 'clip_ratio/region_mean': 0.02423037972766906, 'reward_total_mean': 0.6329219341278076, 'reward_meter_mean': 0.6329219341278076, 'reward_meter_std': 0.4643419682979584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6329219341278076, 'reward_total_composite_std': 0.4643419682979584, 'epoch': 0.04}

 27%|██▋       | 895/3300 [31:52<3:29:27,  5.23s/it]INFO 04-11 23:08:43 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 896/3300 [31:57<3:23:03,  5.07s/it]2026-04-11 23:08:48,060 | INFO | train_grpo_train | metrics_logged mode=train step=896

                                                    
{'loss': 0.0424, 'grad_norm': 17.783985137939453, 'learning_rate': 7.287878787878789e-06, 'num_tokens': 2002459.0, 'completions/mean_length': 63.125, 'completions/min_length': 59.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.926885187625885, 'rewards/meter/std': 0.12749086320400238, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.926885187625885, 'rewards/total_composite/std': 0.12749086320400238, 'reward': 0.926885187625885, 'reward_std': 0.12749086320400238, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04450727999210358, 'sampling/sampling_logp_difference/max': 1.5940380096435547, 'sampling/importance_sampling_ratio/min': 0.20310382544994354, 'sampling/importance_sampling_ratio/mean': 0.9990435838699341, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24534680880606174, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.03242240101099014, 'clip_ratio/high_max': 0.03242240101099014, 'clip_ratio/region_mean': 0.03999815881252289, 'reward_total_mean': 0.926885187625885, 'reward_meter_mean': 0.926885187625885, 'reward_meter_std': 0.12749086320400238, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.926885187625885, 'reward_total_composite_std': 0.12749086320400238, 'epoch': 0.04}

 27%|██▋       | 896/3300 [31:57<3:23:03,  5.07s/it]INFO 04-11 23:08:48 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 897/3300 [32:02<3:22:33,  5.06s/it]2026-04-11 23:08:53,092 | INFO | train_grpo_train | metrics_logged mode=train step=897

                                                    
{'loss': 0.0518, 'grad_norm': 9.401634216308594, 'learning_rate': 7.284848484848486e-06, 'num_tokens': 2004481.0, 'completions/mean_length': 75.75, 'completions/min_length': 72.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9812333583831787, 'rewards/meter/std': 0.03400373458862305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9812333583831787, 'rewards/total_composite/std': 0.03400373458862305, 'reward': 0.9812333583831787, 'reward_std': 0.03400372713804245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04621249809861183, 'sampling/sampling_logp_difference/max': 1.1664955615997314, 'sampling/importance_sampling_ratio/min': 0.31145650148391724, 'sampling/importance_sampling_ratio/mean': 1.0046457052230835, 'sampling/importance_sampling_ratio/max': 1.990276575088501, 'entropy': 0.3472878560423851, 'clip_ratio/low_mean': 0.004629629664123058, 'clip_ratio/low_min': 0.004629629664123058, 'clip_ratio/high_mean': 0.02354571979958564, 'clip_ratio/high_max': 0.02354571979958564, 'clip_ratio/region_mean': 0.0281753494637087, 'reward_total_mean': 0.9812333583831787, 'reward_meter_mean': 0.9812333583831787, 'reward_meter_std': 0.03400373458862305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9812333583831787, 'reward_total_composite_std': 0.03400373458862305, 'epoch': 0.04}

 27%|██▋       | 897/3300 [32:02<3:22:33,  5.06s/it]INFO 04-11 23:08:53 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 898/3300 [32:06<3:17:54,  4.94s/it]2026-04-11 23:08:57,770 | INFO | train_grpo_train | metrics_logged mode=train step=898

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.281818181818182e-06, 'num_tokens': 2006258.0, 'completions/mean_length': 64.125, 'completions/min_length': 63.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9789108037948608, 'rewards/meter/std': 0.013038679957389832, 'rewards/count_adherence/mean': 0.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.0, 'rewards/total_composite/std': 0.0, 'reward': 0.0, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.01310182549059391, 'sampling/sampling_logp_difference/max': 1.4251065254211426, 'sampling/importance_sampling_ratio/min': 0.2404828518629074, 'sampling/importance_sampling_ratio/mean': 1.0009260177612305, 'sampling/importance_sampling_ratio/max': 1.383298635482788, 'entropy': 0.07330859638750553, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.0, 'reward_meter_mean': 0.9789108037948608, 'reward_meter_std': 0.013038679957389832, 'reward_count_adherence_mean': 0.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.0, 'reward_total_composite_std': 0.0, 'epoch': 0.04}

 27%|██▋       | 898/3300 [32:06<3:17:54,  4.94s/it]INFO 04-11 23:08:58 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 899/3300 [32:12<3:22:35,  5.06s/it]2026-04-11 23:09:03,113 | INFO | train_grpo_train | metrics_logged mode=train step=899

                                                    
{'loss': -0.0124, 'grad_norm': 3.2133944034576416, 'learning_rate': 7.2787878787878795e-06, 'num_tokens': 2008460.0, 'completions/mean_length': 111.25, 'completions/min_length': 101.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.25, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.7485231161117554, 'rewards/meter/std': 0.4493827819824219, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.6480754613876343, 'rewards/total_composite/std': 0.3985969126224518, 'reward': 0.6480754613876343, 'reward_std': 0.3985969126224518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021135006099939346, 'sampling/sampling_logp_difference/max': 1.0010976791381836, 'sampling/importance_sampling_ratio/min': 0.36747586727142334, 'sampling/importance_sampling_ratio/mean': 1.003222942352295, 'sampling/importance_sampling_ratio/max': 1.8294382095336914, 'entropy': 0.1524664619937539, 'clip_ratio/low_mean': 0.004611999727785587, 'clip_ratio/low_min': 0.004611999727785587, 'clip_ratio/high_mean': 0.010862766648642719, 'clip_ratio/high_max': 0.010862766648642719, 'clip_ratio/region_mean': 0.015474766376428306, 'reward_total_mean': 0.6480754613876343, 'reward_meter_mean': 0.7485231161117554, 'reward_meter_std': 0.4493827819824219, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.6480754613876343, 'reward_total_composite_std': 0.3985969126224518, 'epoch': 0.04}

 27%|██▋       | 899/3300 [32:12<3:22:35,  5.06s/it]INFO 04-11 23:09:03 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 900/3300 [32:16<3:18:46,  4.97s/it]2026-04-11 23:09:07,869 | INFO | train_grpo_train | metrics_logged mode=train step=900

                                                    
{'loss': 0.0111, 'grad_norm': 6.62265157699585, 'learning_rate': 7.275757575757576e-06, 'num_tokens': 2010132.0, 'completions/mean_length': 54.0, 'completions/min_length': 49.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.47188466787338257, 'rewards/meter/std': 0.3173547685146332, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.47188466787338257, 'rewards/total_composite/std': 0.3173547685146332, 'reward': 0.47188466787338257, 'reward_std': 0.3173547685146332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05849804729223251, 'sampling/sampling_logp_difference/max': 1.0531864166259766, 'sampling/importance_sampling_ratio/min': 0.34882447123527527, 'sampling/importance_sampling_ratio/mean': 1.0136620998382568, 'sampling/importance_sampling_ratio/max': 1.9393072128295898, 'entropy': 0.4540216512978077, 'clip_ratio/low_mean': 0.024088865146040916, 'clip_ratio/low_min': 0.024088865146040916, 'clip_ratio/high_mean': 0.02068706788122654, 'clip_ratio/high_max': 0.02068706788122654, 'clip_ratio/region_mean': 0.044775933027267456, 'reward_total_mean': 0.47188466787338257, 'reward_meter_mean': 0.47188466787338257, 'reward_meter_std': 0.3173547685146332, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.47188466787338257, 'reward_total_composite_std': 0.3173547685146332, 'epoch': 0.04}

 27%|██▋       | 900/3300 [32:16<3:18:46,  4.97s/it]INFO 04-11 23:09:08 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:31,  2.88s/it]

 23%|██▎       | 3/13 [00:09<00:31,  3.11s/it]

 31%|███       | 4/13 [00:16<00:41,  4.61s/it]

 38%|███▊      | 5/13 [00:21<00:38,  4.77s/it]

 46%|████▌     | 6/13 [00:26<00:34,  4.97s/it]

 54%|█████▍    | 7/13 [00:31<00:29,  4.89s/it]

 62%|██████▏   | 8/13 [00:37<00:25,  5.18s/it]

 69%|██████▉   | 9/13 [00:43<00:22,  5.60s/it]

 77%|███████▋  | 10/13 [00:51<00:18,  6.19s/it]

 85%|████████▍ | 11/13 [00:56<00:11,  5.93s/it]

 92%|█████████▏| 12/13 [01:03<00:06,  6.33s/it]

100%|██████████| 13/13 [01:09<00:00,  6.19s/it]2026-04-11 23:10:24,266 | INFO | train_grpo_train | metrics_logged mode=eval step=900

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 76.3688, 'eval_samples_per_second': 1.362, 'eval_steps_per_second': 0.17, 'eval_num_tokens': 2010132.0, 'eval_completions/mean_length': 220.43269230769232, 'eval_completions/min_length': 64.38461538461539, 'eval_completions/max_length': 402.61538461538464, 'eval_completions/clipped_ratio': 0.028846153846153848, 'eval_completions/mean_terminated_length': 212.01373877892127, 'eval_completions/min_terminated_length': 64.38461538461539, 'eval_completions/max_terminated_length': 370.38461538461536, 'eval_rewards/meter/mean': 0.607175561097952, 'eval_rewards/meter/std': 0.4337661495575538, 'eval_rewards/count_adherence/mean': 0.8996222019195557, 'eval_rewards/count_adherence/std': 0.13804738968610764, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.6879472640844492, 'eval_rewards/repeat_penalty/std': 0.20650707471829194, 'eval_rewards/total_composite/mean': 0.3643972415190477, 'eval_rewards/total_composite/std': 0.31932372657152325, 'eval_reward': 0.3643972415190477, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.008556847639668446, 'eval_sampling/sampling_logp_difference/max': 0.8173254269819993, 'eval_sampling/importance_sampling_ratio/min': 0.4500999932105725, 'eval_sampling/importance_sampling_ratio/mean': 1.002149930367103, 'eval_sampling/importance_sampling_ratio/max': 1.312718914105342, 'eval_entropy': 0.0762252899316641, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.3643972415190477, 'eval_reward_meter_mean': 0.607175561097952, 'eval_reward_meter_std': 0.4337661495575538, 'eval_reward_count_adherence_mean': 0.8996222019195557, 'eval_reward_count_adherence_std': 0.13804738968610764, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.6879472640844492, 'eval_reward_repeat_penalty_std': 0.20650707471829194, 'eval_reward_total_composite_mean': 0.3643972415190477, 'eval_reward_total_composite_std': 0.31932372657152325, 'epoch': 0.04}

 27%|██▋       | 900/3300 [33:33<3:18:46,  4.97s/it]

100%|██████████| 13/13 [01:09<00:00,  6.19s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 27%|██▋       | 901/3300 [33:41<19:15:20, 28.90s/it]2026-04-11 23:10:32,590 | INFO | train_grpo_train | metrics_logged mode=train step=901

                                                     
{'loss': 0.0221, 'grad_norm': 4.90191650390625, 'learning_rate': 7.272727272727273e-06, 'num_tokens': 2011886.0, 'completions/mean_length': 65.25, 'completions/min_length': 62.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.25, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.2965307831764221, 'rewards/meter/std': 0.3661009669303894, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.2965307831764221, 'rewards/total_composite/std': 0.3661009669303894, 'reward': 0.2965307831764221, 'reward_std': 0.366100937128067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05390980467200279, 'sampling/sampling_logp_difference/max': 1.267104148864746, 'sampling/importance_sampling_ratio/min': 0.28164607286453247, 'sampling/importance_sampling_ratio/mean': 1.0055922269821167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39315336383879185, 'clip_ratio/low_mean': 0.02774310251697898, 'clip_ratio/low_min': 0.02774310251697898, 'clip_ratio/high_mean': 0.00974025996401906, 'clip_ratio/high_max': 0.00974025996401906, 'clip_ratio/region_mean': 0.03748336248099804, 'reward_total_mean': 0.2965307831764221, 'reward_meter_mean': 0.2965307831764221, 'reward_meter_std': 0.3661009669303894, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.2965307831764221, 'reward_total_composite_std': 0.3661009669303894, 'epoch': 0.04}

 27%|██▋       | 901/3300 [33:41<19:15:20, 28.90s/it]INFO 04-11 23:10:32 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 902/3300 [33:47<14:33:32, 21.86s/it]2026-04-11 23:10:38,021 | INFO | train_grpo_train | metrics_logged mode=train step=902

                                                     
{'loss': 0.1103, 'grad_norm': 2.6275370121002197, 'learning_rate': 7.26969696969697e-06, 'num_tokens': 2014158.0, 'completions/mean_length': 108.0, 'completions/min_length': 92.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.0, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9951065182685852, 'rewards/meter/std': 0.0021594627760350704, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7571429014205933, 'rewards/repeat_penalty/std': 0.04581620916724205, 'rewards/total_composite/mean': 0.6347411870956421, 'rewards/total_composite/std': 0.17126502096652985, 'reward': 0.6347411870956421, 'reward_std': 0.17126502096652985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019786102697253227, 'sampling/sampling_logp_difference/max': 0.9821138381958008, 'sampling/importance_sampling_ratio/min': 0.37451860308647156, 'sampling/importance_sampling_ratio/mean': 1.0035145282745361, 'sampling/importance_sampling_ratio/max': 1.7781134843826294, 'entropy': 0.10153948981314898, 'clip_ratio/low_mean': 0.011442428571172059, 'clip_ratio/low_min': 0.011442428571172059, 'clip_ratio/high_mean': 0.016934875398874283, 'clip_ratio/high_max': 0.016934875398874283, 'clip_ratio/region_mean': 0.02837730397004634, 'reward_total_mean': 0.6347411870956421, 'reward_meter_mean': 0.9951065182685852, 'reward_meter_std': 0.0021594627760350704, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7571429014205933, 'reward_repeat_penalty_std': 0.04581620916724205, 'reward_total_composite_mean': 0.6347411870956421, 'reward_total_composite_std': 0.17126502096652985, 'epoch': 0.04}

 27%|██▋       | 902/3300 [33:47<14:33:32, 21.86s/it]INFO 04-11 23:10:38 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 903/3300 [33:52<11:14:51, 16.89s/it]2026-04-11 23:10:43,333 | INFO | train_grpo_train | metrics_logged mode=train step=903

                                                     
{'loss': -0.0025, 'grad_norm': 1.540696620941162, 'learning_rate': 7.266666666666668e-06, 'num_tokens': 2016340.0, 'completions/mean_length': 103.75, 'completions/min_length': 103.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.75, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9977596998214722, 'rewards/meter/std': 0.00013270843192003667, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.773260235786438, 'rewards/total_composite/std': 0.07051651179790497, 'reward': 0.773260235786438, 'reward_std': 0.07051652669906616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012240896932780743, 'sampling/sampling_logp_difference/max': 1.134617805480957, 'sampling/importance_sampling_ratio/min': 0.32154497504234314, 'sampling/importance_sampling_ratio/mean': 1.0004560947418213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05680669890716672, 'clip_ratio/low_mean': 0.0012135922443121672, 'clip_ratio/low_min': 0.0012135922443121672, 'clip_ratio/high_mean': 0.009641934651881456, 'clip_ratio/high_max': 0.009641934651881456, 'clip_ratio/region_mean': 0.010855526896193624, 'reward_total_mean': 0.773260235786438, 'reward_meter_mean': 0.9977596998214722, 'reward_meter_std': 0.00013270843192003667, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.773260235786438, 'reward_total_composite_std': 0.07051651179790497, 'epoch': 0.04}

 27%|██▋       | 903/3300 [33:52<11:14:51, 16.89s/it]INFO 04-11 23:10:43 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 904/3300 [34:02<9:50:30, 14.79s/it] 2026-04-11 23:10:53,208 | INFO | train_grpo_train | metrics_logged mode=train step=904

                                                    
{'loss': -0.0841, 'grad_norm': 2.963177442550659, 'learning_rate': 7.263636363636364e-06, 'num_tokens': 2017997.0, 'completions/mean_length': 97.125, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.85714340209961, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8145281076431274, 'rewards/meter/std': 0.27075159549713135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7298738360404968, 'rewards/total_composite/std': 0.3964882493019104, 'reward': 0.7298738360404968, 'reward_std': 0.3964882493019104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09322468191385269, 'sampling/sampling_logp_difference/max': 1.8810420036315918, 'sampling/importance_sampling_ratio/min': 0.1524311900138855, 'sampling/importance_sampling_ratio/mean': 1.0168981552124023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5050923340022564, 'clip_ratio/low_mean': 0.00657894741743803, 'clip_ratio/low_min': 0.00657894741743803, 'clip_ratio/high_mean': 0.02894315170124173, 'clip_ratio/high_max': 0.02894315170124173, 'clip_ratio/region_mean': 0.03552209911867976, 'reward_total_mean': 0.7298738360404968, 'reward_meter_mean': 0.8145281076431274, 'reward_meter_std': 0.27075159549713135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7298738360404968, 'reward_total_composite_std': 0.3964882493019104, 'epoch': 0.04}

 27%|██▋       | 904/3300 [34:02<9:50:30, 14.79s/it]INFO 04-11 23:10:53 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 905/3300 [34:08<8:03:53, 12.12s/it]2026-04-11 23:10:59,112 | INFO | train_grpo_train | metrics_logged mode=train step=905

                                                    
{'loss': 0.0081, 'grad_norm': 1.4031760692596436, 'learning_rate': 7.260606060606061e-06, 'num_tokens': 2020527.0, 'completions/mean_length': 153.25, 'completions/min_length': 139.0, 'completions/max_length': 162.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 153.25, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.9857839941978455, 'rewards/meter/std': 0.010634319856762886, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6071429252624512, 'rewards/repeat_penalty/std': 0.147871196269989, 'rewards/total_composite/mean': 0.5992619395256042, 'rewards/total_composite/std': 0.14893361926078796, 'reward': 0.5992619395256042, 'reward_std': 0.14893361926078796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008156189695000648, 'sampling/sampling_logp_difference/max': 0.697784423828125, 'sampling/importance_sampling_ratio/min': 0.5704922080039978, 'sampling/importance_sampling_ratio/mean': 1.0026439428329468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.048865064047276974, 'clip_ratio/low_mean': 0.00478401588043198, 'clip_ratio/low_min': 0.00478401588043198, 'clip_ratio/high_mean': 0.004120342840906233, 'clip_ratio/high_max': 0.004120342840906233, 'clip_ratio/region_mean': 0.008904358721338212, 'reward_total_mean': 0.5992619395256042, 'reward_meter_mean': 0.9857839941978455, 'reward_meter_std': 0.010634319856762886, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6071429252624512, 'reward_repeat_penalty_std': 0.147871196269989, 'reward_total_composite_mean': 0.5992619395256042, 'reward_total_composite_std': 0.14893361926078796, 'epoch': 0.04}

 27%|██▋       | 905/3300 [34:08<8:03:53, 12.12s/it]INFO 04-11 23:10:59 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 906/3300 [34:17<7:33:45, 11.37s/it]2026-04-11 23:11:08,735 | INFO | train_grpo_train | metrics_logged mode=train step=906

                                                    
{'loss': -0.0066, 'grad_norm': 3.238330125808716, 'learning_rate': 7.257575757575758e-06, 'num_tokens': 2022039.0, 'completions/mean_length': 98.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.85714340209961, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8625515699386597, 'rewards/meter/std': 0.34917646646499634, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7457597255706787, 'rewards/total_composite/std': 0.4603087902069092, 'reward': 0.7457597255706787, 'reward_std': 0.4603087604045868, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.053011078387498856, 'sampling/sampling_logp_difference/max': 2.6139426231384277, 'sampling/importance_sampling_ratio/min': 0.07324519753456116, 'sampling/importance_sampling_ratio/mean': 0.988756537437439, 'sampling/importance_sampling_ratio/max': 1.4525083303451538, 'entropy': 0.1836453713476658, 'clip_ratio/low_mean': 0.003846153849735856, 'clip_ratio/low_min': 0.003846153849735856, 'clip_ratio/high_mean': 0.03257575840689242, 'clip_ratio/high_max': 0.03257575840689242, 'clip_ratio/region_mean': 0.036421912256628275, 'reward_total_mean': 0.7457597255706787, 'reward_meter_mean': 0.8625515699386597, 'reward_meter_std': 0.34917646646499634, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7457597255706787, 'reward_total_composite_std': 0.4603087902069092, 'epoch': 0.04}

 27%|██▋       | 906/3300 [34:17<7:33:45, 11.37s/it]INFO 04-11 23:11:08 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 907/3300 [34:25<6:52:56, 10.35s/it]2026-04-11 23:11:16,708 | INFO | train_grpo_train | metrics_logged mode=train step=907

                                                    
{'loss': -0.0466, 'grad_norm': 2.8838396072387695, 'learning_rate': 7.254545454545455e-06, 'num_tokens': 2025618.0, 'completions/mean_length': 257.375, 'completions/min_length': 241.0, 'completions/max_length': 293.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 257.375, 'completions/min_terminated_length': 241.0, 'completions/max_terminated_length': 293.0, 'rewards/meter/mean': 0.1675853729248047, 'rewards/meter/std': 0.09037812799215317, 'rewards/count_adherence/mean': 0.9166666269302368, 'rewards/count_adherence/std': 0.05143444612622261, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.31433823704719543, 'rewards/repeat_penalty/std': 0.18327650427818298, 'rewards/total_composite/mean': 0.058102987706661224, 'rewards/total_composite/std': 0.07150173932313919, 'reward': 0.058102987706661224, 'reward_std': 0.0715017318725586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008167757652699947, 'sampling/sampling_logp_difference/max': 1.5336222648620605, 'sampling/importance_sampling_ratio/min': 0.2157527506351471, 'sampling/importance_sampling_ratio/mean': 1.0021930932998657, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.030403183307498693, 'clip_ratio/low_mean': 0.0029962139669805765, 'clip_ratio/low_min': 0.0029962139669805765, 'clip_ratio/high_mean': 0.0027064846362918615, 'clip_ratio/high_max': 0.0027064846362918615, 'clip_ratio/region_mean': 0.005702698603272438, 'reward_total_mean': 0.058102987706661224, 'reward_meter_mean': 0.1675853729248047, 'reward_meter_std': 0.09037812799215317, 'reward_count_adherence_mean': 0.9166666269302368, 'reward_count_adherence_std': 0.05143444612622261, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.31433823704719543, 'reward_repeat_penalty_std': 0.18327650427818298, 'reward_total_composite_mean': 0.058102987706661224, 'reward_total_composite_std': 0.07150173932313919, 'epoch': 0.04}

 27%|██▋       | 907/3300 [34:25<6:52:56, 10.35s/it]INFO 04-11 23:11:16 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 908/3300 [34:35<6:44:56, 10.16s/it]2026-04-11 23:11:26,409 | INFO | train_grpo_train | metrics_logged mode=train step=908

                                                    
{'loss': -0.183, 'grad_norm': 0.40971025824546814, 'learning_rate': 7.251515151515151e-06, 'num_tokens': 2027442.0, 'completions/mean_length': 132.0, 'completions/min_length': 76.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 77.71428680419922, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.8670340180397034, 'rewards/meter/std': 0.3502245545387268, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8669998645782471, 'rewards/total_composite/std': 0.35032105445861816, 'reward': 0.8669998645782471, 'reward_std': 0.35032105445861816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010826216079294682, 'sampling/sampling_logp_difference/max': 0.651118278503418, 'sampling/importance_sampling_ratio/min': 0.5214623212814331, 'sampling/importance_sampling_ratio/mean': 1.0027331113815308, 'sampling/importance_sampling_ratio/max': 1.2967445850372314, 'entropy': 0.07274728454649448, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007976973778568208, 'clip_ratio/high_max': 0.007976973778568208, 'clip_ratio/region_mean': 0.007976973778568208, 'reward_total_mean': 0.8669998645782471, 'reward_meter_mean': 0.8670340180397034, 'reward_meter_std': 0.3502245545387268, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8669998645782471, 'reward_total_composite_std': 0.35032105445861816, 'epoch': 0.04}

 28%|██▊       | 908/3300 [34:35<6:44:56, 10.16s/it]INFO 04-11 23:11:26 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 909/3300 [34:41<5:52:35,  8.85s/it]2026-04-11 23:11:32,202 | INFO | train_grpo_train | metrics_logged mode=train step=909

                                                    
{'loss': 0.0147, 'grad_norm': 2.3503267765045166, 'learning_rate': 7.2484848484848495e-06, 'num_tokens': 2030075.0, 'completions/mean_length': 151.125, 'completions/min_length': 135.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.125, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.8722712993621826, 'rewards/meter/std': 0.33660420775413513, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.6414377689361572, 'rewards/total_composite/std': 0.2504880726337433, 'reward': 0.6414377689361572, 'reward_std': 0.2504880726337433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009546966291964054, 'sampling/sampling_logp_difference/max': 0.9272332191467285, 'sampling/importance_sampling_ratio/min': 0.395646870136261, 'sampling/importance_sampling_ratio/mean': 1.0012683868408203, 'sampling/importance_sampling_ratio/max': 1.8332045078277588, 'entropy': 0.051733002765104175, 'clip_ratio/low_mean': 0.003980891779065132, 'clip_ratio/low_min': 0.003980891779065132, 'clip_ratio/high_mean': 0.002550398523453623, 'clip_ratio/high_max': 0.002550398523453623, 'clip_ratio/region_mean': 0.006531290302518755, 'reward_total_mean': 0.6414377689361572, 'reward_meter_mean': 0.8722712993621826, 'reward_meter_std': 0.33660420775413513, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.6414377689361572, 'reward_total_composite_std': 0.2504880726337433, 'epoch': 0.04}

 28%|██▊       | 909/3300 [34:41<5:52:35,  8.85s/it]INFO 04-11 23:11:32 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 910/3300 [34:49<5:40:25,  8.55s/it]2026-04-11 23:11:40,044 | INFO | train_grpo_train | metrics_logged mode=train step=910

                                                    
{'loss': -0.0173, 'grad_norm': 1.5391802787780762, 'learning_rate': 7.245454545454546e-06, 'num_tokens': 2034379.0, 'completions/mean_length': 297.0, 'completions/min_length': 274.0, 'completions/max_length': 318.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 297.0, 'completions/min_terminated_length': 274.0, 'completions/max_terminated_length': 318.0, 'rewards/meter/mean': 0.10647714138031006, 'rewards/meter/std': 0.29925623536109924, 'rewards/count_adherence/mean': 0.890625, 'rewards/count_adherence/std': 0.08010874688625336, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6176573634147644, 'rewards/repeat_penalty/std': 0.15113945305347443, 'rewards/total_composite/mean': 0.06387612223625183, 'rewards/total_composite/std': 0.1795579344034195, 'reward': 0.06387612223625183, 'reward_std': 0.1795579344034195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02341293916106224, 'sampling/sampling_logp_difference/max': 4.994693756103516, 'sampling/importance_sampling_ratio/min': 0.0067737954668700695, 'sampling/importance_sampling_ratio/mean': 0.999440610408783, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10688875103369355, 'clip_ratio/low_mean': 0.011030830384697765, 'clip_ratio/low_min': 0.011030830384697765, 'clip_ratio/high_mean': 0.005175159312784672, 'clip_ratio/high_max': 0.005175159312784672, 'clip_ratio/region_mean': 0.016205989697482437, 'reward_total_mean': 0.06387612223625183, 'reward_meter_mean': 0.10647714138031006, 'reward_meter_std': 0.29925623536109924, 'reward_count_adherence_mean': 0.890625, 'reward_count_adherence_std': 0.08010874688625336, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6176573634147644, 'reward_repeat_penalty_std': 0.15113945305347443, 'reward_total_composite_mean': 0.06387612223625183, 'reward_total_composite_std': 0.1795579344034195, 'epoch': 0.04}

 28%|██▊       | 910/3300 [34:49<5:40:25,  8.55s/it]INFO 04-11 23:11:40 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 911/3300 [34:58<5:52:23,  8.85s/it]2026-04-11 23:11:49,603 | INFO | train_grpo_train | metrics_logged mode=train step=911

                                                    
{'loss': -0.0319, 'grad_norm': 2.3471314907073975, 'learning_rate': 7.242424242424243e-06, 'num_tokens': 2036087.0, 'completions/mean_length': 130.5, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 76.0, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.00040913556586019695, 'rewards/meter/std': 0.00048293921281583607, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.0003843327867798507, 'rewards/total_composite/std': 0.0004887212999165058, 'reward': 0.0003843327867798507, 'reward_std': 0.0004887212999165058, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03155261650681496, 'sampling/sampling_logp_difference/max': 1.4526575803756714, 'sampling/importance_sampling_ratio/min': 0.2339477390050888, 'sampling/importance_sampling_ratio/mean': 0.9987106919288635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11146194022148848, 'clip_ratio/low_mean': 0.012957202387042344, 'clip_ratio/low_min': 0.012957202387042344, 'clip_ratio/high_mean': 0.003086419776082039, 'clip_ratio/high_max': 0.003086419776082039, 'clip_ratio/region_mean': 0.016043622163124382, 'reward_total_mean': 0.0003843327867798507, 'reward_meter_mean': 0.00040913556586019695, 'reward_meter_std': 0.00048293921281583607, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.0003843327867798507, 'reward_total_composite_std': 0.0004887212999165058, 'epoch': 0.04}

 28%|██▊       | 911/3300 [34:58<5:52:23,  8.85s/it]INFO 04-11 23:11:49 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 912/3300 [35:08<6:02:58,  9.12s/it]2026-04-11 23:11:59,356 | INFO | train_grpo_train | metrics_logged mode=train step=912

                                                    
{'loss': -0.2509, 'grad_norm': 0.514905571937561, 'learning_rate': 7.2393939393939404e-06, 'num_tokens': 2038908.0, 'completions/mean_length': 227.625, 'completions/min_length': 187.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 187.00001525878906, 'completions/min_terminated_length': 187.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.8691276907920837, 'rewards/meter/std': 0.3413867950439453, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.7222222089767456, 'rewards/repeat_penalty/std': 0.11878277361392975, 'rewards/total_composite/mean': 0.5911375284194946, 'rewards/total_composite/std': 0.24190568923950195, 'reward': 0.5911375284194946, 'reward_std': 0.24190570414066315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0038684571627527475, 'sampling/sampling_logp_difference/max': 0.8482571840286255, 'sampling/importance_sampling_ratio/min': 0.42816048860549927, 'sampling/importance_sampling_ratio/mean': 1.0010285377502441, 'sampling/importance_sampling_ratio/max': 1.1373112201690674, 'entropy': 0.026463798945769668, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.002673796727322042, 'clip_ratio/high_max': 0.002673796727322042, 'clip_ratio/region_mean': 0.002673796727322042, 'reward_total_mean': 0.5911375284194946, 'reward_meter_mean': 0.8691276907920837, 'reward_meter_std': 0.3413867950439453, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.7222222089767456, 'reward_repeat_penalty_std': 0.11878277361392975, 'reward_total_composite_mean': 0.5911375284194946, 'reward_total_composite_std': 0.24190568923950195, 'epoch': 0.04}

 28%|██▊       | 912/3300 [35:08<6:02:58,  9.12s/it]INFO 04-11 23:11:59 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 913/3300 [35:18<6:11:32,  9.34s/it]2026-04-11 23:12:09,203 | INFO | train_grpo_train | metrics_logged mode=train step=913

                                                    
{'loss': -0.0935, 'grad_norm': 1.2060858011245728, 'learning_rate': 7.236363636363637e-06, 'num_tokens': 2042728.0, 'completions/mean_length': 378.5, 'completions/min_length': 328.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 334.0, 'completions/min_terminated_length': 328.0, 'completions/max_terminated_length': 343.0, 'rewards/meter/mean': 0.24957457184791565, 'rewards/meter/std': 0.42763951420783997, 'rewards/count_adherence/mean': 0.6477272510528564, 'rewards/count_adherence/std': 0.24021585285663605, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5911239385604858, 'rewards/repeat_penalty/std': 0.21299511194229126, 'rewards/total_composite/mean': 0.08265350759029388, 'rewards/total_composite/std': 0.1630992591381073, 'reward': 0.08265350759029388, 'reward_std': 0.1630992740392685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01384922955185175, 'sampling/sampling_logp_difference/max': 2.1652774810791016, 'sampling/importance_sampling_ratio/min': 0.11471810191869736, 'sampling/importance_sampling_ratio/mean': 1.0017143487930298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04513176158070564, 'clip_ratio/low_mean': 0.004096297547221184, 'clip_ratio/low_min': 0.004096297547221184, 'clip_ratio/high_mean': 0.0011353294248692691, 'clip_ratio/high_max': 0.0011353294248692691, 'clip_ratio/region_mean': 0.005231626972090453, 'reward_total_mean': 0.08265350759029388, 'reward_meter_mean': 0.24957457184791565, 'reward_meter_std': 0.42763951420783997, 'reward_count_adherence_mean': 0.6477272510528564, 'reward_count_adherence_std': 0.24021585285663605, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5911239385604858, 'reward_repeat_penalty_std': 0.21299511194229126, 'reward_total_composite_mean': 0.08265350759029388, 'reward_total_composite_std': 0.1630992591381073, 'epoch': 0.04}

 28%|██▊       | 913/3300 [35:18<6:11:32,  9.34s/it]INFO 04-11 23:12:09 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 914/3300 [35:23<5:17:53,  7.99s/it]2026-04-11 23:12:14,063 | INFO | train_grpo_train | metrics_logged mode=train step=914

                                                    
{'loss': 0.0031, 'grad_norm': 3.0142698287963867, 'learning_rate': 7.233333333333334e-06, 'num_tokens': 2044450.0, 'completions/mean_length': 69.25, 'completions/min_length': 69.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9978233575820923, 'rewards/meter/std': 0.000473748950753361, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978233575820923, 'rewards/total_composite/std': 0.000473748950753361, 'reward': 0.9978233575820923, 'reward_std': 0.0004737446433864534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017529845237731934, 'sampling/sampling_logp_difference/max': 0.8328394889831543, 'sampling/importance_sampling_ratio/min': 0.4556933641433716, 'sampling/importance_sampling_ratio/mean': 1.0037966966629028, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09249463677406311, 'clip_ratio/low_mean': 0.0053571430034935474, 'clip_ratio/low_min': 0.0053571430034935474, 'clip_ratio/high_mean': 0.009057971183210611, 'clip_ratio/high_max': 0.009057971183210611, 'clip_ratio/region_mean': 0.014415114186704159, 'reward_total_mean': 0.9978233575820923, 'reward_meter_mean': 0.9978233575820923, 'reward_meter_std': 0.000473748950753361, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978233575820923, 'reward_total_composite_std': 0.000473748950753361, 'epoch': 0.04}

 28%|██▊       | 914/3300 [35:23<5:17:53,  7.99s/it]INFO 04-11 23:12:14 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 915/3300 [35:33<5:41:21,  8.59s/it]2026-04-11 23:12:24,030 | INFO | train_grpo_train | metrics_logged mode=train step=915

                                                    
{'loss': -0.2438, 'grad_norm': 1.0550512075424194, 'learning_rate': 7.2303030303030305e-06, 'num_tokens': 2047404.0, 'completions/mean_length': 410.25, 'completions/min_length': 296.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 308.5, 'completions/min_terminated_length': 296.0, 'completions/max_terminated_length': 330.0, 'rewards/meter/mean': 0.7131782174110413, 'rewards/meter/std': 0.40048038959503174, 'rewards/count_adherence/mean': 0.4027777910232544, 'rewards/count_adherence/std': 0.36581191420555115, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.8327265977859497, 'rewards/repeat_penalty/std': 0.20120510458946228, 'rewards/total_composite/mean': 0.1930341273546219, 'rewards/total_composite/std': 0.20283441245555878, 'reward': 0.1930341273546219, 'reward_std': 0.20283441245555878, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016473982483148575, 'sampling/sampling_logp_difference/max': 2.0499348640441895, 'sampling/importance_sampling_ratio/min': 0.12874329090118408, 'sampling/importance_sampling_ratio/mean': 1.0016696453094482, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03112065652385354, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.00370671006385237, 'clip_ratio/high_max': 0.00370671006385237, 'clip_ratio/region_mean': 0.005600649514235556, 'reward_total_mean': 0.1930341273546219, 'reward_meter_mean': 0.7131782174110413, 'reward_meter_std': 0.40048038959503174, 'reward_count_adherence_mean': 0.4027777910232544, 'reward_count_adherence_std': 0.36581191420555115, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.8327265977859497, 'reward_repeat_penalty_std': 0.20120510458946228, 'reward_total_composite_mean': 0.1930341273546219, 'reward_total_composite_std': 0.20283441245555878, 'epoch': 0.04}

 28%|██▊       | 915/3300 [35:33<5:41:21,  8.59s/it]INFO 04-11 23:12:24 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 916/3300 [35:37<4:54:41,  7.42s/it]2026-04-11 23:12:28,716 | INFO | train_grpo_train | metrics_logged mode=train step=916

                                                    
{'loss': 0.1148, 'grad_norm': 6.668519973754883, 'learning_rate': 7.227272727272729e-06, 'num_tokens': 2048990.0, 'completions/mean_length': 54.25, 'completions/min_length': 42.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.4157664477825165, 'rewards/meter/std': 0.3722110092639923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4157664477825165, 'rewards/total_composite/std': 0.3722110092639923, 'reward': 0.4157664477825165, 'reward_std': 0.3722110092639923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06838522106409073, 'sampling/sampling_logp_difference/max': 2.055908203125, 'sampling/importance_sampling_ratio/min': 0.12797655165195465, 'sampling/importance_sampling_ratio/mean': 1.011838436126709, 'sampling/importance_sampling_ratio/max': 1.8987327814102173, 'entropy': 0.36291532032191753, 'clip_ratio/low_mean': 0.02030206471681595, 'clip_ratio/low_min': 0.02030206471681595, 'clip_ratio/high_mean': 0.021643032785505056, 'clip_ratio/high_max': 0.021643032785505056, 'clip_ratio/region_mean': 0.041945097502321005, 'reward_total_mean': 0.4157664477825165, 'reward_meter_mean': 0.4157664477825165, 'reward_meter_std': 0.3722110092639923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4157664477825165, 'reward_total_composite_std': 0.3722110092639923, 'epoch': 0.04}

 28%|██▊       | 916/3300 [35:37<4:54:41,  7.42s/it]INFO 04-11 23:12:28 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 917/3300 [35:42<4:26:04,  6.70s/it]2026-04-11 23:12:33,742 | INFO | train_grpo_train | metrics_logged mode=train step=917

                                                    
{'loss': 0.0255, 'grad_norm': 4.49209451675415, 'learning_rate': 7.224242424242425e-06, 'num_tokens': 2050857.0, 'completions/mean_length': 81.375, 'completions/min_length': 76.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.375, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.024357754737138748, 'rewards/meter/std': 0.066401407122612, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.01962619088590145, 'rewards/total_composite/std': 0.053068071603775024, 'reward': 0.01962619088590145, 'reward_std': 0.053068071603775024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06128174066543579, 'sampling/sampling_logp_difference/max': 2.628598213195801, 'sampling/importance_sampling_ratio/min': 0.07217957079410553, 'sampling/importance_sampling_ratio/mean': 1.002654790878296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2575049940496683, 'clip_ratio/low_mean': 0.05021517118439078, 'clip_ratio/low_min': 0.05021517118439078, 'clip_ratio/high_mean': 0.009615384973585606, 'clip_ratio/high_max': 0.009615384973585606, 'clip_ratio/region_mean': 0.05983055615797639, 'reward_total_mean': 0.01962619088590145, 'reward_meter_mean': 0.024357754737138748, 'reward_meter_std': 0.066401407122612, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.01962619088590145, 'reward_total_composite_std': 0.053068071603775024, 'epoch': 0.04}

 28%|██▊       | 917/3300 [35:42<4:26:04,  6.70s/it]INFO 04-11 23:12:33 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 918/3300 [35:47<3:59:28,  6.03s/it]2026-04-11 23:12:38,217 | INFO | train_grpo_train | metrics_logged mode=train step=918

                                                    
{'loss': 0.0014, 'grad_norm': 3.6781203746795654, 'learning_rate': 7.221212121212122e-06, 'num_tokens': 2052361.0, 'completions/mean_length': 40.0, 'completions/min_length': 39.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.25423121452331543, 'rewards/meter/std': 0.42384836077690125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.25423121452331543, 'rewards/total_composite/std': 0.42384836077690125, 'reward': 0.25423121452331543, 'reward_std': 0.42384836077690125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025724250823259354, 'sampling/sampling_logp_difference/max': 0.5552816390991211, 'sampling/importance_sampling_ratio/min': 0.5739105939865112, 'sampling/importance_sampling_ratio/mean': 1.0119025707244873, 'sampling/importance_sampling_ratio/max': 1.5022114515304565, 'entropy': 0.1840200126171112, 'clip_ratio/low_mean': 0.015625000232830644, 'clip_ratio/low_min': 0.015625000232830644, 'clip_ratio/high_mean': 0.01266416534781456, 'clip_ratio/high_max': 0.01266416534781456, 'clip_ratio/region_mean': 0.028289165580645204, 'reward_total_mean': 0.25423121452331543, 'reward_meter_mean': 0.25423121452331543, 'reward_meter_std': 0.42384836077690125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.25423121452331543, 'reward_total_composite_std': 0.42384836077690125, 'epoch': 0.04}

 28%|██▊       | 918/3300 [35:47<3:59:28,  6.03s/it]INFO 04-11 23:12:38 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 919/3300 [35:52<3:45:44,  5.69s/it]2026-04-11 23:12:43,104 | INFO | train_grpo_train | metrics_logged mode=train step=919

                                                    
{'loss': -0.0107, 'grad_norm': 3.6577181816101074, 'learning_rate': 7.218181818181819e-06, 'num_tokens': 2054233.0, 'completions/mean_length': 67.0, 'completions/min_length': 64.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8576065301895142, 'rewards/meter/std': 0.346675843000412, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8576065301895142, 'rewards/total_composite/std': 0.346675843000412, 'reward': 0.8576065301895142, 'reward_std': 0.3466758131980896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024599751457571983, 'sampling/sampling_logp_difference/max': 2.0653536319732666, 'sampling/importance_sampling_ratio/min': 0.12677344679832458, 'sampling/importance_sampling_ratio/mean': 1.0019348859786987, 'sampling/importance_sampling_ratio/max': 1.7233705520629883, 'entropy': 0.11905911657959223, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.023888979107141495, 'clip_ratio/high_max': 0.023888979107141495, 'clip_ratio/region_mean': 0.027795229107141495, 'reward_total_mean': 0.8576065301895142, 'reward_meter_mean': 0.8576065301895142, 'reward_meter_std': 0.346675843000412, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8576065301895142, 'reward_total_composite_std': 0.346675843000412, 'epoch': 0.04}

 28%|██▊       | 919/3300 [35:52<3:45:44,  5.69s/it]INFO 04-11 23:12:43 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 920/3300 [35:59<4:10:55,  6.33s/it]2026-04-11 23:12:50,916 | INFO | train_grpo_train | metrics_logged mode=train step=920

                                                    
{'loss': 0.0055, 'grad_norm': 0.8415653109550476, 'learning_rate': 7.215151515151516e-06, 'num_tokens': 2058494.0, 'completions/mean_length': 330.625, 'completions/min_length': 313.0, 'completions/max_length': 334.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 330.625, 'completions/min_terminated_length': 313.0, 'completions/max_terminated_length': 334.0, 'rewards/meter/mean': 0.7762564420700073, 'rewards/meter/std': 0.19550828635692596, 'rewards/count_adherence/mean': 0.7692307829856873, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.625, 'rewards/repeat_penalty/std': 0.0864252969622612, 'rewards/total_composite/mean': 0.3763912320137024, 'rewards/total_composite/std': 0.11739427596330643, 'reward': 0.3763912320137024, 'reward_std': 0.11739426851272583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007077803369611502, 'sampling/sampling_logp_difference/max': 0.9340815544128418, 'sampling/importance_sampling_ratio/min': 0.39294660091400146, 'sampling/importance_sampling_ratio/mean': 0.9997749328613281, 'sampling/importance_sampling_ratio/max': 1.6070106029510498, 'entropy': 0.03318166173994541, 'clip_ratio/low_mean': 0.0014981299173086882, 'clip_ratio/low_min': 0.0014981299173086882, 'clip_ratio/high_mean': 0.005704862414859235, 'clip_ratio/high_max': 0.005704862414859235, 'clip_ratio/region_mean': 0.0072029923321679235, 'reward_total_mean': 0.3763912320137024, 'reward_meter_mean': 0.7762564420700073, 'reward_meter_std': 0.19550828635692596, 'reward_count_adherence_mean': 0.7692307829856873, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.625, 'reward_repeat_penalty_std': 0.0864252969622612, 'reward_total_composite_mean': 0.3763912320137024, 'reward_total_composite_std': 0.11739427596330643, 'epoch': 0.04}

 28%|██▊       | 920/3300 [35:59<4:10:55,  6.33s/it]INFO 04-11 23:12:51 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 921/3300 [36:10<5:02:26,  7.63s/it]2026-04-11 23:13:01,582 | INFO | train_grpo_train | metrics_logged mode=train step=921

                                                    
{'loss': -0.1723, 'grad_norm': 1.4356052875518799, 'learning_rate': 7.212121212121212e-06, 'num_tokens': 2060102.0, 'completions/mean_length': 313.0, 'completions/min_length': 108.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 114.0, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.842616856098175, 'rewards/meter/std': 0.3094305694103241, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.33034372329711914, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.44831955432891846, 'rewards/total_composite/std': 0.4836740791797638, 'reward': 0.44831955432891846, 'reward_std': 0.4836740791797638, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0745995044708252, 'sampling/sampling_logp_difference/max': 0.8662757873535156, 'sampling/importance_sampling_ratio/min': 0.4205147325992584, 'sampling/importance_sampling_ratio/mean': 1.0298396348953247, 'sampling/importance_sampling_ratio/max': 1.9944467544555664, 'entropy': 0.4247525855898857, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.029582271818071604, 'clip_ratio/high_max': 0.029582271818071604, 'clip_ratio/region_mean': 0.029582271818071604, 'reward_total_mean': 0.44831955432891846, 'reward_meter_mean': 0.842616856098175, 'reward_meter_std': 0.3094305694103241, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.33034372329711914, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.44831955432891846, 'reward_total_composite_std': 0.4836740791797638, 'epoch': 0.04}

 28%|██▊       | 921/3300 [36:10<5:02:26,  7.63s/it]INFO 04-11 23:13:01 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 922/3300 [36:17<4:56:49,  7.49s/it]2026-04-11 23:13:08,752 | INFO | train_grpo_train | metrics_logged mode=train step=922

                                                    
{'loss': -0.0231, 'grad_norm': 1.704238772392273, 'learning_rate': 7.2090909090909104e-06, 'num_tokens': 2063555.0, 'completions/mean_length': 248.625, 'completions/min_length': 238.0, 'completions/max_length': 266.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 248.625, 'completions/min_terminated_length': 238.0, 'completions/max_terminated_length': 266.0, 'rewards/meter/mean': 0.007957300171256065, 'rewards/meter/std': 0.0194723941385746, 'rewards/count_adherence/mean': 0.8392857313156128, 'rewards/count_adherence/std': 0.141575887799263, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6661838293075562, 'rewards/repeat_penalty/std': 0.07680089771747589, 'rewards/total_composite/mean': 0.004727967549115419, 'rewards/total_composite/std': 0.012029902078211308, 'reward': 0.004727967549115419, 'reward_std': 0.012029902078211308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02039029449224472, 'sampling/sampling_logp_difference/max': 2.663630962371826, 'sampling/importance_sampling_ratio/min': 0.06969470530748367, 'sampling/importance_sampling_ratio/mean': 0.9993036985397339, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06539201783016324, 'clip_ratio/low_mean': 0.012246111291460693, 'clip_ratio/low_min': 0.012246111291460693, 'clip_ratio/high_mean': 0.001409774413332343, 'clip_ratio/high_max': 0.001409774413332343, 'clip_ratio/region_mean': 0.013655885704793036, 'reward_total_mean': 0.004727967549115419, 'reward_meter_mean': 0.007957300171256065, 'reward_meter_std': 0.0194723941385746, 'reward_count_adherence_mean': 0.8392857313156128, 'reward_count_adherence_std': 0.141575887799263, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6661838293075562, 'reward_repeat_penalty_std': 0.07680089771747589, 'reward_total_composite_mean': 0.004727967549115419, 'reward_total_composite_std': 0.012029902078211308, 'epoch': 0.04}

 28%|██▊       | 922/3300 [36:17<4:56:49,  7.49s/it]INFO 04-11 23:13:09 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 923/3300 [36:22<4:23:50,  6.66s/it]2026-04-11 23:13:13,476 | INFO | train_grpo_train | metrics_logged mode=train step=923

                                                    
{'loss': -0.059, 'grad_norm': 10.85976791381836, 'learning_rate': 7.206060606060606e-06, 'num_tokens': 2065157.0, 'completions/mean_length': 49.25, 'completions/min_length': 44.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.514995276927948, 'rewards/meter/std': 0.3742930591106415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.514995276927948, 'rewards/total_composite/std': 0.3742930591106415, 'reward': 0.514995276927948, 'reward_std': 0.3742930591106415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04450875520706177, 'sampling/sampling_logp_difference/max': 1.193382740020752, 'sampling/importance_sampling_ratio/min': 0.3031938970088959, 'sampling/importance_sampling_ratio/mean': 1.0013222694396973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17652267403900623, 'clip_ratio/low_mean': 0.02332585956901312, 'clip_ratio/low_min': 0.02332585956901312, 'clip_ratio/high_mean': 0.027541207149624825, 'clip_ratio/high_max': 0.027541207149624825, 'clip_ratio/region_mean': 0.05086706671863794, 'reward_total_mean': 0.514995276927948, 'reward_meter_mean': 0.514995276927948, 'reward_meter_std': 0.3742930591106415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.514995276927948, 'reward_total_composite_std': 0.3742930591106415, 'epoch': 0.04}

 28%|██▊       | 923/3300 [36:22<4:23:50,  6.66s/it]INFO 04-11 23:13:13 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 924/3300 [36:32<5:03:39,  7.67s/it]2026-04-11 23:13:23,504 | INFO | train_grpo_train | metrics_logged mode=train step=924

                                                    
{'loss': -0.127, 'grad_norm': 1.6212074756622314, 'learning_rate': 7.203030303030304e-06, 'num_tokens': 2067180.0, 'completions/mean_length': 214.875, 'completions/min_length': 110.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 115.83333587646484, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.594925582408905, 'rewards/meter/std': 0.46331310272216797, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.40130022168159485, 'rewards/total_composite/std': 0.3983116149902344, 'reward': 0.40130022168159485, 'reward_std': 0.39831164479255676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051524095237255096, 'sampling/sampling_logp_difference/max': 1.1322174072265625, 'sampling/importance_sampling_ratio/min': 0.3223177492618561, 'sampling/importance_sampling_ratio/mean': 1.0083558559417725, 'sampling/importance_sampling_ratio/max': 1.917328119277954, 'entropy': 0.3687305059283972, 'clip_ratio/low_mean': 0.01368181873112917, 'clip_ratio/low_min': 0.01368181873112917, 'clip_ratio/high_mean': 0.019490185426548123, 'clip_ratio/high_max': 0.019490185426548123, 'clip_ratio/region_mean': 0.03317200415767729, 'reward_total_mean': 0.40130022168159485, 'reward_meter_mean': 0.594925582408905, 'reward_meter_std': 0.46331310272216797, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.40130022168159485, 'reward_total_composite_std': 0.3983116149902344, 'epoch': 0.04}

 28%|██▊       | 924/3300 [36:32<5:03:39,  7.67s/it]INFO 04-11 23:13:23 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 925/3300 [36:37<4:33:10,  6.90s/it]2026-04-11 23:13:28,605 | INFO | train_grpo_train | metrics_logged mode=train step=925

                                                    
{'loss': -0.029, 'grad_norm': 5.2026143074035645, 'learning_rate': 7.2000000000000005e-06, 'num_tokens': 2069136.0, 'completions/mean_length': 73.5, 'completions/min_length': 68.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8316957354545593, 'rewards/meter/std': 0.3097537159919739, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7086162567138672, 'rewards/total_composite/std': 0.4172649383544922, 'reward': 0.7086162567138672, 'reward_std': 0.4172649383544922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051565829664468765, 'sampling/sampling_logp_difference/max': 1.1075992584228516, 'sampling/importance_sampling_ratio/min': 0.3303510844707489, 'sampling/importance_sampling_ratio/mean': 1.0118038654327393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36769964918494225, 'clip_ratio/low_mean': 0.010874067898839712, 'clip_ratio/low_min': 0.010874067898839712, 'clip_ratio/high_mean': 0.03658821329008788, 'clip_ratio/high_max': 0.03658821329008788, 'clip_ratio/region_mean': 0.04746228118892759, 'reward_total_mean': 0.7086162567138672, 'reward_meter_mean': 0.8316957354545593, 'reward_meter_std': 0.3097537159919739, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7086162567138672, 'reward_total_composite_std': 0.4172649383544922, 'epoch': 0.04}

 28%|██▊       | 925/3300 [36:37<4:33:10,  6.90s/it]INFO 04-11 23:13:28 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 926/3300 [36:42<4:08:26,  6.28s/it]2026-04-11 23:13:33,441 | INFO | train_grpo_train | metrics_logged mode=train step=926

                                                    
{'loss': 0.001, 'grad_norm': 2.8455419540405273, 'learning_rate': 7.196969696969698e-06, 'num_tokens': 2070956.0, 'completions/mean_length': 69.5, 'completions/min_length': 69.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9984892010688782, 'rewards/meter/std': 8.65640613483265e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984892010688782, 'rewards/total_composite/std': 8.65640613483265e-05, 'reward': 0.9984892010688782, 'reward_std': 8.656938007334247e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015567532740533352, 'sampling/sampling_logp_difference/max': 1.3253982067108154, 'sampling/importance_sampling_ratio/min': 0.2656971514225006, 'sampling/importance_sampling_ratio/mean': 1.0004963874816895, 'sampling/importance_sampling_ratio/max': 1.7046222686767578, 'entropy': 0.06283332780003548, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.00898033136036247, 'clip_ratio/high_max': 0.00898033136036247, 'clip_ratio/region_mean': 0.00898033136036247, 'reward_total_mean': 0.9984892010688782, 'reward_meter_mean': 0.9984892010688782, 'reward_meter_std': 8.65640613483265e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984892010688782, 'reward_total_composite_std': 8.65640613483265e-05, 'epoch': 0.04}

 28%|██▊       | 926/3300 [36:42<4:08:26,  6.28s/it]INFO 04-11 23:13:33 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 927/3300 [36:51<4:45:39,  7.22s/it]2026-04-11 23:13:42,865 | INFO | train_grpo_train | metrics_logged mode=train step=927

                                                    
{'loss': -0.014, 'grad_norm': 0.8467232584953308, 'learning_rate': 7.193939393939394e-06, 'num_tokens': 2075400.0, 'completions/mean_length': 347.5, 'completions/min_length': 341.0, 'completions/max_length': 375.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 347.5, 'completions/min_terminated_length': 341.0, 'completions/max_terminated_length': 375.0, 'rewards/meter/mean': 0.9979950189590454, 'rewards/meter/std': 0.0002653496921993792, 'rewards/count_adherence/mean': 0.7232142686843872, 'rewards/count_adherence/std': 0.02525380253791809, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6365914344787598, 'rewards/repeat_penalty/std': 0.01973436214029789, 'rewards/total_composite/mean': 0.4593149423599243, 'rewards/total_composite/std': 0.01691725291311741, 'reward': 0.4593149423599243, 'reward_std': 0.01691724918782711, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0067790052853524685, 'sampling/sampling_logp_difference/max': 1.7181689739227295, 'sampling/importance_sampling_ratio/min': 0.17939433455467224, 'sampling/importance_sampling_ratio/mean': 1.0001847743988037, 'sampling/importance_sampling_ratio/max': 1.8701292276382446, 'entropy': 0.026407914701849222, 'clip_ratio/low_mean': 0.00428855000063777, 'clip_ratio/low_min': 0.00428855000063777, 'clip_ratio/high_mean': 0.0010664711589924991, 'clip_ratio/high_max': 0.0010664711589924991, 'clip_ratio/region_mean': 0.005355021159630269, 'reward_total_mean': 0.4593149423599243, 'reward_meter_mean': 0.9979950189590454, 'reward_meter_std': 0.0002653496921993792, 'reward_count_adherence_mean': 0.7232142686843872, 'reward_count_adherence_std': 0.02525380253791809, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6365914344787598, 'reward_repeat_penalty_std': 0.01973436214029789, 'reward_total_composite_mean': 0.4593149423599243, 'reward_total_composite_std': 0.01691725291311741, 'epoch': 0.04}

 28%|██▊       | 927/3300 [36:51<4:45:39,  7.22s/it]INFO 04-11 23:13:43 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 928/3300 [36:58<4:35:33,  6.97s/it]2026-04-11 23:13:49,242 | INFO | train_grpo_train | metrics_logged mode=train step=928

                                                    
{'loss': 0.0067, 'grad_norm': 1.9796561002731323, 'learning_rate': 7.1909090909090914e-06, 'num_tokens': 2078027.0, 'completions/mean_length': 154.375, 'completions/min_length': 150.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.375, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.8792700171470642, 'rewards/meter/std': 0.2660437226295471, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.645013689994812, 'rewards/total_composite/std': 0.20092783868312836, 'reward': 0.645013689994812, 'reward_std': 0.20092782378196716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013188260607421398, 'sampling/sampling_logp_difference/max': 1.643843650817871, 'sampling/importance_sampling_ratio/min': 0.19323588907718658, 'sampling/importance_sampling_ratio/mean': 0.9983150959014893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04184746090322733, 'clip_ratio/low_mean': 0.0015723269898444414, 'clip_ratio/low_min': 0.0015723269898444414, 'clip_ratio/high_mean': 0.008021060610190034, 'clip_ratio/high_max': 0.008021060610190034, 'clip_ratio/region_mean': 0.009593387600034475, 'reward_total_mean': 0.645013689994812, 'reward_meter_mean': 0.8792700171470642, 'reward_meter_std': 0.2660437226295471, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.645013689994812, 'reward_total_composite_std': 0.20092783868312836, 'epoch': 0.04}

 28%|██▊       | 928/3300 [36:58<4:35:33,  6.97s/it]INFO 04-11 23:13:49 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 929/3300 [37:04<4:22:34,  6.64s/it]2026-04-11 23:13:55,124 | INFO | train_grpo_train | metrics_logged mode=train step=929

                                                    
{'loss': 0.0038, 'grad_norm': 2.6719958782196045, 'learning_rate': 7.187878787878788e-06, 'num_tokens': 2080954.0, 'completions/mean_length': 170.875, 'completions/min_length': 170.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.875, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.9981188178062439, 'rewards/meter/std': 0.0002052018535323441, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8055555820465088, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.8040474057197571, 'rewards/total_composite/std': 0.051476798951625824, 'reward': 0.8040474057197571, 'reward_std': 0.05147679150104523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008916568011045456, 'sampling/sampling_logp_difference/max': 1.7634490728378296, 'sampling/importance_sampling_ratio/min': 0.17145249247550964, 'sampling/importance_sampling_ratio/mean': 0.9989221692085266, 'sampling/importance_sampling_ratio/max': 1.3845103979110718, 'entropy': 0.03398467996157706, 'clip_ratio/low_mean': 0.003659270762000233, 'clip_ratio/low_min': 0.003659270762000233, 'clip_ratio/high_mean': 0.002923976629972458, 'clip_ratio/high_max': 0.002923976629972458, 'clip_ratio/region_mean': 0.006583247391972691, 'reward_total_mean': 0.8040474057197571, 'reward_meter_mean': 0.9981188178062439, 'reward_meter_std': 0.0002052018535323441, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8055555820465088, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.8040474057197571, 'reward_total_composite_std': 0.051476798951625824, 'epoch': 0.04}

 28%|██▊       | 929/3300 [37:04<4:22:34,  6.64s/it]INFO 04-11 23:13:55 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 930/3300 [37:09<4:02:04,  6.13s/it]2026-04-11 23:14:00,050 | INFO | train_grpo_train | metrics_logged mode=train step=930

                                                    
{'loss': 0.0414, 'grad_norm': 5.791079998016357, 'learning_rate': 7.184848484848486e-06, 'num_tokens': 2082718.0, 'completions/mean_length': 77.5, 'completions/min_length': 73.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9255967140197754, 'rewards/meter/std': 0.06689538806676865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9255967140197754, 'rewards/total_composite/std': 0.06689538806676865, 'reward': 0.9255967140197754, 'reward_std': 0.06689538061618805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04519684612751007, 'sampling/sampling_logp_difference/max': 1.1719346046447754, 'sampling/importance_sampling_ratio/min': 0.30976709723472595, 'sampling/importance_sampling_ratio/mean': 1.0131640434265137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26597489789128304, 'clip_ratio/low_mean': 0.016923486720770597, 'clip_ratio/low_min': 0.016923486720770597, 'clip_ratio/high_mean': 0.016671410761773586, 'clip_ratio/high_max': 0.016671410761773586, 'clip_ratio/region_mean': 0.033594897482544184, 'reward_total_mean': 0.9255967140197754, 'reward_meter_mean': 0.9255967140197754, 'reward_meter_std': 0.06689538806676865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9255967140197754, 'reward_total_composite_std': 0.06689538806676865, 'epoch': 0.04}

 28%|██▊       | 930/3300 [37:09<4:02:04,  6.13s/it]INFO 04-11 23:14:00 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 931/3300 [37:16<4:22:09,  6.64s/it]2026-04-11 23:14:07,882 | INFO | train_grpo_train | metrics_logged mode=train step=931

                                                    
{'loss': -0.0247, 'grad_norm': 0.8847100734710693, 'learning_rate': 7.181818181818182e-06, 'num_tokens': 2086955.0, 'completions/mean_length': 309.625, 'completions/min_length': 289.0, 'completions/max_length': 322.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 309.625, 'completions/min_terminated_length': 289.0, 'completions/max_terminated_length': 322.0, 'rewards/meter/mean': 0.7983001470565796, 'rewards/meter/std': 0.3538459241390228, 'rewards/count_adherence/mean': 0.8888888955116272, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6083333492279053, 'rewards/repeat_penalty/std': 0.0235702246427536, 'rewards/total_composite/mean': 0.4331192076206207, 'rewards/total_composite/std': 0.19441884756088257, 'reward': 0.4331192076206207, 'reward_std': 0.19441883265972137, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008012884296476841, 'sampling/sampling_logp_difference/max': 1.079960823059082, 'sampling/importance_sampling_ratio/min': 0.3396088182926178, 'sampling/importance_sampling_ratio/mean': 0.9997681379318237, 'sampling/importance_sampling_ratio/max': 1.7735785245895386, 'entropy': 0.031651133904233575, 'clip_ratio/low_mean': 0.00041946308920159936, 'clip_ratio/low_min': 0.00041946308920159936, 'clip_ratio/high_mean': 0.009181037603411824, 'clip_ratio/high_max': 0.009181037603411824, 'clip_ratio/region_mean': 0.009600500692613423, 'reward_total_mean': 0.4331192076206207, 'reward_meter_mean': 0.7983001470565796, 'reward_meter_std': 0.3538459241390228, 'reward_count_adherence_mean': 0.8888888955116272, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6083333492279053, 'reward_repeat_penalty_std': 0.0235702246427536, 'reward_total_composite_mean': 0.4331192076206207, 'reward_total_composite_std': 0.19441884756088257, 'epoch': 0.04}

 28%|██▊       | 931/3300 [37:16<4:22:09,  6.64s/it]INFO 04-11 23:14:08 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 932/3300 [37:22<4:10:40,  6.35s/it]2026-04-11 23:14:13,565 | INFO | train_grpo_train | metrics_logged mode=train step=932

                                                    
{'loss': -0.0003, 'grad_norm': 2.1663684844970703, 'learning_rate': 7.17878787878788e-06, 'num_tokens': 2089791.0, 'completions/mean_length': 148.5, 'completions/min_length': 146.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.5, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.7308982610702515, 'rewards/meter/std': 0.4457210898399353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.5223252773284912, 'rewards/total_composite/std': 0.317903995513916, 'reward': 0.5223252773284912, 'reward_std': 0.317903995513916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012427828274667263, 'sampling/sampling_logp_difference/max': 1.0588593482971191, 'sampling/importance_sampling_ratio/min': 0.3468512296676636, 'sampling/importance_sampling_ratio/mean': 0.9969629049301147, 'sampling/importance_sampling_ratio/max': 1.8277068138122559, 'entropy': 0.037923737429082394, 'clip_ratio/low_mean': 0.0017123287543654442, 'clip_ratio/low_min': 0.0017123287543654442, 'clip_ratio/high_mean': 0.010114155360497534, 'clip_ratio/high_max': 0.010114155360497534, 'clip_ratio/region_mean': 0.011826484114862978, 'reward_total_mean': 0.5223252773284912, 'reward_meter_mean': 0.7308982610702515, 'reward_meter_std': 0.4457210898399353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.5223252773284912, 'reward_total_composite_std': 0.317903995513916, 'epoch': 0.04}

 28%|██▊       | 932/3300 [37:22<4:10:40,  6.35s/it]INFO 04-11 23:14:13 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 933/3300 [37:27<3:51:46,  5.88s/it]2026-04-11 23:14:18,330 | INFO | train_grpo_train | metrics_logged mode=train step=933

                                                    
{'loss': -0.0034, 'grad_norm': 8.870975494384766, 'learning_rate': 7.175757575757576e-06, 'num_tokens': 2091401.0, 'completions/mean_length': 53.25, 'completions/min_length': 52.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.25, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9159232378005981, 'rewards/meter/std': 0.05563880503177643, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9159232378005981, 'rewards/total_composite/std': 0.05563880503177643, 'reward': 0.9159232378005981, 'reward_std': 0.05563879758119583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023871900513768196, 'sampling/sampling_logp_difference/max': 1.3186838626861572, 'sampling/importance_sampling_ratio/min': 0.26748713850975037, 'sampling/importance_sampling_ratio/mean': 1.0049976110458374, 'sampling/importance_sampling_ratio/max': 1.4143028259277344, 'entropy': 0.1050838534720242, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.009302935097366571, 'clip_ratio/high_max': 0.009302935097366571, 'clip_ratio/region_mean': 0.009302935097366571, 'reward_total_mean': 0.9159232378005981, 'reward_meter_mean': 0.9159232378005981, 'reward_meter_std': 0.05563880503177643, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9159232378005981, 'reward_total_composite_std': 0.05563880503177643, 'epoch': 0.04}

 28%|██▊       | 933/3300 [37:27<3:51:46,  5.88s/it]INFO 04-11 23:14:18 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 934/3300 [37:32<3:40:17,  5.59s/it]2026-04-11 23:14:23,236 | INFO | train_grpo_train | metrics_logged mode=train step=934

                                                    
{'loss': 0.0017, 'grad_norm': 6.214112281799316, 'learning_rate': 7.172727272727273e-06, 'num_tokens': 2093250.0, 'completions/mean_length': 69.125, 'completions/min_length': 66.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.40722712874412537, 'rewards/meter/std': 0.28478607535362244, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.40722712874412537, 'rewards/total_composite/std': 0.28478607535362244, 'reward': 0.40722712874412537, 'reward_std': 0.28478607535362244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03458065912127495, 'sampling/sampling_logp_difference/max': 1.8944740295410156, 'sampling/importance_sampling_ratio/min': 0.1503974199295044, 'sampling/importance_sampling_ratio/mean': 0.9984850883483887, 'sampling/importance_sampling_ratio/max': 1.4551042318344116, 'entropy': 0.1644162703305483, 'clip_ratio/low_mean': 0.01640054234303534, 'clip_ratio/low_min': 0.01640054234303534, 'clip_ratio/high_mean': 0.016192146576941013, 'clip_ratio/high_max': 0.016192146576941013, 'clip_ratio/region_mean': 0.032592688919976354, 'reward_total_mean': 0.40722712874412537, 'reward_meter_mean': 0.40722712874412537, 'reward_meter_std': 0.28478607535362244, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.40722712874412537, 'reward_total_composite_std': 0.28478607535362244, 'epoch': 0.04}

 28%|██▊       | 934/3300 [37:32<3:40:17,  5.59s/it]INFO 04-11 23:14:23 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 935/3300 [37:36<3:24:57,  5.20s/it]2026-04-11 23:14:27,536 | INFO | train_grpo_train | metrics_logged mode=train step=935

                                                    
{'loss': 0.0203, 'grad_norm': 7.902743816375732, 'learning_rate': 7.16969696969697e-06, 'num_tokens': 2094727.0, 'completions/mean_length': 32.625, 'completions/min_length': 32.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9710537195205688, 'rewards/meter/std': 0.003800545586273074, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9710537195205688, 'rewards/total_composite/std': 0.003800545586273074, 'reward': 0.9710537195205688, 'reward_std': 0.003800547681748867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009311332367360592, 'sampling/sampling_logp_difference/max': 0.41676831245422363, 'sampling/importance_sampling_ratio/min': 0.7209804058074951, 'sampling/importance_sampling_ratio/mean': 1.0037425756454468, 'sampling/importance_sampling_ratio/max': 1.517050862312317, 'entropy': 0.051071187015622854, 'clip_ratio/low_mean': 0.011363636702299118, 'clip_ratio/low_min': 0.011363636702299118, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.015269886702299118, 'reward_total_mean': 0.9710537195205688, 'reward_meter_mean': 0.9710537195205688, 'reward_meter_std': 0.003800545586273074, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9710537195205688, 'reward_total_composite_std': 0.003800545586273074, 'epoch': 0.04}

 28%|██▊       | 935/3300 [37:36<3:24:57,  5.20s/it]INFO 04-11 23:14:27 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 936/3300 [37:41<3:22:16,  5.13s/it]2026-04-11 23:14:32,514 | INFO | train_grpo_train | metrics_logged mode=train step=936

                                                    
{'loss': 0.008, 'grad_norm': 12.27761173248291, 'learning_rate': 7.166666666666667e-06, 'num_tokens': 2096787.0, 'completions/mean_length': 77.5, 'completions/min_length': 76.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9931102991104126, 'rewards/meter/std': 0.006781714037060738, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9931102991104126, 'rewards/total_composite/std': 0.006781714037060738, 'reward': 0.9931102991104126, 'reward_std': 0.006781699601560831, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017138954252004623, 'sampling/sampling_logp_difference/max': 0.852900505065918, 'sampling/importance_sampling_ratio/min': 0.4686746895313263, 'sampling/importance_sampling_ratio/mean': 1.001671314239502, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09636378474533558, 'clip_ratio/low_mean': 0.0015625000232830644, 'clip_ratio/low_min': 0.0015625000232830644, 'clip_ratio/high_mean': 0.0048290317645296454, 'clip_ratio/high_max': 0.0048290317645296454, 'clip_ratio/region_mean': 0.00639153178781271, 'reward_total_mean': 0.9931102991104126, 'reward_meter_mean': 0.9931102991104126, 'reward_meter_std': 0.006781714037060738, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9931102991104126, 'reward_total_composite_std': 0.006781714037060738, 'epoch': 0.04}

 28%|██▊       | 936/3300 [37:41<3:22:16,  5.13s/it]INFO 04-11 23:14:32 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 937/3300 [37:46<3:19:12,  5.06s/it]2026-04-11 23:14:37,398 | INFO | train_grpo_train | metrics_logged mode=train step=937

                                                    
{'loss': 0.0106, 'grad_norm': 9.313652992248535, 'learning_rate': 7.163636363636363e-06, 'num_tokens': 2098529.0, 'completions/mean_length': 68.75, 'completions/min_length': 67.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9961028099060059, 'rewards/meter/std': 0.0017646643100306392, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961028099060059, 'rewards/total_composite/std': 0.0017646643100306392, 'reward': 0.9961028099060059, 'reward_std': 0.001764659769833088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028187252581119537, 'sampling/sampling_logp_difference/max': 1.3503470420837402, 'sampling/importance_sampling_ratio/min': 0.25915029644966125, 'sampling/importance_sampling_ratio/mean': 1.0019060373306274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0943189668469131, 'clip_ratio/low_mean': 0.0036231884732842445, 'clip_ratio/low_min': 0.0036231884732842445, 'clip_ratio/high_mean': 0.014654986094683409, 'clip_ratio/high_max': 0.014654986094683409, 'clip_ratio/region_mean': 0.018278174567967653, 'reward_total_mean': 0.9961028099060059, 'reward_meter_mean': 0.9961028099060059, 'reward_meter_std': 0.0017646643100306392, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961028099060059, 'reward_total_composite_std': 0.0017646643100306392, 'epoch': 0.04}

 28%|██▊       | 937/3300 [37:46<3:19:12,  5.06s/it]INFO 04-11 23:14:37 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 938/3300 [37:52<3:27:31,  5.27s/it]2026-04-11 23:14:43,170 | INFO | train_grpo_train | metrics_logged mode=train step=938

                                                    
{'loss': -0.0077, 'grad_norm': 2.60717511177063, 'learning_rate': 7.1606060606060615e-06, 'num_tokens': 2101075.0, 'completions/mean_length': 150.25, 'completions/min_length': 144.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.25, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.9614237546920776, 'rewards/meter/std': 0.02061135321855545, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7142857313156128, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.686731219291687, 'rewards/total_composite/std': 0.014722409658133984, 'reward': 0.686731219291687, 'reward_std': 0.014722414314746857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005851758643984795, 'sampling/sampling_logp_difference/max': 1.237354040145874, 'sampling/importance_sampling_ratio/min': 0.2901509404182434, 'sampling/importance_sampling_ratio/mean': 1.0006721019744873, 'sampling/importance_sampling_ratio/max': 1.5528877973556519, 'entropy': 0.02918590500485152, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.004902798566035926, 'clip_ratio/high_max': 0.004902798566035926, 'clip_ratio/region_mean': 0.004902798566035926, 'reward_total_mean': 0.686731219291687, 'reward_meter_mean': 0.9614237546920776, 'reward_meter_std': 0.02061135321855545, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7142857313156128, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.686731219291687, 'reward_total_composite_std': 0.014722409658133984, 'epoch': 0.04}

 28%|██▊       | 938/3300 [37:52<3:27:31,  5.27s/it]INFO 04-11 23:14:43 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 939/3300 [37:56<3:19:01,  5.06s/it]2026-04-11 23:14:47,726 | INFO | train_grpo_train | metrics_logged mode=train step=939

                                                    
{'loss': -0.0037, 'grad_norm': 12.418314933776855, 'learning_rate': 7.157575757575758e-06, 'num_tokens': 2102575.0, 'completions/mean_length': 36.5, 'completions/min_length': 34.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.21266832947731018, 'rewards/meter/std': 0.20291058719158173, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.21266832947731018, 'rewards/total_composite/std': 0.20291058719158173, 'reward': 0.21266832947731018, 'reward_std': 0.20291060209274292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03699643164873123, 'sampling/sampling_logp_difference/max': 1.311621904373169, 'sampling/importance_sampling_ratio/min': 0.2693828046321869, 'sampling/importance_sampling_ratio/mean': 1.0057456493377686, 'sampling/importance_sampling_ratio/max': 1.8581043481826782, 'entropy': 0.13651060685515404, 'clip_ratio/low_mean': 0.017079579178243876, 'clip_ratio/low_min': 0.017079579178243876, 'clip_ratio/high_mean': 0.013518452877178788, 'clip_ratio/high_max': 0.013518452877178788, 'clip_ratio/region_mean': 0.030598032055422664, 'reward_total_mean': 0.21266832947731018, 'reward_meter_mean': 0.21266832947731018, 'reward_meter_std': 0.20291058719158173, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.21266832947731018, 'reward_total_composite_std': 0.20291058719158173, 'epoch': 0.04}

 28%|██▊       | 939/3300 [37:56<3:19:01,  5.06s/it]INFO 04-11 23:14:47 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 940/3300 [38:02<3:23:16,  5.17s/it]2026-04-11 23:14:53,156 | INFO | train_grpo_train | metrics_logged mode=train step=940

                                                    
{'loss': 0.023, 'grad_norm': 4.8500847816467285, 'learning_rate': 7.154545454545455e-06, 'num_tokens': 2104942.0, 'completions/mean_length': 115.875, 'completions/min_length': 111.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.875, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.968498945236206, 'rewards/meter/std': 0.029964042827486992, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7747991681098938, 'rewards/total_composite/std': 0.02397123910486698, 'reward': 0.7747991681098938, 'reward_std': 0.023971248418092728, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0124904103577137, 'sampling/sampling_logp_difference/max': 1.0763840675354004, 'sampling/importance_sampling_ratio/min': 0.3408257067203522, 'sampling/importance_sampling_ratio/mean': 0.9974149465560913, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.044212615583091974, 'clip_ratio/low_mean': 0.0020325202494859695, 'clip_ratio/low_min': 0.0020325202494859695, 'clip_ratio/high_mean': 0.016339467372745275, 'clip_ratio/high_max': 0.016339467372745275, 'clip_ratio/region_mean': 0.018371987622231245, 'reward_total_mean': 0.7747991681098938, 'reward_meter_mean': 0.968498945236206, 'reward_meter_std': 0.029964042827486992, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7747991681098938, 'reward_total_composite_std': 0.02397123910486698, 'epoch': 0.04}

 28%|██▊       | 940/3300 [38:02<3:23:16,  5.17s/it]INFO 04-11 23:14:53 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 941/3300 [38:06<3:19:38,  5.08s/it]2026-04-11 23:14:58,016 | INFO | train_grpo_train | metrics_logged mode=train step=941

                                                    
{'loss': 0.0096, 'grad_norm': 3.194761276245117, 'learning_rate': 7.151515151515152e-06, 'num_tokens': 2106780.0, 'completions/mean_length': 75.75, 'completions/min_length': 72.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.7696533203125, 'rewards/meter/std': 0.264241486787796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.7389180660247803, 'rewards/total_composite/std': 0.2821867763996124, 'reward': 0.7389180660247803, 'reward_std': 0.2821867763996124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02568567357957363, 'sampling/sampling_logp_difference/max': 1.0589550733566284, 'sampling/importance_sampling_ratio/min': 0.3468180000782013, 'sampling/importance_sampling_ratio/mean': 1.0013823509216309, 'sampling/importance_sampling_ratio/max': 1.8680615425109863, 'entropy': 0.08911147527396679, 'clip_ratio/low_mean': 0.011406926438212395, 'clip_ratio/low_min': 0.011406926438212395, 'clip_ratio/high_mean': 0.008275165455415845, 'clip_ratio/high_max': 0.008275165455415845, 'clip_ratio/region_mean': 0.01968209189362824, 'reward_total_mean': 0.7389180660247803, 'reward_meter_mean': 0.7696533203125, 'reward_meter_std': 0.264241486787796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.7389180660247803, 'reward_total_composite_std': 0.2821867763996124, 'epoch': 0.04}

 29%|██▊       | 941/3300 [38:07<3:19:38,  5.08s/it]INFO 04-11 23:14:58 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 942/3300 [38:11<3:11:51,  4.88s/it]2026-04-11 23:15:02,442 | INFO | train_grpo_train | metrics_logged mode=train step=942

                                                    
{'loss': -0.0086, 'grad_norm': 11.929455757141113, 'learning_rate': 7.148484848484849e-06, 'num_tokens': 2108257.0, 'completions/mean_length': 38.625, 'completions/min_length': 37.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7800159454345703, 'rewards/meter/std': 0.17752352356910706, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7800159454345703, 'rewards/total_composite/std': 0.17752352356910706, 'reward': 0.7800159454345703, 'reward_std': 0.17752352356910706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03380492329597473, 'sampling/sampling_logp_difference/max': 1.4989476203918457, 'sampling/importance_sampling_ratio/min': 0.5866755843162537, 'sampling/importance_sampling_ratio/mean': 1.0109392404556274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15884944330900908, 'clip_ratio/low_mean': 0.006667852168902755, 'clip_ratio/low_min': 0.006667852168902755, 'clip_ratio/high_mean': 0.016215793788433075, 'clip_ratio/high_max': 0.016215793788433075, 'clip_ratio/region_mean': 0.02288364595733583, 'reward_total_mean': 0.7800159454345703, 'reward_meter_mean': 0.7800159454345703, 'reward_meter_std': 0.17752352356910706, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7800159454345703, 'reward_total_composite_std': 0.17752352356910706, 'epoch': 0.04}

 29%|██▊       | 942/3300 [38:11<3:11:51,  4.88s/it]INFO 04-11 23:15:02 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 943/3300 [38:15<3:06:38,  4.75s/it]2026-04-11 23:15:06,889 | INFO | train_grpo_train | metrics_logged mode=train step=943

                                                    
{'loss': -0.0148, 'grad_norm': 6.928620338439941, 'learning_rate': 7.145454545454547e-06, 'num_tokens': 2109890.0, 'completions/mean_length': 41.125, 'completions/min_length': 39.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9855058193206787, 'rewards/meter/std': 0.010752059519290924, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9855058193206787, 'rewards/total_composite/std': 0.010752059519290924, 'reward': 0.9855058193206787, 'reward_std': 0.010752064175903797, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01714111864566803, 'sampling/sampling_logp_difference/max': 1.1181397438049316, 'sampling/importance_sampling_ratio/min': 0.326887309551239, 'sampling/importance_sampling_ratio/mean': 0.9965353012084961, 'sampling/importance_sampling_ratio/max': 1.2362208366394043, 'entropy': 0.09372959332540631, 'clip_ratio/low_mean': 0.006253908621147275, 'clip_ratio/low_min': 0.006253908621147275, 'clip_ratio/high_mean': 0.015178571688011289, 'clip_ratio/high_max': 0.015178571688011289, 'clip_ratio/region_mean': 0.021432480309158564, 'reward_total_mean': 0.9855058193206787, 'reward_meter_mean': 0.9855058193206787, 'reward_meter_std': 0.010752059519290924, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9855058193206787, 'reward_total_composite_std': 0.010752059519290924, 'epoch': 0.04}

 29%|██▊       | 943/3300 [38:15<3:06:38,  4.75s/it]INFO 04-11 23:15:07 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 944/3300 [38:21<3:14:54,  4.96s/it]2026-04-11 23:15:12,355 | INFO | train_grpo_train | metrics_logged mode=train step=944

                                                    
{'loss': 0.0057, 'grad_norm': 4.29167366027832, 'learning_rate': 7.142424242424243e-06, 'num_tokens': 2112370.0, 'completions/mean_length': 123.0, 'completions/min_length': 117.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.0, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.8062102794647217, 'rewards/meter/std': 0.18132120370864868, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.5850111842155457, 'rewards/total_composite/std': 0.11425027996301651, 'reward': 0.5850111842155457, 'reward_std': 0.11425027996301651, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026136090978980064, 'sampling/sampling_logp_difference/max': 1.7183735370635986, 'sampling/importance_sampling_ratio/min': 0.1793576329946518, 'sampling/importance_sampling_ratio/mean': 0.994314432144165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06403588084504008, 'clip_ratio/low_mean': 0.004050420364364982, 'clip_ratio/low_min': 0.004050420364364982, 'clip_ratio/high_mean': 0.01517849147785455, 'clip_ratio/high_max': 0.01517849147785455, 'clip_ratio/region_mean': 0.01922891184221953, 'reward_total_mean': 0.5850111842155457, 'reward_meter_mean': 0.8062102794647217, 'reward_meter_std': 0.18132120370864868, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.5850111842155457, 'reward_total_composite_std': 0.11425027996301651, 'epoch': 0.04}

 29%|██▊       | 944/3300 [38:21<3:14:54,  4.96s/it]INFO 04-11 23:15:12 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 945/3300 [38:27<3:24:44,  5.22s/it]2026-04-11 23:15:18,158 | INFO | train_grpo_train | metrics_logged mode=train step=945

                                                    
{'loss': 0.0177, 'grad_norm': 4.3193440437316895, 'learning_rate': 7.1393939393939405e-06, 'num_tokens': 2114655.0, 'completions/mean_length': 120.625, 'completions/min_length': 119.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.625, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9766536355018616, 'rewards/meter/std': 0.021832915022969246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7142857313156128, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6976097822189331, 'rewards/total_composite/std': 0.015594935044646263, 'reward': 0.6976097822189331, 'reward_std': 0.015594931319355965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015354371629655361, 'sampling/sampling_logp_difference/max': 1.3104934692382812, 'sampling/importance_sampling_ratio/min': 0.2696869373321533, 'sampling/importance_sampling_ratio/mean': 0.9987999200820923, 'sampling/importance_sampling_ratio/max': 1.6762388944625854, 'entropy': 0.05396859138272703, 'clip_ratio/low_mean': 0.002040850231423974, 'clip_ratio/low_min': 0.002040850231423974, 'clip_ratio/high_mean': 0.011334090260788798, 'clip_ratio/high_max': 0.011334090260788798, 'clip_ratio/region_mean': 0.013374940492212772, 'reward_total_mean': 0.6976097822189331, 'reward_meter_mean': 0.9766536355018616, 'reward_meter_std': 0.021832915022969246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7142857313156128, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6976097822189331, 'reward_total_composite_std': 0.015594935044646263, 'epoch': 0.04}

 29%|██▊       | 945/3300 [38:27<3:24:44,  5.22s/it]INFO 04-11 23:15:18 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 946/3300 [38:32<3:32:18,  5.41s/it]2026-04-11 23:15:24,021 | INFO | train_grpo_train | metrics_logged mode=train step=946

                                                    
{'loss': 0.0042, 'grad_norm': 1.7848657369613647, 'learning_rate': 7.136363636363637e-06, 'num_tokens': 2117343.0, 'completions/mean_length': 166.0, 'completions/min_length': 159.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.0, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.9958733320236206, 'rewards/meter/std': 0.00040963603532873094, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6428571939468384, 'rewards/repeat_penalty/std': 0.1322600245475769, 'rewards/total_composite/mean': 0.6401993036270142, 'rewards/total_composite/std': 0.13169293105602264, 'reward': 0.6401993036270142, 'reward_std': 0.13169293105602264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009576267562806606, 'sampling/sampling_logp_difference/max': 1.0388219356536865, 'sampling/importance_sampling_ratio/min': 0.35387131571769714, 'sampling/importance_sampling_ratio/mean': 0.9997925162315369, 'sampling/importance_sampling_ratio/max': 1.6956576108932495, 'entropy': 0.03645319910719991, 'clip_ratio/low_mean': 0.00222885946277529, 'clip_ratio/low_min': 0.00222885946277529, 'clip_ratio/high_mean': 0.008308789343573153, 'clip_ratio/high_max': 0.008308789343573153, 'clip_ratio/region_mean': 0.010537648806348443, 'reward_total_mean': 0.6401993036270142, 'reward_meter_mean': 0.9958733320236206, 'reward_meter_std': 0.00040963603532873094, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6428571939468384, 'reward_repeat_penalty_std': 0.1322600245475769, 'reward_total_composite_mean': 0.6401993036270142, 'reward_total_composite_std': 0.13169293105602264, 'epoch': 0.04}

 29%|██▊       | 946/3300 [38:33<3:32:18,  5.41s/it]INFO 04-11 23:15:24 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 947/3300 [38:40<3:54:16,  5.97s/it]2026-04-11 23:15:31,305 | INFO | train_grpo_train | metrics_logged mode=train step=947

                                                    
{'loss': -0.004, 'grad_norm': 1.0996932983398438, 'learning_rate': 7.133333333333334e-06, 'num_tokens': 2121363.0, 'completions/mean_length': 293.5, 'completions/min_length': 287.0, 'completions/max_length': 306.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 293.5, 'completions/min_terminated_length': 287.0, 'completions/max_terminated_length': 306.0, 'rewards/meter/mean': 0.9976366758346558, 'rewards/meter/std': 0.0012180638732388616, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.39835166931152344, 'rewards/repeat_penalty/std': 0.22086545825004578, 'rewards/total_composite/mean': 0.37909868359565735, 'rewards/total_composite/std': 0.22086751461029053, 'reward': 0.37909868359565735, 'reward_std': 0.22086749970912933, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0038092564791440964, 'sampling/sampling_logp_difference/max': 0.7576923370361328, 'sampling/importance_sampling_ratio/min': 0.4687469005584717, 'sampling/importance_sampling_ratio/mean': 1.0007612705230713, 'sampling/importance_sampling_ratio/max': 1.6547387838363647, 'entropy': 0.021893693367019296, 'clip_ratio/low_mean': 0.0021205995872151107, 'clip_ratio/low_min': 0.0021205995872151107, 'clip_ratio/high_mean': 0.0017141569405794144, 'clip_ratio/high_max': 0.0017141569405794144, 'clip_ratio/region_mean': 0.003834756527794525, 'reward_total_mean': 0.37909868359565735, 'reward_meter_mean': 0.9976366758346558, 'reward_meter_std': 0.0012180638732388616, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.39835166931152344, 'reward_repeat_penalty_std': 0.22086545825004578, 'reward_total_composite_mean': 0.37909868359565735, 'reward_total_composite_std': 0.22086751461029053, 'epoch': 0.04}

 29%|██▊       | 947/3300 [38:40<3:54:16,  5.97s/it]INFO 04-11 23:15:31 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 948/3300 [38:44<3:36:52,  5.53s/it]2026-04-11 23:15:35,808 | INFO | train_grpo_train | metrics_logged mode=train step=948

                                                    
{'loss': -0.012, 'grad_norm': 6.672044277191162, 'learning_rate': 7.130303030303031e-06, 'num_tokens': 2123173.0, 'completions/mean_length': 70.25, 'completions/min_length': 67.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9970372915267944, 'rewards/meter/std': 0.0017233057878911495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970372915267944, 'rewards/total_composite/std': 0.0017233057878911495, 'reward': 0.9970372915267944, 'reward_std': 0.001723314169794321, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02216244488954544, 'sampling/sampling_logp_difference/max': 0.823767364025116, 'sampling/importance_sampling_ratio/min': 0.43877550959587097, 'sampling/importance_sampling_ratio/mean': 1.0034829378128052, 'sampling/importance_sampling_ratio/max': 1.8399385213851929, 'entropy': 0.1040426753461361, 'clip_ratio/low_mean': 0.0072228144854307175, 'clip_ratio/low_min': 0.0072228144854307175, 'clip_ratio/high_mean': 0.008802816737443209, 'clip_ratio/high_max': 0.008802816737443209, 'clip_ratio/region_mean': 0.016025631222873926, 'reward_total_mean': 0.9970372915267944, 'reward_meter_mean': 0.9970372915267944, 'reward_meter_std': 0.0017233057878911495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970372915267944, 'reward_total_composite_std': 0.0017233057878911495, 'epoch': 0.04}

 29%|██▊       | 948/3300 [38:44<3:36:52,  5.53s/it]INFO 04-11 23:15:36 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 949/3300 [38:49<3:24:00,  5.21s/it]2026-04-11 23:15:40,254 | INFO | train_grpo_train | metrics_logged mode=train step=949

                                                    
{'loss': 0.0108, 'grad_norm': 1.9708746671676636, 'learning_rate': 7.127272727272728e-06, 'num_tokens': 2125026.0, 'completions/mean_length': 60.625, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7747896909713745, 'rewards/meter/std': 0.38887178897857666, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.5285885334014893, 'rewards/total_composite/std': 0.23691311478614807, 'reward': 0.5285885334014893, 'reward_std': 0.23691309988498688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0069820573553442955, 'sampling/sampling_logp_difference/max': 0.7136227488517761, 'sampling/importance_sampling_ratio/min': 0.48986631631851196, 'sampling/importance_sampling_ratio/mean': 1.0037261247634888, 'sampling/importance_sampling_ratio/max': 1.2806702852249146, 'entropy': 0.03217482636682689, 'clip_ratio/low_mean': 0.006048386916518211, 'clip_ratio/low_min': 0.006048386916518211, 'clip_ratio/high_mean': 0.004098360426723957, 'clip_ratio/high_max': 0.004098360426723957, 'clip_ratio/region_mean': 0.010146747343242168, 'reward_total_mean': 0.5285885334014893, 'reward_meter_mean': 0.7747896909713745, 'reward_meter_std': 0.38887178897857666, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.5285885334014893, 'reward_total_composite_std': 0.23691311478614807, 'epoch': 0.04}

 29%|██▉       | 949/3300 [38:49<3:24:00,  5.21s/it]INFO 04-11 23:15:40 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 950/3300 [38:54<3:22:56,  5.18s/it]2026-04-11 23:15:45,378 | INFO | train_grpo_train | metrics_logged mode=train step=950

                                                    
{'loss': 0.0008, 'grad_norm': 3.3005993366241455, 'learning_rate': 7.124242424242424e-06, 'num_tokens': 2127299.0, 'completions/mean_length': 120.125, 'completions/min_length': 117.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.125, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9637161493301392, 'rewards/meter/std': 0.08138205856084824, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.7464566826820374, 'rewards/total_composite/std': 0.09103824943304062, 'reward': 0.7464566826820374, 'reward_std': 0.09103825688362122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010762141086161137, 'sampling/sampling_logp_difference/max': 1.548635721206665, 'sampling/importance_sampling_ratio/min': 0.2125377207994461, 'sampling/importance_sampling_ratio/mean': 0.996985137462616, 'sampling/importance_sampling_ratio/max': 1.577369213104248, 'entropy': 0.030512073542922735, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007258509169332683, 'clip_ratio/high_max': 0.007258509169332683, 'clip_ratio/region_mean': 0.007258509169332683, 'reward_total_mean': 0.7464566826820374, 'reward_meter_mean': 0.9637161493301392, 'reward_meter_std': 0.08138205856084824, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.7464566826820374, 'reward_total_composite_std': 0.09103824943304062, 'epoch': 0.04}

 29%|██▉       | 950/3300 [38:54<3:22:56,  5.18s/it]INFO 04-11 23:15:45 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:31,  2.86s/it]

 23%|██▎       | 3/13 [00:09<00:31,  3.16s/it]

 31%|███       | 4/13 [00:14<00:35,  3.96s/it]

 38%|███▊      | 5/13 [00:19<00:34,  4.30s/it]

 46%|████▌     | 6/13 [00:25<00:34,  4.93s/it]

 54%|█████▍    | 7/13 [00:30<00:29,  4.92s/it]

 62%|██████▏   | 8/13 [00:36<00:26,  5.29s/it]

 69%|██████▉   | 9/13 [00:43<00:23,  5.86s/it]

 77%|███████▋  | 10/13 [00:48<00:16,  5.45s/it]

 85%|████████▍ | 11/13 [00:53<00:10,  5.40s/it]

 92%|█████████▏| 12/13 [00:59<00:05,  5.44s/it]

100%|██████████| 13/13 [01:04<00:00,  5.52s/it]2026-04-11 23:16:56,665 | INFO | train_grpo_train | metrics_logged mode=eval step=950

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 71.2662, 'eval_samples_per_second': 1.459, 'eval_steps_per_second': 0.182, 'eval_num_tokens': 2127299.0, 'eval_completions/mean_length': 205.85576923076923, 'eval_completions/min_length': 61.53846153846154, 'eval_completions/max_length': 376.0769230769231, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 205.85576923076923, 'eval_completions/min_terminated_length': 61.53846153846154, 'eval_completions/max_terminated_length': 376.0769230769231, 'eval_rewards/meter/mean': 0.6758117584081796, 'eval_rewards/meter/std': 0.3830513243491833, 'eval_rewards/count_adherence/mean': 0.9363291034331689, 'eval_rewards/count_adherence/std': 0.09608005073208076, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.6863612211667575, 'eval_rewards/repeat_penalty/std': 0.2006635952454347, 'eval_rewards/total_composite/mean': 0.41787450359417844, 'eval_rewards/total_composite/std': 0.3079184889793396, 'eval_reward': 0.41787450359417844, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0055299800498267776, 'eval_sampling/sampling_logp_difference/max': 0.7134297077472394, 'eval_sampling/importance_sampling_ratio/min': 0.5113821442310627, 'eval_sampling/importance_sampling_ratio/mean': 1.0011087380922759, 'eval_sampling/importance_sampling_ratio/max': 1.2753138267076933, 'eval_entropy': 0.04716356213276203, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.41787450359417844, 'eval_reward_meter_mean': 0.6758117584081796, 'eval_reward_meter_std': 0.3830513243491833, 'eval_reward_count_adherence_mean': 0.9363291034331689, 'eval_reward_count_adherence_std': 0.09608005073208076, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.6863612211667575, 'eval_reward_repeat_penalty_std': 0.2006635952454347, 'eval_reward_total_composite_mean': 0.41787450359417844, 'eval_reward_total_composite_std': 0.3079184889793396, 'epoch': 0.04}

 29%|██▉       | 950/3300 [40:05<3:22:56,  5.18s/it]

100%|██████████| 13/13 [01:04<00:00,  5.52s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 29%|██▉       | 951/3300 [40:12<17:44:19, 27.19s/it]2026-04-11 23:17:03,908 | INFO | train_grpo_train | metrics_logged mode=train step=951

                                                     
{'loss': 0.0095, 'grad_norm': 4.512762069702148, 'learning_rate': 7.121212121212122e-06, 'num_tokens': 2129201.0, 'completions/mean_length': 66.75, 'completions/min_length': 63.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.2129497528076172, 'rewards/meter/std': 0.24503695964813232, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.2129497528076172, 'rewards/total_composite/std': 0.24503695964813232, 'reward': 0.2129497528076172, 'reward_std': 0.24503694474697113, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02744590863585472, 'sampling/sampling_logp_difference/max': 1.3395442962646484, 'sampling/importance_sampling_ratio/min': 0.2619650065898895, 'sampling/importance_sampling_ratio/mean': 1.001453161239624, 'sampling/importance_sampling_ratio/max': 1.650530457496643, 'entropy': 0.21755263209342957, 'clip_ratio/low_mean': 0.00940205657389015, 'clip_ratio/low_min': 0.00940205657389015, 'clip_ratio/high_mean': 0.011278833262622356, 'clip_ratio/high_max': 0.011278833262622356, 'clip_ratio/region_mean': 0.020680889836512506, 'reward_total_mean': 0.2129497528076172, 'reward_meter_mean': 0.2129497528076172, 'reward_meter_std': 0.24503695964813232, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.2129497528076172, 'reward_total_composite_std': 0.24503695964813232, 'epoch': 0.04}

 29%|██▉       | 951/3300 [40:12<17:44:19, 27.19s/it]INFO 04-11 23:17:04 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 952/3300 [40:22<14:18:48, 21.95s/it]2026-04-11 23:17:13,628 | INFO | train_grpo_train | metrics_logged mode=train step=952

                                                     
{'loss': -0.0019, 'grad_norm': 0.9411554336547852, 'learning_rate': 7.118181818181819e-06, 'num_tokens': 2134537.0, 'completions/mean_length': 451.0, 'completions/min_length': 446.0, 'completions/max_length': 470.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 451.0, 'completions/min_terminated_length': 446.0, 'completions/max_terminated_length': 470.0, 'rewards/meter/mean': 0.49129414558410645, 'rewards/meter/std': 0.5206316113471985, 'rewards/count_adherence/mean': 0.6315789222717285, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5652173757553101, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.17538189888000488, 'rewards/total_composite/std': 0.18585476279258728, 'reward': 0.17538189888000488, 'reward_std': 0.18585477769374847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004929275717586279, 'sampling/sampling_logp_difference/max': 3.009263277053833, 'sampling/importance_sampling_ratio/min': 0.04932800680398941, 'sampling/importance_sampling_ratio/mean': 1.0005483627319336, 'sampling/importance_sampling_ratio/max': 1.997017741203308, 'entropy': 0.013566843350417912, 'clip_ratio/low_mean': 0.0016632305341772735, 'clip_ratio/low_min': 0.0016632305341772735, 'clip_ratio/high_mean': 0.0016673028003424406, 'clip_ratio/high_max': 0.0016673028003424406, 'clip_ratio/region_mean': 0.003330533334519714, 'reward_total_mean': 0.17538189888000488, 'reward_meter_mean': 0.49129414558410645, 'reward_meter_std': 0.5206316113471985, 'reward_count_adherence_mean': 0.6315789222717285, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5652173757553101, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.17538189888000488, 'reward_total_composite_std': 0.18585476279258728, 'epoch': 0.04}

 29%|██▉       | 952/3300 [40:22<14:18:48, 21.95s/it]INFO 04-11 23:17:13 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 953/3300 [40:27<10:58:51, 16.84s/it]2026-04-11 23:17:18,566 | INFO | train_grpo_train | metrics_logged mode=train step=953

                                                     
{'loss': 0.0166, 'grad_norm': 7.106107711791992, 'learning_rate': 7.115151515151516e-06, 'num_tokens': 2136613.0, 'completions/mean_length': 80.5, 'completions/min_length': 78.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.5, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9900895357131958, 'rewards/meter/std': 0.006374072283506393, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8412514925003052, 'rewards/total_composite/std': 0.08799233287572861, 'reward': 0.8412514925003052, 'reward_std': 0.08799233287572861, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03556607663631439, 'sampling/sampling_logp_difference/max': 2.8349223136901855, 'sampling/importance_sampling_ratio/min': 0.0587230883538723, 'sampling/importance_sampling_ratio/mean': 1.0007519721984863, 'sampling/importance_sampling_ratio/max': 1.8997235298156738, 'entropy': 0.09662660211324692, 'clip_ratio/low_mean': 0.02908552496228367, 'clip_ratio/low_min': 0.02908552496228367, 'clip_ratio/high_mean': 0.00769954826682806, 'clip_ratio/high_max': 0.00769954826682806, 'clip_ratio/region_mean': 0.03678507322911173, 'reward_total_mean': 0.8412514925003052, 'reward_meter_mean': 0.9900895357131958, 'reward_meter_std': 0.006374072283506393, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8412514925003052, 'reward_total_composite_std': 0.08799233287572861, 'epoch': 0.04}

 29%|██▉       | 953/3300 [40:27<10:58:51, 16.84s/it]INFO 04-11 23:17:18 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 954/3300 [40:32<8:42:35, 13.37s/it] 2026-04-11 23:17:23,815 | INFO | train_grpo_train | metrics_logged mode=train step=954

                                                    
{'loss': 0.0104, 'grad_norm': 3.721968650817871, 'learning_rate': 7.1121212121212125e-06, 'num_tokens': 2138876.0, 'completions/mean_length': 104.875, 'completions/min_length': 99.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.875, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.28197556734085083, 'rewards/meter/std': 0.30284935235977173, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.22144606709480286, 'rewards/total_composite/std': 0.24437181651592255, 'reward': 0.22144606709480286, 'reward_std': 0.24437181651592255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029477477073669434, 'sampling/sampling_logp_difference/max': 1.4413394927978516, 'sampling/importance_sampling_ratio/min': 0.23661062121391296, 'sampling/importance_sampling_ratio/mean': 1.00920569896698, 'sampling/importance_sampling_ratio/max': 1.7154247760772705, 'entropy': 0.17342047207057476, 'clip_ratio/low_mean': 0.0235287812538445, 'clip_ratio/low_min': 0.0235287812538445, 'clip_ratio/high_mean': 0.009797494392842054, 'clip_ratio/high_max': 0.009797494392842054, 'clip_ratio/region_mean': 0.033326275646686554, 'reward_total_mean': 0.22144606709480286, 'reward_meter_mean': 0.28197556734085083, 'reward_meter_std': 0.30284935235977173, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.22144606709480286, 'reward_total_composite_std': 0.24437181651592255, 'epoch': 0.04}

 29%|██▉       | 954/3300 [40:32<8:42:35, 13.37s/it]INFO 04-11 23:17:24 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 955/3300 [40:38<7:07:30, 10.94s/it]2026-04-11 23:17:29,113 | INFO | train_grpo_train | metrics_logged mode=train step=955

                                                    
{'loss': 0.009, 'grad_norm': 0.7482571005821228, 'learning_rate': 7.10909090909091e-06, 'num_tokens': 2141066.0, 'completions/mean_length': 113.75, 'completions/min_length': 113.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.75, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.969186544418335, 'rewards/meter/std': 0.005740889813750982, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7753492593765259, 'rewards/total_composite/std': 0.0045927222818136215, 'reward': 0.7753492593765259, 'reward_std': 0.0045927101746201515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003761922474950552, 'sampling/sampling_logp_difference/max': 0.406435489654541, 'sampling/importance_sampling_ratio/min': 0.729997992515564, 'sampling/importance_sampling_ratio/mean': 1.0022902488708496, 'sampling/importance_sampling_ratio/max': 1.501456379890442, 'entropy': 0.02124447817914188, 'clip_ratio/low_mean': 0.005445147980935872, 'clip_ratio/low_min': 0.005445147980935872, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.005445147980935872, 'reward_total_mean': 0.7753492593765259, 'reward_meter_mean': 0.969186544418335, 'reward_meter_std': 0.005740889813750982, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7753492593765259, 'reward_total_composite_std': 0.0045927222818136215, 'epoch': 0.04}

 29%|██▉       | 955/3300 [40:38<7:07:30, 10.94s/it]INFO 04-11 23:17:29 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 956/3300 [40:43<5:59:52,  9.21s/it]2026-04-11 23:17:34,277 | INFO | train_grpo_train | metrics_logged mode=train step=956

                                                    
{'loss': 0.014, 'grad_norm': 3.8508245944976807, 'learning_rate': 7.106060606060606e-06, 'num_tokens': 2143130.0, 'completions/mean_length': 104.0, 'completions/min_length': 101.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.0, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.9977437853813171, 'rewards/meter/std': 0.0006820790586061776, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8979662656784058, 'rewards/total_composite/std': 0.10663474351167679, 'reward': 0.8979662656784058, 'reward_std': 0.10663473606109619, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026750473305583, 'sampling/sampling_logp_difference/max': 2.033867359161377, 'sampling/importance_sampling_ratio/min': 0.1308285892009735, 'sampling/importance_sampling_ratio/mean': 1.0018119812011719, 'sampling/importance_sampling_ratio/max': 1.998322606086731, 'entropy': 0.07065040059387684, 'clip_ratio/low_mean': 0.0059523810632526875, 'clip_ratio/low_min': 0.0059523810632526875, 'clip_ratio/high_mean': 0.0036183277843520045, 'clip_ratio/high_max': 0.0036183277843520045, 'clip_ratio/region_mean': 0.009570708847604692, 'reward_total_mean': 0.8979662656784058, 'reward_meter_mean': 0.9977437853813171, 'reward_meter_std': 0.0006820790586061776, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8979662656784058, 'reward_total_composite_std': 0.10663474351167679, 'epoch': 0.04}

 29%|██▉       | 956/3300 [40:43<5:59:52,  9.21s/it]INFO 04-11 23:17:34 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 957/3300 [40:47<5:06:11,  7.84s/it]2026-04-11 23:17:38,927 | INFO | train_grpo_train | metrics_logged mode=train step=957

                                                    
{'loss': 0.033, 'grad_norm': 6.492129802703857, 'learning_rate': 7.103030303030304e-06, 'num_tokens': 2145169.0, 'completions/mean_length': 60.875, 'completions/min_length': 60.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9347407817840576, 'rewards/meter/std': 0.14110930263996124, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7083333730697632, 'rewards/repeat_penalty/std': 0.11785111576318741, 'rewards/total_composite/mean': 0.6475568413734436, 'rewards/total_composite/std': 0.025069493800401688, 'reward': 0.6475568413734436, 'reward_std': 0.02506948821246624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003160211257636547, 'sampling/sampling_logp_difference/max': 0.3890615701675415, 'sampling/importance_sampling_ratio/min': 0.6776925325393677, 'sampling/importance_sampling_ratio/mean': 0.9995921850204468, 'sampling/importance_sampling_ratio/max': 1.1547554731369019, 'entropy': 0.01487105863634497, 'clip_ratio/low_mean': 0.0037313431967049837, 'clip_ratio/low_min': 0.0037313431967049837, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0037313431967049837, 'reward_total_mean': 0.6475568413734436, 'reward_meter_mean': 0.9347407817840576, 'reward_meter_std': 0.14110930263996124, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7083333730697632, 'reward_repeat_penalty_std': 0.11785111576318741, 'reward_total_composite_mean': 0.6475568413734436, 'reward_total_composite_std': 0.025069493800401688, 'epoch': 0.04}

 29%|██▉       | 957/3300 [40:47<5:06:11,  7.84s/it]INFO 04-11 23:17:39 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 958/3300 [40:52<4:24:38,  6.78s/it]2026-04-11 23:17:43,219 | INFO | train_grpo_train | metrics_logged mode=train step=958

                                                    
{'loss': -0.0039, 'grad_norm': 10.342998504638672, 'learning_rate': 7.100000000000001e-06, 'num_tokens': 2146649.0, 'completions/mean_length': 35.0, 'completions/min_length': 34.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.4885302782058716, 'rewards/meter/std': 0.35837680101394653, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4885302782058716, 'rewards/total_composite/std': 0.35837680101394653, 'reward': 0.4885302782058716, 'reward_std': 0.35837680101394653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03277422487735748, 'sampling/sampling_logp_difference/max': 1.3705358505249023, 'sampling/importance_sampling_ratio/min': 0.25397083163261414, 'sampling/importance_sampling_ratio/mean': 0.9996507167816162, 'sampling/importance_sampling_ratio/max': 1.507129430770874, 'entropy': 0.10543786082416773, 'clip_ratio/low_mean': 0.010714286006987095, 'clip_ratio/low_min': 0.010714286006987095, 'clip_ratio/high_mean': 0.010615079430863261, 'clip_ratio/high_max': 0.010615079430863261, 'clip_ratio/region_mean': 0.021329365437850356, 'reward_total_mean': 0.4885302782058716, 'reward_meter_mean': 0.4885302782058716, 'reward_meter_std': 0.35837680101394653, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4885302782058716, 'reward_total_composite_std': 0.35837680101394653, 'epoch': 0.04}

 29%|██▉       | 958/3300 [40:52<4:24:38,  6.78s/it]INFO 04-11 23:17:43 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 959/3300 [40:59<4:34:11,  7.03s/it]2026-04-11 23:17:50,835 | INFO | train_grpo_train | metrics_logged mode=train step=959

                                                    
{'loss': -0.0351, 'grad_norm': 1.4002562761306763, 'learning_rate': 7.096969696969698e-06, 'num_tokens': 2150789.0, 'completions/mean_length': 294.5, 'completions/min_length': 274.0, 'completions/max_length': 310.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 294.5, 'completions/min_terminated_length': 274.0, 'completions/max_terminated_length': 310.0, 'rewards/meter/mean': 0.9977169036865234, 'rewards/meter/std': 0.0006972014671191573, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4833333492279053, 'rewards/repeat_penalty/std': 0.216024711728096, 'rewards/total_composite/mean': 0.4822729825973511, 'rewards/total_composite/std': 0.21558886766433716, 'reward': 0.4822729825973511, 'reward_std': 0.21558886766433716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0083305099979043, 'sampling/sampling_logp_difference/max': 1.0299842357635498, 'sampling/importance_sampling_ratio/min': 0.3570125997066498, 'sampling/importance_sampling_ratio/mean': 0.9994634389877319, 'sampling/importance_sampling_ratio/max': 1.689242959022522, 'entropy': 0.04552340344525874, 'clip_ratio/low_mean': 0.001365295291179791, 'clip_ratio/low_min': 0.001365295291179791, 'clip_ratio/high_mean': 0.006645367917371914, 'clip_ratio/high_max': 0.006645367917371914, 'clip_ratio/region_mean': 0.008010663208551705, 'reward_total_mean': 0.4822729825973511, 'reward_meter_mean': 0.9977169036865234, 'reward_meter_std': 0.0006972014671191573, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4833333492279053, 'reward_repeat_penalty_std': 0.216024711728096, 'reward_total_composite_mean': 0.4822729825973511, 'reward_total_composite_std': 0.21558886766433716, 'epoch': 0.04}

 29%|██▉       | 959/3300 [40:59<4:34:11,  7.03s/it]INFO 04-11 23:17:51 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 960/3300 [41:05<4:19:27,  6.65s/it]2026-04-11 23:17:56,620 | INFO | train_grpo_train | metrics_logged mode=train step=960

                                                    
{'loss': -0.0083, 'grad_norm': 6.993642807006836, 'learning_rate': 7.093939393939394e-06, 'num_tokens': 2153531.0, 'completions/mean_length': 160.75, 'completions/min_length': 156.0, 'completions/max_length': 165.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.75, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.36679452657699585, 'rewards/meter/std': 0.16856136918067932, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7222222089767456, 'rewards/repeat_penalty/std': 0.08399210125207901, 'rewards/total_composite/mean': 0.2650865912437439, 'rewards/total_composite/std': 0.1350317746400833, 'reward': 0.2650865912437439, 'reward_std': 0.1350317746400833, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02235065959393978, 'sampling/sampling_logp_difference/max': 4.669411659240723, 'sampling/importance_sampling_ratio/min': 0.00937778502702713, 'sampling/importance_sampling_ratio/mean': 0.9979313015937805, 'sampling/importance_sampling_ratio/max': 1.8480039834976196, 'entropy': 0.05065230838954449, 'clip_ratio/low_mean': 0.009524673456326127, 'clip_ratio/low_min': 0.009524673456326127, 'clip_ratio/high_mean': 0.009170350269414485, 'clip_ratio/high_max': 0.009170350269414485, 'clip_ratio/region_mean': 0.01869502372574061, 'reward_total_mean': 0.2650865912437439, 'reward_meter_mean': 0.36679452657699585, 'reward_meter_std': 0.16856136918067932, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7222222089767456, 'reward_repeat_penalty_std': 0.08399210125207901, 'reward_total_composite_mean': 0.2650865912437439, 'reward_total_composite_std': 0.1350317746400833, 'epoch': 0.04}

 29%|██▉       | 960/3300 [41:05<4:19:27,  6.65s/it]INFO 04-11 23:17:56 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 961/3300 [41:10<3:56:08,  6.06s/it]2026-04-11 23:18:01,282 | INFO | train_grpo_train | metrics_logged mode=train step=961

                                                    
{'loss': 0.0198, 'grad_norm': 5.430922508239746, 'learning_rate': 7.0909090909090916e-06, 'num_tokens': 2155316.0, 'completions/mean_length': 65.125, 'completions/min_length': 62.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9629546403884888, 'rewards/meter/std': 0.05961780995130539, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9629546403884888, 'rewards/total_composite/std': 0.05961780995130539, 'reward': 0.9629546403884888, 'reward_std': 0.05961783230304718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024385379627346992, 'sampling/sampling_logp_difference/max': 1.6775139570236206, 'sampling/importance_sampling_ratio/min': 0.18683788180351257, 'sampling/importance_sampling_ratio/mean': 0.9952221512794495, 'sampling/importance_sampling_ratio/max': 1.2267403602600098, 'entropy': 0.0778669067658484, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.01166338287293911, 'clip_ratio/high_max': 0.01166338287293911, 'clip_ratio/region_mean': 0.01166338287293911, 'reward_total_mean': 0.9629546403884888, 'reward_meter_mean': 0.9629546403884888, 'reward_meter_std': 0.05961780995130539, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9629546403884888, 'reward_total_composite_std': 0.05961780995130539, 'epoch': 0.04}

 29%|██▉       | 961/3300 [41:10<3:56:08,  6.06s/it]INFO 04-11 23:18:01 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 962/3300 [41:14<3:40:04,  5.65s/it]2026-04-11 23:18:05,963 | INFO | train_grpo_train | metrics_logged mode=train step=962

                                                    
{'loss': 0.0499, 'grad_norm': 8.99122142791748, 'learning_rate': 7.087878787878788e-06, 'num_tokens': 2157171.0, 'completions/mean_length': 65.875, 'completions/min_length': 62.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.32342612743377686, 'rewards/meter/std': 0.291739284992218, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.32342612743377686, 'rewards/total_composite/std': 0.291739284992218, 'reward': 0.32342612743377686, 'reward_std': 0.291739284992218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04449179768562317, 'sampling/sampling_logp_difference/max': 1.642319679260254, 'sampling/importance_sampling_ratio/min': 0.1935305893421173, 'sampling/importance_sampling_ratio/mean': 1.0009416341781616, 'sampling/importance_sampling_ratio/max': 1.8445723056793213, 'entropy': 0.24257362261414528, 'clip_ratio/low_mean': 0.01819333794992417, 'clip_ratio/low_min': 0.01819333794992417, 'clip_ratio/high_mean': 0.015785872004926205, 'clip_ratio/high_max': 0.015785872004926205, 'clip_ratio/region_mean': 0.033979209954850376, 'reward_total_mean': 0.32342612743377686, 'reward_meter_mean': 0.32342612743377686, 'reward_meter_std': 0.291739284992218, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.32342612743377686, 'reward_total_composite_std': 0.291739284992218, 'epoch': 0.04}

 29%|██▉       | 962/3300 [41:14<3:40:04,  5.65s/it]INFO 04-11 23:18:06 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 963/3300 [41:24<4:25:04,  6.81s/it]2026-04-11 23:18:15,471 | INFO | train_grpo_train | metrics_logged mode=train step=963

                                                    
{'loss': 0.0162, 'grad_norm': 1.0686217546463013, 'learning_rate': 7.084848484848485e-06, 'num_tokens': 2162491.0, 'completions/mean_length': 445.0, 'completions/min_length': 435.0, 'completions/max_length': 468.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 445.0, 'completions/min_terminated_length': 435.0, 'completions/max_terminated_length': 468.0, 'rewards/meter/mean': 0.8052209615707397, 'rewards/meter/std': 0.37245726585388184, 'rewards/count_adherence/mean': 0.7291666269302368, 'rewards/count_adherence/std': 0.01964186504483223, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6346794962882996, 'rewards/repeat_penalty/std': 0.046481478959321976, 'rewards/total_composite/mean': 0.3684366047382355, 'rewards/total_composite/std': 0.16659744083881378, 'reward': 0.3684366047382355, 'reward_std': 0.1665974259376526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010675698518753052, 'sampling/sampling_logp_difference/max': 2.1844980716705322, 'sampling/importance_sampling_ratio/min': 0.11253420263528824, 'sampling/importance_sampling_ratio/mean': 0.9988970160484314, 'sampling/importance_sampling_ratio/max': 1.6414300203323364, 'entropy': 0.03673372324556112, 'clip_ratio/low_mean': 0.002477022586390376, 'clip_ratio/low_min': 0.002477022586390376, 'clip_ratio/high_mean': 0.0073644123040139675, 'clip_ratio/high_max': 0.0073644123040139675, 'clip_ratio/region_mean': 0.009841434890404344, 'reward_total_mean': 0.3684366047382355, 'reward_meter_mean': 0.8052209615707397, 'reward_meter_std': 0.37245726585388184, 'reward_count_adherence_mean': 0.7291666269302368, 'reward_count_adherence_std': 0.01964186504483223, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6346794962882996, 'reward_repeat_penalty_std': 0.046481478959321976, 'reward_total_composite_mean': 0.3684366047382355, 'reward_total_composite_std': 0.16659744083881378, 'epoch': 0.04}

 29%|██▉       | 963/3300 [41:24<4:25:04,  6.81s/it]INFO 04-11 23:18:15 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 964/3300 [41:31<4:26:15,  6.84s/it]2026-04-11 23:18:22,389 | INFO | train_grpo_train | metrics_logged mode=train step=964

                                                    
{'loss': -0.005, 'grad_norm': 2.381939649581909, 'learning_rate': 7.081818181818182e-06, 'num_tokens': 2165977.0, 'completions/mean_length': 240.75, 'completions/min_length': 234.0, 'completions/max_length': 261.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 240.75, 'completions/min_terminated_length': 234.0, 'completions/max_terminated_length': 261.0, 'rewards/meter/mean': 0.9859161376953125, 'rewards/meter/std': 0.0027882950380444527, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6057692766189575, 'rewards/repeat_penalty/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.5972245931625366, 'rewards/total_composite/std': 0.026564769446849823, 'reward': 0.5972245931625366, 'reward_std': 0.026564771309494972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004197314847260714, 'sampling/sampling_logp_difference/max': 1.4283943176269531, 'sampling/importance_sampling_ratio/min': 0.23969349265098572, 'sampling/importance_sampling_ratio/mean': 1.0001258850097656, 'sampling/importance_sampling_ratio/max': 1.6210657358169556, 'entropy': 0.012361971399514005, 'clip_ratio/low_mean': 0.0005319148767739534, 'clip_ratio/low_min': 0.0005319148767739534, 'clip_ratio/high_mean': 0.0026147099561057985, 'clip_ratio/high_max': 0.0026147099561057985, 'clip_ratio/region_mean': 0.003146624832879752, 'reward_total_mean': 0.5972245931625366, 'reward_meter_mean': 0.9859161376953125, 'reward_meter_std': 0.0027882950380444527, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6057692766189575, 'reward_repeat_penalty_std': 0.027196412906050682, 'reward_total_composite_mean': 0.5972245931625366, 'reward_total_composite_std': 0.026564769446849823, 'epoch': 0.04}

 29%|██▉       | 964/3300 [41:31<4:26:15,  6.84s/it]INFO 04-11 23:18:22 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 965/3300 [41:36<4:02:27,  6.23s/it]2026-04-11 23:18:27,196 | INFO | train_grpo_train | metrics_logged mode=train step=965

                                                    
{'loss': 0.0083, 'grad_norm': 6.1013593673706055, 'learning_rate': 7.07878787878788e-06, 'num_tokens': 2168080.0, 'completions/mean_length': 71.875, 'completions/min_length': 68.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9859936833381653, 'rewards/meter/std': 0.00813794881105423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9859936833381653, 'rewards/total_composite/std': 0.00813794881105423, 'reward': 0.9859936833381653, 'reward_std': 0.008137939497828484, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008848754689097404, 'sampling/sampling_logp_difference/max': 0.8346943855285645, 'sampling/importance_sampling_ratio/min': 0.6168099045753479, 'sampling/importance_sampling_ratio/mean': 1.001380443572998, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03642029734328389, 'clip_ratio/low_mean': 0.0036498295376077294, 'clip_ratio/low_min': 0.0036498295376077294, 'clip_ratio/high_mean': 0.005208333372138441, 'clip_ratio/high_max': 0.005208333372138441, 'clip_ratio/region_mean': 0.00885816290974617, 'reward_total_mean': 0.9859936833381653, 'reward_meter_mean': 0.9859936833381653, 'reward_meter_std': 0.00813794881105423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9859936833381653, 'reward_total_composite_std': 0.00813794881105423, 'epoch': 0.04}

 29%|██▉       | 965/3300 [41:36<4:02:27,  6.23s/it]INFO 04-11 23:18:27 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 966/3300 [41:40<3:43:40,  5.75s/it]2026-04-11 23:18:31,827 | INFO | train_grpo_train | metrics_logged mode=train step=966

                                                    
{'loss': 0.0052, 'grad_norm': 10.40634822845459, 'learning_rate': 7.075757575757576e-06, 'num_tokens': 2169886.0, 'completions/mean_length': 64.75, 'completions/min_length': 62.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7862817645072937, 'rewards/meter/std': 0.3110516667366028, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7862817645072937, 'rewards/total_composite/std': 0.3110516667366028, 'reward': 0.7862817645072937, 'reward_std': 0.3110516369342804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05011482164263725, 'sampling/sampling_logp_difference/max': 1.9868323802947998, 'sampling/importance_sampling_ratio/min': 0.3482150137424469, 'sampling/importance_sampling_ratio/mean': 1.0035208463668823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18688618578016758, 'clip_ratio/low_mean': 0.005769230891019106, 'clip_ratio/low_min': 0.005769230891019106, 'clip_ratio/high_mean': 0.048127192771062255, 'clip_ratio/high_max': 0.048127192771062255, 'clip_ratio/region_mean': 0.05389642366208136, 'reward_total_mean': 0.7862817645072937, 'reward_meter_mean': 0.7862817645072937, 'reward_meter_std': 0.3110516667366028, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7862817645072937, 'reward_total_composite_std': 0.3110516667366028, 'epoch': 0.04}

 29%|██▉       | 966/3300 [41:40<3:43:40,  5.75s/it]INFO 04-11 23:18:32 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 967/3300 [41:47<3:51:57,  5.97s/it]2026-04-11 23:18:38,296 | INFO | train_grpo_train | metrics_logged mode=train step=967

                                                    
{'loss': -0.0072, 'grad_norm': 2.4970195293426514, 'learning_rate': 7.072727272727273e-06, 'num_tokens': 2172905.0, 'completions/mean_length': 194.375, 'completions/min_length': 184.0, 'completions/max_length': 206.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 194.375, 'completions/min_terminated_length': 184.0, 'completions/max_terminated_length': 206.0, 'rewards/meter/mean': 0.5914201736450195, 'rewards/meter/std': 0.4835823178291321, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6805555820465088, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.39438170194625854, 'rewards/total_composite/std': 0.32224810123443604, 'reward': 0.39438170194625854, 'reward_std': 0.32224807143211365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008972376585006714, 'sampling/sampling_logp_difference/max': 1.8997198343276978, 'sampling/importance_sampling_ratio/min': 0.1496105194091797, 'sampling/importance_sampling_ratio/mean': 1.000235676765442, 'sampling/importance_sampling_ratio/max': 1.642437219619751, 'entropy': 0.03619017521850765, 'clip_ratio/low_mean': 0.005190499185118824, 'clip_ratio/low_min': 0.005190499185118824, 'clip_ratio/high_mean': 0.004360174119938165, 'clip_ratio/high_max': 0.004360174119938165, 'clip_ratio/region_mean': 0.00955067330505699, 'reward_total_mean': 0.39438170194625854, 'reward_meter_mean': 0.5914201736450195, 'reward_meter_std': 0.4835823178291321, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6805555820465088, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.39438170194625854, 'reward_total_composite_std': 0.32224810123443604, 'epoch': 0.04}

 29%|██▉       | 967/3300 [41:47<3:51:57,  5.97s/it]INFO 04-11 23:18:38 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 968/3300 [41:52<3:39:21,  5.64s/it]2026-04-11 23:18:43,190 | INFO | train_grpo_train | metrics_logged mode=train step=968

                                                    
{'loss': -0.0085, 'grad_norm': 5.905569076538086, 'learning_rate': 7.06969696969697e-06, 'num_tokens': 2174629.0, 'completions/mean_length': 60.5, 'completions/min_length': 58.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.1817864626646042, 'rewards/meter/std': 0.293150931596756, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.1817864626646042, 'rewards/total_composite/std': 0.293150931596756, 'reward': 0.1817864626646042, 'reward_std': 0.293150931596756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026289377361536026, 'sampling/sampling_logp_difference/max': 0.7234911918640137, 'sampling/importance_sampling_ratio/min': 0.4889802932739258, 'sampling/importance_sampling_ratio/mean': 1.0088768005371094, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0875935684889555, 'clip_ratio/low_mean': 0.03520799265243113, 'clip_ratio/low_min': 0.03520799265243113, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.03520799265243113, 'reward_total_mean': 0.1817864626646042, 'reward_meter_mean': 0.1817864626646042, 'reward_meter_std': 0.293150931596756, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.1817864626646042, 'reward_total_composite_std': 0.293150931596756, 'epoch': 0.04}

 29%|██▉       | 968/3300 [41:52<3:39:21,  5.64s/it]INFO 04-11 23:18:43 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 969/3300 [41:57<3:35:01,  5.53s/it]2026-04-11 23:18:48,470 | INFO | train_grpo_train | metrics_logged mode=train step=969

                                                    
{'loss': -0.0153, 'grad_norm': 3.71621036529541, 'learning_rate': 7.066666666666667e-06, 'num_tokens': 2176683.0, 'completions/mean_length': 93.75, 'completions/min_length': 87.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.75, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.5728106498718262, 'rewards/meter/std': 0.2251601219177246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.824999988079071, 'rewards/repeat_penalty/std': 0.12817399203777313, 'rewards/total_composite/mean': 0.4796490967273712, 'rewards/total_composite/std': 0.21056687831878662, 'reward': 0.4796490967273712, 'reward_std': 0.21056686341762543, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024638189002871513, 'sampling/sampling_logp_difference/max': 1.6777713298797607, 'sampling/importance_sampling_ratio/min': 0.18678979575634003, 'sampling/importance_sampling_ratio/mean': 0.9970004558563232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07108194520696998, 'clip_ratio/low_mean': 0.009395235683768988, 'clip_ratio/low_min': 0.009395235683768988, 'clip_ratio/high_mean': 0.01694070769008249, 'clip_ratio/high_max': 0.01694070769008249, 'clip_ratio/region_mean': 0.026335943373851478, 'reward_total_mean': 0.4796490967273712, 'reward_meter_mean': 0.5728106498718262, 'reward_meter_std': 0.2251601219177246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.824999988079071, 'reward_repeat_penalty_std': 0.12817399203777313, 'reward_total_composite_mean': 0.4796490967273712, 'reward_total_composite_std': 0.21056687831878662, 'epoch': 0.04}

 29%|██▉       | 969/3300 [41:57<3:35:01,  5.53s/it]INFO 04-11 23:18:48 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 970/3300 [42:02<3:26:28,  5.32s/it]2026-04-11 23:18:53,277 | INFO | train_grpo_train | metrics_logged mode=train step=970

                                                    
{'loss': -0.0005, 'grad_norm': 2.2865495681762695, 'learning_rate': 7.063636363636365e-06, 'num_tokens': 2178721.0, 'completions/mean_length': 78.75, 'completions/min_length': 77.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.75, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9977834224700928, 'rewards/meter/std': 0.00039716452010907233, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977834224700928, 'rewards/total_composite/std': 0.00039716452010907233, 'reward': 0.9977834224700928, 'reward_std': 0.0003971673722844571, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019730689004063606, 'sampling/sampling_logp_difference/max': 0.7925920486450195, 'sampling/importance_sampling_ratio/min': 0.4526699483394623, 'sampling/importance_sampling_ratio/mean': 1.0018593072891235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09748060069978237, 'clip_ratio/low_mean': 0.011199243483133614, 'clip_ratio/low_min': 0.011199243483133614, 'clip_ratio/high_mean': 0.006389971007592976, 'clip_ratio/high_max': 0.006389971007592976, 'clip_ratio/region_mean': 0.01758921449072659, 'reward_total_mean': 0.9977834224700928, 'reward_meter_mean': 0.9977834224700928, 'reward_meter_std': 0.00039716452010907233, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977834224700928, 'reward_total_composite_std': 0.00039716452010907233, 'epoch': 0.04}

 29%|██▉       | 970/3300 [42:02<3:26:28,  5.32s/it]INFO 04-11 23:18:53 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 971/3300 [42:08<3:35:21,  5.55s/it]2026-04-11 23:18:59,365 | INFO | train_grpo_train | metrics_logged mode=train step=971

                                                    
{'loss': -0.0062, 'grad_norm': 1.4529743194580078, 'learning_rate': 7.060606060606061e-06, 'num_tokens': 2181793.0, 'completions/mean_length': 189.0, 'completions/min_length': 183.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 189.0, 'completions/min_terminated_length': 183.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.9985735416412354, 'rewards/meter/std': 0.0002435932110529393, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6527777910232544, 'rewards/repeat_penalty/std': 0.13849149644374847, 'rewards/total_composite/mean': 0.6518542766571045, 'rewards/total_composite/std': 0.13831683993339539, 'reward': 0.6518542766571045, 'reward_std': 0.13831683993339539, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016793714836239815, 'sampling/sampling_logp_difference/max': 5.936721324920654, 'sampling/importance_sampling_ratio/min': 0.0026406734250485897, 'sampling/importance_sampling_ratio/mean': 0.998883068561554, 'sampling/importance_sampling_ratio/max': 1.8489798307418823, 'entropy': 0.06829563481733203, 'clip_ratio/low_mean': 0.002027027076110244, 'clip_ratio/low_min': 0.002027027076110244, 'clip_ratio/high_mean': 0.010603155242279172, 'clip_ratio/high_max': 0.010603155242279172, 'clip_ratio/region_mean': 0.012630182318389416, 'reward_total_mean': 0.6518542766571045, 'reward_meter_mean': 0.9985735416412354, 'reward_meter_std': 0.0002435932110529393, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6527777910232544, 'reward_repeat_penalty_std': 0.13849149644374847, 'reward_total_composite_mean': 0.6518542766571045, 'reward_total_composite_std': 0.13831683993339539, 'epoch': 0.04}

 29%|██▉       | 971/3300 [42:08<3:35:21,  5.55s/it]INFO 04-11 23:18:59 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 972/3300 [42:15<3:53:18,  6.01s/it]2026-04-11 23:19:06,465 | INFO | train_grpo_train | metrics_logged mode=train step=972

                                                    
{'loss': -0.01, 'grad_norm': 1.3656582832336426, 'learning_rate': 7.057575757575759e-06, 'num_tokens': 2185373.0, 'completions/mean_length': 263.5, 'completions/min_length': 248.0, 'completions/max_length': 273.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 263.5, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 273.0, 'rewards/meter/mean': 0.6416140794754028, 'rewards/meter/std': 0.4567407965660095, 'rewards/count_adherence/mean': 0.859375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4212246239185333, 'rewards/repeat_penalty/std': 0.21046686172485352, 'rewards/total_composite/mean': 0.19842997193336487, 'rewards/total_composite/std': 0.19151148200035095, 'reward': 0.19842997193336487, 'reward_std': 0.19151146709918976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017121732234954834, 'sampling/sampling_logp_difference/max': 7.143291473388672, 'sampling/importance_sampling_ratio/min': 0.0007901470526121557, 'sampling/importance_sampling_ratio/mean': 1.0005074739456177, 'sampling/importance_sampling_ratio/max': 1.6406102180480957, 'entropy': 0.05923895724117756, 'clip_ratio/low_mean': 0.006205761310411617, 'clip_ratio/low_min': 0.006205761310411617, 'clip_ratio/high_mean': 0.0037591225700452924, 'clip_ratio/high_max': 0.0037591225700452924, 'clip_ratio/region_mean': 0.00996488388045691, 'reward_total_mean': 0.19842997193336487, 'reward_meter_mean': 0.6416140794754028, 'reward_meter_std': 0.4567407965660095, 'reward_count_adherence_mean': 0.859375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4212246239185333, 'reward_repeat_penalty_std': 0.21046686172485352, 'reward_total_composite_mean': 0.19842997193336487, 'reward_total_composite_std': 0.19151148200035095, 'epoch': 0.04}

 29%|██▉       | 972/3300 [42:15<3:53:18,  6.01s/it]INFO 04-11 23:19:06 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 973/3300 [42:20<3:36:38,  5.59s/it]2026-04-11 23:19:11,053 | INFO | train_grpo_train | metrics_logged mode=train step=973

                                                    
{'loss': 0.0171, 'grad_norm': 13.225342750549316, 'learning_rate': 7.054545454545455e-06, 'num_tokens': 2187067.0, 'completions/mean_length': 60.75, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.75, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5853145122528076, 'rewards/meter/std': 0.1321393847465515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5853145122528076, 'rewards/total_composite/std': 0.1321393847465515, 'reward': 0.5853145122528076, 'reward_std': 0.13213936984539032, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022791719064116478, 'sampling/sampling_logp_difference/max': 1.2052631378173828, 'sampling/importance_sampling_ratio/min': 0.2996131479740143, 'sampling/importance_sampling_ratio/mean': 0.9999557733535767, 'sampling/importance_sampling_ratio/max': 1.8551063537597656, 'entropy': 0.09659606032073498, 'clip_ratio/low_mean': 0.008165820967406034, 'clip_ratio/low_min': 0.008165820967406034, 'clip_ratio/high_mean': 0.010416667209938169, 'clip_ratio/high_max': 0.010416667209938169, 'clip_ratio/region_mean': 0.018582488177344203, 'reward_total_mean': 0.5853145122528076, 'reward_meter_mean': 0.5853145122528076, 'reward_meter_std': 0.1321393847465515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5853145122528076, 'reward_total_composite_std': 0.1321393847465515, 'epoch': 0.04}

 29%|██▉       | 973/3300 [42:20<3:36:38,  5.59s/it]INFO 04-11 23:19:11 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 974/3300 [42:24<3:27:20,  5.35s/it]2026-04-11 23:19:15,848 | INFO | train_grpo_train | metrics_logged mode=train step=974

                                                    
{'loss': 0.0027, 'grad_norm': 2.125204563140869, 'learning_rate': 7.0515151515151525e-06, 'num_tokens': 2188986.0, 'completions/mean_length': 78.875, 'completions/min_length': 78.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.875, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9978776574134827, 'rewards/meter/std': 0.0004215049266349524, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978776574134827, 'rewards/total_composite/std': 0.0004215049266349524, 'reward': 0.9978776574134827, 'reward_std': 0.0004215273365844041, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013205158524215221, 'sampling/sampling_logp_difference/max': 0.6857354640960693, 'sampling/importance_sampling_ratio/min': 0.503719687461853, 'sampling/importance_sampling_ratio/mean': 1.0038025379180908, 'sampling/importance_sampling_ratio/max': 1.3348405361175537, 'entropy': 0.08664691727608442, 'clip_ratio/low_mean': 0.006329114083200693, 'clip_ratio/low_min': 0.006329114083200693, 'clip_ratio/high_mean': 0.004807692370377481, 'clip_ratio/high_max': 0.004807692370377481, 'clip_ratio/region_mean': 0.011136806453578174, 'reward_total_mean': 0.9978776574134827, 'reward_meter_mean': 0.9978776574134827, 'reward_meter_std': 0.0004215049266349524, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978776574134827, 'reward_total_composite_std': 0.0004215049266349524, 'epoch': 0.04}

 30%|██▉       | 974/3300 [42:24<3:27:20,  5.35s/it]INFO 04-11 23:19:16 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 975/3300 [42:29<3:22:44,  5.23s/it]2026-04-11 23:19:20,807 | INFO | train_grpo_train | metrics_logged mode=train step=975

                                                    
{'loss': 0.0108, 'grad_norm': 2.8000571727752686, 'learning_rate': 7.048484848484849e-06, 'num_tokens': 2191193.0, 'completions/mean_length': 100.875, 'completions/min_length': 97.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.996698796749115, 'rewards/meter/std': 0.0009329493041150272, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.921911358833313, 'rewards/total_composite/std': 0.10283299535512924, 'reward': 0.921911358833313, 'reward_std': 0.10283298790454865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027565594762563705, 'sampling/sampling_logp_difference/max': 1.987597942352295, 'sampling/importance_sampling_ratio/min': 0.1370241641998291, 'sampling/importance_sampling_ratio/mean': 0.9971323609352112, 'sampling/importance_sampling_ratio/max': 1.8002057075500488, 'entropy': 0.08865811070427299, 'clip_ratio/low_mean': 0.003676470718346536, 'clip_ratio/low_min': 0.003676470718346536, 'clip_ratio/high_mean': 0.019024619832634926, 'clip_ratio/high_max': 0.019024619832634926, 'clip_ratio/region_mean': 0.022701090550981462, 'reward_total_mean': 0.921911358833313, 'reward_meter_mean': 0.996698796749115, 'reward_meter_std': 0.0009329493041150272, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.921911358833313, 'reward_total_composite_std': 0.10283299535512924, 'epoch': 0.04}

 30%|██▉       | 975/3300 [42:29<3:22:44,  5.23s/it]INFO 04-11 23:19:21 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 976/3300 [42:34<3:14:48,  5.03s/it]2026-04-11 23:19:25,364 | INFO | train_grpo_train | metrics_logged mode=train step=976

                                                    
{'loss': -0.0053, 'grad_norm': 8.02316665649414, 'learning_rate': 7.045454545454546e-06, 'num_tokens': 2193051.0, 'completions/mean_length': 60.25, 'completions/min_length': 60.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9850078821182251, 'rewards/meter/std': 0.0014023992698639631, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6566718816757202, 'rewards/total_composite/std': 0.0009349193423986435, 'reward': 0.6566718816757202, 'reward_std': 0.0009349363390356302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009783357381820679, 'sampling/sampling_logp_difference/max': 0.7054405212402344, 'sampling/importance_sampling_ratio/min': 0.4981825351715088, 'sampling/importance_sampling_ratio/mean': 1.0041868686676025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.029942457331344485, 'clip_ratio/low_mean': 0.006250000325962901, 'clip_ratio/low_min': 0.006250000325962901, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.006250000325962901, 'reward_total_mean': 0.6566718816757202, 'reward_meter_mean': 0.9850078821182251, 'reward_meter_std': 0.0014023992698639631, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6566718816757202, 'reward_total_composite_std': 0.0009349193423986435, 'epoch': 0.04}

 30%|██▉       | 976/3300 [42:34<3:14:48,  5.03s/it]INFO 04-11 23:19:25 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 977/3300 [42:39<3:12:00,  4.96s/it]2026-04-11 23:19:30,161 | INFO | train_grpo_train | metrics_logged mode=train step=977

                                                    
{'loss': -0.0065, 'grad_norm': 1.635718584060669, 'learning_rate': 7.0424242424242426e-06, 'num_tokens': 2194984.0, 'completions/mean_length': 79.625, 'completions/min_length': 79.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.625, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9982393980026245, 'rewards/meter/std': 0.0004721701261587441, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982393980026245, 'rewards/total_composite/std': 0.0004721701261587441, 'reward': 0.9982393980026245, 'reward_std': 0.00047218104009516537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013445981778204441, 'sampling/sampling_logp_difference/max': 0.5942535400390625, 'sampling/importance_sampling_ratio/min': 0.5519744157791138, 'sampling/importance_sampling_ratio/mean': 1.0005873441696167, 'sampling/importance_sampling_ratio/max': 1.298281192779541, 'entropy': 0.09024440124630928, 'clip_ratio/low_mean': 0.006289557088166475, 'clip_ratio/low_min': 0.006289557088166475, 'clip_ratio/high_mean': 0.007735339575447142, 'clip_ratio/high_max': 0.007735339575447142, 'clip_ratio/region_mean': 0.014024896663613617, 'reward_total_mean': 0.9982393980026245, 'reward_meter_mean': 0.9982393980026245, 'reward_meter_std': 0.0004721701261587441, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982393980026245, 'reward_total_composite_std': 0.0004721701261587441, 'epoch': 0.04}

 30%|██▉       | 977/3300 [42:39<3:12:00,  4.96s/it]INFO 04-11 23:19:30 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 978/3300 [42:44<3:10:58,  4.93s/it]2026-04-11 23:19:35,038 | INFO | train_grpo_train | metrics_logged mode=train step=978

                                                    
{'loss': 0.0049, 'grad_norm': 3.430257558822632, 'learning_rate': 7.039393939393941e-06, 'num_tokens': 2196933.0, 'completions/mean_length': 84.625, 'completions/min_length': 82.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.625, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.99366694688797, 'rewards/meter/std': 0.0010994295589625835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7949336171150208, 'rewards/total_composite/std': 0.0008795479079708457, 'reward': 0.7949336171150208, 'reward_std': 0.0008795478497631848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016848498955368996, 'sampling/sampling_logp_difference/max': 1.4610930681228638, 'sampling/importance_sampling_ratio/min': 0.23198257386684418, 'sampling/importance_sampling_ratio/mean': 0.9972406029701233, 'sampling/importance_sampling_ratio/max': 1.2917805910110474, 'entropy': 0.06632238999009132, 'clip_ratio/low_mean': 0.0029248768696561456, 'clip_ratio/low_min': 0.0029248768696561456, 'clip_ratio/high_mean': 0.008984935469925404, 'clip_ratio/high_max': 0.008984935469925404, 'clip_ratio/region_mean': 0.01190981233958155, 'reward_total_mean': 0.7949336171150208, 'reward_meter_mean': 0.99366694688797, 'reward_meter_std': 0.0010994295589625835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7949336171150208, 'reward_total_composite_std': 0.0008795479079708457, 'epoch': 0.04}

 30%|██▉       | 978/3300 [42:44<3:10:58,  4.93s/it]INFO 04-11 23:19:35 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 979/3300 [42:48<3:09:31,  4.90s/it]2026-04-11 23:19:39,855 | INFO | train_grpo_train | metrics_logged mode=train step=979

                                                    
{'loss': -0.0028, 'grad_norm': 6.046172618865967, 'learning_rate': 7.036363636363637e-06, 'num_tokens': 2198744.0, 'completions/mean_length': 60.375, 'completions/min_length': 59.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.375, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6340106725692749, 'rewards/meter/std': 0.20943842828273773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6340106725692749, 'rewards/total_composite/std': 0.20943842828273773, 'reward': 0.6340106725692749, 'reward_std': 0.20943844318389893, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030584923923015594, 'sampling/sampling_logp_difference/max': 1.2826333045959473, 'sampling/importance_sampling_ratio/min': 0.27730610966682434, 'sampling/importance_sampling_ratio/mean': 0.9989505410194397, 'sampling/importance_sampling_ratio/max': 1.586508870124817, 'entropy': 0.1247731251642108, 'clip_ratio/low_mean': 0.016670140204951167, 'clip_ratio/low_min': 0.016670140204951167, 'clip_ratio/high_mean': 0.020312500651925802, 'clip_ratio/high_max': 0.020312500651925802, 'clip_ratio/region_mean': 0.03698264085687697, 'reward_total_mean': 0.6340106725692749, 'reward_meter_mean': 0.6340106725692749, 'reward_meter_std': 0.20943842828273773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6340106725692749, 'reward_total_composite_std': 0.20943842828273773, 'epoch': 0.04}

 30%|██▉       | 979/3300 [42:48<3:09:31,  4.90s/it]INFO 04-11 23:19:40 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 980/3300 [42:54<3:14:19,  5.03s/it]2026-04-11 23:19:45,176 | INFO | train_grpo_train | metrics_logged mode=train step=980

                                                    
{'loss': -0.0001, 'grad_norm': 1.3390476703643799, 'learning_rate': 7.033333333333334e-06, 'num_tokens': 2201037.0, 'completions/mean_length': 117.625, 'completions/min_length': 116.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.625, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9986953139305115, 'rewards/meter/std': 0.00018651132995728403, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7989562153816223, 'rewards/total_composite/std': 0.00014920807734597474, 'reward': 0.7989562153816223, 'reward_std': 0.00014920474495738745, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012999390251934528, 'sampling/sampling_logp_difference/max': 1.3501521348953247, 'sampling/importance_sampling_ratio/min': 0.2592008113861084, 'sampling/importance_sampling_ratio/mean': 1.001077651977539, 'sampling/importance_sampling_ratio/max': 1.392466425895691, 'entropy': 0.0817764438688755, 'clip_ratio/low_mean': 0.005235925782471895, 'clip_ratio/low_min': 0.005235925782471895, 'clip_ratio/high_mean': 0.0010683761211112142, 'clip_ratio/high_max': 0.0010683761211112142, 'clip_ratio/region_mean': 0.006304301903583109, 'reward_total_mean': 0.7989562153816223, 'reward_meter_mean': 0.9986953139305115, 'reward_meter_std': 0.00018651132995728403, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7989562153816223, 'reward_total_composite_std': 0.00014920807734597474, 'epoch': 0.04}

 30%|██▉       | 980/3300 [42:54<3:14:19,  5.03s/it]INFO 04-11 23:19:45 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 981/3300 [42:58<3:10:03,  4.92s/it]2026-04-11 23:19:49,840 | INFO | train_grpo_train | metrics_logged mode=train step=981

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.030303030303031e-06, 'num_tokens': 2202789.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9872130751609802, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6581420302391052, 'rewards/total_composite/std': 0.0, 'reward': 0.6581420302391052, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0014747538371011615, 'sampling/sampling_logp_difference/max': 0.05223120003938675, 'sampling/importance_sampling_ratio/min': 0.9491093754768372, 'sampling/importance_sampling_ratio/mean': 1.001099944114685, 'sampling/importance_sampling_ratio/max': 1.0427228212356567, 'entropy': 0.012774199014529586, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.6581420302391052, 'reward_meter_mean': 0.9872130751609802, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6581420302391052, 'reward_total_composite_std': 0.0, 'epoch': 0.04}

 30%|██▉       | 981/3300 [42:58<3:10:03,  4.92s/it]INFO 04-11 23:19:50 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 982/3300 [43:03<3:05:24,  4.80s/it]2026-04-11 23:19:54,364 | INFO | train_grpo_train | metrics_logged mode=train step=982

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.027272727272728e-06, 'num_tokens': 2204653.0, 'completions/mean_length': 52.0, 'completions/min_length': 52.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.961655855178833, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.961655855178833, 'rewards/total_composite/std': 0.0, 'reward': 0.961655855178833, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.007805653847754002, 'sampling/sampling_logp_difference/max': 0.8052540421485901, 'sampling/importance_sampling_ratio/min': 0.44697439670562744, 'sampling/importance_sampling_ratio/mean': 1.0008008480072021, 'sampling/importance_sampling_ratio/max': 1.1538965702056885, 'entropy': 0.04045943822711706, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.961655855178833, 'reward_meter_mean': 0.961655855178833, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.961655855178833, 'reward_total_composite_std': 0.0, 'epoch': 0.04}

 30%|██▉       | 982/3300 [43:03<3:05:24,  4.80s/it]INFO 04-11 23:19:54 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 983/3300 [43:08<3:06:46,  4.84s/it]2026-04-11 23:19:59,289 | INFO | train_grpo_train | metrics_logged mode=train step=983

                                                    
{'loss': -0.0039, 'grad_norm': 2.655090093612671, 'learning_rate': 7.024242424242424e-06, 'num_tokens': 2206569.0, 'completions/mean_length': 80.5, 'completions/min_length': 79.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.5, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.998711347579956, 'rewards/meter/std': 0.0004273413505870849, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998711347579956, 'rewards/total_composite/std': 0.0004273413505870849, 'reward': 0.998711347579956, 'reward_std': 0.000427335879066959, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01396770216524601, 'sampling/sampling_logp_difference/max': 0.902503490447998, 'sampling/importance_sampling_ratio/min': 0.40555310249328613, 'sampling/importance_sampling_ratio/mean': 1.0012972354888916, 'sampling/importance_sampling_ratio/max': 1.6649136543273926, 'entropy': 0.07731548231095076, 'clip_ratio/low_mean': 0.0015625000232830644, 'clip_ratio/low_min': 0.0015625000232830644, 'clip_ratio/high_mean': 0.0030487803742289543, 'clip_ratio/high_max': 0.0030487803742289543, 'clip_ratio/region_mean': 0.004611280397512019, 'reward_total_mean': 0.998711347579956, 'reward_meter_mean': 0.998711347579956, 'reward_meter_std': 0.0004273413505870849, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998711347579956, 'reward_total_composite_std': 0.0004273413505870849, 'epoch': 0.04}

 30%|██▉       | 983/3300 [43:08<3:06:46,  4.84s/it]INFO 04-11 23:19:59 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 984/3300 [43:12<3:01:27,  4.70s/it]2026-04-11 23:20:03,674 | INFO | train_grpo_train | metrics_logged mode=train step=984

                                                    
{'loss': -0.0007, 'grad_norm': 4.302158832550049, 'learning_rate': 7.021212121212122e-06, 'num_tokens': 2208124.0, 'completions/mean_length': 31.375, 'completions/min_length': 31.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9924632906913757, 'rewards/meter/std': 0.0021250685676932335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924632906913757, 'rewards/total_composite/std': 0.0021250685676932335, 'reward': 0.9924632906913757, 'reward_std': 0.0021250564604997635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014777499251067638, 'sampling/sampling_logp_difference/max': 0.5788311958312988, 'sampling/importance_sampling_ratio/min': 0.7517499923706055, 'sampling/importance_sampling_ratio/mean': 1.0098501443862915, 'sampling/importance_sampling_ratio/max': 1.7839521169662476, 'entropy': 0.08878939691931009, 'clip_ratio/low_mean': 0.016003023833036423, 'clip_ratio/low_min': 0.016003023833036423, 'clip_ratio/high_mean': 0.007938507944345474, 'clip_ratio/high_max': 0.007938507944345474, 'clip_ratio/region_mean': 0.023941531777381897, 'reward_total_mean': 0.9924632906913757, 'reward_meter_mean': 0.9924632906913757, 'reward_meter_std': 0.0021250685676932335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924632906913757, 'reward_total_composite_std': 0.0021250685676932335, 'epoch': 0.04}

 30%|██▉       | 984/3300 [43:12<3:01:27,  4.70s/it]INFO 04-11 23:20:03 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 985/3300 [43:17<3:01:44,  4.71s/it]2026-04-11 23:20:08,412 | INFO | train_grpo_train | metrics_logged mode=train step=985

                                                    
{'loss': -0.0119, 'grad_norm': 8.724225997924805, 'learning_rate': 7.018181818181818e-06, 'num_tokens': 2209934.0, 'completions/mean_length': 61.25, 'completions/min_length': 61.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9870957136154175, 'rewards/meter/std': 0.00033180107129737735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7083333730697632, 'rewards/repeat_penalty/std': 0.11785111576318741, 'rewards/total_composite/mean': 0.6991585493087769, 'rewards/total_composite/std': 0.11601238697767258, 'reward': 0.6991585493087769, 'reward_std': 0.11601236462593079, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012023198418319225, 'sampling/sampling_logp_difference/max': 1.0977474451065063, 'sampling/importance_sampling_ratio/min': 0.3336217403411865, 'sampling/importance_sampling_ratio/mean': 1.0018260478973389, 'sampling/importance_sampling_ratio/max': 1.6728845834732056, 'entropy': 0.04339231993071735, 'clip_ratio/low_mean': 0.020491802133619785, 'clip_ratio/low_min': 0.020491802133619785, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.020491802133619785, 'reward_total_mean': 0.6991585493087769, 'reward_meter_mean': 0.9870957136154175, 'reward_meter_std': 0.00033180107129737735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7083333730697632, 'reward_repeat_penalty_std': 0.11785111576318741, 'reward_total_composite_mean': 0.6991585493087769, 'reward_total_composite_std': 0.11601238697767258, 'epoch': 0.04}

 30%|██▉       | 985/3300 [43:17<3:01:44,  4.71s/it]INFO 04-11 23:20:08 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 986/3300 [43:22<3:09:01,  4.90s/it]2026-04-11 23:20:13,749 | INFO | train_grpo_train | metrics_logged mode=train step=986

                                                    
{'loss': -0.0012, 'grad_norm': 10.352254867553711, 'learning_rate': 7.015151515151516e-06, 'num_tokens': 2212176.0, 'completions/mean_length': 99.25, 'completions/min_length': 94.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.25, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9944287538528442, 'rewards/meter/std': 0.0026359334588050842, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9695063829421997, 'rewards/total_composite/std': 0.06953759491443634, 'reward': 0.9695063829421997, 'reward_std': 0.06953759491443634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044116295874118805, 'sampling/sampling_logp_difference/max': 2.196290969848633, 'sampling/importance_sampling_ratio/min': 0.11121489107608795, 'sampling/importance_sampling_ratio/mean': 1.007062315940857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17434827517718077, 'clip_ratio/low_mean': 0.003989361692219973, 'clip_ratio/low_min': 0.003989361692219973, 'clip_ratio/high_mean': 0.037494323682039976, 'clip_ratio/high_max': 0.037494323682039976, 'clip_ratio/region_mean': 0.04148368537425995, 'reward_total_mean': 0.9695063829421997, 'reward_meter_mean': 0.9944287538528442, 'reward_meter_std': 0.0026359334588050842, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9695063829421997, 'reward_total_composite_std': 0.06953759491443634, 'epoch': 0.04}

 30%|██▉       | 986/3300 [43:22<3:09:01,  4.90s/it]INFO 04-11 23:20:14 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 987/3300 [43:27<3:08:25,  4.89s/it]2026-04-11 23:20:18,608 | INFO | train_grpo_train | metrics_logged mode=train step=987

                                                    
{'loss': -0.0152, 'grad_norm': 4.8408732414245605, 'learning_rate': 7.0121212121212126e-06, 'num_tokens': 2213841.0, 'completions/mean_length': 61.125, 'completions/min_length': 58.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9843782782554626, 'rewards/meter/std': 0.00654714647680521, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.7787973880767822, 'rewards/total_composite/std': 0.16700218617916107, 'reward': 0.7787973880767822, 'reward_std': 0.16700218617916107, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02119199000298977, 'sampling/sampling_logp_difference/max': 1.5137114524841309, 'sampling/importance_sampling_ratio/min': 0.22009159624576569, 'sampling/importance_sampling_ratio/mean': 0.9993690252304077, 'sampling/importance_sampling_ratio/max': 1.5299021005630493, 'entropy': 0.08241967670619488, 'clip_ratio/low_mean': 0.022379032103344798, 'clip_ratio/low_min': 0.022379032103344798, 'clip_ratio/high_mean': 0.0040023052133619785, 'clip_ratio/high_max': 0.0040023052133619785, 'clip_ratio/region_mean': 0.026381337316706777, 'reward_total_mean': 0.7787973880767822, 'reward_meter_mean': 0.9843782782554626, 'reward_meter_std': 0.00654714647680521, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.7787973880767822, 'reward_total_composite_std': 0.16700218617916107, 'epoch': 0.04}

 30%|██▉       | 987/3300 [43:27<3:08:25,  4.89s/it]INFO 04-11 23:20:18 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 988/3300 [43:32<3:10:02,  4.93s/it]2026-04-11 23:20:23,643 | INFO | train_grpo_train | metrics_logged mode=train step=988

                                                    
{'loss': 0.0136, 'grad_norm': 3.5085859298706055, 'learning_rate': 7.00909090909091e-06, 'num_tokens': 2215857.0, 'completions/mean_length': 89.0, 'completions/min_length': 87.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.0, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9934116005897522, 'rewards/meter/std': 0.0031344336457550526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8443044424057007, 'rewards/total_composite/std': 0.09088363498449326, 'reward': 0.8443044424057007, 'reward_std': 0.09088362008333206, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023283695802092552, 'sampling/sampling_logp_difference/max': 1.422020673751831, 'sampling/importance_sampling_ratio/min': 0.24122609198093414, 'sampling/importance_sampling_ratio/mean': 1.0044476985931396, 'sampling/importance_sampling_ratio/max': 1.6077567338943481, 'entropy': 0.10643400624394417, 'clip_ratio/low_mean': 0.019569998257793486, 'clip_ratio/low_min': 0.019569998257793486, 'clip_ratio/high_mean': 0.007151253987103701, 'clip_ratio/high_max': 0.007151253987103701, 'clip_ratio/region_mean': 0.026721252244897187, 'reward_total_mean': 0.8443044424057007, 'reward_meter_mean': 0.9934116005897522, 'reward_meter_std': 0.0031344336457550526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8443044424057007, 'reward_total_composite_std': 0.09088363498449326, 'epoch': 0.04}

 30%|██▉       | 988/3300 [43:32<3:10:02,  4.93s/it]INFO 04-11 23:20:23 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 989/3300 [43:37<3:07:55,  4.88s/it]2026-04-11 23:20:28,398 | INFO | train_grpo_train | metrics_logged mode=train step=989

                                                    
{'loss': 0.0136, 'grad_norm': 4.891083717346191, 'learning_rate': 7.006060606060606e-06, 'num_tokens': 2217566.0, 'completions/mean_length': 61.625, 'completions/min_length': 61.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9817750453948975, 'rewards/meter/std': 0.007981428876519203, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9817750453948975, 'rewards/total_composite/std': 0.007981428876519203, 'reward': 0.9817750453948975, 'reward_std': 0.00798142608255148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02616567350924015, 'sampling/sampling_logp_difference/max': 1.2843780517578125, 'sampling/importance_sampling_ratio/min': 0.27682268619537354, 'sampling/importance_sampling_ratio/mean': 1.0027326345443726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14261625241488218, 'clip_ratio/low_mean': 0.01203381922096014, 'clip_ratio/low_min': 0.01203381922096014, 'clip_ratio/high_mean': 0.014212056528776884, 'clip_ratio/high_max': 0.014212056528776884, 'clip_ratio/region_mean': 0.026245875749737024, 'reward_total_mean': 0.9817750453948975, 'reward_meter_mean': 0.9817750453948975, 'reward_meter_std': 0.007981428876519203, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9817750453948975, 'reward_total_composite_std': 0.007981428876519203, 'epoch': 0.04}

 30%|██▉       | 989/3300 [43:37<3:07:55,  4.88s/it]INFO 04-11 23:20:28 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 990/3300 [43:42<3:05:48,  4.83s/it]2026-04-11 23:20:33,101 | INFO | train_grpo_train | metrics_logged mode=train step=990

                                                    
{'loss': 0.0012, 'grad_norm': 5.835471153259277, 'learning_rate': 7.0030303030303035e-06, 'num_tokens': 2219326.0, 'completions/mean_length': 60.0, 'completions/min_length': 59.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.0, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9907355308532715, 'rewards/meter/std': 0.002091656206175685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9495816826820374, 'rewards/total_composite/std': 0.11764256656169891, 'reward': 0.9495816826820374, 'reward_std': 0.11764255911111832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024150043725967407, 'sampling/sampling_logp_difference/max': 1.4593958854675293, 'sampling/importance_sampling_ratio/min': 0.2323766052722931, 'sampling/importance_sampling_ratio/mean': 1.0002789497375488, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09698301088064909, 'clip_ratio/low_mean': 0.0042372881434857845, 'clip_ratio/low_min': 0.0042372881434857845, 'clip_ratio/high_mean': 0.022955394815653563, 'clip_ratio/high_max': 0.022955394815653563, 'clip_ratio/region_mean': 0.027192682959139347, 'reward_total_mean': 0.9495816826820374, 'reward_meter_mean': 0.9907355308532715, 'reward_meter_std': 0.002091656206175685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9495816826820374, 'reward_total_composite_std': 0.11764256656169891, 'epoch': 0.04}

 30%|███       | 990/3300 [43:42<3:05:48,  4.83s/it]INFO 04-11 23:20:33 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 991/3300 [43:51<3:53:33,  6.07s/it]2026-04-11 23:20:42,071 | INFO | train_grpo_train | metrics_logged mode=train step=991

                                                    
{'loss': 0.0039, 'grad_norm': 1.7695684432983398, 'learning_rate': 7e-06, 'num_tokens': 2224258.0, 'completions/mean_length': 407.5, 'completions/min_length': 394.0, 'completions/max_length': 411.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 407.5, 'completions/min_terminated_length': 394.0, 'completions/max_terminated_length': 411.0, 'rewards/meter/mean': 0.9920412302017212, 'rewards/meter/std': 0.003196289064362645, 'rewards/count_adherence/mean': 0.7272727489471436, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.46064817905426025, 'rewards/repeat_penalty/std': 0.15420734882354736, 'rewards/total_composite/mean': 0.3326081335544586, 'rewards/total_composite/std': 0.11179135739803314, 'reward': 0.3326081335544586, 'reward_std': 0.11179134249687195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01644131913781166, 'sampling/sampling_logp_difference/max': 3.4832382202148438, 'sampling/importance_sampling_ratio/min': 0.030707810074090958, 'sampling/importance_sampling_ratio/mean': 0.9987042546272278, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.027644895017147064, 'clip_ratio/low_mean': 0.0006097560981288552, 'clip_ratio/low_min': 0.0006097560981288552, 'clip_ratio/high_mean': 0.008049219497479498, 'clip_ratio/high_max': 0.008049219497479498, 'clip_ratio/region_mean': 0.008658975595608354, 'reward_total_mean': 0.3326081335544586, 'reward_meter_mean': 0.9920412302017212, 'reward_meter_std': 0.003196289064362645, 'reward_count_adherence_mean': 0.7272727489471436, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.46064817905426025, 'reward_repeat_penalty_std': 0.15420734882354736, 'reward_total_composite_mean': 0.3326081335544586, 'reward_total_composite_std': 0.11179135739803314, 'epoch': 0.04}

 30%|███       | 991/3300 [43:51<3:53:33,  6.07s/it]INFO 04-11 23:20:42 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 992/3300 [44:00<4:29:59,  7.02s/it]2026-04-11 23:20:51,306 | INFO | train_grpo_train | metrics_logged mode=train step=992

                                                    
{'loss': 0.0087, 'grad_norm': 1.6267327070236206, 'learning_rate': 6.996969696969698e-06, 'num_tokens': 2229368.0, 'completions/mean_length': 419.75, 'completions/min_length': 409.0, 'completions/max_length': 428.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 419.75, 'completions/min_terminated_length': 409.0, 'completions/max_terminated_length': 428.0, 'rewards/meter/mean': 0.9450836181640625, 'rewards/meter/std': 0.07978925108909607, 'rewards/count_adherence/mean': 0.7954545021057129, 'rewards/count_adherence/std': 0.04208271950483322, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.62321937084198, 'rewards/repeat_penalty/std': 0.04523347690701485, 'rewards/total_composite/mean': 0.4672620892524719, 'rewards/total_composite/std': 0.044083379209041595, 'reward': 0.4672620892524719, 'reward_std': 0.04408339038491249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01929199881851673, 'sampling/sampling_logp_difference/max': 2.7198145389556885, 'sampling/importance_sampling_ratio/min': 0.0658869743347168, 'sampling/importance_sampling_ratio/mean': 1.000062108039856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07821797719225287, 'clip_ratio/low_mean': 0.0047512390883639455, 'clip_ratio/low_min': 0.0047512390883639455, 'clip_ratio/high_mean': 0.01013427460566163, 'clip_ratio/high_max': 0.01013427460566163, 'clip_ratio/region_mean': 0.014885513694025576, 'reward_total_mean': 0.4672620892524719, 'reward_meter_mean': 0.9450836181640625, 'reward_meter_std': 0.07978925108909607, 'reward_count_adherence_mean': 0.7954545021057129, 'reward_count_adherence_std': 0.04208271950483322, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.62321937084198, 'reward_repeat_penalty_std': 0.04523347690701485, 'reward_total_composite_mean': 0.4672620892524719, 'reward_total_composite_std': 0.044083379209041595, 'epoch': 0.04}

 30%|███       | 992/3300 [44:00<4:29:59,  7.02s/it]INFO 04-11 23:20:51 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 993/3300 [44:04<4:03:02,  6.32s/it]2026-04-11 23:20:55,999 | INFO | train_grpo_train | metrics_logged mode=train step=993

                                                    
{'loss': -0.0038, 'grad_norm': 5.988306522369385, 'learning_rate': 6.993939393939394e-06, 'num_tokens': 2231169.0, 'completions/mean_length': 63.125, 'completions/min_length': 61.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.125, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6225343942642212, 'rewards/meter/std': 0.3685222268104553, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6225343942642212, 'rewards/total_composite/std': 0.3685222268104553, 'reward': 0.6225343942642212, 'reward_std': 0.3685222268104553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02788170613348484, 'sampling/sampling_logp_difference/max': 1.1369147300720215, 'sampling/importance_sampling_ratio/min': 0.3208072781562805, 'sampling/importance_sampling_ratio/mean': 1.0002118349075317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1402796907350421, 'clip_ratio/low_mean': 0.01408805197570473, 'clip_ratio/low_min': 0.01408805197570473, 'clip_ratio/high_mean': 0.021799394860863686, 'clip_ratio/high_max': 0.021799394860863686, 'clip_ratio/region_mean': 0.035887446836568415, 'reward_total_mean': 0.6225343942642212, 'reward_meter_mean': 0.6225343942642212, 'reward_meter_std': 0.3685222268104553, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6225343942642212, 'reward_total_composite_std': 0.3685222268104553, 'epoch': 0.04}

 30%|███       | 993/3300 [44:04<4:03:02,  6.32s/it]INFO 04-11 23:20:56 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 994/3300 [44:15<4:48:05,  7.50s/it]2026-04-11 23:21:06,235 | INFO | train_grpo_train | metrics_logged mode=train step=994

                                                    
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.990909090909092e-06, 'num_tokens': 2232881.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9936738014221191, 'rewards/meter/std': 0.003524529282003641, 'rewards/count_adherence/mean': 0.8088235855102539, 'rewards/count_adherence/std': 0.027230001986026764, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5536096096038818, 'rewards/repeat_penalty/std': 0.012604749761521816, 'rewards/total_composite/mean': 0.44484955072402954, 'rewards/total_composite/std': 0.015336754731833935, 'reward': 0.44484955072402954, 'reward_std': 0.015336750075221062, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.44484955072402954, 'reward_meter_mean': 0.9936738014221191, 'reward_meter_std': 0.003524529282003641, 'reward_count_adherence_mean': 0.8088235855102539, 'reward_count_adherence_std': 0.027230001986026764, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5536096096038818, 'reward_repeat_penalty_std': 0.012604749761521816, 'reward_total_composite_mean': 0.44484955072402954, 'reward_total_composite_std': 0.015336754731833935, 'epoch': 0.04}

 30%|███       | 994/3300 [44:15<4:48:05,  7.50s/it]INFO 04-11 23:21:06 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 995/3300 [44:20<4:17:46,  6.71s/it]2026-04-11 23:21:11,117 | INFO | train_grpo_train | metrics_logged mode=train step=995

                                                    
{'loss': 0.0319, 'grad_norm': 5.63449764251709, 'learning_rate': 6.987878787878788e-06, 'num_tokens': 2234891.0, 'completions/mean_length': 76.25, 'completions/min_length': 72.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.03228817135095596, 'rewards/meter/std': 0.06514540314674377, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.03228817135095596, 'rewards/total_composite/std': 0.06514540314674377, 'reward': 0.03228817135095596, 'reward_std': 0.06514539569616318, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0636926218867302, 'sampling/sampling_logp_difference/max': 2.1695494651794434, 'sampling/importance_sampling_ratio/min': 0.11422906816005707, 'sampling/importance_sampling_ratio/mean': 1.0099796056747437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28803054243326187, 'clip_ratio/low_mean': 0.03413326805457473, 'clip_ratio/low_min': 0.03413326805457473, 'clip_ratio/high_mean': 0.010277777910232544, 'clip_ratio/high_max': 0.010277777910232544, 'clip_ratio/region_mean': 0.04441104596480727, 'reward_total_mean': 0.03228817135095596, 'reward_meter_mean': 0.03228817135095596, 'reward_meter_std': 0.06514540314674377, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.03228817135095596, 'reward_total_composite_std': 0.06514540314674377, 'epoch': 0.04}

 30%|███       | 995/3300 [44:20<4:17:46,  6.71s/it]INFO 04-11 23:21:11 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 996/3300 [44:24<3:56:22,  6.16s/it]2026-04-11 23:21:15,975 | INFO | train_grpo_train | metrics_logged mode=train step=996

                                                    
{'loss': 0.0295, 'grad_norm': 6.409210205078125, 'learning_rate': 6.984848484848485e-06, 'num_tokens': 2236814.0, 'completions/mean_length': 76.375, 'completions/min_length': 73.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.010214247740805149, 'rewards/meter/std': 0.007614050526171923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.010214247740805149, 'rewards/total_composite/std': 0.007614050526171923, 'reward': 0.010214247740805149, 'reward_std': 0.007614050526171923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0485016293823719, 'sampling/sampling_logp_difference/max': 1.5220353603363037, 'sampling/importance_sampling_ratio/min': 0.21826720237731934, 'sampling/importance_sampling_ratio/mean': 1.0045915842056274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23532197065651417, 'clip_ratio/low_mean': 0.019090469810180366, 'clip_ratio/low_min': 0.019090469810180366, 'clip_ratio/high_mean': 0.01975645322818309, 'clip_ratio/high_max': 0.01975645322818309, 'clip_ratio/region_mean': 0.03884692303836346, 'reward_total_mean': 0.010214247740805149, 'reward_meter_mean': 0.010214247740805149, 'reward_meter_std': 0.007614050526171923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.010214247740805149, 'reward_total_composite_std': 0.007614050526171923, 'epoch': 0.04}

 30%|███       | 996/3300 [44:24<3:56:22,  6.16s/it]INFO 04-11 23:21:16 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 997/3300 [44:29<3:42:39,  5.80s/it]2026-04-11 23:21:20,949 | INFO | train_grpo_train | metrics_logged mode=train step=997

                                                    
{'loss': 0.0187, 'grad_norm': 5.538641929626465, 'learning_rate': 6.981818181818183e-06, 'num_tokens': 2238755.0, 'completions/mean_length': 77.625, 'completions/min_length': 72.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.12371634691953659, 'rewards/meter/std': 0.3405037224292755, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.002925167791545391, 'rewards/total_composite/std': 0.005130887497216463, 'reward': 0.002925167791545391, 'reward_std': 0.005130887031555176, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06538589298725128, 'sampling/sampling_logp_difference/max': 4.094723701477051, 'sampling/importance_sampling_ratio/min': 0.016660349443554878, 'sampling/importance_sampling_ratio/mean': 0.9927118420600891, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15790804475545883, 'clip_ratio/low_mean': 0.03696049621794373, 'clip_ratio/low_min': 0.03696049621794373, 'clip_ratio/high_mean': 0.009801336331292987, 'clip_ratio/high_max': 0.009801336331292987, 'clip_ratio/region_mean': 0.046761832549236715, 'reward_total_mean': 0.002925167791545391, 'reward_meter_mean': 0.12371634691953659, 'reward_meter_std': 0.3405037224292755, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.002925167791545391, 'reward_total_composite_std': 0.005130887497216463, 'epoch': 0.04}

 30%|███       | 997/3300 [44:29<3:42:39,  5.80s/it]INFO 04-11 23:21:21 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 998/3300 [44:34<3:26:13,  5.38s/it]2026-04-11 23:21:25,331 | INFO | train_grpo_train | metrics_logged mode=train step=998

                                                    
{'loss': -0.0461, 'grad_norm': 7.545363903045654, 'learning_rate': 6.978787878787879e-06, 'num_tokens': 2240295.0, 'completions/mean_length': 42.5, 'completions/min_length': 37.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9984815120697021, 'rewards/meter/std': 0.001035381923429668, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984815120697021, 'rewards/total_composite/std': 0.001035381923429668, 'reward': 0.9984815120697021, 'reward_std': 0.0010353842517361045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043179791420698166, 'sampling/sampling_logp_difference/max': 1.6201910972595215, 'sampling/importance_sampling_ratio/min': 0.19786088168621063, 'sampling/importance_sampling_ratio/mean': 0.9990119934082031, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2211740417405963, 'clip_ratio/low_mean': 0.009268068009987473, 'clip_ratio/low_min': 0.009268068009987473, 'clip_ratio/high_mean': 0.020238095661625266, 'clip_ratio/high_max': 0.020238095661625266, 'clip_ratio/region_mean': 0.02950616367161274, 'reward_total_mean': 0.9984815120697021, 'reward_meter_mean': 0.9984815120697021, 'reward_meter_std': 0.001035381923429668, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984815120697021, 'reward_total_composite_std': 0.001035381923429668, 'epoch': 0.04}

 30%|███       | 998/3300 [44:34<3:26:13,  5.38s/it]INFO 04-11 23:21:25 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 999/3300 [44:39<3:21:53,  5.26s/it]2026-04-11 23:21:30,338 | INFO | train_grpo_train | metrics_logged mode=train step=999

                                                    
{'loss': 0.0238, 'grad_norm': 7.812422752380371, 'learning_rate': 6.975757575757577e-06, 'num_tokens': 2242408.0, 'completions/mean_length': 99.125, 'completions/min_length': 97.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.125, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.8570353984832764, 'rewards/meter/std': 0.18506589531898499, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8570353984832764, 'rewards/total_composite/std': 0.18506589531898499, 'reward': 0.8570353984832764, 'reward_std': 0.18506589531898499, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039642542600631714, 'sampling/sampling_logp_difference/max': 1.2252435684204102, 'sampling/importance_sampling_ratio/min': 0.2936861515045166, 'sampling/importance_sampling_ratio/mean': 1.0009180307388306, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22251357324421406, 'clip_ratio/low_mean': 0.003737623686902225, 'clip_ratio/low_min': 0.003737623686902225, 'clip_ratio/high_mean': 0.026683697244152427, 'clip_ratio/high_max': 0.026683697244152427, 'clip_ratio/region_mean': 0.030421320931054652, 'reward_total_mean': 0.8570353984832764, 'reward_meter_mean': 0.8570353984832764, 'reward_meter_std': 0.18506589531898499, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8570353984832764, 'reward_total_composite_std': 0.18506589531898499, 'epoch': 0.04}

 30%|███       | 999/3300 [44:39<3:21:53,  5.26s/it]INFO 04-11 23:21:30 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1000/3300 [44:44<3:18:38,  5.18s/it]2026-04-11 23:21:35,326 | INFO | train_grpo_train | metrics_logged mode=train step=1000

                                                     
{'loss': 0.016, 'grad_norm': 2.9554319381713867, 'learning_rate': 6.9727272727272735e-06, 'num_tokens': 2244489.0, 'completions/mean_length': 89.125, 'completions/min_length': 86.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.125, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9947108030319214, 'rewards/meter/std': 0.004218456335365772, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8204019665718079, 'rewards/total_composite/std': 0.06665877252817154, 'reward': 0.8204019665718079, 'reward_std': 0.06665877997875214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015147675760090351, 'sampling/sampling_logp_difference/max': 1.3051207065582275, 'sampling/importance_sampling_ratio/min': 0.27113983035087585, 'sampling/importance_sampling_ratio/mean': 1.0028539896011353, 'sampling/importance_sampling_ratio/max': 1.6871954202651978, 'entropy': 0.0689164032228291, 'clip_ratio/low_mean': 0.01114646252244711, 'clip_ratio/low_min': 0.01114646252244711, 'clip_ratio/high_mean': 0.0014534883666783571, 'clip_ratio/high_max': 0.0014534883666783571, 'clip_ratio/region_mean': 0.012599950889125466, 'reward_total_mean': 0.8204019665718079, 'reward_meter_mean': 0.9947108030319214, 'reward_meter_std': 0.004218456335365772, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8204019665718079, 'reward_total_composite_std': 0.06665877252817154, 'epoch': 0.04}

 30%|███       | 1000/3300 [44:44<3:18:38,  5.18s/it]INFO 04-11 23:21:35 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:39,  3.55s/it]

 23%|██▎       | 3/13 [00:10<00:36,  3.62s/it]

 31%|███       | 4/13 [00:17<00:41,  4.63s/it]

 38%|███▊      | 5/13 [00:23<00:40,  5.12s/it]

 46%|████▌     | 6/13 [00:30<00:40,  5.78s/it]

 54%|█████▍    | 7/13 [00:36<00:34,  5.76s/it]

 62%|██████▏   | 8/13 [00:43<00:31,  6.29s/it]

 69%|██████▉   | 9/13 [00:50<00:26,  6.64s/it]

 77%|███████▋  | 10/13 [00:56<00:18,  6.32s/it]

 85%|████████▍ | 11/13 [01:02<00:12,  6.29s/it]

 92%|█████████▏| 12/13 [01:09<00:06,  6.39s/it]

100%|██████████| 13/13 [01:16<00:00,  6.63s/it]2026-04-11 23:22:59,015 | INFO | train_grpo_train | metrics_logged mode=eval step=1000

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 83.6634, 'eval_samples_per_second': 1.243, 'eval_steps_per_second': 0.155, 'eval_num_tokens': 2244489.0, 'eval_completions/mean_length': 233.69230769230768, 'eval_completions/min_length': 60.0, 'eval_completions/max_length': 449.38461538461536, 'eval_completions/clipped_ratio': 0.04807692307692308, 'eval_completions/mean_terminated_length': 219.62088364821213, 'eval_completions/min_terminated_length': 60.0, 'eval_completions/max_terminated_length': 413.2307692307692, 'eval_rewards/meter/mean': 0.644585329752702, 'eval_rewards/meter/std': 0.44141573172349197, 'eval_rewards/count_adherence/mean': 0.9180091665341303, 'eval_rewards/count_adherence/std': 0.09919005589416394, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.7226548378284161, 'eval_rewards/repeat_penalty/std': 0.17840037838770792, 'eval_rewards/total_composite/mean': 0.42624477927501386, 'eval_rewards/total_composite/std': 0.3310052202298091, 'eval_reward': 0.42624477927501386, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.00861521104637247, 'eval_sampling/sampling_logp_difference/max': 0.8023634048608633, 'eval_sampling/importance_sampling_ratio/min': 0.4674220451941857, 'eval_sampling/importance_sampling_ratio/mean': 1.0016976503225474, 'eval_sampling/importance_sampling_ratio/max': 1.3331910921977117, 'eval_entropy': 0.07349483840740643, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.42624477927501386, 'eval_reward_meter_mean': 0.644585329752702, 'eval_reward_meter_std': 0.44141573172349197, 'eval_reward_count_adherence_mean': 0.9180091665341303, 'eval_reward_count_adherence_std': 0.09919005589416394, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.7226548378284161, 'eval_reward_repeat_penalty_std': 0.17840037838770792, 'eval_reward_total_composite_mean': 0.42624477927501386, 'eval_reward_total_composite_std': 0.3310052202298091, 'epoch': 0.04}

 30%|███       | 1000/3300 [46:08<3:18:38,  5.18s/it]

100%|██████████| 13/13 [01:16<00:00,  6.63s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 30%|███       | 1001/3300 [46:15<19:46:40, 30.97s/it]2026-04-11 23:23:06,473 | INFO | train_grpo_train | metrics_logged mode=train step=1001

                                                      
{'loss': 0.0233, 'grad_norm': 8.061637878417969, 'learning_rate': 6.969696969696971e-06, 'num_tokens': 2246187.0, 'completions/mean_length': 62.25, 'completions/min_length': 62.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.25, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8264527320861816, 'rewards/meter/std': 0.1885237991809845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8264527320861816, 'rewards/total_composite/std': 0.1885237991809845, 'reward': 0.8264527320861816, 'reward_std': 0.1885237991809845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02593499980866909, 'sampling/sampling_logp_difference/max': 1.3253092765808105, 'sampling/importance_sampling_ratio/min': 0.38382700085639954, 'sampling/importance_sampling_ratio/mean': 1.0087472200393677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13137870281934738, 'clip_ratio/low_mean': 0.005859375, 'clip_ratio/low_min': 0.005859375, 'clip_ratio/high_mean': 0.02217741869390011, 'clip_ratio/high_max': 0.02217741869390011, 'clip_ratio/region_mean': 0.02803679369390011, 'reward_total_mean': 0.8264527320861816, 'reward_meter_mean': 0.8264527320861816, 'reward_meter_std': 0.1885237991809845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8264527320861816, 'reward_total_composite_std': 0.1885237991809845, 'epoch': 0.04}

 30%|███       | 1001/3300 [46:15<19:46:40, 30.97s/it]INFO 04-11 23:23:06 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1002/3300 [46:20<14:44:59, 23.11s/it]2026-04-11 23:23:11,231 | INFO | train_grpo_train | metrics_logged mode=train step=1002

                                                      
{'loss': 0.0292, 'grad_norm': 9.637553215026855, 'learning_rate': 6.966666666666667e-06, 'num_tokens': 2247834.0, 'completions/mean_length': 65.875, 'completions/min_length': 63.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.8884119987487793, 'rewards/meter/std': 0.23295524716377258, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8884119987487793, 'rewards/total_composite/std': 0.23295524716377258, 'reward': 0.8884119987487793, 'reward_std': 0.2329552173614502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06217576563358307, 'sampling/sampling_logp_difference/max': 1.78690767288208, 'sampling/importance_sampling_ratio/min': 0.16747726500034332, 'sampling/importance_sampling_ratio/mean': 1.0028692483901978, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23851481638848782, 'clip_ratio/low_mean': 0.012928195297718048, 'clip_ratio/low_min': 0.012928195297718048, 'clip_ratio/high_mean': 0.03800246538594365, 'clip_ratio/high_max': 0.03800246538594365, 'clip_ratio/region_mean': 0.0509306606836617, 'reward_total_mean': 0.8884119987487793, 'reward_meter_mean': 0.8884119987487793, 'reward_meter_std': 0.23295524716377258, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8884119987487793, 'reward_total_composite_std': 0.23295524716377258, 'epoch': 0.04}

 30%|███       | 1002/3300 [46:20<14:44:59, 23.11s/it]INFO 04-11 23:23:11 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1003/3300 [46:25<11:19:58, 17.76s/it]2026-04-11 23:23:16,517 | INFO | train_grpo_train | metrics_logged mode=train step=1003

                                                      
{'loss': 0.0045, 'grad_norm': 5.540070056915283, 'learning_rate': 6.963636363636364e-06, 'num_tokens': 2250215.0, 'completions/mean_length': 114.625, 'completions/min_length': 112.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.625, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.781272828578949, 'rewards/meter/std': 0.39195698499679565, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8214285373687744, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.634056806564331, 'rewards/total_composite/std': 0.31964385509490967, 'reward': 0.634056806564331, 'reward_std': 0.31964385509490967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014822704717516899, 'sampling/sampling_logp_difference/max': 1.5009106397628784, 'sampling/importance_sampling_ratio/min': 0.22292707860469818, 'sampling/importance_sampling_ratio/mean': 0.9998385906219482, 'sampling/importance_sampling_ratio/max': 1.965267539024353, 'entropy': 0.05840234411880374, 'clip_ratio/low_mean': 0.004366895416751504, 'clip_ratio/low_min': 0.004366895416751504, 'clip_ratio/high_mean': 0.005464061861857772, 'clip_ratio/high_max': 0.005464061861857772, 'clip_ratio/region_mean': 0.009830957278609276, 'reward_total_mean': 0.634056806564331, 'reward_meter_mean': 0.781272828578949, 'reward_meter_std': 0.39195698499679565, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8214285373687744, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.634056806564331, 'reward_total_composite_std': 0.31964385509490967, 'epoch': 0.04}

 30%|███       | 1003/3300 [46:25<11:19:58, 17.76s/it]INFO 04-11 23:23:16 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1004/3300 [46:30<8:49:12, 13.83s/it] 2026-04-11 23:23:21,176 | INFO | train_grpo_train | metrics_logged mode=train step=1004

                                                     
{'loss': 0.0036, 'grad_norm': 7.136742115020752, 'learning_rate': 6.960606060606061e-06, 'num_tokens': 2251916.0, 'completions/mean_length': 59.625, 'completions/min_length': 58.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9974349737167358, 'rewards/meter/std': 0.0002844088012352586, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974349737167358, 'rewards/total_composite/std': 0.0002844088012352586, 'reward': 0.9974349737167358, 'reward_std': 0.0002844203554559499, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018237588927149773, 'sampling/sampling_logp_difference/max': 1.3813776969909668, 'sampling/importance_sampling_ratio/min': 0.25123220682144165, 'sampling/importance_sampling_ratio/mean': 0.9971798062324524, 'sampling/importance_sampling_ratio/max': 1.5021758079528809, 'entropy': 0.06448704563081264, 'clip_ratio/low_mean': 0.004273816477507353, 'clip_ratio/low_min': 0.004273816477507353, 'clip_ratio/high_mean': 0.014583333861082792, 'clip_ratio/high_max': 0.014583333861082792, 'clip_ratio/region_mean': 0.018857150338590145, 'reward_total_mean': 0.9974349737167358, 'reward_meter_mean': 0.9974349737167358, 'reward_meter_std': 0.0002844088012352586, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974349737167358, 'reward_total_composite_std': 0.0002844088012352586, 'epoch': 0.04}

 30%|███       | 1004/3300 [46:30<8:49:12, 13.83s/it]INFO 04-11 23:23:21 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1005/3300 [46:36<7:23:28, 11.59s/it]2026-04-11 23:23:27,553 | INFO | train_grpo_train | metrics_logged mode=train step=1005

                                                     
{'loss': 0.0226, 'grad_norm': 2.455113649368286, 'learning_rate': 6.957575757575759e-06, 'num_tokens': 2255101.0, 'completions/mean_length': 197.125, 'completions/min_length': 190.0, 'completions/max_length': 203.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.125, 'completions/min_terminated_length': 190.0, 'completions/max_terminated_length': 203.0, 'rewards/meter/mean': 0.9832211136817932, 'rewards/meter/std': 0.004388164728879929, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7840908765792847, 'rewards/repeat_penalty/std': 0.06763852387666702, 'rewards/total_composite/mean': 0.6168004274368286, 'rewards/total_composite/std': 0.054140087217092514, 'reward': 0.6168004274368286, 'reward_std': 0.054140083491802216, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023318301886320114, 'sampling/sampling_logp_difference/max': 1.2863155603408813, 'sampling/importance_sampling_ratio/min': 0.27628687024116516, 'sampling/importance_sampling_ratio/mean': 1.0025383234024048, 'sampling/importance_sampling_ratio/max': 1.8359591960906982, 'entropy': 0.11270508635789156, 'clip_ratio/low_mean': 0.005632697488181293, 'clip_ratio/low_min': 0.005632697488181293, 'clip_ratio/high_mean': 0.014221564109902829, 'clip_ratio/high_max': 0.014221564109902829, 'clip_ratio/region_mean': 0.019854261598084122, 'reward_total_mean': 0.6168004274368286, 'reward_meter_mean': 0.9832211136817932, 'reward_meter_std': 0.004388164728879929, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7840908765792847, 'reward_repeat_penalty_std': 0.06763852387666702, 'reward_total_composite_mean': 0.6168004274368286, 'reward_total_composite_std': 0.054140087217092514, 'epoch': 0.04}

 30%|███       | 1005/3300 [46:36<7:23:28, 11.59s/it]INFO 04-11 23:23:27 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1006/3300 [46:41<6:08:08,  9.63s/it]2026-04-11 23:23:32,597 | INFO | train_grpo_train | metrics_logged mode=train step=1006

                                                     
{'loss': -0.0394, 'grad_norm': 7.012386798858643, 'learning_rate': 6.954545454545455e-06, 'num_tokens': 2257080.0, 'completions/mean_length': 91.375, 'completions/min_length': 80.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.375, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9163761734962463, 'rewards/meter/std': 0.1686118096113205, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7331008911132812, 'rewards/total_composite/std': 0.13488943874835968, 'reward': 0.7331008911132812, 'reward_std': 0.1348894238471985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024793440476059914, 'sampling/sampling_logp_difference/max': 1.819629192352295, 'sampling/importance_sampling_ratio/min': 0.16208584606647491, 'sampling/importance_sampling_ratio/mean': 0.9956486225128174, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09509005583822727, 'clip_ratio/low_mean': 0.0031250000465661287, 'clip_ratio/low_min': 0.0031250000465661287, 'clip_ratio/high_mean': 0.010740866768173873, 'clip_ratio/high_max': 0.010740866768173873, 'clip_ratio/region_mean': 0.013865866814740002, 'reward_total_mean': 0.7331008911132812, 'reward_meter_mean': 0.9163761734962463, 'reward_meter_std': 0.1686118096113205, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7331008911132812, 'reward_total_composite_std': 0.13488943874835968, 'epoch': 0.04}

 30%|███       | 1006/3300 [46:41<6:08:08,  9.63s/it]INFO 04-11 23:23:32 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1007/3300 [46:46<5:11:15,  8.14s/it]2026-04-11 23:23:37,276 | INFO | train_grpo_train | metrics_logged mode=train step=1007

                                                     
{'loss': -0.0001, 'grad_norm': 2.0443665981292725, 'learning_rate': 6.951515151515153e-06, 'num_tokens': 2258850.0, 'completions/mean_length': 60.25, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9974480867385864, 'rewards/meter/std': 0.00012538061127997935, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974480867385864, 'rewards/total_composite/std': 0.00012538061127997935, 'reward': 0.9974480867385864, 'reward_std': 0.00012537713337223977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013044199906289577, 'sampling/sampling_logp_difference/max': 0.7166652679443359, 'sampling/importance_sampling_ratio/min': 0.4883781373500824, 'sampling/importance_sampling_ratio/mean': 0.9988527894020081, 'sampling/importance_sampling_ratio/max': 1.6381621360778809, 'entropy': 0.04919788311235607, 'clip_ratio/low_mean': 0.004166666883975267, 'clip_ratio/low_min': 0.004166666883975267, 'clip_ratio/high_mean': 0.008266129298135638, 'clip_ratio/high_max': 0.008266129298135638, 'clip_ratio/region_mean': 0.012432796182110906, 'reward_total_mean': 0.9974480867385864, 'reward_meter_mean': 0.9974480867385864, 'reward_meter_std': 0.00012538061127997935, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974480867385864, 'reward_total_composite_std': 0.00012538061127997935, 'epoch': 0.04}

 31%|███       | 1007/3300 [46:46<5:11:15,  8.14s/it]INFO 04-11 23:23:37 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1008/3300 [46:55<5:28:31,  8.60s/it]2026-04-11 23:23:46,939 | INFO | train_grpo_train | metrics_logged mode=train step=1008

                                                     
{'loss': 0.0083, 'grad_norm': 2.1715681552886963, 'learning_rate': 6.948484848484849e-06, 'num_tokens': 2263802.0, 'completions/mean_length': 423.0, 'completions/min_length': 336.0, 'completions/max_length': 471.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 423.0, 'completions/min_terminated_length': 336.0, 'completions/max_terminated_length': 471.0, 'rewards/meter/mean': 0.029939208179712296, 'rewards/meter/std': 0.05622902140021324, 'rewards/count_adherence/mean': 0.9107142686843872, 'rewards/count_adherence/std': 0.0832117572426796, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5447744727134705, 'rewards/repeat_penalty/std': 0.19969096779823303, 'rewards/total_composite/mean': 0.017225507646799088, 'rewards/total_composite/std': 0.03621317073702812, 'reward': 0.017225507646799088, 'reward_std': 0.036213167011737823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014329123310744762, 'sampling/sampling_logp_difference/max': 2.782862424850464, 'sampling/importance_sampling_ratio/min': 0.061861179769039154, 'sampling/importance_sampling_ratio/mean': 1.0038172006607056, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08116073720157146, 'clip_ratio/low_mean': 0.010143639519810677, 'clip_ratio/low_min': 0.010143639519810677, 'clip_ratio/high_mean': 0.003859857562929392, 'clip_ratio/high_max': 0.003859857562929392, 'clip_ratio/region_mean': 0.014003497082740068, 'reward_total_mean': 0.017225507646799088, 'reward_meter_mean': 0.029939208179712296, 'reward_meter_std': 0.05622902140021324, 'reward_count_adherence_mean': 0.9107142686843872, 'reward_count_adherence_std': 0.0832117572426796, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5447744727134705, 'reward_repeat_penalty_std': 0.19969096779823303, 'reward_total_composite_mean': 0.017225507646799088, 'reward_total_composite_std': 0.03621317073702812, 'epoch': 0.04}

 31%|███       | 1008/3300 [46:55<5:28:31,  8.60s/it]INFO 04-11 23:23:47 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1009/3300 [47:01<4:48:42,  7.56s/it]2026-04-11 23:23:52,080 | INFO | train_grpo_train | metrics_logged mode=train step=1009

                                                     
{'loss': 0.0182, 'grad_norm': 6.43966007232666, 'learning_rate': 6.945454545454546e-06, 'num_tokens': 2265904.0, 'completions/mean_length': 88.75, 'completions/min_length': 88.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.75, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9968655705451965, 'rewards/meter/std': 0.0008970692288130522, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.7726212739944458, 'rewards/total_composite/std': 0.07099877297878265, 'reward': 0.7726212739944458, 'reward_std': 0.07099878042936325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006405611056834459, 'sampling/sampling_logp_difference/max': 0.8346314430236816, 'sampling/importance_sampling_ratio/min': 0.6927711367607117, 'sampling/importance_sampling_ratio/mean': 1.0016840696334839, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.02403299231082201, 'clip_ratio/low_mean': 0.0027173913549631834, 'clip_ratio/low_min': 0.0027173913549631834, 'clip_ratio/high_mean': 0.004261363763362169, 'clip_ratio/high_max': 0.004261363763362169, 'clip_ratio/region_mean': 0.006978755118325353, 'reward_total_mean': 0.7726212739944458, 'reward_meter_mean': 0.9968655705451965, 'reward_meter_std': 0.0008970692288130522, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.7726212739944458, 'reward_total_composite_std': 0.07099877297878265, 'epoch': 0.04}

 31%|███       | 1009/3300 [47:01<4:48:42,  7.56s/it]INFO 04-11 23:23:52 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1010/3300 [47:05<4:16:44,  6.73s/it]2026-04-11 23:23:56,857 | INFO | train_grpo_train | metrics_logged mode=train step=1010

                                                     
{'loss': 0.0216, 'grad_norm': 7.395992755889893, 'learning_rate': 6.942424242424243e-06, 'num_tokens': 2267676.0, 'completions/mean_length': 69.5, 'completions/min_length': 66.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9437720775604248, 'rewards/meter/std': 0.10555455833673477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9021741151809692, 'rewards/total_composite/std': 0.14069606363773346, 'reward': 0.9021741151809692, 'reward_std': 0.14069606363773346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0496804304420948, 'sampling/sampling_logp_difference/max': 1.3186415433883667, 'sampling/importance_sampling_ratio/min': 0.2674984335899353, 'sampling/importance_sampling_ratio/mean': 1.0084481239318848, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22077207453548908, 'clip_ratio/low_mean': 0.01056338008493185, 'clip_ratio/low_min': 0.01056338008493185, 'clip_ratio/high_mean': 0.023642246145755053, 'clip_ratio/high_max': 0.023642246145755053, 'clip_ratio/region_mean': 0.0342056262306869, 'reward_total_mean': 0.9021741151809692, 'reward_meter_mean': 0.9437720775604248, 'reward_meter_std': 0.10555455833673477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9021741151809692, 'reward_total_composite_std': 0.14069606363773346, 'epoch': 0.04}

 31%|███       | 1010/3300 [47:05<4:16:44,  6.73s/it]INFO 04-11 23:23:57 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1011/3300 [47:10<3:54:40,  6.15s/it]2026-04-11 23:24:01,667 | INFO | train_grpo_train | metrics_logged mode=train step=1011

                                                     
{'loss': 0.0142, 'grad_norm': 5.829395771026611, 'learning_rate': 6.93939393939394e-06, 'num_tokens': 2269520.0, 'completions/mean_length': 68.5, 'completions/min_length': 63.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9850254058837891, 'rewards/meter/std': 0.01548534631729126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9850254058837891, 'rewards/total_composite/std': 0.01548534631729126, 'reward': 0.9850254058837891, 'reward_std': 0.015485321171581745, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06408417969942093, 'sampling/sampling_logp_difference/max': 1.6191701889038086, 'sampling/importance_sampling_ratio/min': 0.1980629861354828, 'sampling/importance_sampling_ratio/mean': 0.9924511313438416, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2325905654579401, 'clip_ratio/low_mean': 0.018277311231940985, 'clip_ratio/low_min': 0.018277311231940985, 'clip_ratio/high_mean': 0.047181503381580114, 'clip_ratio/high_max': 0.047181503381580114, 'clip_ratio/region_mean': 0.0654588146135211, 'reward_total_mean': 0.9850254058837891, 'reward_meter_mean': 0.9850254058837891, 'reward_meter_std': 0.01548534631729126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9850254058837891, 'reward_total_composite_std': 0.01548534631729126, 'epoch': 0.04}

 31%|███       | 1011/3300 [47:10<3:54:40,  6.15s/it]INFO 04-11 23:24:01 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1012/3300 [47:15<3:43:04,  5.85s/it]2026-04-11 23:24:06,814 | INFO | train_grpo_train | metrics_logged mode=train step=1012

                                                     
{'loss': 0.0315, 'grad_norm': 5.250216960906982, 'learning_rate': 6.936363636363636e-06, 'num_tokens': 2271718.0, 'completions/mean_length': 102.75, 'completions/min_length': 91.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.75, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9896959066390991, 'rewards/meter/std': 0.009748869575560093, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9896959066390991, 'rewards/total_composite/std': 0.009748869575560093, 'reward': 0.9896959066390991, 'reward_std': 0.009748890995979309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05327339842915535, 'sampling/sampling_logp_difference/max': 1.0836901664733887, 'sampling/importance_sampling_ratio/min': 0.33834466338157654, 'sampling/importance_sampling_ratio/mean': 1.0056629180908203, 'sampling/importance_sampling_ratio/max': 1.915394902229309, 'entropy': 0.30692350305616856, 'clip_ratio/low_mean': 0.01664550113491714, 'clip_ratio/low_min': 0.01664550113491714, 'clip_ratio/high_mean': 0.03817800944671035, 'clip_ratio/high_max': 0.03817800944671035, 'clip_ratio/region_mean': 0.05482351058162749, 'reward_total_mean': 0.9896959066390991, 'reward_meter_mean': 0.9896959066390991, 'reward_meter_std': 0.009748869575560093, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9896959066390991, 'reward_total_composite_std': 0.009748869575560093, 'epoch': 0.04}

 31%|███       | 1012/3300 [47:15<3:43:04,  5.85s/it]INFO 04-11 23:24:07 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1013/3300 [47:21<3:36:59,  5.69s/it]2026-04-11 23:24:12,139 | INFO | train_grpo_train | metrics_logged mode=train step=1013

                                                     
{'loss': -0.0071, 'grad_norm': 5.079646110534668, 'learning_rate': 6.9333333333333344e-06, 'num_tokens': 2273974.0, 'completions/mean_length': 114.0, 'completions/min_length': 104.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.0, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.004421628080308437, 'rewards/meter/std': 0.004275813698768616, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.0039320518262684345, 'rewards/total_composite/std': 0.004037159029394388, 'reward': 0.0039320518262684345, 'reward_std': 0.0040371594950556755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035947732627391815, 'sampling/sampling_logp_difference/max': 1.231058120727539, 'sampling/importance_sampling_ratio/min': 0.2919834554195404, 'sampling/importance_sampling_ratio/mean': 1.0108870267868042, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.25786815397441387, 'clip_ratio/low_mean': 0.01979458425194025, 'clip_ratio/low_min': 0.01979458425194025, 'clip_ratio/high_mean': 0.00749256182461977, 'clip_ratio/high_max': 0.00749256182461977, 'clip_ratio/region_mean': 0.02728714607656002, 'reward_total_mean': 0.0039320518262684345, 'reward_meter_mean': 0.004421628080308437, 'reward_meter_std': 0.004275813698768616, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.0039320518262684345, 'reward_total_composite_std': 0.004037159029394388, 'epoch': 0.04}

 31%|███       | 1013/3300 [47:21<3:36:59,  5.69s/it]INFO 04-11 23:24:12 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1014/3300 [47:26<3:27:53,  5.46s/it]2026-04-11 23:24:17,043 | INFO | train_grpo_train | metrics_logged mode=train step=1014

                                                     
{'loss': 0.006, 'grad_norm': 3.9192073345184326, 'learning_rate': 6.930303030303031e-06, 'num_tokens': 2275905.0, 'completions/mean_length': 74.375, 'completions/min_length': 70.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9924027919769287, 'rewards/meter/std': 0.005311480723321438, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924027919769287, 'rewards/total_composite/std': 0.005311480723321438, 'reward': 0.9924027919769287, 'reward_std': 0.005311472807079554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024386199191212654, 'sampling/sampling_logp_difference/max': 0.6070661544799805, 'sampling/importance_sampling_ratio/min': 0.5449473261833191, 'sampling/importance_sampling_ratio/mean': 1.0100598335266113, 'sampling/importance_sampling_ratio/max': 1.5236577987670898, 'entropy': 0.16714145429432392, 'clip_ratio/low_mean': 0.005090707214549184, 'clip_ratio/low_min': 0.005090707214549184, 'clip_ratio/high_mean': 0.020599488052539527, 'clip_ratio/high_max': 0.020599488052539527, 'clip_ratio/region_mean': 0.02569019526708871, 'reward_total_mean': 0.9924027919769287, 'reward_meter_mean': 0.9924027919769287, 'reward_meter_std': 0.005311480723321438, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924027919769287, 'reward_total_composite_std': 0.005311480723321438, 'epoch': 0.04}

 31%|███       | 1014/3300 [47:26<3:27:53,  5.46s/it]INFO 04-11 23:24:17 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1015/3300 [47:30<3:15:42,  5.14s/it]2026-04-11 23:24:21,443 | INFO | train_grpo_train | metrics_logged mode=train step=1015

                                                     
{'loss': 0.0154, 'grad_norm': 6.082520008087158, 'learning_rate': 6.927272727272728e-06, 'num_tokens': 2277383.0, 'completions/mean_length': 33.75, 'completions/min_length': 32.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9623132944107056, 'rewards/meter/std': 0.02709752693772316, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9623132944107056, 'rewards/total_composite/std': 0.02709752693772316, 'reward': 0.9623132944107056, 'reward_std': 0.027097532525658607, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03504392132163048, 'sampling/sampling_logp_difference/max': 1.1891875267028809, 'sampling/importance_sampling_ratio/min': 0.3405103385448456, 'sampling/importance_sampling_ratio/mean': 1.01416015625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22182095050811768, 'clip_ratio/low_mean': 0.007352941203862429, 'clip_ratio/low_min': 0.007352941203862429, 'clip_ratio/high_mean': 0.007247899193316698, 'clip_ratio/high_max': 0.007247899193316698, 'clip_ratio/region_mean': 0.014600840397179127, 'reward_total_mean': 0.9623132944107056, 'reward_meter_mean': 0.9623132944107056, 'reward_meter_std': 0.02709752693772316, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9623132944107056, 'reward_total_composite_std': 0.02709752693772316, 'epoch': 0.04}

 31%|███       | 1015/3300 [47:30<3:15:42,  5.14s/it]INFO 04-11 23:24:21 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1016/3300 [47:40<4:13:01,  6.65s/it]2026-04-11 23:24:31,610 | INFO | train_grpo_train | metrics_logged mode=train step=1016

                                                     
{'loss': -0.1996, 'grad_norm': 1.2893778085708618, 'learning_rate': 6.9242424242424245e-06, 'num_tokens': 2281936.0, 'completions/mean_length': 404.125, 'completions/min_length': 374.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 388.71429443359375, 'completions/min_terminated_length': 374.0, 'completions/max_terminated_length': 400.0, 'rewards/meter/mean': 0.8926552534103394, 'rewards/meter/std': 0.2755105197429657, 'rewards/count_adherence/mean': 0.8083332777023315, 'rewards/count_adherence/std': 0.27472931146621704, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.5423148274421692, 'rewards/repeat_penalty/std': 0.26660624146461487, 'rewards/total_composite/mean': 0.37017643451690674, 'rewards/total_composite/std': 0.22237010300159454, 'reward': 0.37017643451690674, 'reward_std': 0.22237010300159454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015432414598762989, 'sampling/sampling_logp_difference/max': 3.34275484085083, 'sampling/importance_sampling_ratio/min': 0.035339467227458954, 'sampling/importance_sampling_ratio/mean': 1.002998948097229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04754339391365647, 'clip_ratio/low_mean': 0.002274398400913924, 'clip_ratio/low_min': 0.002274398400913924, 'clip_ratio/high_mean': 0.00996917171869427, 'clip_ratio/high_max': 0.00996917171869427, 'clip_ratio/region_mean': 0.012243570119608194, 'reward_total_mean': 0.37017643451690674, 'reward_meter_mean': 0.8926552534103394, 'reward_meter_std': 0.2755105197429657, 'reward_count_adherence_mean': 0.8083332777023315, 'reward_count_adherence_std': 0.27472931146621704, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.5423148274421692, 'reward_repeat_penalty_std': 0.26660624146461487, 'reward_total_composite_mean': 0.37017643451690674, 'reward_total_composite_std': 0.22237010300159454, 'epoch': 0.04}

 31%|███       | 1016/3300 [47:40<4:13:01,  6.65s/it]INFO 04-11 23:24:31 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1017/3300 [47:45<3:53:00,  6.12s/it]2026-04-11 23:24:36,509 | INFO | train_grpo_train | metrics_logged mode=train step=1017

                                                     
{'loss': -0.0018, 'grad_norm': 4.289333343505859, 'learning_rate': 6.921212121212122e-06, 'num_tokens': 2283872.0, 'completions/mean_length': 75.0, 'completions/min_length': 73.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.0, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9958159327507019, 'rewards/meter/std': 0.0007772990502417088, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9958159327507019, 'rewards/total_composite/std': 0.0007772990502417088, 'reward': 0.9958159327507019, 'reward_std': 0.0007773060933686793, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05131428688764572, 'sampling/sampling_logp_difference/max': 1.3117103576660156, 'sampling/importance_sampling_ratio/min': 0.26935896277427673, 'sampling/importance_sampling_ratio/mean': 1.0068724155426025, 'sampling/importance_sampling_ratio/max': 1.9063295125961304, 'entropy': 0.2947879508137703, 'clip_ratio/low_mean': 0.013492224738001823, 'clip_ratio/low_min': 0.013492224738001823, 'clip_ratio/high_mean': 0.02297498215921223, 'clip_ratio/high_max': 0.02297498215921223, 'clip_ratio/region_mean': 0.036467206897214055, 'reward_total_mean': 0.9958159327507019, 'reward_meter_mean': 0.9958159327507019, 'reward_meter_std': 0.0007772990502417088, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9958159327507019, 'reward_total_composite_std': 0.0007772990502417088, 'epoch': 0.04}

 31%|███       | 1017/3300 [47:45<3:53:00,  6.12s/it]INFO 04-11 23:24:36 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1018/3300 [47:50<3:36:29,  5.69s/it]2026-04-11 23:24:41,195 | INFO | train_grpo_train | metrics_logged mode=train step=1018

                                                     
{'loss': 0.0122, 'grad_norm': 6.446331977844238, 'learning_rate': 6.918181818181818e-06, 'num_tokens': 2285685.0, 'completions/mean_length': 69.625, 'completions/min_length': 67.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9972782135009766, 'rewards/meter/std': 0.0013893006835132837, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972782135009766, 'rewards/total_composite/std': 0.0013893006835132837, 'reward': 0.9972782135009766, 'reward_std': 0.0013893075520172715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09435711801052094, 'sampling/sampling_logp_difference/max': 6.117408752441406, 'sampling/importance_sampling_ratio/min': 0.0022041599731892347, 'sampling/importance_sampling_ratio/mean': 1.0028208494186401, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4668789766728878, 'clip_ratio/low_mean': 0.02322303969413042, 'clip_ratio/low_min': 0.02322303969413042, 'clip_ratio/high_mean': 0.057335917837917805, 'clip_ratio/high_max': 0.057335917837917805, 'clip_ratio/region_mean': 0.08055895753204823, 'reward_total_mean': 0.9972782135009766, 'reward_meter_mean': 0.9972782135009766, 'reward_meter_std': 0.0013893006835132837, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972782135009766, 'reward_total_composite_std': 0.0013893006835132837, 'epoch': 0.04}

 31%|███       | 1018/3300 [47:50<3:36:29,  5.69s/it]INFO 04-11 23:24:41 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1019/3300 [47:55<3:28:16,  5.48s/it]2026-04-11 23:24:46,172 | INFO | train_grpo_train | metrics_logged mode=train step=1019

                                                     
{'loss': -0.0074, 'grad_norm': 8.041248321533203, 'learning_rate': 6.915151515151515e-06, 'num_tokens': 2287596.0, 'completions/mean_length': 82.875, 'completions/min_length': 77.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.875, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.42031192779541016, 'rewards/meter/std': 0.37385284900665283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.42031192779541016, 'rewards/total_composite/std': 0.37385284900665283, 'reward': 0.42031192779541016, 'reward_std': 0.37385284900665283, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04477884620428085, 'sampling/sampling_logp_difference/max': 1.506969928741455, 'sampling/importance_sampling_ratio/min': 0.221580371260643, 'sampling/importance_sampling_ratio/mean': 0.9995813965797424, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2028534710407257, 'clip_ratio/low_mean': 0.02015177276916802, 'clip_ratio/low_min': 0.02015177276916802, 'clip_ratio/high_mean': 0.023812503553926945, 'clip_ratio/high_max': 0.023812503553926945, 'clip_ratio/region_mean': 0.043964276323094964, 'reward_total_mean': 0.42031192779541016, 'reward_meter_mean': 0.42031192779541016, 'reward_meter_std': 0.37385284900665283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.42031192779541016, 'reward_total_composite_std': 0.37385284900665283, 'epoch': 0.04}

 31%|███       | 1019/3300 [47:55<3:28:16,  5.48s/it]INFO 04-11 23:24:46 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1020/3300 [48:00<3:30:45,  5.55s/it]2026-04-11 23:24:51,892 | INFO | train_grpo_train | metrics_logged mode=train step=1020

                                                     
{'loss': 0.0039, 'grad_norm': 3.4062325954437256, 'learning_rate': 6.912121212121212e-06, 'num_tokens': 2290034.0, 'completions/mean_length': 115.75, 'completions/min_length': 109.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.75, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9905360341072083, 'rewards/meter/std': 0.011200688779354095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8668707013130188, 'rewards/total_composite/std': 0.10446701943874359, 'reward': 0.8668707013130188, 'reward_std': 0.10446702688932419, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031655244529247284, 'sampling/sampling_logp_difference/max': 1.1195735931396484, 'sampling/importance_sampling_ratio/min': 0.3525734841823578, 'sampling/importance_sampling_ratio/mean': 1.0034315586090088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20292328391224146, 'clip_ratio/low_mean': 0.0065684852888807654, 'clip_ratio/low_min': 0.0065684852888807654, 'clip_ratio/high_mean': 0.01801690785214305, 'clip_ratio/high_max': 0.01801690785214305, 'clip_ratio/region_mean': 0.024585393141023815, 'reward_total_mean': 0.8668707013130188, 'reward_meter_mean': 0.9905360341072083, 'reward_meter_std': 0.011200688779354095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8668707013130188, 'reward_total_composite_std': 0.10446701943874359, 'epoch': 0.04}

 31%|███       | 1020/3300 [48:00<3:30:45,  5.55s/it]INFO 04-11 23:24:52 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1021/3300 [48:05<3:25:36,  5.41s/it]2026-04-11 23:24:56,979 | INFO | train_grpo_train | metrics_logged mode=train step=1021

                                                     
{'loss': -0.0207, 'grad_norm': 5.57641077041626, 'learning_rate': 6.90909090909091e-06, 'num_tokens': 2291779.0, 'completions/mean_length': 69.125, 'completions/min_length': 63.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.1570926308631897, 'rewards/meter/std': 0.11832874268293381, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.1570926308631897, 'rewards/total_composite/std': 0.11832874268293381, 'reward': 0.1570926308631897, 'reward_std': 0.1183287501335144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038187529891729355, 'sampling/sampling_logp_difference/max': 0.8104653358459473, 'sampling/importance_sampling_ratio/min': 0.44465112686157227, 'sampling/importance_sampling_ratio/mean': 1.0030486583709717, 'sampling/importance_sampling_ratio/max': 1.6118385791778564, 'entropy': 0.2529278863221407, 'clip_ratio/low_mean': 0.022817460587248206, 'clip_ratio/low_min': 0.022817460587248206, 'clip_ratio/high_mean': 0.023018090520054102, 'clip_ratio/high_max': 0.023018090520054102, 'clip_ratio/region_mean': 0.04583555110730231, 'reward_total_mean': 0.1570926308631897, 'reward_meter_mean': 0.1570926308631897, 'reward_meter_std': 0.11832874268293381, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.1570926308631897, 'reward_total_composite_std': 0.11832874268293381, 'epoch': 0.04}

 31%|███       | 1021/3300 [48:05<3:25:36,  5.41s/it]INFO 04-11 23:24:57 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1022/3300 [48:10<3:19:11,  5.25s/it]2026-04-11 23:25:01,839 | INFO | train_grpo_train | metrics_logged mode=train step=1022

                                                     
{'loss': 0.0374, 'grad_norm': 3.1528494358062744, 'learning_rate': 6.906060606060606e-06, 'num_tokens': 2293742.0, 'completions/mean_length': 82.375, 'completions/min_length': 75.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.375, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9961868524551392, 'rewards/meter/std': 0.0018421995919197798, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961868524551392, 'rewards/total_composite/std': 0.0018421995919197798, 'reward': 0.9961868524551392, 'reward_std': 0.0018422063440084457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036886442452669144, 'sampling/sampling_logp_difference/max': 1.0753283500671387, 'sampling/importance_sampling_ratio/min': 0.34118568897247314, 'sampling/importance_sampling_ratio/mean': 1.0033493041992188, 'sampling/importance_sampling_ratio/max': 1.9688695669174194, 'entropy': 0.21936733089387417, 'clip_ratio/low_mean': 0.008588795084506273, 'clip_ratio/low_min': 0.008588795084506273, 'clip_ratio/high_mean': 0.02277943748049438, 'clip_ratio/high_max': 0.02277943748049438, 'clip_ratio/region_mean': 0.03136823256500065, 'reward_total_mean': 0.9961868524551392, 'reward_meter_mean': 0.9961868524551392, 'reward_meter_std': 0.0018421995919197798, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961868524551392, 'reward_total_composite_std': 0.0018421995919197798, 'epoch': 0.04}

 31%|███       | 1022/3300 [48:10<3:19:11,  5.25s/it]INFO 04-11 23:25:02 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1023/3300 [48:15<3:15:37,  5.15s/it]2026-04-11 23:25:06,781 | INFO | train_grpo_train | metrics_logged mode=train step=1023

                                                     
{'loss': 0.0171, 'grad_norm': 5.055918216705322, 'learning_rate': 6.903030303030304e-06, 'num_tokens': 2295652.0, 'completions/mean_length': 68.75, 'completions/min_length': 66.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.2434961199760437, 'rewards/meter/std': 0.1991458386182785, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.2434961199760437, 'rewards/total_composite/std': 0.1991458386182785, 'reward': 0.2434961199760437, 'reward_std': 0.1991458386182785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05065210163593292, 'sampling/sampling_logp_difference/max': 1.4522520303726196, 'sampling/importance_sampling_ratio/min': 0.23404262959957123, 'sampling/importance_sampling_ratio/mean': 1.0007877349853516, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2538141794502735, 'clip_ratio/low_mean': 0.019885645247995853, 'clip_ratio/low_min': 0.019885645247995853, 'clip_ratio/high_mean': 0.023702416568994522, 'clip_ratio/high_max': 0.023702416568994522, 'clip_ratio/region_mean': 0.043588061816990376, 'reward_total_mean': 0.2434961199760437, 'reward_meter_mean': 0.2434961199760437, 'reward_meter_std': 0.1991458386182785, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.2434961199760437, 'reward_total_composite_std': 0.1991458386182785, 'epoch': 0.04}

 31%|███       | 1023/3300 [48:15<3:15:37,  5.15s/it]INFO 04-11 23:25:07 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1024/3300 [48:24<3:59:05,  6.30s/it]2026-04-11 23:25:15,764 | INFO | train_grpo_train | metrics_logged mode=train step=1024

                                                     
{'loss': -0.0145, 'grad_norm': 0.9608837962150574, 'learning_rate': 6.9e-06, 'num_tokens': 2300311.0, 'completions/mean_length': 388.375, 'completions/min_length': 351.0, 'completions/max_length': 406.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 388.375, 'completions/min_terminated_length': 351.0, 'completions/max_terminated_length': 406.0, 'rewards/meter/mean': 0.988775372505188, 'rewards/meter/std': 0.01991611160337925, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6315789222717285, 'rewards/repeat_penalty/std': 0.05626552551984787, 'rewards/total_composite/mean': 0.44603413343429565, 'rewards/total_composite/std': 0.0405924916267395, 'reward': 0.44603413343429565, 'reward_std': 0.0405924953520298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012534739449620247, 'sampling/sampling_logp_difference/max': 1.506673812866211, 'sampling/importance_sampling_ratio/min': 0.22164598107337952, 'sampling/importance_sampling_ratio/mean': 1.0023077726364136, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08917687484063208, 'clip_ratio/low_mean': 0.004947002336848527, 'clip_ratio/low_min': 0.004947002336848527, 'clip_ratio/high_mean': 0.006947804824449122, 'clip_ratio/high_max': 0.006947804824449122, 'clip_ratio/region_mean': 0.01189480716129765, 'reward_total_mean': 0.44603413343429565, 'reward_meter_mean': 0.988775372505188, 'reward_meter_std': 0.01991611160337925, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6315789222717285, 'reward_repeat_penalty_std': 0.05626552551984787, 'reward_total_composite_mean': 0.44603413343429565, 'reward_total_composite_std': 0.0405924916267395, 'epoch': 0.04}

 31%|███       | 1024/3300 [48:24<3:59:05,  6.30s/it]INFO 04-11 23:25:16 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1025/3300 [48:29<3:45:04,  5.94s/it]2026-04-11 23:25:20,841 | INFO | train_grpo_train | metrics_logged mode=train step=1025

                                                     
{'loss': 0.0116, 'grad_norm': 4.088324069976807, 'learning_rate': 6.896969696969697e-06, 'num_tokens': 2302372.0, 'completions/mean_length': 97.625, 'completions/min_length': 92.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.625, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.6049685478210449, 'rewards/meter/std': 0.33388465642929077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.1511857807636261, 'rewards/total_composite/mean': 0.4717435836791992, 'rewards/total_composite/std': 0.2634972631931305, 'reward': 0.4717435836791992, 'reward_std': 0.2634972631931305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03984498605132103, 'sampling/sampling_logp_difference/max': 1.4417171478271484, 'sampling/importance_sampling_ratio/min': 0.23652127385139465, 'sampling/importance_sampling_ratio/mean': 1.010300636291504, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2695270273834467, 'clip_ratio/low_mean': 0.01788918604142964, 'clip_ratio/low_min': 0.01788918604142964, 'clip_ratio/high_mean': 0.011576289543882012, 'clip_ratio/high_max': 0.011576289543882012, 'clip_ratio/region_mean': 0.02946547558531165, 'reward_total_mean': 0.4717435836791992, 'reward_meter_mean': 0.6049685478210449, 'reward_meter_std': 0.33388465642929077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.1511857807636261, 'reward_total_composite_mean': 0.4717435836791992, 'reward_total_composite_std': 0.2634972631931305, 'epoch': 0.04}

 31%|███       | 1025/3300 [48:29<3:45:04,  5.94s/it]INFO 04-11 23:25:21 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1026/3300 [48:34<3:29:14,  5.52s/it]2026-04-11 23:25:25,396 | INFO | train_grpo_train | metrics_logged mode=train step=1026

                                                     
{'loss': -0.0083, 'grad_norm': 4.054835796356201, 'learning_rate': 6.893939393939395e-06, 'num_tokens': 2304286.0, 'completions/mean_length': 61.25, 'completions/min_length': 60.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.990164577960968, 'rewards/meter/std': 0.004598760511726141, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.990164577960968, 'rewards/total_composite/std': 0.004598760511726141, 'reward': 0.990164577960968, 'reward_std': 0.004598761908710003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02800033800303936, 'sampling/sampling_logp_difference/max': 1.0160877704620361, 'sampling/importance_sampling_ratio/min': 0.3620084226131439, 'sampling/importance_sampling_ratio/mean': 1.0041474103927612, 'sampling/importance_sampling_ratio/max': 1.7953885793685913, 'entropy': 0.15393794886767864, 'clip_ratio/low_mean': 0.020427904091775417, 'clip_ratio/low_min': 0.020427904091775417, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.022381029091775417, 'reward_total_mean': 0.990164577960968, 'reward_meter_mean': 0.990164577960968, 'reward_meter_std': 0.004598760511726141, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.990164577960968, 'reward_total_composite_std': 0.004598760511726141, 'epoch': 0.04}

 31%|███       | 1026/3300 [48:34<3:29:14,  5.52s/it]INFO 04-11 23:25:25 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1027/3300 [48:39<3:22:44,  5.35s/it]2026-04-11 23:25:30,353 | INFO | train_grpo_train | metrics_logged mode=train step=1027

                                                     
{'loss': -0.0138, 'grad_norm': 3.0363378524780273, 'learning_rate': 6.890909090909092e-06, 'num_tokens': 2306411.0, 'completions/mean_length': 90.625, 'completions/min_length': 89.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.625, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9952998757362366, 'rewards/meter/std': 0.0028919558972120285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8707464337348938, 'rewards/total_composite/std': 0.10161302238702774, 'reward': 0.8707464337348938, 'reward_std': 0.10161302983760834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030290186405181885, 'sampling/sampling_logp_difference/max': 1.4388771057128906, 'sampling/importance_sampling_ratio/min': 0.23719395697116852, 'sampling/importance_sampling_ratio/mean': 0.9983088374137878, 'sampling/importance_sampling_ratio/max': 1.4683204889297485, 'entropy': 0.18594534322619438, 'clip_ratio/low_mean': 0.012531211483292282, 'clip_ratio/low_min': 0.012531211483292282, 'clip_ratio/high_mean': 0.0187515887664631, 'clip_ratio/high_max': 0.0187515887664631, 'clip_ratio/region_mean': 0.03128280024975538, 'reward_total_mean': 0.8707464337348938, 'reward_meter_mean': 0.9952998757362366, 'reward_meter_std': 0.0028919558972120285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8707464337348938, 'reward_total_composite_std': 0.10161302238702774, 'epoch': 0.04}

 31%|███       | 1027/3300 [48:39<3:22:44,  5.35s/it]INFO 04-11 23:25:30 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1028/3300 [48:43<3:13:50,  5.12s/it]2026-04-11 23:25:34,928 | INFO | train_grpo_train | metrics_logged mode=train step=1028

                                                     
{'loss': 0.0248, 'grad_norm': 12.675787925720215, 'learning_rate': 6.887878787878789e-06, 'num_tokens': 2308182.0, 'completions/mean_length': 66.375, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9956432580947876, 'rewards/meter/std': 0.003459326224401593, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9956432580947876, 'rewards/total_composite/std': 0.003459326224401593, 'reward': 0.9956432580947876, 'reward_std': 0.003459327155724168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08104093372821808, 'sampling/sampling_logp_difference/max': 2.1169023513793945, 'sampling/importance_sampling_ratio/min': 0.12040401995182037, 'sampling/importance_sampling_ratio/mean': 1.0011348724365234, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39760861173272133, 'clip_ratio/low_mean': 0.011456389911472797, 'clip_ratio/low_min': 0.011456389911472797, 'clip_ratio/high_mean': 0.06568031571805477, 'clip_ratio/high_max': 0.06568031571805477, 'clip_ratio/region_mean': 0.07713670562952757, 'reward_total_mean': 0.9956432580947876, 'reward_meter_mean': 0.9956432580947876, 'reward_meter_std': 0.003459326224401593, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9956432580947876, 'reward_total_composite_std': 0.003459326224401593, 'epoch': 0.04}

 31%|███       | 1028/3300 [48:43<3:13:50,  5.12s/it]INFO 04-11 23:25:35 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1029/3300 [48:48<3:05:23,  4.90s/it]2026-04-11 23:25:39,311 | INFO | train_grpo_train | metrics_logged mode=train step=1029

                                                     
{'loss': 0.0126, 'grad_norm': 5.34241247177124, 'learning_rate': 6.8848484848484854e-06, 'num_tokens': 2309806.0, 'completions/mean_length': 47.0, 'completions/min_length': 46.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.12917472422122955, 'rewards/meter/std': 0.3293028473854065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.12917472422122955, 'rewards/total_composite/std': 0.3293028473854065, 'reward': 0.12917472422122955, 'reward_std': 0.3293028175830841, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07030355930328369, 'sampling/sampling_logp_difference/max': 2.201859712600708, 'sampling/importance_sampling_ratio/min': 0.11059728264808655, 'sampling/importance_sampling_ratio/mean': 0.9958956837654114, 'sampling/importance_sampling_ratio/max': 1.494503378868103, 'entropy': 0.34705137088894844, 'clip_ratio/low_mean': 0.026484928792342544, 'clip_ratio/low_min': 0.026484928792342544, 'clip_ratio/high_mean': 0.007978723384439945, 'clip_ratio/high_max': 0.007978723384439945, 'clip_ratio/region_mean': 0.03446365217678249, 'reward_total_mean': 0.12917472422122955, 'reward_meter_mean': 0.12917472422122955, 'reward_meter_std': 0.3293028473854065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.12917472422122955, 'reward_total_composite_std': 0.3293028473854065, 'epoch': 0.04}

 31%|███       | 1029/3300 [48:48<3:05:23,  4.90s/it]INFO 04-11 23:25:39 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1030/3300 [48:52<3:02:55,  4.83s/it]2026-04-11 23:25:43,999 | INFO | train_grpo_train | metrics_logged mode=train step=1030

                                                     
{'loss': -0.0076, 'grad_norm': 8.219462394714355, 'learning_rate': 6.881818181818183e-06, 'num_tokens': 2311551.0, 'completions/mean_length': 70.125, 'completions/min_length': 67.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.12689465284347534, 'rewards/meter/std': 0.10111062228679657, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.12689465284347534, 'rewards/total_composite/std': 0.10111062228679657, 'reward': 0.12689465284347534, 'reward_std': 0.10111062228679657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07496663182973862, 'sampling/sampling_logp_difference/max': 1.6190156936645508, 'sampling/importance_sampling_ratio/min': 0.19809359312057495, 'sampling/importance_sampling_ratio/mean': 1.0122308731079102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4643949382007122, 'clip_ratio/low_mean': 0.02179320773575455, 'clip_ratio/low_min': 0.02179320773575455, 'clip_ratio/high_mean': 0.032955562230199575, 'clip_ratio/high_max': 0.032955562230199575, 'clip_ratio/region_mean': 0.054748769965954125, 'reward_total_mean': 0.12689465284347534, 'reward_meter_mean': 0.12689465284347534, 'reward_meter_std': 0.10111062228679657, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.12689465284347534, 'reward_total_composite_std': 0.10111062228679657, 'epoch': 0.04}

 31%|███       | 1030/3300 [48:52<3:02:55,  4.83s/it]INFO 04-11 23:25:44 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1031/3300 [49:02<3:56:29,  6.25s/it]2026-04-11 23:25:53,564 | INFO | train_grpo_train | metrics_logged mode=train step=1031

                                                     
{'loss': -0.2286, 'grad_norm': 1.1240724325180054, 'learning_rate': 6.878787878787879e-06, 'num_tokens': 2313782.0, 'completions/mean_length': 224.875, 'completions/min_length': 122.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 129.1666717529297, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.6844993233680725, 'rewards/meter/std': 0.36398857831954956, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.2777460217475891, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.8611111044883728, 'rewards/repeat_penalty/std': 0.09848947077989578, 'rewards/total_composite/mean': 0.5306993722915649, 'rewards/total_composite/std': 0.33074015378952026, 'reward': 0.5306993722915649, 'reward_std': 0.33074015378952026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03939226269721985, 'sampling/sampling_logp_difference/max': 1.3999316692352295, 'sampling/importance_sampling_ratio/min': 0.24661383032798767, 'sampling/importance_sampling_ratio/mean': 1.0042911767959595, 'sampling/importance_sampling_ratio/max': 1.6992144584655762, 'entropy': 0.16609951201826334, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.027175352559424937, 'clip_ratio/high_max': 0.027175352559424937, 'clip_ratio/region_mean': 0.027175352559424937, 'reward_total_mean': 0.5306993722915649, 'reward_meter_mean': 0.6844993233680725, 'reward_meter_std': 0.36398857831954956, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.2777460217475891, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.8611111044883728, 'reward_repeat_penalty_std': 0.09848947077989578, 'reward_total_composite_mean': 0.5306993722915649, 'reward_total_composite_std': 0.33074015378952026, 'epoch': 0.04}

 31%|███       | 1031/3300 [49:02<3:56:29,  6.25s/it]INFO 04-11 23:25:53 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1032/3300 [49:12<4:37:43,  7.35s/it]2026-04-11 23:26:03,461 | INFO | train_grpo_train | metrics_logged mode=train step=1032

                                                     
{'loss': -0.1879, 'grad_norm': 1.4397751092910767, 'learning_rate': 6.875757575757576e-06, 'num_tokens': 2317508.0, 'completions/mean_length': 322.75, 'completions/min_length': 279.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 295.71429443359375, 'completions/min_terminated_length': 279.0, 'completions/max_terminated_length': 312.0, 'rewards/meter/mean': 0.8027582168579102, 'rewards/meter/std': 0.345333069562912, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.2070196568965912, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.630974292755127, 'rewards/repeat_penalty/std': 0.19635772705078125, 'rewards/total_composite/mean': 0.3974968194961548, 'rewards/total_composite/std': 0.2327551245689392, 'reward': 0.3974968194961548, 'reward_std': 0.23275509476661682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022777464240789413, 'sampling/sampling_logp_difference/max': 3.0765089988708496, 'sampling/importance_sampling_ratio/min': 0.04611998051404953, 'sampling/importance_sampling_ratio/mean': 1.0034023523330688, 'sampling/importance_sampling_ratio/max': 1.966111660003662, 'entropy': 0.12301667220890522, 'clip_ratio/low_mean': 0.005716652231058106, 'clip_ratio/low_min': 0.005716652231058106, 'clip_ratio/high_mean': 0.007109183439752087, 'clip_ratio/high_max': 0.007109183439752087, 'clip_ratio/region_mean': 0.012825835670810193, 'reward_total_mean': 0.3974968194961548, 'reward_meter_mean': 0.8027582168579102, 'reward_meter_std': 0.345333069562912, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.2070196568965912, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.630974292755127, 'reward_repeat_penalty_std': 0.19635772705078125, 'reward_total_composite_mean': 0.3974968194961548, 'reward_total_composite_std': 0.2327551245689392, 'epoch': 0.04}

 31%|███▏      | 1032/3300 [49:12<4:37:43,  7.35s/it]INFO 04-11 23:26:03 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1033/3300 [49:16<4:05:46,  6.51s/it]2026-04-11 23:26:08,004 | INFO | train_grpo_train | metrics_logged mode=train step=1033

                                                     
{'loss': 0.0392, 'grad_norm': 16.486671447753906, 'learning_rate': 6.872727272727273e-06, 'num_tokens': 2319195.0, 'completions/mean_length': 50.875, 'completions/min_length': 48.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8368604779243469, 'rewards/meter/std': 0.27409422397613525, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8368604779243469, 'rewards/total_composite/std': 0.27409422397613525, 'reward': 0.8368604779243469, 'reward_std': 0.27409422397613525, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04724201560020447, 'sampling/sampling_logp_difference/max': 1.2481775283813477, 'sampling/importance_sampling_ratio/min': 0.28702741861343384, 'sampling/importance_sampling_ratio/mean': 1.001241683959961, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21698110736906528, 'clip_ratio/low_mean': 0.002358490601181984, 'clip_ratio/low_min': 0.002358490601181984, 'clip_ratio/high_mean': 0.026730769779533148, 'clip_ratio/high_max': 0.026730769779533148, 'clip_ratio/region_mean': 0.029089260380715132, 'reward_total_mean': 0.8368604779243469, 'reward_meter_mean': 0.8368604779243469, 'reward_meter_std': 0.27409422397613525, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8368604779243469, 'reward_total_composite_std': 0.27409422397613525, 'epoch': 0.04}

 31%|███▏      | 1033/3300 [49:17<4:05:46,  6.51s/it]INFO 04-11 23:26:08 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1034/3300 [49:21<3:45:49,  5.98s/it]2026-04-11 23:26:12,756 | INFO | train_grpo_train | metrics_logged mode=train step=1034

                                                     
{'loss': 0.003, 'grad_norm': 5.985336780548096, 'learning_rate': 6.869696969696971e-06, 'num_tokens': 2321016.0, 'completions/mean_length': 67.625, 'completions/min_length': 63.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.053763557225465775, 'rewards/meter/std': 0.056538958102464676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.053763557225465775, 'rewards/total_composite/std': 0.056538958102464676, 'reward': 0.053763557225465775, 'reward_std': 0.05653895437717438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07932179421186447, 'sampling/sampling_logp_difference/max': 1.892876386642456, 'sampling/importance_sampling_ratio/min': 0.1506378948688507, 'sampling/importance_sampling_ratio/mean': 1.0089935064315796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5639077946543694, 'clip_ratio/low_mean': 0.02950768917798996, 'clip_ratio/low_min': 0.02950768917798996, 'clip_ratio/high_mean': 0.022404347429983318, 'clip_ratio/high_max': 0.022404347429983318, 'clip_ratio/region_mean': 0.05191203660797328, 'reward_total_mean': 0.053763557225465775, 'reward_meter_mean': 0.053763557225465775, 'reward_meter_std': 0.056538958102464676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.053763557225465775, 'reward_total_composite_std': 0.056538958102464676, 'epoch': 0.04}

 31%|███▏      | 1034/3300 [49:21<3:45:49,  5.98s/it]INFO 04-11 23:26:12 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1035/3300 [49:26<3:34:17,  5.68s/it]2026-04-11 23:26:17,723 | INFO | train_grpo_train | metrics_logged mode=train step=1035

                                                     
{'loss': 0.0078, 'grad_norm': 4.859825134277344, 'learning_rate': 6.866666666666667e-06, 'num_tokens': 2323151.0, 'completions/mean_length': 95.875, 'completions/min_length': 93.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.875, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.8145841360092163, 'rewards/meter/std': 0.33935561776161194, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.7163466215133667, 'rewards/total_composite/std': 0.304019570350647, 'reward': 0.7163466215133667, 'reward_std': 0.304019570350647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08979999274015427, 'sampling/sampling_logp_difference/max': 24.219942092895508, 'sampling/importance_sampling_ratio/min': 3.0297920422528435e-11, 'sampling/importance_sampling_ratio/mean': 0.9907196164131165, 'sampling/importance_sampling_ratio/max': 1.6885980367660522, 'entropy': 0.18867953680455685, 'clip_ratio/low_mean': 0.006497558439150453, 'clip_ratio/low_min': 0.006497558439150453, 'clip_ratio/high_mean': 0.03407761920243502, 'clip_ratio/high_max': 0.03407761920243502, 'clip_ratio/region_mean': 0.04057517764158547, 'reward_total_mean': 0.7163466215133667, 'reward_meter_mean': 0.8145841360092163, 'reward_meter_std': 0.33935561776161194, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.7163466215133667, 'reward_total_composite_std': 0.304019570350647, 'epoch': 0.04}

 31%|███▏      | 1035/3300 [49:26<3:34:17,  5.68s/it]INFO 04-11 23:26:17 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1036/3300 [49:30<3:18:01,  5.25s/it]2026-04-11 23:26:21,974 | INFO | train_grpo_train | metrics_logged mode=train step=1036

                                                     
{'loss': 0.0157, 'grad_norm': 9.299543380737305, 'learning_rate': 6.8636363636363645e-06, 'num_tokens': 2324686.0, 'completions/mean_length': 34.875, 'completions/min_length': 33.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9909951686859131, 'rewards/meter/std': 0.0066036987118422985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9909951686859131, 'rewards/total_composite/std': 0.0066036987118422985, 'reward': 0.9909951686859131, 'reward_std': 0.006603698246181011, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047876570373773575, 'sampling/sampling_logp_difference/max': 0.7040905952453613, 'sampling/importance_sampling_ratio/min': 0.49455809593200684, 'sampling/importance_sampling_ratio/mean': 1.0070772171020508, 'sampling/importance_sampling_ratio/max': 1.8759939670562744, 'entropy': 0.272716524079442, 'clip_ratio/low_mean': 0.021446609403938055, 'clip_ratio/low_min': 0.021446609403938055, 'clip_ratio/high_mean': 0.021130952751263976, 'clip_ratio/high_max': 0.021130952751263976, 'clip_ratio/region_mean': 0.04257756215520203, 'reward_total_mean': 0.9909951686859131, 'reward_meter_mean': 0.9909951686859131, 'reward_meter_std': 0.0066036987118422985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9909951686859131, 'reward_total_composite_std': 0.0066036987118422985, 'epoch': 0.04}

 31%|███▏      | 1036/3300 [49:30<3:18:01,  5.25s/it]INFO 04-11 23:26:22 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1037/3300 [49:36<3:23:07,  5.39s/it]2026-04-11 23:26:27,680 | INFO | train_grpo_train | metrics_logged mode=train step=1037

                                                     
{'loss': -0.0406, 'grad_norm': 13.22099494934082, 'learning_rate': 6.860606060606061e-06, 'num_tokens': 2327237.0, 'completions/mean_length': 133.875, 'completions/min_length': 109.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.875, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.9935377836227417, 'rewards/meter/std': 0.007430072408169508, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8857142925262451, 'rewards/repeat_penalty/std': 0.07324227690696716, 'rewards/total_composite/mean': 0.8550522923469543, 'rewards/total_composite/std': 0.1225212812423706, 'reward': 0.8550522923469543, 'reward_std': 0.1225212812423706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04958852007985115, 'sampling/sampling_logp_difference/max': 2.24063777923584, 'sampling/importance_sampling_ratio/min': 0.10639062523841858, 'sampling/importance_sampling_ratio/mean': 1.0073775053024292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2784906532615423, 'clip_ratio/low_mean': 0.01997046614997089, 'clip_ratio/low_min': 0.01997046614997089, 'clip_ratio/high_mean': 0.018199163547251374, 'clip_ratio/high_max': 0.018199163547251374, 'clip_ratio/region_mean': 0.03816962969722226, 'reward_total_mean': 0.8550522923469543, 'reward_meter_mean': 0.9935377836227417, 'reward_meter_std': 0.007430072408169508, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8857142925262451, 'reward_repeat_penalty_std': 0.07324227690696716, 'reward_total_composite_mean': 0.8550522923469543, 'reward_total_composite_std': 0.1225212812423706, 'epoch': 0.04}

 31%|███▏      | 1037/3300 [49:36<3:23:07,  5.39s/it]INFO 04-11 23:26:27 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1038/3300 [49:40<3:11:01,  5.07s/it]2026-04-11 23:26:32,002 | INFO | train_grpo_train | metrics_logged mode=train step=1038

                                                     
{'loss': 0.0376, 'grad_norm': 9.27010726928711, 'learning_rate': 6.857575757575758e-06, 'num_tokens': 2328768.0, 'completions/mean_length': 36.375, 'completions/min_length': 33.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.24713511765003204, 'rewards/meter/std': 0.3414146900177002, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.24713511765003204, 'rewards/total_composite/std': 0.3414146900177002, 'reward': 0.24713511765003204, 'reward_std': 0.3414146900177002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09440852701663971, 'sampling/sampling_logp_difference/max': 1.0471820831298828, 'sampling/importance_sampling_ratio/min': 0.3509252369403839, 'sampling/importance_sampling_ratio/mean': 1.0138072967529297, 'sampling/importance_sampling_ratio/max': 1.7845203876495361, 'entropy': 0.8697552978992462, 'clip_ratio/low_mean': 0.05519176600500941, 'clip_ratio/low_min': 0.05519176600500941, 'clip_ratio/high_mean': 0.017664092825725675, 'clip_ratio/high_max': 0.017664092825725675, 'clip_ratio/region_mean': 0.07285585883073509, 'reward_total_mean': 0.24713511765003204, 'reward_meter_mean': 0.24713511765003204, 'reward_meter_std': 0.3414146900177002, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.24713511765003204, 'reward_total_composite_std': 0.3414146900177002, 'epoch': 0.04}

 31%|███▏      | 1038/3300 [49:40<3:11:01,  5.07s/it]INFO 04-11 23:26:32 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1039/3300 [49:45<3:06:52,  4.96s/it]2026-04-11 23:26:36,712 | INFO | train_grpo_train | metrics_logged mode=train step=1039

                                                     
{'loss': 0.0252, 'grad_norm': 4.874062538146973, 'learning_rate': 6.854545454545455e-06, 'num_tokens': 2330669.0, 'completions/mean_length': 69.625, 'completions/min_length': 67.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.19316673278808594, 'rewards/meter/std': 0.19894513487815857, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.19316673278808594, 'rewards/total_composite/std': 0.19894513487815857, 'reward': 0.19316673278808594, 'reward_std': 0.19894511997699738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0636194571852684, 'sampling/sampling_logp_difference/max': 1.560652732849121, 'sampling/importance_sampling_ratio/min': 0.2099989503622055, 'sampling/importance_sampling_ratio/mean': 1.0039423704147339, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47740989923477173, 'clip_ratio/low_mean': 0.03564694127999246, 'clip_ratio/low_min': 0.03564694127999246, 'clip_ratio/high_mean': 0.024171422701328993, 'clip_ratio/high_max': 0.024171422701328993, 'clip_ratio/region_mean': 0.059818363981321454, 'reward_total_mean': 0.19316673278808594, 'reward_meter_mean': 0.19316673278808594, 'reward_meter_std': 0.19894513487815857, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.19316673278808594, 'reward_total_composite_std': 0.19894513487815857, 'epoch': 0.04}

 31%|███▏      | 1039/3300 [49:45<3:06:52,  4.96s/it]INFO 04-11 23:26:36 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1040/3300 [49:51<3:18:45,  5.28s/it]2026-04-11 23:26:42,729 | INFO | train_grpo_train | metrics_logged mode=train step=1040

                                                     
{'loss': -0.0641, 'grad_norm': 5.022459030151367, 'learning_rate': 6.851515151515153e-06, 'num_tokens': 2333405.0, 'completions/mean_length': 160.0, 'completions/min_length': 131.0, 'completions/max_length': 182.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.0, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.9006245136260986, 'rewards/meter/std': 0.21294330060482025, 'rewards/count_adherence/mean': 0.925000011920929, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7440476417541504, 'rewards/repeat_penalty/std': 0.17643596231937408, 'rewards/total_composite/mean': 0.6102761030197144, 'rewards/total_composite/std': 0.19687247276306152, 'reward': 0.6102761030197144, 'reward_std': 0.19687248766422272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05596721172332764, 'sampling/sampling_logp_difference/max': 1.3793213367462158, 'sampling/importance_sampling_ratio/min': 0.25236961245536804, 'sampling/importance_sampling_ratio/mean': 1.0185436010360718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39750632736831903, 'clip_ratio/low_mean': 0.016255452763289213, 'clip_ratio/low_min': 0.016255452763289213, 'clip_ratio/high_mean': 0.025824516778811812, 'clip_ratio/high_max': 0.025824516778811812, 'clip_ratio/region_mean': 0.042079969542101026, 'reward_total_mean': 0.6102761030197144, 'reward_meter_mean': 0.9006245136260986, 'reward_meter_std': 0.21294330060482025, 'reward_count_adherence_mean': 0.925000011920929, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7440476417541504, 'reward_repeat_penalty_std': 0.17643596231937408, 'reward_total_composite_mean': 0.6102761030197144, 'reward_total_composite_std': 0.19687247276306152, 'epoch': 0.04}

 32%|███▏      | 1040/3300 [49:51<3:18:45,  5.28s/it]INFO 04-11 23:26:42 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1041/3300 [49:55<3:06:35,  4.96s/it]2026-04-11 23:26:46,939 | INFO | train_grpo_train | metrics_logged mode=train step=1041

                                                     
{'loss': 0.0176, 'grad_norm': 16.1435604095459, 'learning_rate': 6.848484848484849e-06, 'num_tokens': 2334958.0, 'completions/mean_length': 40.125, 'completions/min_length': 39.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9823716878890991, 'rewards/meter/std': 0.015363764949142933, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9823716878890991, 'rewards/total_composite/std': 0.015363764949142933, 'reward': 0.9823716878890991, 'reward_std': 0.015363766811788082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03611617907881737, 'sampling/sampling_logp_difference/max': 1.3837116956710815, 'sampling/importance_sampling_ratio/min': 0.2506465017795563, 'sampling/importance_sampling_ratio/mean': 1.0097277164459229, 'sampling/importance_sampling_ratio/max': 1.6625901460647583, 'entropy': 0.2032633163034916, 'clip_ratio/low_mean': 0.0031250000465661287, 'clip_ratio/low_min': 0.0031250000465661287, 'clip_ratio/high_mean': 0.012427689041942358, 'clip_ratio/high_max': 0.012427689041942358, 'clip_ratio/region_mean': 0.015552689088508487, 'reward_total_mean': 0.9823716878890991, 'reward_meter_mean': 0.9823716878890991, 'reward_meter_std': 0.015363764949142933, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9823716878890991, 'reward_total_composite_std': 0.015363764949142933, 'epoch': 0.04}

 32%|███▏      | 1041/3300 [49:55<3:06:35,  4.96s/it]INFO 04-11 23:26:47 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1042/3300 [50:01<3:12:54,  5.13s/it]2026-04-11 23:26:52,468 | INFO | train_grpo_train | metrics_logged mode=train step=1042

                                                     
{'loss': 0.0581, 'grad_norm': 3.8245151042938232, 'learning_rate': 6.845454545454546e-06, 'num_tokens': 2337338.0, 'completions/mean_length': 126.5, 'completions/min_length': 115.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.5, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.7728732824325562, 'rewards/meter/std': 0.3359214961528778, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.10101525485515594, 'rewards/total_composite/mean': 0.5652507543563843, 'rewards/total_composite/std': 0.22422268986701965, 'reward': 0.5652507543563843, 'reward_std': 0.22422267496585846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031546179205179214, 'sampling/sampling_logp_difference/max': 1.852052092552185, 'sampling/importance_sampling_ratio/min': 0.1569148302078247, 'sampling/importance_sampling_ratio/mean': 1.0046995878219604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1586785214021802, 'clip_ratio/low_mean': 0.007275132229551673, 'clip_ratio/low_min': 0.007275132229551673, 'clip_ratio/high_mean': 0.023097628843970597, 'clip_ratio/high_max': 0.023097628843970597, 'clip_ratio/region_mean': 0.03037276107352227, 'reward_total_mean': 0.5652507543563843, 'reward_meter_mean': 0.7728732824325562, 'reward_meter_std': 0.3359214961528778, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.10101525485515594, 'reward_total_composite_mean': 0.5652507543563843, 'reward_total_composite_std': 0.22422268986701965, 'epoch': 0.04}

 32%|███▏      | 1042/3300 [50:01<3:12:54,  5.13s/it]INFO 04-11 23:26:52 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1043/3300 [50:06<3:07:51,  4.99s/it]2026-04-11 23:26:57,146 | INFO | train_grpo_train | metrics_logged mode=train step=1043

                                                     
{'loss': 0.0128, 'grad_norm': 5.958546161651611, 'learning_rate': 6.842424242424243e-06, 'num_tokens': 2339119.0, 'completions/mean_length': 67.625, 'completions/min_length': 63.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.5403692722320557, 'rewards/meter/std': 0.4386324882507324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5403692722320557, 'rewards/total_composite/std': 0.4386324882507324, 'reward': 0.5403692722320557, 'reward_std': 0.43863245844841003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047015149146318436, 'sampling/sampling_logp_difference/max': 1.6448850631713867, 'sampling/importance_sampling_ratio/min': 0.19303473830223083, 'sampling/importance_sampling_ratio/mean': 1.0055934190750122, 'sampling/importance_sampling_ratio/max': 1.7478914260864258, 'entropy': 0.3001182395964861, 'clip_ratio/low_mean': 0.02028265898115933, 'clip_ratio/low_min': 0.02028265898115933, 'clip_ratio/high_mean': 0.018411532044410706, 'clip_ratio/high_max': 0.018411532044410706, 'clip_ratio/region_mean': 0.038694191025570035, 'reward_total_mean': 0.5403692722320557, 'reward_meter_mean': 0.5403692722320557, 'reward_meter_std': 0.4386324882507324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5403692722320557, 'reward_total_composite_std': 0.4386324882507324, 'epoch': 0.04}

 32%|███▏      | 1043/3300 [50:06<3:07:51,  4.99s/it]INFO 04-11 23:26:57 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1044/3300 [50:11<3:09:57,  5.05s/it]2026-04-11 23:27:02,338 | INFO | train_grpo_train | metrics_logged mode=train step=1044

                                                     
{'loss': -0.0019, 'grad_norm': 3.107640266418457, 'learning_rate': 6.83939393939394e-06, 'num_tokens': 2341333.0, 'completions/mean_length': 98.75, 'completions/min_length': 93.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.75, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.4054219126701355, 'rewards/meter/std': 0.4075765311717987, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4054219126701355, 'rewards/total_composite/std': 0.4075765311717987, 'reward': 0.4054219126701355, 'reward_std': 0.4075765013694763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031202714890241623, 'sampling/sampling_logp_difference/max': 0.8987345695495605, 'sampling/importance_sampling_ratio/min': 0.40708449482917786, 'sampling/importance_sampling_ratio/mean': 1.0041404962539673, 'sampling/importance_sampling_ratio/max': 1.5978652238845825, 'entropy': 0.1880979547277093, 'clip_ratio/low_mean': 0.020476194215007126, 'clip_ratio/low_min': 0.020476194215007126, 'clip_ratio/high_mean': 0.006274061510339379, 'clip_ratio/high_max': 0.006274061510339379, 'clip_ratio/region_mean': 0.026750255725346506, 'reward_total_mean': 0.4054219126701355, 'reward_meter_mean': 0.4054219126701355, 'reward_meter_std': 0.4075765311717987, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4054219126701355, 'reward_total_composite_std': 0.4075765311717987, 'epoch': 0.04}

 32%|███▏      | 1044/3300 [50:11<3:09:57,  5.05s/it]INFO 04-11 23:27:02 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1045/3300 [50:16<3:14:16,  5.17s/it]2026-04-11 23:27:07,775 | INFO | train_grpo_train | metrics_logged mode=train step=1045

                                                     
{'loss': -0.0095, 'grad_norm': 6.213621616363525, 'learning_rate': 6.8363636363636364e-06, 'num_tokens': 2343378.0, 'completions/mean_length': 93.625, 'completions/min_length': 89.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.625, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.745398998260498, 'rewards/meter/std': 0.21442271769046783, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.745398998260498, 'rewards/total_composite/std': 0.21442271769046783, 'reward': 0.745398998260498, 'reward_std': 0.21442271769046783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036333851516246796, 'sampling/sampling_logp_difference/max': 1.0359835624694824, 'sampling/importance_sampling_ratio/min': 0.35487717390060425, 'sampling/importance_sampling_ratio/mean': 1.0055217742919922, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20175535418093204, 'clip_ratio/low_mean': 0.015083948150277138, 'clip_ratio/low_min': 0.015083948150277138, 'clip_ratio/high_mean': 0.021053530042991042, 'clip_ratio/high_max': 0.021053530042991042, 'clip_ratio/region_mean': 0.03613747819326818, 'reward_total_mean': 0.745398998260498, 'reward_meter_mean': 0.745398998260498, 'reward_meter_std': 0.21442271769046783, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.745398998260498, 'reward_total_composite_std': 0.21442271769046783, 'epoch': 0.04}

 32%|███▏      | 1045/3300 [50:16<3:14:16,  5.17s/it]INFO 04-11 23:27:08 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1046/3300 [50:21<3:08:31,  5.02s/it]2026-04-11 23:27:12,443 | INFO | train_grpo_train | metrics_logged mode=train step=1046

                                                     
{'loss': 0.0232, 'grad_norm': 8.28740406036377, 'learning_rate': 6.833333333333334e-06, 'num_tokens': 2345073.0, 'completions/mean_length': 68.875, 'completions/min_length': 66.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5793697834014893, 'rewards/meter/std': 0.473821759223938, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5793697834014893, 'rewards/total_composite/std': 0.473821759223938, 'reward': 0.5793697834014893, 'reward_std': 0.473821759223938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058582451194524765, 'sampling/sampling_logp_difference/max': 1.9549522399902344, 'sampling/importance_sampling_ratio/min': 0.14157122373580933, 'sampling/importance_sampling_ratio/mean': 1.0050064325332642, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3025083411484957, 'clip_ratio/low_mean': 0.026830808725208044, 'clip_ratio/low_min': 0.026830808725208044, 'clip_ratio/high_mean': 0.031451608054339886, 'clip_ratio/high_max': 0.031451608054339886, 'clip_ratio/region_mean': 0.05828241677954793, 'reward_total_mean': 0.5793697834014893, 'reward_meter_mean': 0.5793697834014893, 'reward_meter_std': 0.473821759223938, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5793697834014893, 'reward_total_composite_std': 0.473821759223938, 'epoch': 0.04}

 32%|███▏      | 1046/3300 [50:21<3:08:31,  5.02s/it]INFO 04-11 23:27:12 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1047/3300 [50:31<4:00:18,  6.40s/it]2026-04-11 23:27:22,064 | INFO | train_grpo_train | metrics_logged mode=train step=1047

                                                     
{'loss': -0.2118, 'grad_norm': 1.4744713306427002, 'learning_rate': 6.83030303030303e-06, 'num_tokens': 2347417.0, 'completions/mean_length': 181.0, 'completions/min_length': 128.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 133.71429443359375, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.9518733024597168, 'rewards/meter/std': 0.08209281414747238, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.7358262538909912, 'rewards/total_composite/std': 0.31345105171203613, 'reward': 0.7358262538909912, 'reward_std': 0.31345105171203613, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05928022786974907, 'sampling/sampling_logp_difference/max': 1.9391239881515503, 'sampling/importance_sampling_ratio/min': 0.14382989704608917, 'sampling/importance_sampling_ratio/mean': 1.0017281770706177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30693456530570984, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0447791040642187, 'clip_ratio/high_max': 0.0447791040642187, 'clip_ratio/region_mean': 0.0447791040642187, 'reward_total_mean': 0.7358262538909912, 'reward_meter_mean': 0.9518733024597168, 'reward_meter_std': 0.08209281414747238, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.7358262538909912, 'reward_total_composite_std': 0.31345105171203613, 'epoch': 0.04}

 32%|███▏      | 1047/3300 [50:31<4:00:18,  6.40s/it]INFO 04-11 23:27:22 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1048/3300 [50:35<3:41:35,  5.90s/it]2026-04-11 23:27:26,811 | INFO | train_grpo_train | metrics_logged mode=train step=1048

                                                     
{'loss': -0.0116, 'grad_norm': 8.561767578125, 'learning_rate': 6.827272727272728e-06, 'num_tokens': 2349002.0, 'completions/mean_length': 47.125, 'completions/min_length': 44.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6955418586730957, 'rewards/meter/std': 0.4201207160949707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6955418586730957, 'rewards/total_composite/std': 0.4201207160949707, 'reward': 0.6955418586730957, 'reward_std': 0.4201207160949707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051052458584308624, 'sampling/sampling_logp_difference/max': 2.202850818634033, 'sampling/importance_sampling_ratio/min': 0.11048772186040878, 'sampling/importance_sampling_ratio/mean': 1.009395718574524, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.374115826562047, 'clip_ratio/low_mean': 0.014204545877873898, 'clip_ratio/low_min': 0.014204545877873898, 'clip_ratio/high_mean': 0.026004510698840022, 'clip_ratio/high_max': 0.026004510698840022, 'clip_ratio/region_mean': 0.04020905657671392, 'reward_total_mean': 0.6955418586730957, 'reward_meter_mean': 0.6955418586730957, 'reward_meter_std': 0.4201207160949707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6955418586730957, 'reward_total_composite_std': 0.4201207160949707, 'epoch': 0.04}

 32%|███▏      | 1048/3300 [50:35<3:41:35,  5.90s/it]INFO 04-11 23:27:27 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1049/3300 [50:41<3:37:23,  5.79s/it]2026-04-11 23:27:32,348 | INFO | train_grpo_train | metrics_logged mode=train step=1049

                                                     
{'loss': 0.0138, 'grad_norm': 8.349677085876465, 'learning_rate': 6.824242424242425e-06, 'num_tokens': 2351424.0, 'completions/mean_length': 144.75, 'completions/min_length': 139.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 144.75, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.716080904006958, 'rewards/meter/std': 0.3239102363586426, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.6298381090164185, 'rewards/total_composite/std': 0.2847732603549957, 'reward': 0.6298381090164185, 'reward_std': 0.2847732603549957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020672447979450226, 'sampling/sampling_logp_difference/max': 1.2336063385009766, 'sampling/importance_sampling_ratio/min': 0.2912403643131256, 'sampling/importance_sampling_ratio/mean': 1.0015819072723389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1037906464189291, 'clip_ratio/low_mean': 0.012227151426486671, 'clip_ratio/low_min': 0.012227151426486671, 'clip_ratio/high_mean': 0.012093263794668019, 'clip_ratio/high_max': 0.012093263794668019, 'clip_ratio/region_mean': 0.02432041522115469, 'reward_total_mean': 0.6298381090164185, 'reward_meter_mean': 0.716080904006958, 'reward_meter_std': 0.3239102363586426, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.6298381090164185, 'reward_total_composite_std': 0.2847732603549957, 'epoch': 0.04}

 32%|███▏      | 1049/3300 [50:41<3:37:23,  5.79s/it]INFO 04-11 23:27:32 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1050/3300 [50:46<3:26:42,  5.51s/it]2026-04-11 23:27:37,202 | INFO | train_grpo_train | metrics_logged mode=train step=1050

                                                     
{'loss': 0.0067, 'grad_norm': 9.696529388427734, 'learning_rate': 6.821212121212122e-06, 'num_tokens': 2353303.0, 'completions/mean_length': 72.875, 'completions/min_length': 70.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.7290558815002441, 'rewards/meter/std': 0.3074507713317871, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7290558815002441, 'rewards/total_composite/std': 0.3074507713317871, 'reward': 0.7290558815002441, 'reward_std': 0.3074507415294647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07516960054636002, 'sampling/sampling_logp_difference/max': 3.8051562309265137, 'sampling/importance_sampling_ratio/min': 0.02225572057068348, 'sampling/importance_sampling_ratio/mean': 0.9999740719795227, 'sampling/importance_sampling_ratio/max': 1.8488932847976685, 'entropy': 0.3899807743728161, 'clip_ratio/low_mean': 0.008656773250550032, 'clip_ratio/low_min': 0.008656773250550032, 'clip_ratio/high_mean': 0.020683051785454154, 'clip_ratio/high_max': 0.020683051785454154, 'clip_ratio/region_mean': 0.029339825036004186, 'reward_total_mean': 0.7290558815002441, 'reward_meter_mean': 0.7290558815002441, 'reward_meter_std': 0.3074507713317871, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7290558815002441, 'reward_total_composite_std': 0.3074507713317871, 'epoch': 0.04}

 32%|███▏      | 1050/3300 [50:46<3:26:42,  5.51s/it]INFO 04-11 23:27:37 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:28,  2.55s/it]

 23%|██▎       | 3/13 [00:08<00:30,  3.04s/it]

 31%|███       | 4/13 [00:14<00:36,  4.10s/it]

 38%|███▊      | 5/13 [00:20<00:36,  4.53s/it]

 46%|████▌     | 6/13 [00:26<00:35,  5.13s/it]

 54%|█████▍    | 7/13 [00:31<00:31,  5.25s/it]

 62%|██████▏   | 8/13 [00:38<00:28,  5.67s/it]

 69%|██████▉   | 9/13 [00:45<00:23,  5.95s/it]

 77%|███████▋  | 10/13 [00:50<00:17,  5.67s/it]

 85%|████████▍ | 11/13 [00:55<00:11,  5.60s/it]

 92%|█████████▏| 12/13 [01:01<00:05,  5.74s/it]

100%|██████████| 13/13 [01:08<00:00,  6.17s/it]2026-04-11 23:28:52,133 | INFO | train_grpo_train | metrics_logged mode=eval step=1050

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 74.9052, 'eval_samples_per_second': 1.388, 'eval_steps_per_second': 0.174, 'eval_num_tokens': 2353303.0, 'eval_completions/mean_length': 226.30769230769232, 'eval_completions/min_length': 73.84615384615384, 'eval_completions/max_length': 395.0769230769231, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 223.2651108961839, 'eval_completions/min_terminated_length': 73.84615384615384, 'eval_completions/max_terminated_length': 383.3076923076923, 'eval_rewards/meter/mean': 0.5502180938537304, 'eval_rewards/meter/std': 0.40295018599583554, 'eval_rewards/count_adherence/mean': 0.8679342178198007, 'eval_rewards/count_adherence/std': 0.11943456530570984, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.724188873401055, 'eval_rewards/repeat_penalty/std': 0.2071983149418464, 'eval_rewards/total_composite/mean': 0.33965545204969555, 'eval_rewards/total_composite/std': 0.3024227091899285, 'eval_reward': 0.33965545204969555, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.015186995840989627, 'eval_sampling/sampling_logp_difference/max': 1.0244657076322115, 'eval_sampling/importance_sampling_ratio/min': 0.3627954079554631, 'eval_sampling/importance_sampling_ratio/mean': 1.0033980562136724, 'eval_sampling/importance_sampling_ratio/max': 1.4160706263322096, 'eval_entropy': 0.13593376485201028, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.33965545204969555, 'eval_reward_meter_mean': 0.5502180938537304, 'eval_reward_meter_std': 0.40295018599583554, 'eval_reward_count_adherence_mean': 0.8679342178198007, 'eval_reward_count_adherence_std': 0.11943456530570984, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.724188873401055, 'eval_reward_repeat_penalty_std': 0.2071983149418464, 'eval_reward_total_composite_mean': 0.33965545204969555, 'eval_reward_total_composite_std': 0.3024227091899285, 'epoch': 0.04}

 32%|███▏      | 1050/3300 [52:01<3:26:42,  5.51s/it]

100%|██████████| 13/13 [01:08<00:00,  6.17s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 32%|███▏      | 1051/3300 [52:10<18:13:34, 29.17s/it]2026-04-11 23:29:01,593 | INFO | train_grpo_train | metrics_logged mode=train step=1051

                                                      
{'loss': 0.0139, 'grad_norm': 2.1154820919036865, 'learning_rate': 6.818181818181818e-06, 'num_tokens': 2356570.0, 'completions/mean_length': 212.375, 'completions/min_length': 183.0, 'completions/max_length': 235.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 212.375, 'completions/min_terminated_length': 183.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.802075982093811, 'rewards/meter/std': 0.3460130989551544, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7899305820465088, 'rewards/repeat_penalty/std': 0.090753473341465, 'rewards/total_composite/mean': 0.5123095512390137, 'rewards/total_composite/std': 0.2536679208278656, 'reward': 0.5123095512390137, 'reward_std': 0.2536678910255432, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038277287036180496, 'sampling/sampling_logp_difference/max': 1.5595903396606445, 'sampling/importance_sampling_ratio/min': 0.21022215485572815, 'sampling/importance_sampling_ratio/mean': 1.0035831928253174, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31745208986103535, 'clip_ratio/low_mean': 0.005197861581109464, 'clip_ratio/low_min': 0.005197861581109464, 'clip_ratio/high_mean': 0.026332546956837177, 'clip_ratio/high_max': 0.026332546956837177, 'clip_ratio/region_mean': 0.03153040853794664, 'reward_total_mean': 0.5123095512390137, 'reward_meter_mean': 0.802075982093811, 'reward_meter_std': 0.3460130989551544, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7899305820465088, 'reward_repeat_penalty_std': 0.090753473341465, 'reward_total_composite_mean': 0.5123095512390137, 'reward_total_composite_std': 0.2536679208278656, 'epoch': 0.04}

 32%|███▏      | 1051/3300 [52:10<18:13:34, 29.17s/it]INFO 04-11 23:29:01 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1052/3300 [52:15<13:42:09, 21.94s/it]2026-04-11 23:29:06,661 | INFO | train_grpo_train | metrics_logged mode=train step=1052

                                                      
{'loss': -0.0141, 'grad_norm': 4.994956970214844, 'learning_rate': 6.8151515151515155e-06, 'num_tokens': 2358608.0, 'completions/mean_length': 89.75, 'completions/min_length': 84.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.75, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9309325218200684, 'rewards/meter/std': 0.17771661281585693, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9309325218200684, 'rewards/total_composite/std': 0.17771661281585693, 'reward': 0.9309325218200684, 'reward_std': 0.17771658301353455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07444296777248383, 'sampling/sampling_logp_difference/max': 1.3812413215637207, 'sampling/importance_sampling_ratio/min': 0.2512664496898651, 'sampling/importance_sampling_ratio/mean': 1.0161553621292114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.586686696857214, 'clip_ratio/low_mean': 0.0028735632076859474, 'clip_ratio/low_min': 0.0028735632076859474, 'clip_ratio/high_mean': 0.05841635470278561, 'clip_ratio/high_max': 0.05841635470278561, 'clip_ratio/region_mean': 0.06128991791047156, 'reward_total_mean': 0.9309325218200684, 'reward_meter_mean': 0.9309325218200684, 'reward_meter_std': 0.17771661281585693, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9309325218200684, 'reward_total_composite_std': 0.17771661281585693, 'epoch': 0.04}

 32%|███▏      | 1052/3300 [52:15<13:42:09, 21.94s/it]INFO 04-11 23:29:06 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1053/3300 [52:21<10:37:13, 17.02s/it]2026-04-11 23:29:12,178 | INFO | train_grpo_train | metrics_logged mode=train step=1053

                                                      
{'loss': 0.0063, 'grad_norm': 4.2257232666015625, 'learning_rate': 6.812121212121212e-06, 'num_tokens': 2360466.0, 'completions/mean_length': 69.25, 'completions/min_length': 67.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9586122035980225, 'rewards/meter/std': 0.00875465851277113, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9586122035980225, 'rewards/total_composite/std': 0.00875465851277113, 'reward': 0.9586122035980225, 'reward_std': 0.00875465851277113, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020308073610067368, 'sampling/sampling_logp_difference/max': 1.0632991790771484, 'sampling/importance_sampling_ratio/min': 0.3453146815299988, 'sampling/importance_sampling_ratio/mean': 0.9992114901542664, 'sampling/importance_sampling_ratio/max': 1.3229237794876099, 'entropy': 0.13288897648453712, 'clip_ratio/low_mean': 0.01263241225387901, 'clip_ratio/low_min': 0.01263241225387901, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.016363755450583994, 'reward_total_mean': 0.9586122035980225, 'reward_meter_mean': 0.9586122035980225, 'reward_meter_std': 0.00875465851277113, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9586122035980225, 'reward_total_composite_std': 0.00875465851277113, 'epoch': 0.04}

 32%|███▏      | 1053/3300 [52:21<10:37:13, 17.02s/it]INFO 04-11 23:29:12 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1054/3300 [52:26<8:27:30, 13.56s/it] 2026-04-11 23:29:17,666 | INFO | train_grpo_train | metrics_logged mode=train step=1054

                                                     
{'loss': 0.0069, 'grad_norm': 2.932860851287842, 'learning_rate': 6.80909090909091e-06, 'num_tokens': 2362481.0, 'completions/mean_length': 96.875, 'completions/min_length': 93.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.875, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9400673508644104, 'rewards/meter/std': 0.03255482763051987, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9002121686935425, 'rewards/total_composite/std': 0.11076170206069946, 'reward': 0.9002121686935425, 'reward_std': 0.11076170206069946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0384301133453846, 'sampling/sampling_logp_difference/max': 1.6087217330932617, 'sampling/importance_sampling_ratio/min': 0.20014327764511108, 'sampling/importance_sampling_ratio/mean': 0.9991159439086914, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17869562469422817, 'clip_ratio/low_mean': 0.008994319243356586, 'clip_ratio/low_min': 0.008994319243356586, 'clip_ratio/high_mean': 0.02460848237387836, 'clip_ratio/high_max': 0.02460848237387836, 'clip_ratio/region_mean': 0.033602801617234945, 'reward_total_mean': 0.9002121686935425, 'reward_meter_mean': 0.9400673508644104, 'reward_meter_std': 0.03255482763051987, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9002121686935425, 'reward_total_composite_std': 0.11076170206069946, 'epoch': 0.04}

 32%|███▏      | 1054/3300 [52:26<8:27:30, 13.56s/it]INFO 04-11 23:29:17 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1055/3300 [52:32<7:02:46, 11.30s/it]2026-04-11 23:29:23,697 | INFO | train_grpo_train | metrics_logged mode=train step=1055

                                                     
{'loss': 0.1006, 'grad_norm': 3.7541143894195557, 'learning_rate': 6.806060606060607e-06, 'num_tokens': 2364800.0, 'completions/mean_length': 137.875, 'completions/min_length': 98.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.875, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.25745514035224915, 'rewards/meter/std': 0.2247081845998764, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6964285969734192, 'rewards/repeat_penalty/std': 0.2466983199119568, 'rewards/total_composite/mean': 0.14893272519111633, 'rewards/total_composite/std': 0.16499118506908417, 'reward': 0.14893272519111633, 'reward_std': 0.16499118506908417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05157129094004631, 'sampling/sampling_logp_difference/max': 1.8216896057128906, 'sampling/importance_sampling_ratio/min': 0.16175222396850586, 'sampling/importance_sampling_ratio/mean': 1.0185996294021606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4246663171797991, 'clip_ratio/low_mean': 0.014537818904500455, 'clip_ratio/low_min': 0.014537818904500455, 'clip_ratio/high_mean': 0.01200975279789418, 'clip_ratio/high_max': 0.01200975279789418, 'clip_ratio/region_mean': 0.026547571702394634, 'reward_total_mean': 0.14893272519111633, 'reward_meter_mean': 0.25745514035224915, 'reward_meter_std': 0.2247081845998764, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6964285969734192, 'reward_repeat_penalty_std': 0.2466983199119568, 'reward_total_composite_mean': 0.14893272519111633, 'reward_total_composite_std': 0.16499118506908417, 'epoch': 0.04}

 32%|███▏      | 1055/3300 [52:32<7:02:46, 11.30s/it]INFO 04-11 23:29:23 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1056/3300 [52:37<5:50:33,  9.37s/it]2026-04-11 23:29:28,576 | INFO | train_grpo_train | metrics_logged mode=train step=1056

                                                     
{'loss': 0.0447, 'grad_norm': 12.402931213378906, 'learning_rate': 6.803030303030304e-06, 'num_tokens': 2366750.0, 'completions/mean_length': 70.75, 'completions/min_length': 68.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.75, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.30388349294662476, 'rewards/meter/std': 0.2288217395544052, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.30388349294662476, 'rewards/total_composite/std': 0.2288217395544052, 'reward': 0.30388349294662476, 'reward_std': 0.22882172465324402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05324329808354378, 'sampling/sampling_logp_difference/max': 2.614670753479004, 'sampling/importance_sampling_ratio/min': 0.07319188117980957, 'sampling/importance_sampling_ratio/mean': 0.9882052540779114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17993433214724064, 'clip_ratio/low_mean': 0.019135613925755024, 'clip_ratio/low_min': 0.019135613925755024, 'clip_ratio/high_mean': 0.02171629574149847, 'clip_ratio/high_max': 0.02171629574149847, 'clip_ratio/region_mean': 0.040851909667253494, 'reward_total_mean': 0.30388349294662476, 'reward_meter_mean': 0.30388349294662476, 'reward_meter_std': 0.2288217395544052, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.30388349294662476, 'reward_total_composite_std': 0.2288217395544052, 'epoch': 0.04}

 32%|███▏      | 1056/3300 [52:37<5:50:33,  9.37s/it]INFO 04-11 23:29:28 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1057/3300 [52:46<5:40:47,  9.12s/it]2026-04-11 23:29:37,092 | INFO | train_grpo_train | metrics_logged mode=train step=1057

                                                     
{'loss': 0.0021, 'grad_norm': 1.8229259252548218, 'learning_rate': 6.800000000000001e-06, 'num_tokens': 2371070.0, 'completions/mean_length': 349.0, 'completions/min_length': 337.0, 'completions/max_length': 377.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 349.0, 'completions/min_terminated_length': 337.0, 'completions/max_terminated_length': 377.0, 'rewards/meter/mean': 0.3180350661277771, 'rewards/meter/std': 0.4220852255821228, 'rewards/count_adherence/mean': 0.762499988079071, 'rewards/count_adherence/std': 0.05175492912530899, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6007440090179443, 'rewards/repeat_penalty/std': 0.11201495677232742, 'rewards/total_composite/mean': 0.14668160676956177, 'rewards/total_composite/std': 0.19722160696983337, 'reward': 0.14668160676956177, 'reward_std': 0.197221577167511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020703818649053574, 'sampling/sampling_logp_difference/max': 1.6899352073669434, 'sampling/importance_sampling_ratio/min': 0.18453148007392883, 'sampling/importance_sampling_ratio/mean': 1.001380205154419, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14950765296816826, 'clip_ratio/low_mean': 0.013276972225867212, 'clip_ratio/low_min': 0.013276972225867212, 'clip_ratio/high_mean': 0.0039429860189557076, 'clip_ratio/high_max': 0.0039429860189557076, 'clip_ratio/region_mean': 0.01721995824482292, 'reward_total_mean': 0.14668160676956177, 'reward_meter_mean': 0.3180350661277771, 'reward_meter_std': 0.4220852255821228, 'reward_count_adherence_mean': 0.762499988079071, 'reward_count_adherence_std': 0.05175492912530899, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6007440090179443, 'reward_repeat_penalty_std': 0.11201495677232742, 'reward_total_composite_mean': 0.14668160676956177, 'reward_total_composite_std': 0.19722160696983337, 'epoch': 0.04}

 32%|███▏      | 1057/3300 [52:46<5:40:47,  9.12s/it]INFO 04-11 23:29:37 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1058/3300 [52:54<5:31:27,  8.87s/it]2026-04-11 23:29:45,389 | INFO | train_grpo_train | metrics_logged mode=train step=1058

                                                     
{'loss': -0.0193, 'grad_norm': 2.157697916030884, 'learning_rate': 6.796969696969697e-06, 'num_tokens': 2375539.0, 'completions/mean_length': 347.625, 'completions/min_length': 326.0, 'completions/max_length': 367.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 347.625, 'completions/min_terminated_length': 326.0, 'completions/max_terminated_length': 367.0, 'rewards/meter/mean': 0.7610747814178467, 'rewards/meter/std': 0.25479352474212646, 'rewards/count_adherence/mean': 0.7222222089767456, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4871794581413269, 'rewards/repeat_penalty/std': 0.16510966420173645, 'rewards/total_composite/mean': 0.26854199171066284, 'rewards/total_composite/std': 0.14636212587356567, 'reward': 0.26854199171066284, 'reward_std': 0.14636212587356567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017443938180804253, 'sampling/sampling_logp_difference/max': 2.1108410358428955, 'sampling/importance_sampling_ratio/min': 0.12113604694604874, 'sampling/importance_sampling_ratio/mean': 1.0041615962982178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07590485410764813, 'clip_ratio/low_mean': 0.008206001890357584, 'clip_ratio/low_min': 0.008206001890357584, 'clip_ratio/high_mean': 0.0041363348718732595, 'clip_ratio/high_max': 0.0041363348718732595, 'clip_ratio/region_mean': 0.012342336762230843, 'reward_total_mean': 0.26854199171066284, 'reward_meter_mean': 0.7610747814178467, 'reward_meter_std': 0.25479352474212646, 'reward_count_adherence_mean': 0.7222222089767456, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4871794581413269, 'reward_repeat_penalty_std': 0.16510966420173645, 'reward_total_composite_mean': 0.26854199171066284, 'reward_total_composite_std': 0.14636212587356567, 'epoch': 0.04}

 32%|███▏      | 1058/3300 [52:54<5:31:27,  8.87s/it]INFO 04-11 23:29:45 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1059/3300 [52:59<4:46:19,  7.67s/it]2026-04-11 23:29:50,243 | INFO | train_grpo_train | metrics_logged mode=train step=1059

                                                     
{'loss': 0.0103, 'grad_norm': 3.8445513248443604, 'learning_rate': 6.793939393939395e-06, 'num_tokens': 2377463.0, 'completions/mean_length': 73.5, 'completions/min_length': 71.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.1600387543439865, 'rewards/meter/std': 0.3069753348827362, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.1600387543439865, 'rewards/total_composite/std': 0.3069753348827362, 'reward': 0.1600387543439865, 'reward_std': 0.3069753646850586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0510094054043293, 'sampling/sampling_logp_difference/max': 1.574995994567871, 'sampling/importance_sampling_ratio/min': 0.20700837671756744, 'sampling/importance_sampling_ratio/mean': 1.008049726486206, 'sampling/importance_sampling_ratio/max': 1.8910245895385742, 'entropy': 0.32050950825214386, 'clip_ratio/low_mean': 0.02205127221532166, 'clip_ratio/low_min': 0.02205127221532166, 'clip_ratio/high_mean': 0.006493506487458944, 'clip_ratio/high_max': 0.006493506487458944, 'clip_ratio/region_mean': 0.028544778702780604, 'reward_total_mean': 0.1600387543439865, 'reward_meter_mean': 0.1600387543439865, 'reward_meter_std': 0.3069753348827362, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.1600387543439865, 'reward_total_composite_std': 0.3069753348827362, 'epoch': 0.04}

 32%|███▏      | 1059/3300 [52:59<4:46:19,  7.67s/it]INFO 04-11 23:29:50 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1060/3300 [53:09<5:11:07,  8.33s/it]2026-04-11 23:30:00,136 | INFO | train_grpo_train | metrics_logged mode=train step=1060

                                                     
{'loss': -0.3504, 'grad_norm': 1.0387495756149292, 'learning_rate': 6.790909090909091e-06, 'num_tokens': 2381179.0, 'completions/mean_length': 367.5, 'completions/min_length': 299.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 319.3333435058594, 'completions/min_terminated_length': 299.0, 'completions/max_terminated_length': 347.0, 'rewards/meter/mean': 0.8225868344306946, 'rewards/meter/std': 0.18334050476551056, 'rewards/count_adherence/mean': 0.7613636255264282, 'rewards/count_adherence/std': 0.12798961997032166, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.7348039150238037, 'rewards/repeat_penalty/std': 0.1173701137304306, 'rewards/total_composite/mean': 0.3815041184425354, 'rewards/total_composite/std': 0.23795181512832642, 'reward': 0.3815041184425354, 'reward_std': 0.23795180022716522, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028801359236240387, 'sampling/sampling_logp_difference/max': 1.8230382204055786, 'sampling/importance_sampling_ratio/min': 0.16153423488140106, 'sampling/importance_sampling_ratio/mean': 1.0062330961227417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1975250532850623, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.01580244250362739, 'clip_ratio/high_max': 0.01580244250362739, 'clip_ratio/region_mean': 0.01580244250362739, 'reward_total_mean': 0.3815041184425354, 'reward_meter_mean': 0.8225868344306946, 'reward_meter_std': 0.18334050476551056, 'reward_count_adherence_mean': 0.7613636255264282, 'reward_count_adherence_std': 0.12798961997032166, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.7348039150238037, 'reward_repeat_penalty_std': 0.1173701137304306, 'reward_total_composite_mean': 0.3815041184425354, 'reward_total_composite_std': 0.23795181512832642, 'epoch': 0.04}

 32%|███▏      | 1060/3300 [53:09<5:11:07,  8.33s/it]INFO 04-11 23:30:00 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1061/3300 [53:16<5:04:49,  8.17s/it]2026-04-11 23:30:07,922 | INFO | train_grpo_train | metrics_logged mode=train step=1061

                                                     
{'loss': -0.0089, 'grad_norm': 2.5813870429992676, 'learning_rate': 6.787878787878789e-06, 'num_tokens': 2385246.0, 'completions/mean_length': 300.375, 'completions/min_length': 289.0, 'completions/max_length': 313.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 300.375, 'completions/min_terminated_length': 289.0, 'completions/max_terminated_length': 313.0, 'rewards/meter/mean': 0.560438871383667, 'rewards/meter/std': 0.4554891288280487, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7440122365951538, 'rewards/repeat_penalty/std': 0.10529463738203049, 'rewards/total_composite/mean': 0.4012376070022583, 'rewards/total_composite/std': 0.3423055410385132, 'reward': 0.4012376070022583, 'reward_std': 0.3423055112361908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032200075685977936, 'sampling/sampling_logp_difference/max': 1.681164264678955, 'sampling/importance_sampling_ratio/min': 0.18615710735321045, 'sampling/importance_sampling_ratio/mean': 1.005553126335144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21212731674313545, 'clip_ratio/low_mean': 0.015588806010782719, 'clip_ratio/low_min': 0.015588806010782719, 'clip_ratio/high_mean': 0.01467954705003649, 'clip_ratio/high_max': 0.01467954705003649, 'clip_ratio/region_mean': 0.03026835306081921, 'reward_total_mean': 0.4012376070022583, 'reward_meter_mean': 0.560438871383667, 'reward_meter_std': 0.4554891288280487, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7440122365951538, 'reward_repeat_penalty_std': 0.10529463738203049, 'reward_total_composite_mean': 0.4012376070022583, 'reward_total_composite_std': 0.3423055410385132, 'epoch': 0.04}

 32%|███▏      | 1061/3300 [53:16<5:04:49,  8.17s/it]INFO 04-11 23:30:08 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1062/3300 [53:23<4:44:05,  7.62s/it]2026-04-11 23:30:14,247 | INFO | train_grpo_train | metrics_logged mode=train step=1062

                                                     
{'loss': 0.0723, 'grad_norm': 2.7553000450134277, 'learning_rate': 6.7848484848484855e-06, 'num_tokens': 2387921.0, 'completions/mean_length': 177.375, 'completions/min_length': 154.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 177.375, 'completions/min_terminated_length': 154.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.3804289400577545, 'rewards/meter/std': 0.4567885398864746, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5972222089767456, 'rewards/repeat_penalty/std': 0.25845491886138916, 'rewards/total_composite/mean': 0.1175825372338295, 'rewards/total_composite/std': 0.16052623093128204, 'reward': 0.1175825372338295, 'reward_std': 0.16052624583244324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02786877565085888, 'sampling/sampling_logp_difference/max': 1.2727034091949463, 'sampling/importance_sampling_ratio/min': 0.34907352924346924, 'sampling/importance_sampling_ratio/mean': 1.005418062210083, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1391224949620664, 'clip_ratio/low_mean': 0.016825482714921236, 'clip_ratio/low_min': 0.016825482714921236, 'clip_ratio/high_mean': 0.004816017230041325, 'clip_ratio/high_max': 0.004816017230041325, 'clip_ratio/region_mean': 0.02164149994496256, 'reward_total_mean': 0.1175825372338295, 'reward_meter_mean': 0.3804289400577545, 'reward_meter_std': 0.4567885398864746, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5972222089767456, 'reward_repeat_penalty_std': 0.25845491886138916, 'reward_total_composite_mean': 0.1175825372338295, 'reward_total_composite_std': 0.16052623093128204, 'epoch': 0.04}

 32%|███▏      | 1062/3300 [53:23<4:44:05,  7.62s/it]INFO 04-11 23:30:14 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1063/3300 [53:28<4:15:56,  6.86s/it]2026-04-11 23:30:19,359 | INFO | train_grpo_train | metrics_logged mode=train step=1063

                                                     
{'loss': 0.0052, 'grad_norm': 7.346792697906494, 'learning_rate': 6.781818181818183e-06, 'num_tokens': 2389705.0, 'completions/mean_length': 66.0, 'completions/min_length': 61.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.4735042452812195, 'rewards/meter/std': 0.4612884223461151, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4735042452812195, 'rewards/total_composite/std': 0.4612884223461151, 'reward': 0.4735042452812195, 'reward_std': 0.4612884521484375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04212275519967079, 'sampling/sampling_logp_difference/max': 1.0822772979736328, 'sampling/importance_sampling_ratio/min': 0.3388230502605438, 'sampling/importance_sampling_ratio/mean': 1.0069684982299805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15965153090655804, 'clip_ratio/low_mean': 0.014632937265560031, 'clip_ratio/low_min': 0.014632937265560031, 'clip_ratio/high_mean': 0.02170370938256383, 'clip_ratio/high_max': 0.02170370938256383, 'clip_ratio/region_mean': 0.03633664664812386, 'reward_total_mean': 0.4735042452812195, 'reward_meter_mean': 0.4735042452812195, 'reward_meter_std': 0.4612884223461151, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4735042452812195, 'reward_total_composite_std': 0.4612884223461151, 'epoch': 0.04}

 32%|███▏      | 1063/3300 [53:28<4:15:56,  6.86s/it]INFO 04-11 23:30:19 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1064/3300 [53:34<4:05:02,  6.58s/it]2026-04-11 23:30:25,259 | INFO | train_grpo_train | metrics_logged mode=train step=1064

                                                     
{'loss': 0.0144, 'grad_norm': 2.248720407485962, 'learning_rate': 6.778787878787879e-06, 'num_tokens': 2392392.0, 'completions/mean_length': 167.875, 'completions/min_length': 163.0, 'completions/max_length': 174.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 167.875, 'completions/min_terminated_length': 163.0, 'completions/max_terminated_length': 174.0, 'rewards/meter/mean': 0.7978705167770386, 'rewards/meter/std': 0.26100829243659973, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7222222089767456, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.4253278374671936, 'rewards/total_composite/std': 0.12737736105918884, 'reward': 0.4253278374671936, 'reward_std': 0.12737736105918884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026998618617653847, 'sampling/sampling_logp_difference/max': 4.974412441253662, 'sampling/importance_sampling_ratio/min': 0.006912579294294119, 'sampling/importance_sampling_ratio/mean': 1.0040228366851807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15179980965331197, 'clip_ratio/low_mean': 0.005747126415371895, 'clip_ratio/low_min': 0.005747126415371895, 'clip_ratio/high_mean': 0.002241053676698357, 'clip_ratio/high_max': 0.002241053676698357, 'clip_ratio/region_mean': 0.007988180092070252, 'reward_total_mean': 0.4253278374671936, 'reward_meter_mean': 0.7978705167770386, 'reward_meter_std': 0.26100829243659973, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7222222089767456, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.4253278374671936, 'reward_total_composite_std': 0.12737736105918884, 'epoch': 0.04}

 32%|███▏      | 1064/3300 [53:34<4:05:02,  6.58s/it]INFO 04-11 23:30:25 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1065/3300 [53:38<3:42:36,  5.98s/it]2026-04-11 23:30:29,838 | INFO | train_grpo_train | metrics_logged mode=train step=1065

                                                     
{'loss': 0.0031, 'grad_norm': 7.699649810791016, 'learning_rate': 6.7757575757575765e-06, 'num_tokens': 2393979.0, 'completions/mean_length': 51.375, 'completions/min_length': 51.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.375, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.3520212769508362, 'rewards/meter/std': 0.2575727701187134, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.3520212769508362, 'rewards/total_composite/std': 0.2575727701187134, 'reward': 0.3520212769508362, 'reward_std': 0.2575727701187134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03394267335534096, 'sampling/sampling_logp_difference/max': 1.1649155616760254, 'sampling/importance_sampling_ratio/min': 0.34641963243484497, 'sampling/importance_sampling_ratio/mean': 1.0043641328811646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17064102366566658, 'clip_ratio/low_mean': 0.02187028736807406, 'clip_ratio/low_min': 0.02187028736807406, 'clip_ratio/high_mean': 0.014428413240239024, 'clip_ratio/high_max': 0.014428413240239024, 'clip_ratio/region_mean': 0.036298700608313084, 'reward_total_mean': 0.3520212769508362, 'reward_meter_mean': 0.3520212769508362, 'reward_meter_std': 0.2575727701187134, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.3520212769508362, 'reward_total_composite_std': 0.2575727701187134, 'epoch': 0.04}

 32%|███▏      | 1065/3300 [53:38<3:42:36,  5.98s/it]INFO 04-11 23:30:30 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1066/3300 [53:43<3:29:20,  5.62s/it]2026-04-11 23:30:34,635 | INFO | train_grpo_train | metrics_logged mode=train step=1066

                                                     
{'loss': 0.0134, 'grad_norm': 4.696002006530762, 'learning_rate': 6.772727272727273e-06, 'num_tokens': 2395725.0, 'completions/mean_length': 67.25, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7684643864631653, 'rewards/meter/std': 0.3259983956813812, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7684643864631653, 'rewards/total_composite/std': 0.3259983956813812, 'reward': 0.7684643864631653, 'reward_std': 0.3259983956813812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04000062495470047, 'sampling/sampling_logp_difference/max': 1.5092440843582153, 'sampling/importance_sampling_ratio/min': 0.22107702493667603, 'sampling/importance_sampling_ratio/mean': 0.9985710978507996, 'sampling/importance_sampling_ratio/max': 1.9387584924697876, 'entropy': 0.23792543355375528, 'clip_ratio/low_mean': 0.00916612590663135, 'clip_ratio/low_min': 0.00916612590663135, 'clip_ratio/high_mean': 0.022364818840287626, 'clip_ratio/high_max': 0.022364818840287626, 'clip_ratio/region_mean': 0.031530944746918976, 'reward_total_mean': 0.7684643864631653, 'reward_meter_mean': 0.7684643864631653, 'reward_meter_std': 0.3259983956813812, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7684643864631653, 'reward_total_composite_std': 0.3259983956813812, 'epoch': 0.04}

 32%|███▏      | 1066/3300 [53:43<3:29:20,  5.62s/it]INFO 04-11 23:30:34 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1067/3300 [53:48<3:16:07,  5.27s/it]2026-04-11 23:30:39,081 | INFO | train_grpo_train | metrics_logged mode=train step=1067

                                                     
{'loss': 0.0273, 'grad_norm': 10.04726791381836, 'learning_rate': 6.76969696969697e-06, 'num_tokens': 2397159.0, 'completions/mean_length': 35.25, 'completions/min_length': 33.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.40233293175697327, 'rewards/meter/std': 0.35556212067604065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.40233293175697327, 'rewards/total_composite/std': 0.35556212067604065, 'reward': 0.40233293175697327, 'reward_std': 0.35556209087371826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0461297444999218, 'sampling/sampling_logp_difference/max': 1.1315374374389648, 'sampling/importance_sampling_ratio/min': 0.32253700494766235, 'sampling/importance_sampling_ratio/mean': 0.9973819255828857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2819352522492409, 'clip_ratio/low_mean': 0.010416666977107525, 'clip_ratio/low_min': 0.010416666977107525, 'clip_ratio/high_mean': 0.01831344375386834, 'clip_ratio/high_max': 0.01831344375386834, 'clip_ratio/region_mean': 0.028730110730975866, 'reward_total_mean': 0.40233293175697327, 'reward_meter_mean': 0.40233293175697327, 'reward_meter_std': 0.35556212067604065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.40233293175697327, 'reward_total_composite_std': 0.35556212067604065, 'epoch': 0.04}

 32%|███▏      | 1067/3300 [53:48<3:16:07,  5.27s/it]INFO 04-11 23:30:39 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1068/3300 [53:52<3:07:30,  5.04s/it]2026-04-11 23:30:43,586 | INFO | train_grpo_train | metrics_logged mode=train step=1068

                                                     
{'loss': 0.0118, 'grad_norm': 3.635596752166748, 'learning_rate': 6.7666666666666665e-06, 'num_tokens': 2398762.0, 'completions/mean_length': 47.375, 'completions/min_length': 45.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7638416290283203, 'rewards/meter/std': 0.2724026143550873, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7638416290283203, 'rewards/total_composite/std': 0.2724026143550873, 'reward': 0.7638416290283203, 'reward_std': 0.2724026143550873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026672368869185448, 'sampling/sampling_logp_difference/max': 1.1027817726135254, 'sampling/importance_sampling_ratio/min': 0.33194637298583984, 'sampling/importance_sampling_ratio/mean': 1.0070512294769287, 'sampling/importance_sampling_ratio/max': 1.4794321060180664, 'entropy': 0.15641734562814236, 'clip_ratio/low_mean': 0.0076530613005161285, 'clip_ratio/low_min': 0.0076530613005161285, 'clip_ratio/high_mean': 0.02335231169126928, 'clip_ratio/high_max': 0.02335231169126928, 'clip_ratio/region_mean': 0.031005372991785407, 'reward_total_mean': 0.7638416290283203, 'reward_meter_mean': 0.7638416290283203, 'reward_meter_std': 0.2724026143550873, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7638416290283203, 'reward_total_composite_std': 0.2724026143550873, 'epoch': 0.04}

 32%|███▏      | 1068/3300 [53:52<3:07:30,  5.04s/it]INFO 04-11 23:30:43 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1069/3300 [54:02<4:04:05,  6.56s/it]2026-04-11 23:30:53,708 | INFO | train_grpo_train | metrics_logged mode=train step=1069

                                                     
{'loss': -0.1635, 'grad_norm': 1.4912827014923096, 'learning_rate': 6.763636363636365e-06, 'num_tokens': 2403084.0, 'completions/mean_length': 370.25, 'completions/min_length': 328.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 350.0000305175781, 'completions/min_terminated_length': 328.0, 'completions/max_terminated_length': 370.0, 'rewards/meter/mean': 0.26017892360687256, 'rewards/meter/std': 0.2312571257352829, 'rewards/count_adherence/mean': 0.7291666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.6733911633491516, 'rewards/repeat_penalty/std': 0.18306629359722137, 'rewards/total_composite/mean': 0.10642074048519135, 'rewards/total_composite/std': 0.06904294341802597, 'reward': 0.10642074048519135, 'reward_std': 0.06904293596744537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02436882257461548, 'sampling/sampling_logp_difference/max': 1.8471393585205078, 'sampling/importance_sampling_ratio/min': 0.15768760442733765, 'sampling/importance_sampling_ratio/mean': 1.0042506456375122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11531824059784412, 'clip_ratio/low_mean': 0.007008010521531105, 'clip_ratio/low_min': 0.007008010521531105, 'clip_ratio/high_mean': 0.009160695597529411, 'clip_ratio/high_max': 0.009160695597529411, 'clip_ratio/region_mean': 0.016168706119060516, 'reward_total_mean': 0.10642074048519135, 'reward_meter_mean': 0.26017892360687256, 'reward_meter_std': 0.2312571257352829, 'reward_count_adherence_mean': 0.7291666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.6733911633491516, 'reward_repeat_penalty_std': 0.18306629359722137, 'reward_total_composite_mean': 0.10642074048519135, 'reward_total_composite_std': 0.06904294341802597, 'epoch': 0.04}

 32%|███▏      | 1069/3300 [54:02<4:04:05,  6.56s/it]INFO 04-11 23:30:53 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1070/3300 [54:07<3:41:02,  5.95s/it]2026-04-11 23:30:58,215 | INFO | train_grpo_train | metrics_logged mode=train step=1070

                                                     
{'loss': 0.0141, 'grad_norm': 7.106749534606934, 'learning_rate': 6.760606060606061e-06, 'num_tokens': 2404519.0, 'completions/mean_length': 36.375, 'completions/min_length': 34.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9458709359169006, 'rewards/meter/std': 0.06306823343038559, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9458709359169006, 'rewards/total_composite/std': 0.06306823343038559, 'reward': 0.9458709359169006, 'reward_std': 0.0630682110786438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043692056089639664, 'sampling/sampling_logp_difference/max': 1.2659907341003418, 'sampling/importance_sampling_ratio/min': 0.28195980191230774, 'sampling/importance_sampling_ratio/mean': 1.0076913833618164, 'sampling/importance_sampling_ratio/max': 1.6269347667694092, 'entropy': 0.26383113488554955, 'clip_ratio/low_mean': 0.02384992502629757, 'clip_ratio/low_min': 0.02384992502629757, 'clip_ratio/high_mean': 0.006850600708276033, 'clip_ratio/high_max': 0.006850600708276033, 'clip_ratio/region_mean': 0.030700525734573603, 'reward_total_mean': 0.9458709359169006, 'reward_meter_mean': 0.9458709359169006, 'reward_meter_std': 0.06306823343038559, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9458709359169006, 'reward_total_composite_std': 0.06306823343038559, 'epoch': 0.04}

 32%|███▏      | 1070/3300 [54:07<3:41:02,  5.95s/it]INFO 04-11 23:30:58 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1071/3300 [54:14<3:55:10,  6.33s/it]2026-04-11 23:31:05,439 | INFO | train_grpo_train | metrics_logged mode=train step=1071

                                                     
{'loss': 0.0028, 'grad_norm': 1.8987226486206055, 'learning_rate': 6.757575757575758e-06, 'num_tokens': 2408287.0, 'completions/mean_length': 257.0, 'completions/min_length': 249.0, 'completions/max_length': 262.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 257.0, 'completions/min_terminated_length': 249.0, 'completions/max_terminated_length': 262.0, 'rewards/meter/mean': 0.8534860610961914, 'rewards/meter/std': 0.08555062115192413, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6499999761581421, 'rewards/repeat_penalty/std': 0.05909368395805359, 'rewards/total_composite/mean': 0.4727606773376465, 'rewards/total_composite/std': 0.03482650965452194, 'reward': 0.4727606773376465, 'reward_std': 0.03482651710510254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011769634671509266, 'sampling/sampling_logp_difference/max': 1.9065320491790771, 'sampling/importance_sampling_ratio/min': 0.29330602288246155, 'sampling/importance_sampling_ratio/mean': 1.001491665840149, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04622745281085372, 'clip_ratio/low_mean': 0.006310773896984756, 'clip_ratio/low_min': 0.006310773896984756, 'clip_ratio/high_mean': 0.0009578543831594288, 'clip_ratio/high_max': 0.0009578543831594288, 'clip_ratio/region_mean': 0.007268628280144185, 'reward_total_mean': 0.4727606773376465, 'reward_meter_mean': 0.8534860610961914, 'reward_meter_std': 0.08555062115192413, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6499999761581421, 'reward_repeat_penalty_std': 0.05909368395805359, 'reward_total_composite_mean': 0.4727606773376465, 'reward_total_composite_std': 0.03482650965452194, 'epoch': 0.04}

 32%|███▏      | 1071/3300 [54:14<3:55:10,  6.33s/it]INFO 04-11 23:31:05 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1072/3300 [54:19<3:41:44,  5.97s/it]2026-04-11 23:31:10,573 | INFO | train_grpo_train | metrics_logged mode=train step=1072

                                                     
{'loss': -0.009, 'grad_norm': 4.010406494140625, 'learning_rate': 6.754545454545455e-06, 'num_tokens': 2410412.0, 'completions/mean_length': 96.625, 'completions/min_length': 92.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.625, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.7025290727615356, 'rewards/meter/std': 0.24330702424049377, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.6105661392211914, 'rewards/total_composite/std': 0.21741104125976562, 'reward': 0.6105661392211914, 'reward_std': 0.21741104125976562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02641892619431019, 'sampling/sampling_logp_difference/max': 1.0504131317138672, 'sampling/importance_sampling_ratio/min': 0.3497931957244873, 'sampling/importance_sampling_ratio/mean': 1.0027921199798584, 'sampling/importance_sampling_ratio/max': 1.5932246446609497, 'entropy': 0.15043141692876816, 'clip_ratio/low_mean': 0.003919956274330616, 'clip_ratio/low_min': 0.003919956274330616, 'clip_ratio/high_mean': 0.01675201370380819, 'clip_ratio/high_max': 0.01675201370380819, 'clip_ratio/region_mean': 0.020671969978138804, 'reward_total_mean': 0.6105661392211914, 'reward_meter_mean': 0.7025290727615356, 'reward_meter_std': 0.24330702424049377, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.6105661392211914, 'reward_total_composite_std': 0.21741104125976562, 'epoch': 0.04}

 32%|███▏      | 1072/3300 [54:19<3:41:44,  5.97s/it]INFO 04-11 23:31:10 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1073/3300 [54:25<3:36:41,  5.84s/it]2026-04-11 23:31:16,099 | INFO | train_grpo_train | metrics_logged mode=train step=1073

                                                     
{'loss': -0.0119, 'grad_norm': 4.956683158874512, 'learning_rate': 6.751515151515152e-06, 'num_tokens': 2412879.0, 'completions/mean_length': 122.375, 'completions/min_length': 117.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.375, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.8654184341430664, 'rewards/meter/std': 0.3119131624698639, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8155930042266846, 'rewards/total_composite/std': 0.30145007371902466, 'reward': 0.8155930042266846, 'reward_std': 0.30145007371902466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04843366891145706, 'sampling/sampling_logp_difference/max': 1.2980079650878906, 'sampling/importance_sampling_ratio/min': 0.2730752229690552, 'sampling/importance_sampling_ratio/mean': 1.0030038356781006, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3304913900792599, 'clip_ratio/low_mean': 0.012625776929780841, 'clip_ratio/low_min': 0.012625776929780841, 'clip_ratio/high_mean': 0.03231395175680518, 'clip_ratio/high_max': 0.03231395175680518, 'clip_ratio/region_mean': 0.04493972868658602, 'reward_total_mean': 0.8155930042266846, 'reward_meter_mean': 0.8654184341430664, 'reward_meter_std': 0.3119131624698639, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8155930042266846, 'reward_total_composite_std': 0.30145007371902466, 'epoch': 0.04}

 33%|███▎      | 1073/3300 [54:25<3:36:41,  5.84s/it]INFO 04-11 23:31:16 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1074/3300 [54:32<3:53:40,  6.30s/it]2026-04-11 23:31:23,473 | INFO | train_grpo_train | metrics_logged mode=train step=1074

                                                     
{'loss': 0.0024, 'grad_norm': 2.035898208618164, 'learning_rate': 6.748484848484848e-06, 'num_tokens': 2416593.0, 'completions/mean_length': 276.25, 'completions/min_length': 261.0, 'completions/max_length': 289.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 276.25, 'completions/min_terminated_length': 261.0, 'completions/max_terminated_length': 289.0, 'rewards/meter/mean': 0.9123198986053467, 'rewards/meter/std': 0.15345944464206696, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625821024179459, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6886509656906128, 'rewards/repeat_penalty/std': 0.08202779293060303, 'rewards/total_composite/mean': 0.5867777466773987, 'rewards/total_composite/std': 0.12737217545509338, 'reward': 0.5867777466773987, 'reward_std': 0.12737217545509338, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01869548298418522, 'sampling/sampling_logp_difference/max': 1.542731761932373, 'sampling/importance_sampling_ratio/min': 0.21379627287387848, 'sampling/importance_sampling_ratio/mean': 0.9994420409202576, 'sampling/importance_sampling_ratio/max': 1.5070128440856934, 'entropy': 0.1006718729622662, 'clip_ratio/low_mean': 0.0050951789889950305, 'clip_ratio/low_min': 0.0050951789889950305, 'clip_ratio/high_mean': 0.0071886846562847495, 'clip_ratio/high_max': 0.0071886846562847495, 'clip_ratio/region_mean': 0.01228386364527978, 'reward_total_mean': 0.5867777466773987, 'reward_meter_mean': 0.9123198986053467, 'reward_meter_std': 0.15345944464206696, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625821024179459, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6886509656906128, 'reward_repeat_penalty_std': 0.08202779293060303, 'reward_total_composite_mean': 0.5867777466773987, 'reward_total_composite_std': 0.12737217545509338, 'epoch': 0.04}

 33%|███▎      | 1074/3300 [54:32<3:53:40,  6.30s/it]INFO 04-11 23:31:23 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1075/3300 [54:40<4:08:34,  6.70s/it]2026-04-11 23:31:31,119 | INFO | train_grpo_train | metrics_logged mode=train step=1075

                                                     
{'loss': 0.012, 'grad_norm': 1.7712578773498535, 'learning_rate': 6.7454545454545465e-06, 'num_tokens': 2420319.0, 'completions/mean_length': 290.75, 'completions/min_length': 268.0, 'completions/max_length': 315.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 290.75, 'completions/min_terminated_length': 268.0, 'completions/max_terminated_length': 315.0, 'rewards/meter/mean': 0.9649861454963684, 'rewards/meter/std': 0.05583993345499039, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.634615421295166, 'rewards/repeat_penalty/std': 0.03560846298933029, 'rewards/total_composite/mean': 0.6113426089286804, 'rewards/total_composite/std': 0.03276226669549942, 'reward': 0.6113426089286804, 'reward_std': 0.032762277871370316, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012581315822899342, 'sampling/sampling_logp_difference/max': 1.3584985733032227, 'sampling/importance_sampling_ratio/min': 0.2570464313030243, 'sampling/importance_sampling_ratio/mean': 0.9999510049819946, 'sampling/importance_sampling_ratio/max': 1.738738775253296, 'entropy': 0.08378143096342683, 'clip_ratio/low_mean': 0.0065517745388206095, 'clip_ratio/low_min': 0.0065517745388206095, 'clip_ratio/high_mean': 0.001742160296998918, 'clip_ratio/high_max': 0.001742160296998918, 'clip_ratio/region_mean': 0.008293934835819528, 'reward_total_mean': 0.6113426089286804, 'reward_meter_mean': 0.9649861454963684, 'reward_meter_std': 0.05583993345499039, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.634615421295166, 'reward_repeat_penalty_std': 0.03560846298933029, 'reward_total_composite_mean': 0.6113426089286804, 'reward_total_composite_std': 0.03276226669549942, 'epoch': 0.04}

 33%|███▎      | 1075/3300 [54:40<4:08:34,  6.70s/it]INFO 04-11 23:31:31 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1076/3300 [54:46<4:02:50,  6.55s/it]2026-04-11 23:31:37,317 | INFO | train_grpo_train | metrics_logged mode=train step=1076

                                                     
{'loss': -0.0057, 'grad_norm': 4.413632869720459, 'learning_rate': 6.742424242424243e-06, 'num_tokens': 2423119.0, 'completions/mean_length': 170.0, 'completions/min_length': 156.0, 'completions/max_length': 186.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.0, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 186.0, 'rewards/meter/mean': 0.7671191692352295, 'rewards/meter/std': 0.2540895640850067, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7361111044883728, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.4233958125114441, 'rewards/total_composite/std': 0.14756236970424652, 'reward': 0.4233958125114441, 'reward_std': 0.14756235480308533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04489843547344208, 'sampling/sampling_logp_difference/max': 8.813060760498047, 'sampling/importance_sampling_ratio/min': 0.0001487771951360628, 'sampling/importance_sampling_ratio/mean': 0.993308961391449, 'sampling/importance_sampling_ratio/max': 1.7489451169967651, 'entropy': 0.14643115364015102, 'clip_ratio/low_mean': 0.012700201012194157, 'clip_ratio/low_min': 0.012700201012194157, 'clip_ratio/high_mean': 0.013822466367855668, 'clip_ratio/high_max': 0.013822466367855668, 'clip_ratio/region_mean': 0.026522667380049825, 'reward_total_mean': 0.4233958125114441, 'reward_meter_mean': 0.7671191692352295, 'reward_meter_std': 0.2540895640850067, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7361111044883728, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.4233958125114441, 'reward_total_composite_std': 0.14756236970424652, 'epoch': 0.04}

 33%|███▎      | 1076/3300 [54:46<4:02:50,  6.55s/it]INFO 04-11 23:31:37 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1077/3300 [54:50<3:40:37,  5.95s/it]2026-04-11 23:31:41,878 | INFO | train_grpo_train | metrics_logged mode=train step=1077

                                                     
{'loss': 0.012, 'grad_norm': 3.8939266204833984, 'learning_rate': 6.73939393939394e-06, 'num_tokens': 2424867.0, 'completions/mean_length': 57.5, 'completions/min_length': 55.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9874081611633301, 'rewards/meter/std': 0.004490252584218979, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6582721471786499, 'rewards/total_composite/std': 0.0029935124330222607, 'reward': 0.6582721471786499, 'reward_std': 0.002993511501699686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010735772550106049, 'sampling/sampling_logp_difference/max': 0.34990429878234863, 'sampling/importance_sampling_ratio/min': 0.7047555446624756, 'sampling/importance_sampling_ratio/mean': 1.0056201219558716, 'sampling/importance_sampling_ratio/max': 1.406072735786438, 'entropy': 0.07559043914079666, 'clip_ratio/low_mean': 0.0021186440717428923, 'clip_ratio/low_min': 0.0021186440717428923, 'clip_ratio/high_mean': 0.0065830720122903585, 'clip_ratio/high_max': 0.0065830720122903585, 'clip_ratio/region_mean': 0.00870171608403325, 'reward_total_mean': 0.6582721471786499, 'reward_meter_mean': 0.9874081611633301, 'reward_meter_std': 0.004490252584218979, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6582721471786499, 'reward_total_composite_std': 0.0029935124330222607, 'epoch': 0.04}

 33%|███▎      | 1077/3300 [54:50<3:40:37,  5.95s/it]INFO 04-11 23:31:42 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1078/3300 [54:59<4:15:23,  6.90s/it]2026-04-11 23:31:50,972 | INFO | train_grpo_train | metrics_logged mode=train step=1078

                                                     
{'loss': -0.0117, 'grad_norm': 0.8265510201454163, 'learning_rate': 6.7363636363636365e-06, 'num_tokens': 2429621.0, 'completions/mean_length': 403.25, 'completions/min_length': 384.0, 'completions/max_length': 433.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 403.25, 'completions/min_terminated_length': 384.0, 'completions/max_terminated_length': 433.0, 'rewards/meter/mean': 0.9796521067619324, 'rewards/meter/std': 0.0062021855264902115, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5921052694320679, 'rewards/repeat_penalty/std': 0.037216152995824814, 'rewards/total_composite/mean': 0.38671886920928955, 'rewards/total_composite/std': 0.02471131458878517, 'reward': 0.38671886920928955, 'reward_std': 0.024711309000849724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005612175911664963, 'sampling/sampling_logp_difference/max': 1.5524296760559082, 'sampling/importance_sampling_ratio/min': 0.2117329090833664, 'sampling/importance_sampling_ratio/mean': 1.0002822875976562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03162188641726971, 'clip_ratio/low_mean': 0.002776339795673266, 'clip_ratio/low_min': 0.002776339795673266, 'clip_ratio/high_mean': 0.0006038647261448205, 'clip_ratio/high_max': 0.0006038647261448205, 'clip_ratio/region_mean': 0.0033802045218180865, 'reward_total_mean': 0.38671886920928955, 'reward_meter_mean': 0.9796521067619324, 'reward_meter_std': 0.0062021855264902115, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5921052694320679, 'reward_repeat_penalty_std': 0.037216152995824814, 'reward_total_composite_mean': 0.38671886920928955, 'reward_total_composite_std': 0.02471131458878517, 'epoch': 0.04}

 33%|███▎      | 1078/3300 [54:59<4:15:23,  6.90s/it]INFO 04-11 23:31:51 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1079/3300 [55:04<3:53:13,  6.30s/it]2026-04-11 23:31:55,883 | INFO | train_grpo_train | metrics_logged mode=train step=1079

                                                     
{'loss': 0.0067, 'grad_norm': 4.684656143188477, 'learning_rate': 6.733333333333334e-06, 'num_tokens': 2431573.0, 'completions/mean_length': 84.0, 'completions/min_length': 81.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.0, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9432131052017212, 'rewards/meter/std': 0.005848580971360207, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9432131052017212, 'rewards/total_composite/std': 0.005848580971360207, 'reward': 0.9432131052017212, 'reward_std': 0.0058485823683440685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00942947156727314, 'sampling/sampling_logp_difference/max': 1.2586157321929932, 'sampling/importance_sampling_ratio/min': 0.2840469479560852, 'sampling/importance_sampling_ratio/mean': 0.9998735785484314, 'sampling/importance_sampling_ratio/max': 1.180408239364624, 'entropy': 0.04214879055507481, 'clip_ratio/low_mean': 0.0014367816038429737, 'clip_ratio/low_min': 0.0014367816038429737, 'clip_ratio/high_mean': 0.005959982983767986, 'clip_ratio/high_max': 0.005959982983767986, 'clip_ratio/region_mean': 0.00739676458761096, 'reward_total_mean': 0.9432131052017212, 'reward_meter_mean': 0.9432131052017212, 'reward_meter_std': 0.005848580971360207, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9432131052017212, 'reward_total_composite_std': 0.005848580971360207, 'epoch': 0.04}

 33%|███▎      | 1079/3300 [55:04<3:53:13,  6.30s/it]INFO 04-11 23:31:56 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1080/3300 [55:09<3:35:59,  5.84s/it]2026-04-11 23:32:00,640 | INFO | train_grpo_train | metrics_logged mode=train step=1080

                                                     
{'loss': -0.088, 'grad_norm': 7.663031578063965, 'learning_rate': 6.73030303030303e-06, 'num_tokens': 2433443.0, 'completions/mean_length': 57.75, 'completions/min_length': 51.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.31589365005493164, 'rewards/meter/std': 0.41653454303741455, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.31589365005493164, 'rewards/total_composite/std': 0.41653454303741455, 'reward': 0.31589365005493164, 'reward_std': 0.41653454303741455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038866013288497925, 'sampling/sampling_logp_difference/max': 1.4967455863952637, 'sampling/importance_sampling_ratio/min': 0.22385750710964203, 'sampling/importance_sampling_ratio/mean': 1.0007740259170532, 'sampling/importance_sampling_ratio/max': 1.726487159729004, 'entropy': 0.14285273849964142, 'clip_ratio/low_mean': 0.022774446289986372, 'clip_ratio/low_min': 0.022774446289986372, 'clip_ratio/high_mean': 0.014718614984303713, 'clip_ratio/high_max': 0.014718614984303713, 'clip_ratio/region_mean': 0.037493061274290085, 'reward_total_mean': 0.31589365005493164, 'reward_meter_mean': 0.31589365005493164, 'reward_meter_std': 0.41653454303741455, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.31589365005493164, 'reward_total_composite_std': 0.41653454303741455, 'epoch': 0.04}

 33%|███▎      | 1080/3300 [55:09<3:35:59,  5.84s/it]INFO 04-11 23:32:00 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1081/3300 [55:14<3:25:43,  5.56s/it]2026-04-11 23:32:05,562 | INFO | train_grpo_train | metrics_logged mode=train step=1081

                                                     
{'loss': 0.0001, 'grad_norm': 2.3848185539245605, 'learning_rate': 6.7272727272727275e-06, 'num_tokens': 2435373.0, 'completions/mean_length': 84.25, 'completions/min_length': 83.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.25, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9907380938529968, 'rewards/meter/std': 0.0005972905782982707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8173696994781494, 'rewards/total_composite/std': 0.07022644579410553, 'reward': 0.8173696994781494, 'reward_std': 0.07022644579410553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018476024270057678, 'sampling/sampling_logp_difference/max': 0.8972263336181641, 'sampling/importance_sampling_ratio/min': 0.40769892930984497, 'sampling/importance_sampling_ratio/mean': 0.9998801946640015, 'sampling/importance_sampling_ratio/max': 1.8647558689117432, 'entropy': 0.06444978155195713, 'clip_ratio/low_mean': 0.010382481734268367, 'clip_ratio/low_min': 0.010382481734268367, 'clip_ratio/high_mean': 0.0029411765281111, 'clip_ratio/high_max': 0.0029411765281111, 'clip_ratio/region_mean': 0.013323658262379467, 'reward_total_mean': 0.8173696994781494, 'reward_meter_mean': 0.9907380938529968, 'reward_meter_std': 0.0005972905782982707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8173696994781494, 'reward_total_composite_std': 0.07022644579410553, 'epoch': 0.04}

 33%|███▎      | 1081/3300 [55:14<3:25:43,  5.56s/it]INFO 04-11 23:32:05 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1082/3300 [55:19<3:20:22,  5.42s/it]2026-04-11 23:32:10,651 | INFO | train_grpo_train | metrics_logged mode=train step=1082

                                                     
{'loss': -0.0216, 'grad_norm': 3.134387969970703, 'learning_rate': 6.724242424242424e-06, 'num_tokens': 2437378.0, 'completions/mean_length': 77.625, 'completions/min_length': 72.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9952170252799988, 'rewards/meter/std': 0.005060417577624321, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952170252799988, 'rewards/total_composite/std': 0.005060417577624321, 'reward': 0.9952170252799988, 'reward_std': 0.005060393828898668, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023099998012185097, 'sampling/sampling_logp_difference/max': 1.7406330108642578, 'sampling/importance_sampling_ratio/min': 0.17540933191776276, 'sampling/importance_sampling_ratio/mean': 1.0053282976150513, 'sampling/importance_sampling_ratio/max': 1.727461814880371, 'entropy': 0.11937720235437155, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.009705353993922472, 'clip_ratio/high_max': 0.009705353993922472, 'clip_ratio/region_mean': 0.011441465117968619, 'reward_total_mean': 0.9952170252799988, 'reward_meter_mean': 0.9952170252799988, 'reward_meter_std': 0.005060417577624321, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952170252799988, 'reward_total_composite_std': 0.005060417577624321, 'epoch': 0.04}

 33%|███▎      | 1082/3300 [55:19<3:20:22,  5.42s/it]INFO 04-11 23:32:10 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1083/3300 [55:25<3:23:41,  5.51s/it]2026-04-11 23:32:16,377 | INFO | train_grpo_train | metrics_logged mode=train step=1083

                                                     
{'loss': -0.0062, 'grad_norm': 5.954423427581787, 'learning_rate': 6.721212121212122e-06, 'num_tokens': 2439776.0, 'completions/mean_length': 127.75, 'completions/min_length': 122.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.75, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.7971546053886414, 'rewards/meter/std': 0.2483818531036377, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7361111640930176, 'rewards/repeat_penalty/std': 0.13197055459022522, 'rewards/total_composite/mean': 0.43424269556999207, 'rewards/total_composite/std': 0.14866070449352264, 'reward': 0.43424269556999207, 'reward_std': 0.14866070449352264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02712804079055786, 'sampling/sampling_logp_difference/max': 1.9316282272338867, 'sampling/importance_sampling_ratio/min': 0.14491204917430878, 'sampling/importance_sampling_ratio/mean': 1.0028880834579468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10621183086186647, 'clip_ratio/low_mean': 0.0030412437627092004, 'clip_ratio/low_min': 0.0030412437627092004, 'clip_ratio/high_mean': 0.01804604707285762, 'clip_ratio/high_max': 0.01804604707285762, 'clip_ratio/region_mean': 0.02108729083556682, 'reward_total_mean': 0.43424269556999207, 'reward_meter_mean': 0.7971546053886414, 'reward_meter_std': 0.2483818531036377, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7361111640930176, 'reward_repeat_penalty_std': 0.13197055459022522, 'reward_total_composite_mean': 0.43424269556999207, 'reward_total_composite_std': 0.14866070449352264, 'epoch': 0.04}

 33%|███▎      | 1083/3300 [55:25<3:23:41,  5.51s/it]INFO 04-11 23:32:16 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1084/3300 [55:30<3:21:07,  5.45s/it]2026-04-11 23:32:21,678 | INFO | train_grpo_train | metrics_logged mode=train step=1084

                                                     
{'loss': 0.005, 'grad_norm': 7.175000190734863, 'learning_rate': 6.718181818181819e-06, 'num_tokens': 2441759.0, 'completions/mean_length': 80.875, 'completions/min_length': 77.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.875, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9879165291786194, 'rewards/meter/std': 0.011460079811513424, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9879165291786194, 'rewards/total_composite/std': 0.011460079811513424, 'reward': 0.9879165291786194, 'reward_std': 0.011460077948868275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02571716159582138, 'sampling/sampling_logp_difference/max': 1.5991357564926147, 'sampling/importance_sampling_ratio/min': 0.20207108557224274, 'sampling/importance_sampling_ratio/mean': 0.9987949132919312, 'sampling/importance_sampling_ratio/max': 1.9328581094741821, 'entropy': 0.10348887462168932, 'clip_ratio/low_mean': 0.009149131015874445, 'clip_ratio/low_min': 0.009149131015874445, 'clip_ratio/high_mean': 0.018721832893788815, 'clip_ratio/high_max': 0.018721832893788815, 'clip_ratio/region_mean': 0.02787096390966326, 'reward_total_mean': 0.9879165291786194, 'reward_meter_mean': 0.9879165291786194, 'reward_meter_std': 0.011460079811513424, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9879165291786194, 'reward_total_composite_std': 0.011460079811513424, 'epoch': 0.04}

 33%|███▎      | 1084/3300 [55:30<3:21:07,  5.45s/it]INFO 04-11 23:32:22 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1085/3300 [55:39<4:00:35,  6.52s/it]2026-04-11 23:32:30,685 | INFO | train_grpo_train | metrics_logged mode=train step=1085

                                                     
{'loss': -0.03, 'grad_norm': 0.6841692924499512, 'learning_rate': 6.715151515151516e-06, 'num_tokens': 2446044.0, 'completions/mean_length': 335.625, 'completions/min_length': 318.0, 'completions/max_length': 355.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 335.625, 'completions/min_terminated_length': 318.0, 'completions/max_terminated_length': 355.0, 'rewards/meter/mean': 0.9889903664588928, 'rewards/meter/std': 0.017724184319376945, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.05750546231865883, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5962929129600525, 'rewards/repeat_penalty/std': 0.030314533039927483, 'rewards/total_composite/mean': 0.5644394755363464, 'rewards/total_composite/std': 0.03291511908173561, 'reward': 0.5644394755363464, 'reward_std': 0.03291511535644531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008025525137782097, 'sampling/sampling_logp_difference/max': 1.0485243797302246, 'sampling/importance_sampling_ratio/min': 0.3504544794559479, 'sampling/importance_sampling_ratio/mean': 0.9996162056922913, 'sampling/importance_sampling_ratio/max': 1.9612756967544556, 'entropy': 0.034420924610458314, 'clip_ratio/low_mean': 0.0023487245198339224, 'clip_ratio/low_min': 0.0023487245198339224, 'clip_ratio/high_mean': 0.004307354771299288, 'clip_ratio/high_max': 0.004307354771299288, 'clip_ratio/region_mean': 0.00665607929113321, 'reward_total_mean': 0.5644394755363464, 'reward_meter_mean': 0.9889903664588928, 'reward_meter_std': 0.017724184319376945, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.05750546231865883, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5962929129600525, 'reward_repeat_penalty_std': 0.030314533039927483, 'reward_total_composite_mean': 0.5644394755363464, 'reward_total_composite_std': 0.03291511908173561, 'epoch': 0.04}

 33%|███▎      | 1085/3300 [55:39<4:00:35,  6.52s/it]INFO 04-11 23:32:30 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1086/3300 [55:48<4:29:29,  7.30s/it]2026-04-11 23:32:39,822 | INFO | train_grpo_train | metrics_logged mode=train step=1086

                                                     
{'loss': -0.0088, 'grad_norm': 0.8804117441177368, 'learning_rate': 6.712121212121213e-06, 'num_tokens': 2450552.0, 'completions/mean_length': 350.5, 'completions/min_length': 344.0, 'completions/max_length': 369.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 350.5, 'completions/min_terminated_length': 344.0, 'completions/max_terminated_length': 369.0, 'rewards/meter/mean': 0.9969491958618164, 'rewards/meter/std': 0.0016905681695789099, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5882353186607361, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5277966260910034, 'rewards/total_composite/std': 0.0008950114133767784, 'reward': 0.5277966260910034, 'reward_std': 0.00089502043556422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005859900265932083, 'sampling/sampling_logp_difference/max': 1.0715160369873047, 'sampling/importance_sampling_ratio/min': 0.3424888849258423, 'sampling/importance_sampling_ratio/mean': 0.999923050403595, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.025596523424610496, 'clip_ratio/low_mean': 0.0010880097979679704, 'clip_ratio/low_min': 0.0010880097979679704, 'clip_ratio/high_mean': 0.0027567246870603412, 'clip_ratio/high_max': 0.0027567246870603412, 'clip_ratio/region_mean': 0.0038447344850283116, 'reward_total_mean': 0.5277966260910034, 'reward_meter_mean': 0.9969491958618164, 'reward_meter_std': 0.0016905681695789099, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5882353186607361, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5277966260910034, 'reward_total_composite_std': 0.0008950114133767784, 'epoch': 0.04}

 33%|███▎      | 1086/3300 [55:48<4:29:29,  7.30s/it]INFO 04-11 23:32:40 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1087/3300 [55:53<3:59:53,  6.50s/it]2026-04-11 23:32:44,463 | INFO | train_grpo_train | metrics_logged mode=train step=1087

                                                     
{'loss': -0.0062, 'grad_norm': 3.9621143341064453, 'learning_rate': 6.709090909090909e-06, 'num_tokens': 2452523.0, 'completions/mean_length': 79.375, 'completions/min_length': 74.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.375, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.7115691900253296, 'rewards/meter/std': 0.38112178444862366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7115691900253296, 'rewards/total_composite/std': 0.38112178444862366, 'reward': 0.7115691900253296, 'reward_std': 0.38112178444862366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034753963351249695, 'sampling/sampling_logp_difference/max': 2.0210351943969727, 'sampling/importance_sampling_ratio/min': 0.1325182169675827, 'sampling/importance_sampling_ratio/mean': 1.0051707029342651, 'sampling/importance_sampling_ratio/max': 1.846152663230896, 'entropy': 0.2076272815465927, 'clip_ratio/low_mean': 0.006756756920367479, 'clip_ratio/low_min': 0.006756756920367479, 'clip_ratio/high_mean': 0.01255706837400794, 'clip_ratio/high_max': 0.01255706837400794, 'clip_ratio/region_mean': 0.01931382529437542, 'reward_total_mean': 0.7115691900253296, 'reward_meter_mean': 0.7115691900253296, 'reward_meter_std': 0.38112178444862366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7115691900253296, 'reward_total_composite_std': 0.38112178444862366, 'epoch': 0.04}

 33%|███▎      | 1087/3300 [55:53<3:59:53,  6.50s/it]INFO 04-11 23:32:44 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1088/3300 [55:57<3:33:24,  5.79s/it]2026-04-11 23:32:48,581 | INFO | train_grpo_train | metrics_logged mode=train step=1088

                                                     
{'loss': -0.0029, 'grad_norm': 6.578573226928711, 'learning_rate': 6.706060606060607e-06, 'num_tokens': 2454071.0, 'completions/mean_length': 33.5, 'completions/min_length': 32.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.7622437477111816, 'rewards/meter/std': 0.3664873540401459, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7622437477111816, 'rewards/total_composite/std': 0.3664873540401459, 'reward': 0.7622437477111816, 'reward_std': 0.3664873540401459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040684185922145844, 'sampling/sampling_logp_difference/max': 1.069122314453125, 'sampling/importance_sampling_ratio/min': 0.3433097004890442, 'sampling/importance_sampling_ratio/mean': 1.0093433856964111, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18665132857859135, 'clip_ratio/low_mean': 0.011363636702299118, 'clip_ratio/low_min': 0.011363636702299118, 'clip_ratio/high_mean': 0.04101851023733616, 'clip_ratio/high_max': 0.04101851023733616, 'clip_ratio/region_mean': 0.05238214693963528, 'reward_total_mean': 0.7622437477111816, 'reward_meter_mean': 0.7622437477111816, 'reward_meter_std': 0.3664873540401459, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7622437477111816, 'reward_total_composite_std': 0.3664873540401459, 'epoch': 0.04}

 33%|███▎      | 1088/3300 [55:57<3:33:24,  5.79s/it]INFO 04-11 23:32:48 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1089/3300 [56:02<3:28:19,  5.65s/it]2026-04-11 23:32:53,922 | INFO | train_grpo_train | metrics_logged mode=train step=1089

                                                     
{'loss': 0.0411, 'grad_norm': 4.875842571258545, 'learning_rate': 6.703030303030304e-06, 'num_tokens': 2455804.0, 'completions/mean_length': 58.625, 'completions/min_length': 54.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8143894672393799, 'rewards/meter/std': 0.22199998795986176, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8143894672393799, 'rewards/total_composite/std': 0.22199998795986176, 'reward': 0.8143894672393799, 'reward_std': 0.22199997305870056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04701100289821625, 'sampling/sampling_logp_difference/max': 1.6473197937011719, 'sampling/importance_sampling_ratio/min': 0.19256533682346344, 'sampling/importance_sampling_ratio/mean': 0.9986065030097961, 'sampling/importance_sampling_ratio/max': 1.836875081062317, 'entropy': 0.1829883987084031, 'clip_ratio/low_mean': 0.009895833441987634, 'clip_ratio/low_min': 0.009895833441987634, 'clip_ratio/high_mean': 0.03449019626714289, 'clip_ratio/high_max': 0.03449019626714289, 'clip_ratio/region_mean': 0.044386029709130526, 'reward_total_mean': 0.8143894672393799, 'reward_meter_mean': 0.8143894672393799, 'reward_meter_std': 0.22199998795986176, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8143894672393799, 'reward_total_composite_std': 0.22199998795986176, 'epoch': 0.04}

 33%|███▎      | 1089/3300 [56:02<3:28:19,  5.65s/it]INFO 04-11 23:32:54 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1090/3300 [56:09<3:44:02,  6.08s/it]2026-04-11 23:33:01,002 | INFO | train_grpo_train | metrics_logged mode=train step=1090

                                                     
{'loss': -0.0213, 'grad_norm': 1.876051902770996, 'learning_rate': 6.700000000000001e-06, 'num_tokens': 2459750.0, 'completions/mean_length': 271.25, 'completions/min_length': 258.0, 'completions/max_length': 288.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 271.25, 'completions/min_terminated_length': 258.0, 'completions/max_terminated_length': 288.0, 'rewards/meter/mean': 0.9913715124130249, 'rewards/meter/std': 0.014886174350976944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5326797366142273, 'rewards/repeat_penalty/std': 0.2576701045036316, 'rewards/total_composite/mean': 0.5263446569442749, 'rewards/total_composite/std': 0.2508489787578583, 'reward': 0.5263446569442749, 'reward_std': 0.2508489787578583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019940270110964775, 'sampling/sampling_logp_difference/max': 1.2876369953155518, 'sampling/importance_sampling_ratio/min': 0.275922030210495, 'sampling/importance_sampling_ratio/mean': 1.0003299713134766, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10771761322394013, 'clip_ratio/low_mean': 0.004208970727631822, 'clip_ratio/low_min': 0.004208970727631822, 'clip_ratio/high_mean': 0.012134980701375753, 'clip_ratio/high_max': 0.012134980701375753, 'clip_ratio/region_mean': 0.016343951429007575, 'reward_total_mean': 0.5263446569442749, 'reward_meter_mean': 0.9913715124130249, 'reward_meter_std': 0.014886174350976944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5326797366142273, 'reward_repeat_penalty_std': 0.2576701045036316, 'reward_total_composite_mean': 0.5263446569442749, 'reward_total_composite_std': 0.2508489787578583, 'epoch': 0.04}

 33%|███▎      | 1090/3300 [56:09<3:44:02,  6.08s/it]INFO 04-11 23:33:01 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1091/3300 [56:14<3:25:26,  5.58s/it]2026-04-11 23:33:05,411 | INFO | train_grpo_train | metrics_logged mode=train step=1091

                                                     
{'loss': -0.03, 'grad_norm': 6.454976558685303, 'learning_rate': 6.6969696969696975e-06, 'num_tokens': 2461500.0, 'completions/mean_length': 64.75, 'completions/min_length': 57.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5685036182403564, 'rewards/meter/std': 0.32084769010543823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5685036182403564, 'rewards/total_composite/std': 0.32084769010543823, 'reward': 0.5685036182403564, 'reward_std': 0.32084769010543823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.046030234545469284, 'sampling/sampling_logp_difference/max': 1.455164909362793, 'sampling/importance_sampling_ratio/min': 0.2333618849515915, 'sampling/importance_sampling_ratio/mean': 1.007980227470398, 'sampling/importance_sampling_ratio/max': 1.7851393222808838, 'entropy': 0.27551514096558094, 'clip_ratio/low_mean': 0.021997954230755568, 'clip_ratio/low_min': 0.021997954230755568, 'clip_ratio/high_mean': 0.005542142200283706, 'clip_ratio/high_max': 0.005542142200283706, 'clip_ratio/region_mean': 0.027540096431039274, 'reward_total_mean': 0.5685036182403564, 'reward_meter_mean': 0.5685036182403564, 'reward_meter_std': 0.32084769010543823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5685036182403564, 'reward_total_composite_std': 0.32084769010543823, 'epoch': 0.04}

 33%|███▎      | 1091/3300 [56:14<3:25:26,  5.58s/it]INFO 04-11 23:33:05 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1092/3300 [56:18<3:13:50,  5.27s/it]2026-04-11 23:33:09,947 | INFO | train_grpo_train | metrics_logged mode=train step=1092

                                                     
{'loss': 0.0095, 'grad_norm': 4.461249351501465, 'learning_rate': 6.693939393939395e-06, 'num_tokens': 2463457.0, 'completions/mean_length': 79.625, 'completions/min_length': 75.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.625, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.8799495697021484, 'rewards/meter/std': 0.1550467163324356, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8799495697021484, 'rewards/total_composite/std': 0.1550467163324356, 'reward': 0.8799495697021484, 'reward_std': 0.1550467312335968, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021693911403417587, 'sampling/sampling_logp_difference/max': 2.3354640007019043, 'sampling/importance_sampling_ratio/min': 0.09676557034254074, 'sampling/importance_sampling_ratio/mean': 1.0008015632629395, 'sampling/importance_sampling_ratio/max': 1.76300048828125, 'entropy': 0.07495173905044794, 'clip_ratio/low_mean': 0.0030868902103975415, 'clip_ratio/low_min': 0.0030868902103975415, 'clip_ratio/high_mean': 0.006234930478967726, 'clip_ratio/high_max': 0.006234930478967726, 'clip_ratio/region_mean': 0.009321820689365268, 'reward_total_mean': 0.8799495697021484, 'reward_meter_mean': 0.8799495697021484, 'reward_meter_std': 0.1550467163324356, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8799495697021484, 'reward_total_composite_std': 0.1550467163324356, 'epoch': 0.04}

 33%|███▎      | 1092/3300 [56:18<3:13:50,  5.27s/it]INFO 04-11 23:33:10 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1093/3300 [56:23<3:04:37,  5.02s/it]2026-04-11 23:33:14,388 | INFO | train_grpo_train | metrics_logged mode=train step=1093

                                                     
{'loss': 0.0004, 'grad_norm': 10.553791046142578, 'learning_rate': 6.690909090909091e-06, 'num_tokens': 2465181.0, 'completions/mean_length': 62.5, 'completions/min_length': 60.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.5, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8936691284179688, 'rewards/meter/std': 0.16734978556632996, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8936691284179688, 'rewards/total_composite/std': 0.16734978556632996, 'reward': 0.8936691284179688, 'reward_std': 0.16734978556632996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03492492809891701, 'sampling/sampling_logp_difference/max': 1.5639657974243164, 'sampling/importance_sampling_ratio/min': 0.20930436253547668, 'sampling/importance_sampling_ratio/mean': 1.0026055574417114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1802924033254385, 'clip_ratio/low_mean': 0.012298387009650469, 'clip_ratio/low_min': 0.012298387009650469, 'clip_ratio/high_mean': 0.021643388201482594, 'clip_ratio/high_max': 0.021643388201482594, 'clip_ratio/region_mean': 0.03394177521113306, 'reward_total_mean': 0.8936691284179688, 'reward_meter_mean': 0.8936691284179688, 'reward_meter_std': 0.16734978556632996, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8936691284179688, 'reward_total_composite_std': 0.16734978556632996, 'epoch': 0.04}

 33%|███▎      | 1093/3300 [56:23<3:04:37,  5.02s/it]INFO 04-11 23:33:14 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1094/3300 [56:28<3:07:47,  5.11s/it]2026-04-11 23:33:19,702 | INFO | train_grpo_train | metrics_logged mode=train step=1094

                                                     
{'loss': 0.0321, 'grad_norm': 2.7043662071228027, 'learning_rate': 6.687878787878788e-06, 'num_tokens': 2467653.0, 'completions/mean_length': 136.0, 'completions/min_length': 125.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.0, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.7940855622291565, 'rewards/meter/std': 0.11631763726472855, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6944444179534912, 'rewards/repeat_penalty/std': 0.05143444612622261, 'rewards/total_composite/mean': 0.41557225584983826, 'rewards/total_composite/std': 0.08242405205965042, 'reward': 0.41557225584983826, 'reward_std': 0.08242405205965042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01717483066022396, 'sampling/sampling_logp_difference/max': 1.4296681880950928, 'sampling/importance_sampling_ratio/min': 0.2393883466720581, 'sampling/importance_sampling_ratio/mean': 0.9989469051361084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07300170278176665, 'clip_ratio/low_mean': 0.0017985611921176314, 'clip_ratio/low_min': 0.0017985611921176314, 'clip_ratio/high_mean': 0.011392419459298253, 'clip_ratio/high_max': 0.011392419459298253, 'clip_ratio/region_mean': 0.013190980651415884, 'reward_total_mean': 0.41557225584983826, 'reward_meter_mean': 0.7940855622291565, 'reward_meter_std': 0.11631763726472855, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6944444179534912, 'reward_repeat_penalty_std': 0.05143444612622261, 'reward_total_composite_mean': 0.41557225584983826, 'reward_total_composite_std': 0.08242405205965042, 'epoch': 0.04}

 33%|███▎      | 1094/3300 [56:28<3:07:47,  5.11s/it]INFO 04-11 23:33:19 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1095/3300 [56:33<2:59:53,  4.90s/it]2026-04-11 23:33:24,112 | INFO | train_grpo_train | metrics_logged mode=train step=1095

                                                     
{'loss': -0.0163, 'grad_norm': 11.521310806274414, 'learning_rate': 6.684848484848485e-06, 'num_tokens': 2469179.0, 'completions/mean_length': 31.75, 'completions/min_length': 30.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8792948722839355, 'rewards/meter/std': 0.25036633014678955, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8792948722839355, 'rewards/total_composite/std': 0.25036633014678955, 'reward': 0.8792948722839355, 'reward_std': 0.25036633014678955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04188957437872887, 'sampling/sampling_logp_difference/max': 1.4607782363891602, 'sampling/importance_sampling_ratio/min': 0.23205561935901642, 'sampling/importance_sampling_ratio/mean': 0.9937551617622375, 'sampling/importance_sampling_ratio/max': 1.482624888420105, 'entropy': 0.16218565311282873, 'clip_ratio/low_mean': 0.008333333767950535, 'clip_ratio/low_min': 0.008333333767950535, 'clip_ratio/high_mean': 0.031376007944345474, 'clip_ratio/high_max': 0.031376007944345474, 'clip_ratio/region_mean': 0.03970934171229601, 'reward_total_mean': 0.8792948722839355, 'reward_meter_mean': 0.8792948722839355, 'reward_meter_std': 0.25036633014678955, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8792948722839355, 'reward_total_composite_std': 0.25036633014678955, 'epoch': 0.04}

 33%|███▎      | 1095/3300 [56:33<2:59:53,  4.90s/it]INFO 04-11 23:33:24 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1096/3300 [56:38<3:06:02,  5.06s/it]2026-04-11 23:33:29,560 | INFO | train_grpo_train | metrics_logged mode=train step=1096

                                                     
{'loss': 0.0349, 'grad_norm': 4.185685634613037, 'learning_rate': 6.681818181818183e-06, 'num_tokens': 2471195.0, 'completions/mean_length': 81.0, 'completions/min_length': 74.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.0, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.8235281705856323, 'rewards/meter/std': 0.1984855979681015, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.7183074951171875, 'rewards/total_composite/std': 0.19789229333400726, 'reward': 0.7183074951171875, 'reward_std': 0.19789229333400726, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03307102620601654, 'sampling/sampling_logp_difference/max': 1.3362131118774414, 'sampling/importance_sampling_ratio/min': 0.262839138507843, 'sampling/importance_sampling_ratio/mean': 0.9967372417449951, 'sampling/importance_sampling_ratio/max': 1.7130012512207031, 'entropy': 0.15196853037923574, 'clip_ratio/low_mean': 0.005780933075584471, 'clip_ratio/low_min': 0.005780933075584471, 'clip_ratio/high_mean': 0.019367096945643425, 'clip_ratio/high_max': 0.019367096945643425, 'clip_ratio/region_mean': 0.025148030021227896, 'reward_total_mean': 0.7183074951171875, 'reward_meter_mean': 0.8235281705856323, 'reward_meter_std': 0.1984855979681015, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.7183074951171875, 'reward_total_composite_std': 0.19789229333400726, 'epoch': 0.04}

 33%|███▎      | 1096/3300 [56:38<3:06:02,  5.06s/it]INFO 04-11 23:33:29 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1097/3300 [56:46<3:34:34,  5.84s/it]2026-04-11 23:33:37,226 | INFO | train_grpo_train | metrics_logged mode=train step=1097

                                                     
{'loss': -0.0005, 'grad_norm': 1.8824577331542969, 'learning_rate': 6.678787878787879e-06, 'num_tokens': 2475359.0, 'completions/mean_length': 304.5, 'completions/min_length': 280.0, 'completions/max_length': 320.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 304.5, 'completions/min_terminated_length': 280.0, 'completions/max_terminated_length': 320.0, 'rewards/meter/mean': 0.953475832939148, 'rewards/meter/std': 0.11951109021902084, 'rewards/count_adherence/mean': 0.7403846383094788, 'rewards/count_adherence/std': 0.039811473339796066, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6966374516487122, 'rewards/repeat_penalty/std': 0.09851117432117462, 'rewards/total_composite/mean': 0.49607303738594055, 'rewards/total_composite/std': 0.11844708025455475, 'reward': 0.49607303738594055, 'reward_std': 0.11844708025455475, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027330797165632248, 'sampling/sampling_logp_difference/max': 2.8147997856140137, 'sampling/importance_sampling_ratio/min': 0.05991671606898308, 'sampling/importance_sampling_ratio/mean': 1.0036851167678833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16177368070930243, 'clip_ratio/low_mean': 0.010694848489947617, 'clip_ratio/low_min': 0.010694848489947617, 'clip_ratio/high_mean': 0.01060027233324945, 'clip_ratio/high_max': 0.01060027233324945, 'clip_ratio/region_mean': 0.021295120823197067, 'reward_total_mean': 0.49607303738594055, 'reward_meter_mean': 0.953475832939148, 'reward_meter_std': 0.11951109021902084, 'reward_count_adherence_mean': 0.7403846383094788, 'reward_count_adherence_std': 0.039811473339796066, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6966374516487122, 'reward_repeat_penalty_std': 0.09851117432117462, 'reward_total_composite_mean': 0.49607303738594055, 'reward_total_composite_std': 0.11844708025455475, 'epoch': 0.04}

 33%|███▎      | 1097/3300 [56:46<3:34:34,  5.84s/it]INFO 04-11 23:33:37 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1098/3300 [56:50<3:20:21,  5.46s/it]2026-04-11 23:33:41,786 | INFO | train_grpo_train | metrics_logged mode=train step=1098

                                                     
{'loss': 0.0173, 'grad_norm': 6.93865966796875, 'learning_rate': 6.6757575757575766e-06, 'num_tokens': 2477124.0, 'completions/mean_length': 61.625, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.876738965511322, 'rewards/meter/std': 0.2505471706390381, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.876738965511322, 'rewards/total_composite/std': 0.2505471706390381, 'reward': 0.876738965511322, 'reward_std': 0.2505471408367157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026354925706982613, 'sampling/sampling_logp_difference/max': 1.0480327606201172, 'sampling/importance_sampling_ratio/min': 0.3506268262863159, 'sampling/importance_sampling_ratio/mean': 1.0015573501586914, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08917439868673682, 'clip_ratio/low_mean': 0.0019841270986944437, 'clip_ratio/low_min': 0.0019841270986944437, 'clip_ratio/high_mean': 0.014381667831912637, 'clip_ratio/high_max': 0.014381667831912637, 'clip_ratio/region_mean': 0.01636579493060708, 'reward_total_mean': 0.876738965511322, 'reward_meter_mean': 0.876738965511322, 'reward_meter_std': 0.2505471706390381, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.876738965511322, 'reward_total_composite_std': 0.2505471706390381, 'epoch': 0.04}

 33%|███▎      | 1098/3300 [56:50<3:20:21,  5.46s/it]INFO 04-11 23:33:42 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1099/3300 [56:57<3:29:17,  5.71s/it]2026-04-11 23:33:48,066 | INFO | train_grpo_train | metrics_logged mode=train step=1099

                                                     
{'loss': -0.0507, 'grad_norm': 3.1013591289520264, 'learning_rate': 6.672727272727273e-06, 'num_tokens': 2480198.0, 'completions/mean_length': 205.25, 'completions/min_length': 171.0, 'completions/max_length': 224.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 205.25, 'completions/min_terminated_length': 171.0, 'completions/max_terminated_length': 224.0, 'rewards/meter/mean': 0.7326644062995911, 'rewards/meter/std': 0.39641380310058594, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6022727489471436, 'rewards/repeat_penalty/std': 0.14527180790901184, 'rewards/total_composite/mean': 0.373015820980072, 'rewards/total_composite/std': 0.22151786088943481, 'reward': 0.373015820980072, 'reward_std': 0.22151786088943481, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017836585640907288, 'sampling/sampling_logp_difference/max': 3.5347535610198975, 'sampling/importance_sampling_ratio/min': 0.02916594408452511, 'sampling/importance_sampling_ratio/mean': 1.0023084878921509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09231163933873177, 'clip_ratio/low_mean': 0.0041160593973472714, 'clip_ratio/low_min': 0.0041160593973472714, 'clip_ratio/high_mean': 0.009313070855569094, 'clip_ratio/high_max': 0.009313070855569094, 'clip_ratio/region_mean': 0.013429130252916366, 'reward_total_mean': 0.373015820980072, 'reward_meter_mean': 0.7326644062995911, 'reward_meter_std': 0.39641380310058594, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6022727489471436, 'reward_repeat_penalty_std': 0.14527180790901184, 'reward_total_composite_mean': 0.373015820980072, 'reward_total_composite_std': 0.22151786088943481, 'epoch': 0.04}

 33%|███▎      | 1099/3300 [56:57<3:29:17,  5.71s/it]INFO 04-11 23:33:48 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1100/3300 [57:04<3:52:37,  6.34s/it]2026-04-11 23:33:55,901 | INFO | train_grpo_train | metrics_logged mode=train step=1100

                                                     
{'loss': -0.0248, 'grad_norm': 0.9629241228103638, 'learning_rate': 6.66969696969697e-06, 'num_tokens': 2484623.0, 'completions/mean_length': 327.125, 'completions/min_length': 302.0, 'completions/max_length': 345.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 327.125, 'completions/min_terminated_length': 302.0, 'completions/max_terminated_length': 345.0, 'rewards/meter/mean': 0.993415355682373, 'rewards/meter/std': 0.007653866894543171, 'rewards/count_adherence/mean': 0.7291666865348816, 'rewards/count_adherence/std': 0.038575831800699234, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5248161554336548, 'rewards/repeat_penalty/std': 0.21153290569782257, 'rewards/total_composite/mean': 0.386793315410614, 'rewards/total_composite/std': 0.16382494568824768, 'reward': 0.386793315410614, 'reward_std': 0.16382494568824768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010836898349225521, 'sampling/sampling_logp_difference/max': 1.045947790145874, 'sampling/importance_sampling_ratio/min': 0.35135865211486816, 'sampling/importance_sampling_ratio/mean': 1.001177430152893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06283333199098706, 'clip_ratio/low_mean': 0.001618139911442995, 'clip_ratio/low_min': 0.001618139911442995, 'clip_ratio/high_mean': 0.010039791552117094, 'clip_ratio/high_max': 0.010039791552117094, 'clip_ratio/region_mean': 0.01165793146356009, 'reward_total_mean': 0.386793315410614, 'reward_meter_mean': 0.993415355682373, 'reward_meter_std': 0.007653866894543171, 'reward_count_adherence_mean': 0.7291666865348816, 'reward_count_adherence_std': 0.038575831800699234, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5248161554336548, 'reward_repeat_penalty_std': 0.21153290569782257, 'reward_total_composite_mean': 0.386793315410614, 'reward_total_composite_std': 0.16382494568824768, 'epoch': 0.04}

 33%|███▎      | 1100/3300 [57:04<3:52:37,  6.34s/it]INFO 04-11 23:33:56 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:23,  2.18s/it]

 23%|██▎       | 3/13 [00:07<00:26,  2.61s/it]

 31%|███       | 4/13 [00:12<00:30,  3.43s/it]

 38%|███▊      | 5/13 [00:16<00:30,  3.83s/it]

 46%|████▌     | 6/13 [00:22<00:31,  4.43s/it]

 54%|█████▍    | 7/13 [00:27<00:26,  4.46s/it]

 62%|██████▏   | 8/13 [00:33<00:24,  4.97s/it]

 69%|██████▉   | 9/13 [00:39<00:20,  5.24s/it]

 77%|███████▋  | 10/13 [00:43<00:15,  5.08s/it]

 85%|████████▍ | 11/13 [00:48<00:09,  4.89s/it]

 92%|█████████▏| 12/13 [00:53<00:04,  5.00s/it]

100%|██████████| 13/13 [00:58<00:00,  5.00s/it]2026-04-11 23:34:59,623 | INFO | train_grpo_train | metrics_logged mode=eval step=1100

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 63.6923, 'eval_samples_per_second': 1.633, 'eval_steps_per_second': 0.204, 'eval_num_tokens': 2484623.0, 'eval_completions/mean_length': 188.08653846153845, 'eval_completions/min_length': 57.92307692307692, 'eval_completions/max_length': 328.61538461538464, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 188.08653846153845, 'eval_completions/min_terminated_length': 57.92307692307692, 'eval_completions/max_terminated_length': 328.61538461538464, 'eval_rewards/meter/mean': 0.6326167858563937, 'eval_rewards/meter/std': 0.35957076343206257, 'eval_rewards/count_adherence/mean': 0.8649956675676199, 'eval_rewards/count_adherence/std': 0.12389520039925209, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.6625011425751907, 'eval_rewards/repeat_penalty/std': 0.24176405255611128, 'eval_rewards/total_composite/mean': 0.385673477099492, 'eval_rewards/total_composite/std': 0.3103876068041875, 'eval_reward': 0.385673477099492, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.008682384119870571, 'eval_sampling/sampling_logp_difference/max': 0.9079764164411105, 'eval_sampling/importance_sampling_ratio/min': 0.4163260803772853, 'eval_sampling/importance_sampling_ratio/mean': 1.0015292717860296, 'eval_sampling/importance_sampling_ratio/max': 1.3779725111447847, 'eval_entropy': 0.07903132902888152, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.385673477099492, 'eval_reward_meter_mean': 0.6326167858563937, 'eval_reward_meter_std': 0.35957076343206257, 'eval_reward_count_adherence_mean': 0.8649956675676199, 'eval_reward_count_adherence_std': 0.12389520039925209, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.6625011425751907, 'eval_reward_repeat_penalty_std': 0.24176405255611128, 'eval_reward_total_composite_mean': 0.385673477099492, 'eval_reward_total_composite_std': 0.3103876068041875, 'epoch': 0.04}

 33%|███▎      | 1100/3300 [58:08<3:52:37,  6.34s/it]

100%|██████████| 13/13 [00:58<00:00,  5.00s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 33%|███▎      | 1101/3300 [58:16<15:45:42, 25.80s/it]2026-04-11 23:35:07,116 | INFO | train_grpo_train | metrics_logged mode=train step=1101

                                                      
{'loss': 0.0008, 'grad_norm': 7.680727958679199, 'learning_rate': 6.666666666666667e-06, 'num_tokens': 2486422.0, 'completions/mean_length': 55.875, 'completions/min_length': 50.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8475104570388794, 'rewards/meter/std': 0.17176933586597443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8475104570388794, 'rewards/total_composite/std': 0.17176933586597443, 'reward': 0.8475104570388794, 'reward_std': 0.17176933586597443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05723501741886139, 'sampling/sampling_logp_difference/max': 2.4088823795318604, 'sampling/importance_sampling_ratio/min': 0.08991573750972748, 'sampling/importance_sampling_ratio/mean': 1.0003966093063354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22917167469859123, 'clip_ratio/low_mean': 0.01142857177183032, 'clip_ratio/low_min': 0.01142857177183032, 'clip_ratio/high_mean': 0.02208862197585404, 'clip_ratio/high_max': 0.02208862197585404, 'clip_ratio/region_mean': 0.03351719374768436, 'reward_total_mean': 0.8475104570388794, 'reward_meter_mean': 0.8475104570388794, 'reward_meter_std': 0.17176933586597443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8475104570388794, 'reward_total_composite_std': 0.17176933586597443, 'epoch': 0.04}

 33%|███▎      | 1101/3300 [58:16<15:45:42, 25.80s/it]INFO 04-11 23:35:07 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1102/3300 [58:24<12:30:04, 20.48s/it]2026-04-11 23:35:15,155 | INFO | train_grpo_train | metrics_logged mode=train step=1102

                                                      
{'loss': -0.0479, 'grad_norm': 9.668932914733887, 'learning_rate': 6.663636363636365e-06, 'num_tokens': 2490583.0, 'completions/mean_length': 296.125, 'completions/min_length': 244.0, 'completions/max_length': 335.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 296.125, 'completions/min_terminated_length': 244.0, 'completions/max_terminated_length': 335.0, 'rewards/meter/mean': 0.7127225399017334, 'rewards/meter/std': 0.37384799122810364, 'rewards/count_adherence/mean': 0.9722222089767456, 'rewards/count_adherence/std': 0.05143444985151291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5183823704719543, 'rewards/repeat_penalty/std': 0.17379139363765717, 'rewards/total_composite/mean': 0.37116703391075134, 'rewards/total_composite/std': 0.24903467297554016, 'reward': 0.37116703391075134, 'reward_std': 0.24903464317321777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014599828980863094, 'sampling/sampling_logp_difference/max': 4.119277000427246, 'sampling/importance_sampling_ratio/min': 0.016256263479590416, 'sampling/importance_sampling_ratio/mean': 1.0006206035614014, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.054934965912252665, 'clip_ratio/low_mean': 0.004965955973602831, 'clip_ratio/low_min': 0.004965955973602831, 'clip_ratio/high_mean': 0.007919449737528339, 'clip_ratio/high_max': 0.007919449737528339, 'clip_ratio/region_mean': 0.01288540571113117, 'reward_total_mean': 0.37116703391075134, 'reward_meter_mean': 0.7127225399017334, 'reward_meter_std': 0.37384799122810364, 'reward_count_adherence_mean': 0.9722222089767456, 'reward_count_adherence_std': 0.05143444985151291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5183823704719543, 'reward_repeat_penalty_std': 0.17379139363765717, 'reward_total_composite_mean': 0.37116703391075134, 'reward_total_composite_std': 0.24903467297554016, 'epoch': 0.04}

 33%|███▎      | 1102/3300 [58:24<12:30:04, 20.48s/it]INFO 04-11 23:35:15 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1103/3300 [58:29<9:39:28, 15.83s/it] 2026-04-11 23:35:20,128 | INFO | train_grpo_train | metrics_logged mode=train step=1103

                                                     
{'loss': -0.0174, 'grad_norm': 3.663283586502075, 'learning_rate': 6.660606060606061e-06, 'num_tokens': 2492366.0, 'completions/mean_length': 66.875, 'completions/min_length': 64.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.49698805809020996, 'rewards/meter/std': 0.368730753660202, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.49698805809020996, 'rewards/total_composite/std': 0.368730753660202, 'reward': 0.49698805809020996, 'reward_std': 0.3687307834625244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02970782294869423, 'sampling/sampling_logp_difference/max': 1.8465352058410645, 'sampling/importance_sampling_ratio/min': 0.25461694598197937, 'sampling/importance_sampling_ratio/mean': 0.9964687824249268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10305365733802319, 'clip_ratio/low_mean': 0.005657031899318099, 'clip_ratio/low_min': 0.005657031899318099, 'clip_ratio/high_mean': 0.012646116432733834, 'clip_ratio/high_max': 0.012646116432733834, 'clip_ratio/region_mean': 0.018303148332051933, 'reward_total_mean': 0.49698805809020996, 'reward_meter_mean': 0.49698805809020996, 'reward_meter_std': 0.368730753660202, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.49698805809020996, 'reward_total_composite_std': 0.368730753660202, 'epoch': 0.04}

 33%|███▎      | 1103/3300 [58:29<9:39:28, 15.83s/it]INFO 04-11 23:35:20 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1104/3300 [58:34<7:44:23, 12.69s/it]2026-04-11 23:35:25,516 | INFO | train_grpo_train | metrics_logged mode=train step=1104

                                                     
{'loss': -0.0068, 'grad_norm': 7.544509410858154, 'learning_rate': 6.657575757575758e-06, 'num_tokens': 2494079.0, 'completions/mean_length': 51.125, 'completions/min_length': 47.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6021931171417236, 'rewards/meter/std': 0.43061062693595886, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6021931171417236, 'rewards/total_composite/std': 0.43061062693595886, 'reward': 0.6021931171417236, 'reward_std': 0.4306105971336365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05377590283751488, 'sampling/sampling_logp_difference/max': 1.3904236555099487, 'sampling/importance_sampling_ratio/min': 0.24896980822086334, 'sampling/importance_sampling_ratio/mean': 0.9914664030075073, 'sampling/importance_sampling_ratio/max': 1.6187946796417236, 'entropy': 0.18264612928032875, 'clip_ratio/low_mean': 0.01027012919075787, 'clip_ratio/low_min': 0.01027012919075787, 'clip_ratio/high_mean': 0.028512453194707632, 'clip_ratio/high_max': 0.028512453194707632, 'clip_ratio/region_mean': 0.0387825823854655, 'reward_total_mean': 0.6021931171417236, 'reward_meter_mean': 0.6021931171417236, 'reward_meter_std': 0.43061062693595886, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6021931171417236, 'reward_total_composite_std': 0.43061062693595886, 'epoch': 0.04}

 33%|███▎      | 1104/3300 [58:34<7:44:23, 12.69s/it]INFO 04-11 23:35:25 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1105/3300 [58:39<6:21:57, 10.44s/it]2026-04-11 23:35:30,708 | INFO | train_grpo_train | metrics_logged mode=train step=1105

                                                     
{'loss': 0.0134, 'grad_norm': 8.104043960571289, 'learning_rate': 6.654545454545455e-06, 'num_tokens': 2495813.0, 'completions/mean_length': 51.75, 'completions/min_length': 48.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8142023086547852, 'rewards/meter/std': 0.21253840625286102, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8142023086547852, 'rewards/total_composite/std': 0.21253840625286102, 'reward': 0.8142023086547852, 'reward_std': 0.21253842115402222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06759097427129745, 'sampling/sampling_logp_difference/max': 1.6661019325256348, 'sampling/importance_sampling_ratio/min': 0.188982293009758, 'sampling/importance_sampling_ratio/mean': 1.0077879428863525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2174948314204812, 'clip_ratio/low_mean': 0.00491898157633841, 'clip_ratio/low_min': 0.00491898157633841, 'clip_ratio/high_mean': 0.035976887214928865, 'clip_ratio/high_max': 0.035976887214928865, 'clip_ratio/region_mean': 0.040895868791267276, 'reward_total_mean': 0.8142023086547852, 'reward_meter_mean': 0.8142023086547852, 'reward_meter_std': 0.21253840625286102, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8142023086547852, 'reward_total_composite_std': 0.21253840625286102, 'epoch': 0.04}

 33%|███▎      | 1105/3300 [58:39<6:21:57, 10.44s/it]INFO 04-11 23:35:30 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1106/3300 [58:47<5:49:16,  9.55s/it]2026-04-11 23:35:38,171 | INFO | train_grpo_train | metrics_logged mode=train step=1106

                                                     
{'loss': -0.0234, 'grad_norm': 1.9001294374465942, 'learning_rate': 6.651515151515152e-06, 'num_tokens': 2499568.0, 'completions/mean_length': 275.375, 'completions/min_length': 256.0, 'completions/max_length': 301.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 275.375, 'completions/min_terminated_length': 256.0, 'completions/max_terminated_length': 301.0, 'rewards/meter/mean': 0.8482787013053894, 'rewards/meter/std': 0.33679211139678955, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6220238208770752, 'rewards/repeat_penalty/std': 0.03127124905586243, 'rewards/total_composite/mean': 0.46502256393432617, 'rewards/total_composite/std': 0.1884276121854782, 'reward': 0.46502256393432617, 'reward_std': 0.1884276121854782, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012011334300041199, 'sampling/sampling_logp_difference/max': 1.7623469829559326, 'sampling/importance_sampling_ratio/min': 0.171641543507576, 'sampling/importance_sampling_ratio/mean': 1.0009040832519531, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05004570330493152, 'clip_ratio/low_mean': 0.00146484375, 'clip_ratio/low_min': 0.00146484375, 'clip_ratio/high_mean': 0.00831845449283719, 'clip_ratio/high_max': 0.00831845449283719, 'clip_ratio/region_mean': 0.00978329824283719, 'reward_total_mean': 0.46502256393432617, 'reward_meter_mean': 0.8482787013053894, 'reward_meter_std': 0.33679211139678955, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6220238208770752, 'reward_repeat_penalty_std': 0.03127124905586243, 'reward_total_composite_mean': 0.46502256393432617, 'reward_total_composite_std': 0.1884276121854782, 'epoch': 0.04}

 34%|███▎      | 1106/3300 [58:47<5:49:16,  9.55s/it]INFO 04-11 23:35:38 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1107/3300 [58:51<4:51:46,  7.98s/it]2026-04-11 23:35:42,493 | INFO | train_grpo_train | metrics_logged mode=train step=1107

                                                     
{'loss': 0.0015, 'grad_norm': 2.1617753505706787, 'learning_rate': 6.6484848484848485e-06, 'num_tokens': 2501018.0, 'completions/mean_length': 38.25, 'completions/min_length': 38.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.92488694190979, 'rewards/meter/std': 0.09670114517211914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.92488694190979, 'rewards/total_composite/std': 0.09670114517211914, 'reward': 0.92488694190979, 'reward_std': 0.09670114517211914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012348960153758526, 'sampling/sampling_logp_difference/max': 0.6579999923706055, 'sampling/importance_sampling_ratio/min': 0.5178860425949097, 'sampling/importance_sampling_ratio/mean': 1.0063523054122925, 'sampling/importance_sampling_ratio/max': 1.7915105819702148, 'entropy': 0.07856028527021408, 'clip_ratio/low_mean': 0.003289473708719015, 'clip_ratio/low_min': 0.003289473708719015, 'clip_ratio/high_mean': 0.009699730202555656, 'clip_ratio/high_max': 0.009699730202555656, 'clip_ratio/region_mean': 0.012989203911274672, 'reward_total_mean': 0.92488694190979, 'reward_meter_mean': 0.92488694190979, 'reward_meter_std': 0.09670114517211914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.92488694190979, 'reward_total_composite_std': 0.09670114517211914, 'epoch': 0.04}

 34%|███▎      | 1107/3300 [58:51<4:51:46,  7.98s/it]INFO 04-11 23:35:42 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1108/3300 [58:56<4:16:42,  7.03s/it]2026-04-11 23:35:47,288 | INFO | train_grpo_train | metrics_logged mode=train step=1108

                                                     
{'loss': -0.0307, 'grad_norm': 6.412082672119141, 'learning_rate': 6.645454545454546e-06, 'num_tokens': 2503004.0, 'completions/mean_length': 79.25, 'completions/min_length': 72.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.4985108971595764, 'rewards/meter/std': 0.402507483959198, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.17728103697299957, 'rewards/total_composite/mean': 0.39024245738983154, 'rewards/total_composite/std': 0.367125540971756, 'reward': 0.39024245738983154, 'reward_std': 0.3671255111694336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043393637984991074, 'sampling/sampling_logp_difference/max': 2.8125345706939697, 'sampling/importance_sampling_ratio/min': 0.06005259230732918, 'sampling/importance_sampling_ratio/mean': 1.0030810832977295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2001291485503316, 'clip_ratio/low_mean': 0.014691225020214915, 'clip_ratio/low_min': 0.014691225020214915, 'clip_ratio/high_mean': 0.015104167046956718, 'clip_ratio/high_max': 0.015104167046956718, 'clip_ratio/region_mean': 0.029795392067171633, 'reward_total_mean': 0.39024245738983154, 'reward_meter_mean': 0.4985108971595764, 'reward_meter_std': 0.402507483959198, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.17728103697299957, 'reward_total_composite_mean': 0.39024245738983154, 'reward_total_composite_std': 0.367125540971756, 'epoch': 0.04}

 34%|███▎      | 1108/3300 [58:56<4:16:42,  7.03s/it]INFO 04-11 23:35:47 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1109/3300 [59:00<3:47:26,  6.23s/it]2026-04-11 23:35:51,654 | INFO | train_grpo_train | metrics_logged mode=train step=1109

                                                     
{'loss': -0.025, 'grad_norm': 8.44660472869873, 'learning_rate': 6.642424242424242e-06, 'num_tokens': 2504551.0, 'completions/mean_length': 47.375, 'completions/min_length': 42.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9353389143943787, 'rewards/meter/std': 0.06214950606226921, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9353389143943787, 'rewards/total_composite/std': 0.06214950606226921, 'reward': 0.9353389143943787, 'reward_std': 0.062149498611688614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043186187744140625, 'sampling/sampling_logp_difference/max': 1.5208110809326172, 'sampling/importance_sampling_ratio/min': 0.21853457391262054, 'sampling/importance_sampling_ratio/mean': 0.9965187907218933, 'sampling/importance_sampling_ratio/max': 1.534562110900879, 'entropy': 0.19667030405253172, 'clip_ratio/low_mean': 0.01064311619848013, 'clip_ratio/low_min': 0.01064311619848013, 'clip_ratio/high_mean': 0.02806122461333871, 'clip_ratio/high_max': 0.02806122461333871, 'clip_ratio/region_mean': 0.03870434081181884, 'reward_total_mean': 0.9353389143943787, 'reward_meter_mean': 0.9353389143943787, 'reward_meter_std': 0.06214950606226921, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9353389143943787, 'reward_total_composite_std': 0.06214950606226921, 'epoch': 0.04}

 34%|███▎      | 1109/3300 [59:00<3:47:26,  6.23s/it]INFO 04-11 23:35:51 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1110/3300 [59:06<3:40:47,  6.05s/it]2026-04-11 23:35:57,284 | INFO | train_grpo_train | metrics_logged mode=train step=1110

                                                     
{'loss': -0.022, 'grad_norm': 4.039949893951416, 'learning_rate': 6.63939393939394e-06, 'num_tokens': 2507175.0, 'completions/mean_length': 150.0, 'completions/min_length': 139.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.0, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9286544322967529, 'rewards/meter/std': 0.16951557993888855, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.5778908729553223, 'rewards/total_composite/std': 0.1044817790389061, 'reward': 0.5778908729553223, 'reward_std': 0.1044817790389061, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038920193910598755, 'sampling/sampling_logp_difference/max': 1.9175208806991577, 'sampling/importance_sampling_ratio/min': 0.14697086811065674, 'sampling/importance_sampling_ratio/mean': 0.9992915987968445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17426198534667492, 'clip_ratio/low_mean': 0.009535853168927133, 'clip_ratio/low_min': 0.009535853168927133, 'clip_ratio/high_mean': 0.024372577434405684, 'clip_ratio/high_max': 0.024372577434405684, 'clip_ratio/region_mean': 0.03390843060333282, 'reward_total_mean': 0.5778908729553223, 'reward_meter_mean': 0.9286544322967529, 'reward_meter_std': 0.16951557993888855, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.5778908729553223, 'reward_total_composite_std': 0.1044817790389061, 'epoch': 0.04}

 34%|███▎      | 1110/3300 [59:06<3:40:47,  6.05s/it]INFO 04-11 23:35:57 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1111/3300 [59:10<3:24:59,  5.62s/it]2026-04-11 23:36:01,900 | INFO | train_grpo_train | metrics_logged mode=train step=1111

                                                     
{'loss': 0.0099, 'grad_norm': 3.789764881134033, 'learning_rate': 6.6363636363636375e-06, 'num_tokens': 2508931.0, 'completions/mean_length': 73.5, 'completions/min_length': 73.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9711061716079712, 'rewards/meter/std': 0.02328452095389366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9711061716079712, 'rewards/total_composite/std': 0.02328452095389366, 'reward': 0.9711061716079712, 'reward_std': 0.023284511640667915, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011047722771763802, 'sampling/sampling_logp_difference/max': 1.3545985221862793, 'sampling/importance_sampling_ratio/min': 0.2580508887767792, 'sampling/importance_sampling_ratio/mean': 1.0022356510162354, 'sampling/importance_sampling_ratio/max': 1.7336740493774414, 'entropy': 0.03971482953056693, 'clip_ratio/low_mean': 0.0050675676902756095, 'clip_ratio/low_min': 0.0050675676902756095, 'clip_ratio/high_mean': 0.0051369862630963326, 'clip_ratio/high_max': 0.0051369862630963326, 'clip_ratio/region_mean': 0.010204553953371942, 'reward_total_mean': 0.9711061716079712, 'reward_meter_mean': 0.9711061716079712, 'reward_meter_std': 0.02328452095389366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9711061716079712, 'reward_total_composite_std': 0.02328452095389366, 'epoch': 0.04}

 34%|███▎      | 1111/3300 [59:10<3:24:59,  5.62s/it]INFO 04-11 23:36:02 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1112/3300 [59:15<3:17:44,  5.42s/it]2026-04-11 23:36:06,885 | INFO | train_grpo_train | metrics_logged mode=train step=1112

                                                     
{'loss': -0.0457, 'grad_norm': 0.8946425914764404, 'learning_rate': 6.633333333333334e-06, 'num_tokens': 2511051.0, 'completions/mean_length': 106.0, 'completions/min_length': 92.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.0, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9727685451507568, 'rewards/meter/std': 0.053206540644168854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7782148122787476, 'rewards/total_composite/std': 0.04256521537899971, 'reward': 0.7782148122787476, 'reward_std': 0.04256521537899971, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0040269712917506695, 'sampling/sampling_logp_difference/max': 0.56025230884552, 'sampling/importance_sampling_ratio/min': 0.5710649490356445, 'sampling/importance_sampling_ratio/mean': 0.9997245073318481, 'sampling/importance_sampling_ratio/max': 1.4477814435958862, 'entropy': 0.01866412186063826, 'clip_ratio/low_mean': 0.0027173913549631834, 'clip_ratio/low_min': 0.0027173913549631834, 'clip_ratio/high_mean': 0.0034722222480922937, 'clip_ratio/high_max': 0.0034722222480922937, 'clip_ratio/region_mean': 0.006189613603055477, 'reward_total_mean': 0.7782148122787476, 'reward_meter_mean': 0.9727685451507568, 'reward_meter_std': 0.053206540644168854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7782148122787476, 'reward_total_composite_std': 0.04256521537899971, 'epoch': 0.04}

 34%|███▎      | 1112/3300 [59:15<3:17:44,  5.42s/it]INFO 04-11 23:36:07 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1113/3300 [59:19<3:03:12,  5.03s/it]2026-04-11 23:36:10,966 | INFO | train_grpo_train | metrics_logged mode=train step=1113

                                                     
{'loss': -0.0282, 'grad_norm': 6.619121074676514, 'learning_rate': 6.630303030303031e-06, 'num_tokens': 2512391.0, 'completions/mean_length': 23.5, 'completions/min_length': 23.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.5, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.867714524269104, 'rewards/meter/std': 0.052471715956926346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.867714524269104, 'rewards/total_composite/std': 0.052471715956926346, 'reward': 0.867714524269104, 'reward_std': 0.052471719682216644, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024627521634101868, 'sampling/sampling_logp_difference/max': 0.767571210861206, 'sampling/importance_sampling_ratio/min': 0.4641389846801758, 'sampling/importance_sampling_ratio/mean': 1.008794903755188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08848860999569297, 'clip_ratio/low_mean': 0.010869565419852734, 'clip_ratio/low_min': 0.010869565419852734, 'clip_ratio/high_mean': 0.010064412374049425, 'clip_ratio/high_max': 0.010064412374049425, 'clip_ratio/region_mean': 0.02093397779390216, 'reward_total_mean': 0.867714524269104, 'reward_meter_mean': 0.867714524269104, 'reward_meter_std': 0.052471715956926346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.867714524269104, 'reward_total_composite_std': 0.052471715956926346, 'epoch': 0.04}

 34%|███▎      | 1113/3300 [59:19<3:03:12,  5.03s/it]INFO 04-11 23:36:11 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1114/3300 [59:24<2:55:32,  4.82s/it]2026-04-11 23:36:15,298 | INFO | train_grpo_train | metrics_logged mode=train step=1114

                                                     
{'loss': -0.0256, 'grad_norm': 11.225849151611328, 'learning_rate': 6.627272727272728e-06, 'num_tokens': 2513999.0, 'completions/mean_length': 43.0, 'completions/min_length': 41.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8079327940940857, 'rewards/meter/std': 0.1427396684885025, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8079327940940857, 'rewards/total_composite/std': 0.1427396684885025, 'reward': 0.8079327940940857, 'reward_std': 0.1427396684885025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04281027242541313, 'sampling/sampling_logp_difference/max': 0.9714531898498535, 'sampling/importance_sampling_ratio/min': 0.3785325586795807, 'sampling/importance_sampling_ratio/mean': 1.0110169649124146, 'sampling/importance_sampling_ratio/max': 1.9035590887069702, 'entropy': 0.15364955179393291, 'clip_ratio/low_mean': 0.014953542733564973, 'clip_ratio/low_min': 0.014953542733564973, 'clip_ratio/high_mean': 0.017787929391488433, 'clip_ratio/high_max': 0.017787929391488433, 'clip_ratio/region_mean': 0.032741472125053406, 'reward_total_mean': 0.8079327940940857, 'reward_meter_mean': 0.8079327940940857, 'reward_meter_std': 0.1427396684885025, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8079327940940857, 'reward_total_composite_std': 0.1427396684885025, 'epoch': 0.04}

 34%|███▍      | 1114/3300 [59:24<2:55:32,  4.82s/it]INFO 04-11 23:36:15 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1115/3300 [59:28<2:53:01,  4.75s/it]2026-04-11 23:36:19,897 | INFO | train_grpo_train | metrics_logged mode=train step=1115

                                                     
{'loss': -0.0173, 'grad_norm': 7.548732757568359, 'learning_rate': 6.624242424242425e-06, 'num_tokens': 2515720.0, 'completions/mean_length': 49.125, 'completions/min_length': 44.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7826473712921143, 'rewards/meter/std': 0.3253748416900635, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7826473712921143, 'rewards/total_composite/std': 0.3253748416900635, 'reward': 0.7826473712921143, 'reward_std': 0.3253748118877411, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051869362592697144, 'sampling/sampling_logp_difference/max': 1.9477629661560059, 'sampling/importance_sampling_ratio/min': 0.14259269833564758, 'sampling/importance_sampling_ratio/mean': 1.001320719718933, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1894742762669921, 'clip_ratio/low_mean': 0.00828598509542644, 'clip_ratio/low_min': 0.00828598509542644, 'clip_ratio/high_mean': 0.029880503891035914, 'clip_ratio/high_max': 0.029880503891035914, 'clip_ratio/region_mean': 0.038166488986462355, 'reward_total_mean': 0.7826473712921143, 'reward_meter_mean': 0.7826473712921143, 'reward_meter_std': 0.3253748416900635, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7826473712921143, 'reward_total_composite_std': 0.3253748416900635, 'epoch': 0.04}

 34%|███▍      | 1115/3300 [59:28<2:53:01,  4.75s/it]INFO 04-11 23:36:20 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1116/3300 [59:33<2:52:32,  4.74s/it]2026-04-11 23:36:24,608 | INFO | train_grpo_train | metrics_logged mode=train step=1116

                                                     
{'loss': 0.0026, 'grad_norm': 8.76555347442627, 'learning_rate': 6.621212121212121e-06, 'num_tokens': 2517481.0, 'completions/mean_length': 73.125, 'completions/min_length': 73.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9871143698692322, 'rewards/meter/std': 0.010886327363550663, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9871143698692322, 'rewards/total_composite/std': 0.010886327363550663, 'reward': 0.9871143698692322, 'reward_std': 0.010886335745453835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011934218928217888, 'sampling/sampling_logp_difference/max': 1.53245210647583, 'sampling/importance_sampling_ratio/min': 0.2160053551197052, 'sampling/importance_sampling_ratio/mean': 0.997620165348053, 'sampling/importance_sampling_ratio/max': 1.276580810546875, 'entropy': 0.03595073730684817, 'clip_ratio/low_mean': 0.0017123287543654442, 'clip_ratio/low_min': 0.0017123287543654442, 'clip_ratio/high_mean': 0.0034246575087308884, 'clip_ratio/high_max': 0.0034246575087308884, 'clip_ratio/region_mean': 0.0051369862630963326, 'reward_total_mean': 0.9871143698692322, 'reward_meter_mean': 0.9871143698692322, 'reward_meter_std': 0.010886327363550663, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9871143698692322, 'reward_total_composite_std': 0.010886327363550663, 'epoch': 0.04}

 34%|███▍      | 1116/3300 [59:33<2:52:32,  4.74s/it]INFO 04-11 23:36:24 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1117/3300 [59:41<3:29:03,  5.75s/it]2026-04-11 23:36:32,701 | INFO | train_grpo_train | metrics_logged mode=train step=1117

                                                     
{'loss': 0.0033, 'grad_norm': 0.39119431376457214, 'learning_rate': 6.618181818181819e-06, 'num_tokens': 2521793.0, 'completions/mean_length': 341.0, 'completions/min_length': 335.0, 'completions/max_length': 360.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 341.0, 'completions/min_terminated_length': 335.0, 'completions/max_terminated_length': 360.0, 'rewards/meter/mean': 0.9725086688995361, 'rewards/meter/std': 0.01511977519840002, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5882353186607361, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.42904794216156006, 'rewards/total_composite/std': 0.006670483388006687, 'reward': 0.42904794216156006, 'reward_std': 0.006670480594038963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00816288124769926, 'sampling/sampling_logp_difference/max': 2.6277527809143066, 'sampling/importance_sampling_ratio/min': 0.07224062830209732, 'sampling/importance_sampling_ratio/mean': 1.0016568899154663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03286444069817662, 'clip_ratio/low_mean': 0.001091206620912999, 'clip_ratio/low_min': 0.001091206620912999, 'clip_ratio/high_mean': 0.002917782054282725, 'clip_ratio/high_max': 0.002917782054282725, 'clip_ratio/region_mean': 0.004008988675195724, 'reward_total_mean': 0.42904794216156006, 'reward_meter_mean': 0.9725086688995361, 'reward_meter_std': 0.01511977519840002, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5882353186607361, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.42904794216156006, 'reward_total_composite_std': 0.006670483388006687, 'epoch': 0.04}

 34%|███▍      | 1117/3300 [59:41<3:29:03,  5.75s/it]INFO 04-11 23:36:32 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1118/3300 [59:47<3:24:50,  5.63s/it]2026-04-11 23:36:38,069 | INFO | train_grpo_train | metrics_logged mode=train step=1118

                                                     
{'loss': -0.0082, 'grad_norm': 4.05780553817749, 'learning_rate': 6.615151515151516e-06, 'num_tokens': 2523775.0, 'completions/mean_length': 86.75, 'completions/min_length': 82.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.75, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.7943093776702881, 'rewards/meter/std': 0.21005631983280182, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.6824886798858643, 'rewards/total_composite/std': 0.2206040620803833, 'reward': 0.6824886798858643, 'reward_std': 0.2206040620803833, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028331898152828217, 'sampling/sampling_logp_difference/max': 1.0872173309326172, 'sampling/importance_sampling_ratio/min': 0.3371534049510956, 'sampling/importance_sampling_ratio/mean': 0.9958962202072144, 'sampling/importance_sampling_ratio/max': 1.9720497131347656, 'entropy': 0.11351043824106455, 'clip_ratio/low_mean': 0.0043270515743643045, 'clip_ratio/low_min': 0.0043270515743643045, 'clip_ratio/high_mean': 0.020123523310758173, 'clip_ratio/high_max': 0.020123523310758173, 'clip_ratio/region_mean': 0.024450574885122478, 'reward_total_mean': 0.6824886798858643, 'reward_meter_mean': 0.7943093776702881, 'reward_meter_std': 0.21005631983280182, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.6824886798858643, 'reward_total_composite_std': 0.2206040620803833, 'epoch': 0.04}

 34%|███▍      | 1118/3300 [59:47<3:24:50,  5.63s/it]INFO 04-11 23:36:38 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1119/3300 [59:51<3:13:42,  5.33s/it]2026-04-11 23:36:42,689 | INFO | train_grpo_train | metrics_logged mode=train step=1119

                                                     
{'loss': 0.0076, 'grad_norm': 4.082438945770264, 'learning_rate': 6.612121212121213e-06, 'num_tokens': 2525659.0, 'completions/mean_length': 60.5, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9949465990066528, 'rewards/meter/std': 0.0008418544312007725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949465990066528, 'rewards/total_composite/std': 0.0008418544312007725, 'reward': 0.9949465990066528, 'reward_std': 0.0008418540237471461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016099898144602776, 'sampling/sampling_logp_difference/max': 1.0732321739196777, 'sampling/importance_sampling_ratio/min': 0.4875824749469757, 'sampling/importance_sampling_ratio/mean': 1.0043621063232422, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.061238054651767015, 'clip_ratio/low_mean': 0.012297285255044699, 'clip_ratio/low_min': 0.012297285255044699, 'clip_ratio/high_mean': 0.006250000325962901, 'clip_ratio/high_max': 0.006250000325962901, 'clip_ratio/region_mean': 0.0185472855810076, 'reward_total_mean': 0.9949465990066528, 'reward_meter_mean': 0.9949465990066528, 'reward_meter_std': 0.0008418544312007725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949465990066528, 'reward_total_composite_std': 0.0008418544312007725, 'epoch': 0.04}

 34%|███▍      | 1119/3300 [59:51<3:13:42,  5.33s/it]INFO 04-11 23:36:42 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1120/3300 [59:56<3:05:11,  5.10s/it]2026-04-11 23:36:47,248 | INFO | train_grpo_train | metrics_logged mode=train step=1120

                                                     
{'loss': -0.0067, 'grad_norm': 3.3688883781433105, 'learning_rate': 6.609090909090909e-06, 'num_tokens': 2527370.0, 'completions/mean_length': 60.875, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9956821203231812, 'rewards/meter/std': 0.0002610879309941083, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9956821203231812, 'rewards/total_composite/std': 0.0002610879309941083, 'reward': 0.9956821203231812, 'reward_std': 0.0002611094678286463, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012183277867734432, 'sampling/sampling_logp_difference/max': 1.1166975498199463, 'sampling/importance_sampling_ratio/min': 0.3273591101169586, 'sampling/importance_sampling_ratio/mean': 1.0035232305526733, 'sampling/importance_sampling_ratio/max': 1.5681954622268677, 'entropy': 0.04563273023813963, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.002016128972172737, 'clip_ratio/high_max': 0.002016128972172737, 'clip_ratio/region_mean': 0.004099462414160371, 'reward_total_mean': 0.9956821203231812, 'reward_meter_mean': 0.9956821203231812, 'reward_meter_std': 0.0002610879309941083, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9956821203231812, 'reward_total_composite_std': 0.0002610879309941083, 'epoch': 0.04}

 34%|███▍      | 1120/3300 [59:56<3:05:11,  5.10s/it]INFO 04-11 23:36:47 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1121/3300 [1:00:00<2:59:42,  4.95s/it]2026-04-11 23:36:51,847 | INFO | train_grpo_train | metrics_logged mode=train step=1121

                                                       
{'loss': -0.0073, 'grad_norm': 6.166923999786377, 'learning_rate': 6.606060606060607e-06, 'num_tokens': 2529160.0, 'completions/mean_length': 69.75, 'completions/min_length': 68.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.75, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7864480018615723, 'rewards/meter/std': 0.26775890588760376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5, 'rewards/repeat_penalty/std': 0.1511857956647873, 'rewards/total_composite/mean': 0.3661602735519409, 'rewards/total_composite/std': 0.13910025358200073, 'reward': 0.3661602735519409, 'reward_std': 0.13910023868083954, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01315419189631939, 'sampling/sampling_logp_difference/max': 1.5172581672668457, 'sampling/importance_sampling_ratio/min': 0.219312384724617, 'sampling/importance_sampling_ratio/mean': 1.0017613172531128, 'sampling/importance_sampling_ratio/max': 1.638248085975647, 'entropy': 0.061232382198795676, 'clip_ratio/low_mean': 0.005462184897623956, 'clip_ratio/low_min': 0.005462184897623956, 'clip_ratio/high_mean': 0.0017361111240461469, 'clip_ratio/high_max': 0.0017361111240461469, 'clip_ratio/region_mean': 0.007198296021670103, 'reward_total_mean': 0.3661602735519409, 'reward_meter_mean': 0.7864480018615723, 'reward_meter_std': 0.26775890588760376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5, 'reward_repeat_penalty_std': 0.1511857956647873, 'reward_total_composite_mean': 0.3661602735519409, 'reward_total_composite_std': 0.13910025358200073, 'epoch': 0.05}

 34%|███▍      | 1121/3300 [1:00:00<2:59:42,  4.95s/it]INFO 04-11 23:36:52 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1122/3300 [1:00:05<2:55:36,  4.84s/it]2026-04-11 23:36:56,426 | INFO | train_grpo_train | metrics_logged mode=train step=1122

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.603030303030303e-06, 'num_tokens': 2530984.0, 'completions/mean_length': 73.0, 'completions/min_length': 73.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.0, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9929870963096619, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9929870963096619, 'rewards/total_composite/std': 0.0, 'reward': 0.9929870963096619, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0027792660985141993, 'sampling/sampling_logp_difference/max': 0.229114830493927, 'sampling/importance_sampling_ratio/min': 0.7952372431755066, 'sampling/importance_sampling_ratio/mean': 1.000472068786621, 'sampling/importance_sampling_ratio/max': 1.0893265008926392, 'entropy': 0.021201700437813997, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9929870963096619, 'reward_meter_mean': 0.9929870963096619, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9929870963096619, 'reward_total_composite_std': 0.0, 'epoch': 0.05}

 34%|███▍      | 1122/3300 [1:00:05<2:55:36,  4.84s/it]INFO 04-11 23:36:56 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1123/3300 [1:00:10<2:53:26,  4.78s/it]2026-04-11 23:37:01,072 | INFO | train_grpo_train | metrics_logged mode=train step=1123

                                                       
{'loss': -0.0032, 'grad_norm': 4.8726959228515625, 'learning_rate': 6.600000000000001e-06, 'num_tokens': 2532779.0, 'completions/mean_length': 66.375, 'completions/min_length': 65.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.996692419052124, 'rewards/meter/std': 0.0006456730188801885, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.996692419052124, 'rewards/total_composite/std': 0.0006456730188801885, 'reward': 0.996692419052124, 'reward_std': 0.0006456732517108321, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04234900325536728, 'sampling/sampling_logp_difference/max': 1.778670310974121, 'sampling/importance_sampling_ratio/min': 0.16886253654956818, 'sampling/importance_sampling_ratio/mean': 0.9999663233757019, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.198346434161067, 'clip_ratio/low_mean': 0.01133623847272247, 'clip_ratio/low_min': 0.01133623847272247, 'clip_ratio/high_mean': 0.0149312699213624, 'clip_ratio/high_max': 0.0149312699213624, 'clip_ratio/region_mean': 0.02626750839408487, 'reward_total_mean': 0.996692419052124, 'reward_meter_mean': 0.996692419052124, 'reward_meter_std': 0.0006456730188801885, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.996692419052124, 'reward_total_composite_std': 0.0006456730188801885, 'epoch': 0.05}

 34%|███▍      | 1123/3300 [1:00:10<2:53:26,  4.78s/it]INFO 04-11 23:37:01 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1124/3300 [1:00:14<2:51:20,  4.72s/it]2026-04-11 23:37:05,665 | INFO | train_grpo_train | metrics_logged mode=train step=1124

                                                       
{'loss': 0.0006, 'grad_norm': 0.6933730244636536, 'learning_rate': 6.596969696969698e-06, 'num_tokens': 2534594.0, 'completions/mean_length': 72.875, 'completions/min_length': 72.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9930577874183655, 'rewards/meter/std': 0.0013781489105895162, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9930577874183655, 'rewards/total_composite/std': 0.0013781489105895162, 'reward': 0.9930577874183655, 'reward_std': 0.001378138200379908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004648053552955389, 'sampling/sampling_logp_difference/max': 0.48731809854507446, 'sampling/importance_sampling_ratio/min': 0.6142716407775879, 'sampling/importance_sampling_ratio/mean': 1.000672459602356, 'sampling/importance_sampling_ratio/max': 1.413206696510315, 'entropy': 0.028838476166129112, 'clip_ratio/low_mean': 0.006849315017461777, 'clip_ratio/low_min': 0.006849315017461777, 'clip_ratio/high_mean': 0.0034722222480922937, 'clip_ratio/high_max': 0.0034722222480922937, 'clip_ratio/region_mean': 0.01032153726555407, 'reward_total_mean': 0.9930577874183655, 'reward_meter_mean': 0.9930577874183655, 'reward_meter_std': 0.0013781489105895162, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9930577874183655, 'reward_total_composite_std': 0.0013781489105895162, 'epoch': 0.05}

 34%|███▍      | 1124/3300 [1:00:14<2:51:20,  4.72s/it]INFO 04-11 23:37:05 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1125/3300 [1:00:20<2:59:44,  4.96s/it]2026-04-11 23:37:11,173 | INFO | train_grpo_train | metrics_logged mode=train step=1125

                                                       
{'loss': -0.0499, 'grad_norm': 3.602660894393921, 'learning_rate': 6.593939393939395e-06, 'num_tokens': 2536357.0, 'completions/mean_length': 70.375, 'completions/min_length': 61.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.375, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9664039611816406, 'rewards/meter/std': 0.08127888292074203, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9664039611816406, 'rewards/total_composite/std': 0.08127888292074203, 'reward': 0.9664039611816406, 'reward_std': 0.08127887547016144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015036996454000473, 'sampling/sampling_logp_difference/max': 1.4699931144714355, 'sampling/importance_sampling_ratio/min': 0.22992707788944244, 'sampling/importance_sampling_ratio/mean': 0.9998224377632141, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.051455921959131956, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010517970658838749, 'clip_ratio/high_max': 0.010517970658838749, 'clip_ratio/region_mean': 0.010517970658838749, 'reward_total_mean': 0.9664039611816406, 'reward_meter_mean': 0.9664039611816406, 'reward_meter_std': 0.08127888292074203, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9664039611816406, 'reward_total_composite_std': 0.08127888292074203, 'epoch': 0.05}

 34%|███▍      | 1125/3300 [1:00:20<2:59:44,  4.96s/it]INFO 04-11 23:37:11 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1126/3300 [1:00:25<2:59:54,  4.97s/it]2026-04-11 23:37:16,153 | INFO | train_grpo_train | metrics_logged mode=train step=1126

                                                       
{'loss': -0.0067, 'grad_norm': 4.485559940338135, 'learning_rate': 6.590909090909091e-06, 'num_tokens': 2538137.0, 'completions/mean_length': 66.5, 'completions/min_length': 65.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9959009289741516, 'rewards/meter/std': 0.0018093610415235162, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9959009289741516, 'rewards/total_composite/std': 0.0018093610415235162, 'reward': 0.9959009289741516, 'reward_std': 0.0018093656981363893, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030639950186014175, 'sampling/sampling_logp_difference/max': 1.1065802574157715, 'sampling/importance_sampling_ratio/min': 0.33068788051605225, 'sampling/importance_sampling_ratio/mean': 0.997355580329895, 'sampling/importance_sampling_ratio/max': 1.4944887161254883, 'entropy': 0.15226891916245222, 'clip_ratio/low_mean': 0.007549191126599908, 'clip_ratio/low_min': 0.007549191126599908, 'clip_ratio/high_mean': 0.02241882192902267, 'clip_ratio/high_max': 0.02241882192902267, 'clip_ratio/region_mean': 0.029968013055622578, 'reward_total_mean': 0.9959009289741516, 'reward_meter_mean': 0.9959009289741516, 'reward_meter_std': 0.0018093610415235162, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9959009289741516, 'reward_total_composite_std': 0.0018093610415235162, 'epoch': 0.05}

 34%|███▍      | 1126/3300 [1:00:25<2:59:54,  4.97s/it]INFO 04-11 23:37:16 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1127/3300 [1:00:29<2:58:04,  4.92s/it]2026-04-11 23:37:20,958 | INFO | train_grpo_train | metrics_logged mode=train step=1127

                                                       
{'loss': -0.0117, 'grad_norm': 3.99946928024292, 'learning_rate': 6.5878787878787885e-06, 'num_tokens': 2539806.0, 'completions/mean_length': 60.625, 'completions/min_length': 58.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9501904249191284, 'rewards/meter/std': 0.06670597940683365, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9501904249191284, 'rewards/total_composite/std': 0.06670597940683365, 'reward': 0.9501904249191284, 'reward_std': 0.06670597940683365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022387670353055, 'sampling/sampling_logp_difference/max': 1.423022985458374, 'sampling/importance_sampling_ratio/min': 0.24098443984985352, 'sampling/importance_sampling_ratio/mean': 0.9991652369499207, 'sampling/importance_sampling_ratio/max': 1.6531167030334473, 'entropy': 0.08485704753547907, 'clip_ratio/low_mean': 0.006355932215228677, 'clip_ratio/low_min': 0.006355932215228677, 'clip_ratio/high_mean': 0.01383362547494471, 'clip_ratio/high_max': 0.01383362547494471, 'clip_ratio/region_mean': 0.020189557690173388, 'reward_total_mean': 0.9501904249191284, 'reward_meter_mean': 0.9501904249191284, 'reward_meter_std': 0.06670597940683365, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9501904249191284, 'reward_total_composite_std': 0.06670597940683365, 'epoch': 0.05}

 34%|███▍      | 1127/3300 [1:00:29<2:58:04,  4.92s/it]INFO 04-11 23:37:21 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1128/3300 [1:00:34<2:53:31,  4.79s/it]2026-04-11 23:37:25,463 | INFO | train_grpo_train | metrics_logged mode=train step=1128

                                                       
{'loss': -0.0059, 'grad_norm': 1.894775390625, 'learning_rate': 6.584848484848485e-06, 'num_tokens': 2541433.0, 'completions/mean_length': 61.375, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9959591627120972, 'rewards/meter/std': 0.0002977726107928902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9959591627120972, 'rewards/total_composite/std': 0.0002977726107928902, 'reward': 0.9959591627120972, 'reward_std': 0.0002977752883452922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008953627198934555, 'sampling/sampling_logp_difference/max': 0.6550607681274414, 'sampling/importance_sampling_ratio/min': 0.5194104909896851, 'sampling/importance_sampling_ratio/mean': 1.0015958547592163, 'sampling/importance_sampling_ratio/max': 1.208601951599121, 'entropy': 0.05297265062108636, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.006048386916518211, 'clip_ratio/high_max': 0.006048386916518211, 'clip_ratio/region_mean': 0.008131720358505845, 'reward_total_mean': 0.9959591627120972, 'reward_meter_mean': 0.9959591627120972, 'reward_meter_std': 0.0002977726107928902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9959591627120972, 'reward_total_composite_std': 0.0002977726107928902, 'epoch': 0.05}

 34%|███▍      | 1128/3300 [1:00:34<2:53:31,  4.79s/it]INFO 04-11 23:37:25 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1129/3300 [1:00:44<3:47:26,  6.29s/it]2026-04-11 23:37:35,230 | INFO | train_grpo_train | metrics_logged mode=train step=1129

                                                       
{'loss': -0.2408, 'grad_norm': 1.491158366203308, 'learning_rate': 6.581818181818182e-06, 'num_tokens': 2544019.0, 'completions/mean_length': 375.25, 'completions/min_length': 221.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 238.5, 'completions/min_terminated_length': 221.0, 'completions/max_terminated_length': 277.0, 'rewards/meter/mean': 0.367766410112381, 'rewards/meter/std': 0.4184606671333313, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2777460217475891, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/repeat_penalty/mean': 0.7755848169326782, 'rewards/repeat_penalty/std': 0.2187419831752777, 'rewards/total_composite/mean': 0.16618868708610535, 'rewards/total_composite/std': 0.2396627515554428, 'reward': 0.16618868708610535, 'reward_std': 0.2396627515554428, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03679489716887474, 'sampling/sampling_logp_difference/max': 5.3318939208984375, 'sampling/importance_sampling_ratio/min': 0.004834904335439205, 'sampling/importance_sampling_ratio/mean': 1.0069185495376587, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09304305911064148, 'clip_ratio/low_mean': 0.005044843070209026, 'clip_ratio/low_min': 0.005044843070209026, 'clip_ratio/high_mean': 0.0071450776886194944, 'clip_ratio/high_max': 0.0071450776886194944, 'clip_ratio/region_mean': 0.01218992075882852, 'reward_total_mean': 0.16618868708610535, 'reward_meter_mean': 0.367766410112381, 'reward_meter_std': 0.4184606671333313, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2777460217475891, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_repeat_penalty_mean': 0.7755848169326782, 'reward_repeat_penalty_std': 0.2187419831752777, 'reward_total_composite_mean': 0.16618868708610535, 'reward_total_composite_std': 0.2396627515554428, 'epoch': 0.05}

 34%|███▍      | 1129/3300 [1:00:44<3:47:26,  6.29s/it]INFO 04-11 23:37:35 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1130/3300 [1:00:53<4:21:25,  7.23s/it]2026-04-11 23:37:44,658 | INFO | train_grpo_train | metrics_logged mode=train step=1130

                                                       
{'loss': -0.0873, 'grad_norm': 3.4096953868865967, 'learning_rate': 6.578787878787879e-06, 'num_tokens': 2545823.0, 'completions/mean_length': 108.5, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.85714340209961, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.5511125326156616, 'rewards/meter/std': 0.42126619815826416, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.47555798292160034, 'rewards/total_composite/std': 0.38399630784988403, 'reward': 0.47555798292160034, 'reward_std': 0.38399630784988403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054451193660497665, 'sampling/sampling_logp_difference/max': 1.399144172668457, 'sampling/importance_sampling_ratio/min': 0.24680811166763306, 'sampling/importance_sampling_ratio/mean': 1.007746696472168, 'sampling/importance_sampling_ratio/max': 1.7676554918289185, 'entropy': 0.25371737964451313, 'clip_ratio/low_mean': 0.015434782486408949, 'clip_ratio/low_min': 0.015434782486408949, 'clip_ratio/high_mean': 0.030356566421687603, 'clip_ratio/high_max': 0.030356566421687603, 'clip_ratio/region_mean': 0.04579134890809655, 'reward_total_mean': 0.47555798292160034, 'reward_meter_mean': 0.5511125326156616, 'reward_meter_std': 0.42126619815826416, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.47555798292160034, 'reward_total_composite_std': 0.38399630784988403, 'epoch': 0.05}

 34%|███▍      | 1130/3300 [1:00:53<4:21:25,  7.23s/it]INFO 04-11 23:37:44 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1131/3300 [1:00:58<3:50:46,  6.38s/it]2026-04-11 23:37:49,070 | INFO | train_grpo_train | metrics_logged mode=train step=1131

                                                       
{'loss': -0.0297, 'grad_norm': 8.209676742553711, 'learning_rate': 6.575757575757577e-06, 'num_tokens': 2547552.0, 'completions/mean_length': 54.125, 'completions/min_length': 50.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9417009353637695, 'rewards/meter/std': 0.008068003691732883, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9417009353637695, 'rewards/total_composite/std': 0.008068003691732883, 'reward': 0.9417009353637695, 'reward_std': 0.008068017661571503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02913379855453968, 'sampling/sampling_logp_difference/max': 2.463469982147217, 'sampling/importance_sampling_ratio/min': 0.08513900637626648, 'sampling/importance_sampling_ratio/mean': 0.9954664707183838, 'sampling/importance_sampling_ratio/max': 1.376886010169983, 'entropy': 0.08877178048714995, 'clip_ratio/low_mean': 0.012500000186264515, 'clip_ratio/low_min': 0.012500000186264515, 'clip_ratio/high_mean': 0.013678450835868716, 'clip_ratio/high_max': 0.013678450835868716, 'clip_ratio/region_mean': 0.02617845102213323, 'reward_total_mean': 0.9417009353637695, 'reward_meter_mean': 0.9417009353637695, 'reward_meter_std': 0.008068003691732883, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9417009353637695, 'reward_total_composite_std': 0.008068003691732883, 'epoch': 0.05}

 34%|███▍      | 1131/3300 [1:00:58<3:50:46,  6.38s/it]INFO 04-11 23:37:49 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1132/3300 [1:01:02<3:27:04,  5.73s/it]2026-04-11 23:37:53,319 | INFO | train_grpo_train | metrics_logged mode=train step=1132

                                                       
{'loss': -0.0471, 'grad_norm': 14.512413024902344, 'learning_rate': 6.572727272727273e-06, 'num_tokens': 2548978.0, 'completions/mean_length': 29.25, 'completions/min_length': 26.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.25, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8743196725845337, 'rewards/meter/std': 0.20709700882434845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8743196725845337, 'rewards/total_composite/std': 0.20709700882434845, 'reward': 0.8743196725845337, 'reward_std': 0.20709700882434845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09784162044525146, 'sampling/sampling_logp_difference/max': 6.468225002288818, 'sampling/importance_sampling_ratio/min': 0.0015519780572503805, 'sampling/importance_sampling_ratio/mean': 1.0037531852722168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42541524581611156, 'clip_ratio/low_mean': 0.014423077460378408, 'clip_ratio/low_min': 0.014423077460378408, 'clip_ratio/high_mean': 0.05365684116259217, 'clip_ratio/high_max': 0.05365684116259217, 'clip_ratio/region_mean': 0.06807991862297058, 'reward_total_mean': 0.8743196725845337, 'reward_meter_mean': 0.8743196725845337, 'reward_meter_std': 0.20709700882434845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8743196725845337, 'reward_total_composite_std': 0.20709700882434845, 'epoch': 0.05}

 34%|███▍      | 1132/3300 [1:01:02<3:27:04,  5.73s/it]INFO 04-11 23:37:53 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1133/3300 [1:01:07<3:21:07,  5.57s/it]2026-04-11 23:37:58,469 | INFO | train_grpo_train | metrics_logged mode=train step=1133

                                                       
{'loss': 0.0098, 'grad_norm': 1.9009183645248413, 'learning_rate': 6.56969696969697e-06, 'num_tokens': 2550664.0, 'completions/mean_length': 59.75, 'completions/min_length': 59.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9870603084564209, 'rewards/meter/std': 0.004516107961535454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9870603084564209, 'rewards/total_composite/std': 0.004516107961535454, 'reward': 0.9870603084564209, 'reward_std': 0.004516112618148327, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013605736196041107, 'sampling/sampling_logp_difference/max': 1.056100606918335, 'sampling/importance_sampling_ratio/min': 0.4231666624546051, 'sampling/importance_sampling_ratio/mean': 1.0032687187194824, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.046739939134567976, 'clip_ratio/low_mean': 0.006215847097337246, 'clip_ratio/low_min': 0.006215847097337246, 'clip_ratio/high_mean': 0.004167824285104871, 'clip_ratio/high_max': 0.004167824285104871, 'clip_ratio/region_mean': 0.010383671382442117, 'reward_total_mean': 0.9870603084564209, 'reward_meter_mean': 0.9870603084564209, 'reward_meter_std': 0.004516107961535454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9870603084564209, 'reward_total_composite_std': 0.004516107961535454, 'epoch': 0.05}

 34%|███▍      | 1133/3300 [1:01:07<3:21:07,  5.57s/it]INFO 04-11 23:37:58 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1134/3300 [1:01:12<3:14:47,  5.40s/it]2026-04-11 23:38:03,461 | INFO | train_grpo_train | metrics_logged mode=train step=1134

                                                       
{'loss': 0.0062, 'grad_norm': 4.714395523071289, 'learning_rate': 6.566666666666667e-06, 'num_tokens': 2552346.0, 'completions/mean_length': 70.25, 'completions/min_length': 68.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9971396923065186, 'rewards/meter/std': 0.0017616376280784607, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971396923065186, 'rewards/total_composite/std': 0.0017616376280784607, 'reward': 0.9971396923065186, 'reward_std': 0.0017616377444937825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030578911304473877, 'sampling/sampling_logp_difference/max': 1.3415107727050781, 'sampling/importance_sampling_ratio/min': 0.2614503800868988, 'sampling/importance_sampling_ratio/mean': 1.0078562498092651, 'sampling/importance_sampling_ratio/max': 1.841591238975525, 'entropy': 0.18766964972019196, 'clip_ratio/low_mean': 0.010638833045959473, 'clip_ratio/low_min': 0.010638833045959473, 'clip_ratio/high_mean': 0.026770169381052256, 'clip_ratio/high_max': 0.026770169381052256, 'clip_ratio/region_mean': 0.03740900242701173, 'reward_total_mean': 0.9971396923065186, 'reward_meter_mean': 0.9971396923065186, 'reward_meter_std': 0.0017616376280784607, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971396923065186, 'reward_total_composite_std': 0.0017616376280784607, 'epoch': 0.05}

 34%|███▍      | 1134/3300 [1:01:12<3:14:47,  5.40s/it]INFO 04-11 23:38:03 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1135/3300 [1:01:17<3:10:33,  5.28s/it]2026-04-11 23:38:08,475 | INFO | train_grpo_train | metrics_logged mode=train step=1135

                                                       
{'loss': 0.0005, 'grad_norm': 3.5454463958740234, 'learning_rate': 6.563636363636364e-06, 'num_tokens': 2554294.0, 'completions/mean_length': 76.5, 'completions/min_length': 73.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9800901412963867, 'rewards/meter/std': 0.01382446475327015, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9800901412963867, 'rewards/total_composite/std': 0.01382446475327015, 'reward': 0.9800901412963867, 'reward_std': 0.013824466615915298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01574569195508957, 'sampling/sampling_logp_difference/max': 0.8164339065551758, 'sampling/importance_sampling_ratio/min': 0.44200506806373596, 'sampling/importance_sampling_ratio/mean': 1.005315899848938, 'sampling/importance_sampling_ratio/max': 1.7477397918701172, 'entropy': 0.10036304593086243, 'clip_ratio/low_mean': 0.00811688310932368, 'clip_ratio/low_min': 0.00811688310932368, 'clip_ratio/high_mean': 0.011680196272209287, 'clip_ratio/high_max': 0.011680196272209287, 'clip_ratio/region_mean': 0.019797079381532967, 'reward_total_mean': 0.9800901412963867, 'reward_meter_mean': 0.9800901412963867, 'reward_meter_std': 0.01382446475327015, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9800901412963867, 'reward_total_composite_std': 0.01382446475327015, 'epoch': 0.05}

 34%|███▍      | 1135/3300 [1:01:17<3:10:33,  5.28s/it]INFO 04-11 23:38:08 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1136/3300 [1:01:26<3:53:40,  6.48s/it]2026-04-11 23:38:17,754 | INFO | train_grpo_train | metrics_logged mode=train step=1136

                                                       
{'loss': -0.019, 'grad_norm': 2.327674627304077, 'learning_rate': 6.56060606060606e-06, 'num_tokens': 2559265.0, 'completions/mean_length': 388.375, 'completions/min_length': 362.0, 'completions/max_length': 427.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 388.375, 'completions/min_terminated_length': 362.0, 'completions/max_terminated_length': 427.0, 'rewards/meter/mean': 0.952599823474884, 'rewards/meter/std': 0.09515554457902908, 'rewards/count_adherence/mean': 0.7410714626312256, 'rewards/count_adherence/std': 0.03696778044104576, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.716478705406189, 'rewards/repeat_penalty/std': 0.12055735290050507, 'rewards/total_composite/mean': 0.49994567036628723, 'rewards/total_composite/std': 0.06088562682271004, 'reward': 0.49994567036628723, 'reward_std': 0.060885630548000336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02755453996360302, 'sampling/sampling_logp_difference/max': 1.6922149658203125, 'sampling/importance_sampling_ratio/min': 0.18411126732826233, 'sampling/importance_sampling_ratio/mean': 1.0076520442962646, 'sampling/importance_sampling_ratio/max': 1.8240277767181396, 'entropy': 0.21599608566612005, 'clip_ratio/low_mean': 0.010929276293609291, 'clip_ratio/low_min': 0.010929276293609291, 'clip_ratio/high_mean': 0.005753153818659484, 'clip_ratio/high_max': 0.005753153818659484, 'clip_ratio/region_mean': 0.016682430112268776, 'reward_total_mean': 0.49994567036628723, 'reward_meter_mean': 0.952599823474884, 'reward_meter_std': 0.09515554457902908, 'reward_count_adherence_mean': 0.7410714626312256, 'reward_count_adherence_std': 0.03696778044104576, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.716478705406189, 'reward_repeat_penalty_std': 0.12055735290050507, 'reward_total_composite_mean': 0.49994567036628723, 'reward_total_composite_std': 0.06088562682271004, 'epoch': 0.05}

 34%|███▍      | 1136/3300 [1:01:26<3:53:40,  6.48s/it]INFO 04-11 23:38:18 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1137/3300 [1:01:32<3:43:17,  6.19s/it]2026-04-11 23:38:23,280 | INFO | train_grpo_train | metrics_logged mode=train step=1137

                                                       
{'loss': 0.0836, 'grad_norm': 3.7329790592193604, 'learning_rate': 6.5575757575757585e-06, 'num_tokens': 2561586.0, 'completions/mean_length': 111.125, 'completions/min_length': 100.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.125, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.9837856292724609, 'rewards/meter/std': 0.012296498753130436, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7892857789993286, 'rewards/repeat_penalty/std': 0.030304575338959694, 'rewards/total_composite/mean': 0.7466674447059631, 'rewards/total_composite/std': 0.11019133776426315, 'reward': 0.7466674447059631, 'reward_std': 0.11019132286310196, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020305845886468887, 'sampling/sampling_logp_difference/max': 2.0209195613861084, 'sampling/importance_sampling_ratio/min': 0.13253353536128998, 'sampling/importance_sampling_ratio/mean': 0.9983324408531189, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08285832032561302, 'clip_ratio/low_mean': 0.0009057971183210611, 'clip_ratio/low_min': 0.0009057971183210611, 'clip_ratio/high_mean': 0.010732458787970245, 'clip_ratio/high_max': 0.010732458787970245, 'clip_ratio/region_mean': 0.011638255906291306, 'reward_total_mean': 0.7466674447059631, 'reward_meter_mean': 0.9837856292724609, 'reward_meter_std': 0.012296498753130436, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7892857789993286, 'reward_repeat_penalty_std': 0.030304575338959694, 'reward_total_composite_mean': 0.7466674447059631, 'reward_total_composite_std': 0.11019133776426315, 'epoch': 0.05}

 34%|███▍      | 1137/3300 [1:01:32<3:43:17,  6.19s/it]INFO 04-11 23:38:23 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1138/3300 [1:01:37<3:32:35,  5.90s/it]2026-04-11 23:38:28,504 | INFO | train_grpo_train | metrics_logged mode=train step=1138

                                                       
{'loss': -0.0078, 'grad_norm': 2.7455623149871826, 'learning_rate': 6.554545454545455e-06, 'num_tokens': 2563284.0, 'completions/mean_length': 59.25, 'completions/min_length': 59.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9892421960830688, 'rewards/meter/std': 0.0002119775745086372, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9892421960830688, 'rewards/total_composite/std': 0.0002119775745086372, 'reward': 0.9892421960830688, 'reward_std': 0.000211986611247994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002105327555909753, 'sampling/sampling_logp_difference/max': 0.09289813041687012, 'sampling/importance_sampling_ratio/min': 0.960770845413208, 'sampling/importance_sampling_ratio/mean': 1.0011262893676758, 'sampling/importance_sampling_ratio/max': 1.097350001335144, 'entropy': 0.020128208212554455, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9892421960830688, 'reward_meter_mean': 0.9892421960830688, 'reward_meter_std': 0.0002119775745086372, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9892421960830688, 'reward_total_composite_std': 0.0002119775745086372, 'epoch': 0.05}

 34%|███▍      | 1138/3300 [1:01:37<3:32:35,  5.90s/it]INFO 04-11 23:38:28 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1139/3300 [1:01:42<3:27:32,  5.76s/it]2026-04-11 23:38:33,938 | INFO | train_grpo_train | metrics_logged mode=train step=1139

                                                       
{'loss': -0.0093, 'grad_norm': 3.7970211505889893, 'learning_rate': 6.551515151515152e-06, 'num_tokens': 2564995.0, 'completions/mean_length': 60.875, 'completions/min_length': 59.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.989832878112793, 'rewards/meter/std': 0.0006110123940743506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.989832878112793, 'rewards/total_composite/std': 0.0006110123940743506, 'reward': 0.989832878112793, 'reward_std': 0.0006110109388828278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011983020231127739, 'sampling/sampling_logp_difference/max': 1.2451696395874023, 'sampling/importance_sampling_ratio/min': 0.28789207339286804, 'sampling/importance_sampling_ratio/mean': 1.0028979778289795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04918731888756156, 'clip_ratio/low_mean': 0.008266184711828828, 'clip_ratio/low_min': 0.008266184711828828, 'clip_ratio/high_mean': 0.002016128972172737, 'clip_ratio/high_max': 0.002016128972172737, 'clip_ratio/region_mean': 0.010282313684001565, 'reward_total_mean': 0.989832878112793, 'reward_meter_mean': 0.989832878112793, 'reward_meter_std': 0.0006110123940743506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.989832878112793, 'reward_total_composite_std': 0.0006110123940743506, 'epoch': 0.05}

 35%|███▍      | 1139/3300 [1:01:42<3:27:32,  5.76s/it]INFO 04-11 23:38:34 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1140/3300 [1:01:47<3:17:06,  5.48s/it]2026-04-11 23:38:38,754 | INFO | train_grpo_train | metrics_logged mode=train step=1140

                                                       
{'loss': -0.0165, 'grad_norm': 5.715131759643555, 'learning_rate': 6.5484848484848494e-06, 'num_tokens': 2566645.0, 'completions/mean_length': 56.25, 'completions/min_length': 54.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.2350359857082367, 'rewards/meter/std': 0.07872093468904495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.2350359857082367, 'rewards/total_composite/std': 0.07872093468904495, 'reward': 0.2350359857082367, 'reward_std': 0.07872093468904495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012172389775514603, 'sampling/sampling_logp_difference/max': 0.5825954675674438, 'sampling/importance_sampling_ratio/min': 0.6520004272460938, 'sampling/importance_sampling_ratio/mean': 1.001919150352478, 'sampling/importance_sampling_ratio/max': 1.7906800508499146, 'entropy': 0.061810399405658245, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.00657894741743803, 'clip_ratio/high_max': 0.00657894741743803, 'clip_ratio/region_mean': 0.00889376224949956, 'reward_total_mean': 0.2350359857082367, 'reward_meter_mean': 0.2350359857082367, 'reward_meter_std': 0.07872093468904495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.2350359857082367, 'reward_total_composite_std': 0.07872093468904495, 'epoch': 0.05}

 35%|███▍      | 1140/3300 [1:01:47<3:17:06,  5.48s/it]INFO 04-11 23:38:39 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1141/3300 [1:01:52<3:11:07,  5.31s/it]2026-04-11 23:38:43,675 | INFO | train_grpo_train | metrics_logged mode=train step=1141

                                                       
{'loss': 0.0143, 'grad_norm': 7.939047813415527, 'learning_rate': 6.545454545454546e-06, 'num_tokens': 2568290.0, 'completions/mean_length': 53.625, 'completions/min_length': 53.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.625, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9498906135559082, 'rewards/meter/std': 0.007521784398704767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9498906135559082, 'rewards/total_composite/std': 0.007521784398704767, 'reward': 0.9498906135559082, 'reward_std': 0.007521784398704767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04453643411397934, 'sampling/sampling_logp_difference/max': 2.4130172729492188, 'sampling/importance_sampling_ratio/min': 0.08954470604658127, 'sampling/importance_sampling_ratio/mean': 0.9988313913345337, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1192027423530817, 'clip_ratio/low_mean': 0.016250555869191885, 'clip_ratio/low_min': 0.016250555869191885, 'clip_ratio/high_mean': 0.02358490601181984, 'clip_ratio/high_max': 0.02358490601181984, 'clip_ratio/region_mean': 0.039835461881011724, 'reward_total_mean': 0.9498906135559082, 'reward_meter_mean': 0.9498906135559082, 'reward_meter_std': 0.007521784398704767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9498906135559082, 'reward_total_composite_std': 0.007521784398704767, 'epoch': 0.05}

 35%|███▍      | 1141/3300 [1:01:52<3:11:07,  5.31s/it]INFO 04-11 23:38:43 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1142/3300 [1:01:57<3:04:46,  5.14s/it]2026-04-11 23:38:48,405 | INFO | train_grpo_train | metrics_logged mode=train step=1142

                                                       
{'loss': 0.0038, 'grad_norm': 4.282461166381836, 'learning_rate': 6.542424242424243e-06, 'num_tokens': 2569755.0, 'completions/mean_length': 39.125, 'completions/min_length': 37.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9843659400939941, 'rewards/meter/std': 0.01807660609483719, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9843659400939941, 'rewards/total_composite/std': 0.01807660609483719, 'reward': 0.9843659400939941, 'reward_std': 0.018076593056321144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03325369209051132, 'sampling/sampling_logp_difference/max': 1.2249364852905273, 'sampling/importance_sampling_ratio/min': 0.29377633333206177, 'sampling/importance_sampling_ratio/mean': 0.9939477443695068, 'sampling/importance_sampling_ratio/max': 1.3938076496124268, 'entropy': 0.17795206978917122, 'clip_ratio/low_mean': 0.0031250000465661287, 'clip_ratio/low_min': 0.0031250000465661287, 'clip_ratio/high_mean': 0.025911615695804358, 'clip_ratio/high_max': 0.025911615695804358, 'clip_ratio/region_mean': 0.029036615742370486, 'reward_total_mean': 0.9843659400939941, 'reward_meter_mean': 0.9843659400939941, 'reward_meter_std': 0.01807660609483719, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9843659400939941, 'reward_total_composite_std': 0.01807660609483719, 'epoch': 0.05}

 35%|███▍      | 1142/3300 [1:01:57<3:04:46,  5.14s/it]INFO 04-11 23:38:48 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1143/3300 [1:02:07<4:03:01,  6.76s/it]2026-04-11 23:38:58,946 | INFO | train_grpo_train | metrics_logged mode=train step=1143

                                                       
{'loss': -0.0509, 'grad_norm': 1.155205488204956, 'learning_rate': 6.5393939393939395e-06, 'num_tokens': 2575500.0, 'completions/mean_length': 441.125, 'completions/min_length': 400.0, 'completions/max_length': 498.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 441.125, 'completions/min_terminated_length': 400.0, 'completions/max_terminated_length': 498.0, 'rewards/meter/mean': 0.9610804319381714, 'rewards/meter/std': 0.050803039222955704, 'rewards/count_adherence/mean': 0.6323529481887817, 'rewards/count_adherence/std': 0.027230001986026764, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5833333134651184, 'rewards/repeat_penalty/std': 0.19633837044239044, 'rewards/total_composite/mean': 0.35644832253456116, 'rewards/total_composite/std': 0.13268794119358063, 'reward': 0.35644832253456116, 'reward_std': 0.13268794119358063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021479658782482147, 'sampling/sampling_logp_difference/max': 1.2659144401550293, 'sampling/importance_sampling_ratio/min': 0.28720539808273315, 'sampling/importance_sampling_ratio/mean': 1.004508376121521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1385105513036251, 'clip_ratio/low_mean': 0.004563539958326146, 'clip_ratio/low_min': 0.004563539958326146, 'clip_ratio/high_mean': 0.011979256058111787, 'clip_ratio/high_max': 0.011979256058111787, 'clip_ratio/region_mean': 0.016542796016437933, 'reward_total_mean': 0.35644832253456116, 'reward_meter_mean': 0.9610804319381714, 'reward_meter_std': 0.050803039222955704, 'reward_count_adherence_mean': 0.6323529481887817, 'reward_count_adherence_std': 0.027230001986026764, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5833333134651184, 'reward_repeat_penalty_std': 0.19633837044239044, 'reward_total_composite_mean': 0.35644832253456116, 'reward_total_composite_std': 0.13268794119358063, 'epoch': 0.05}

 35%|███▍      | 1143/3300 [1:02:07<4:03:01,  6.76s/it]INFO 04-11 23:38:59 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1144/3300 [1:02:17<4:32:56,  7.60s/it]2026-04-11 23:39:08,491 | INFO | train_grpo_train | metrics_logged mode=train step=1144

                                                       
{'loss': -0.0143, 'grad_norm': 0.9518696665763855, 'learning_rate': 6.536363636363638e-06, 'num_tokens': 2580194.0, 'completions/mean_length': 387.75, 'completions/min_length': 354.0, 'completions/max_length': 420.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 387.75, 'completions/min_terminated_length': 354.0, 'completions/max_terminated_length': 420.0, 'rewards/meter/mean': 0.9527665376663208, 'rewards/meter/std': 0.0463380366563797, 'rewards/count_adherence/mean': 0.7053571939468384, 'rewards/count_adherence/std': 0.0707879438996315, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5333124399185181, 'rewards/repeat_penalty/std': 0.15502820909023285, 'rewards/total_composite/mean': 0.36780261993408203, 'rewards/total_composite/std': 0.1357763111591339, 'reward': 0.36780261993408203, 'reward_std': 0.1357763111591339, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011555654928088188, 'sampling/sampling_logp_difference/max': 3.7675018310546875, 'sampling/importance_sampling_ratio/min': 0.02310972288250923, 'sampling/importance_sampling_ratio/mean': 1.0018688440322876, 'sampling/importance_sampling_ratio/max': 1.8800609111785889, 'entropy': 0.06274456530809402, 'clip_ratio/low_mean': 0.0020058397494722158, 'clip_ratio/low_min': 0.0020058397494722158, 'clip_ratio/high_mean': 0.005727101757656783, 'clip_ratio/high_max': 0.005727101757656783, 'clip_ratio/region_mean': 0.007732941507128999, 'reward_total_mean': 0.36780261993408203, 'reward_meter_mean': 0.9527665376663208, 'reward_meter_std': 0.0463380366563797, 'reward_count_adherence_mean': 0.7053571939468384, 'reward_count_adherence_std': 0.0707879438996315, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5333124399185181, 'reward_repeat_penalty_std': 0.15502820909023285, 'reward_total_composite_mean': 0.36780261993408203, 'reward_total_composite_std': 0.1357763111591339, 'epoch': 0.05}

 35%|███▍      | 1144/3300 [1:02:17<4:32:56,  7.60s/it]INFO 04-11 23:39:08 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1145/3300 [1:02:22<4:01:39,  6.73s/it]2026-04-11 23:39:13,197 | INFO | train_grpo_train | metrics_logged mode=train step=1145

                                                       
{'loss': -0.0246, 'grad_norm': 3.6901073455810547, 'learning_rate': 6.533333333333334e-06, 'num_tokens': 2582143.0, 'completions/mean_length': 77.625, 'completions/min_length': 75.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.625, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9917942881584167, 'rewards/meter/std': 0.0048305438831448555, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9917942881584167, 'rewards/total_composite/std': 0.0048305438831448555, 'reward': 0.9917942881584167, 'reward_std': 0.004830518271774054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017006773501634598, 'sampling/sampling_logp_difference/max': 1.3061323165893555, 'sampling/importance_sampling_ratio/min': 0.34659311175346375, 'sampling/importance_sampling_ratio/mean': 1.0097332000732422, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13493561558425426, 'clip_ratio/low_mean': 0.006580086657777429, 'clip_ratio/low_min': 0.006580086657777429, 'clip_ratio/high_mean': 0.0031250000465661287, 'clip_ratio/high_max': 0.0031250000465661287, 'clip_ratio/region_mean': 0.009705086704343557, 'reward_total_mean': 0.9917942881584167, 'reward_meter_mean': 0.9917942881584167, 'reward_meter_std': 0.0048305438831448555, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9917942881584167, 'reward_total_composite_std': 0.0048305438831448555, 'epoch': 0.05}

 35%|███▍      | 1145/3300 [1:02:22<4:01:39,  6.73s/it]INFO 04-11 23:39:13 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1146/3300 [1:02:27<3:41:38,  6.17s/it]2026-04-11 23:39:18,075 | INFO | train_grpo_train | metrics_logged mode=train step=1146

                                                       
{'loss': -0.0049, 'grad_norm': 1.9219002723693848, 'learning_rate': 6.530303030303031e-06, 'num_tokens': 2583831.0, 'completions/mean_length': 72.0, 'completions/min_length': 71.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9963383078575134, 'rewards/meter/std': 0.0003733669000212103, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9548450708389282, 'rewards/total_composite/std': 0.11756306886672974, 'reward': 0.9548450708389282, 'reward_std': 0.11756306141614914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003385355230420828, 'sampling/sampling_logp_difference/max': 0.539313793182373, 'sampling/importance_sampling_ratio/min': 0.5831482410430908, 'sampling/importance_sampling_ratio/mean': 1.000481128692627, 'sampling/importance_sampling_ratio/max': 1.0752372741699219, 'entropy': 0.016070799436420202, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9548450708389282, 'reward_meter_mean': 0.9963383078575134, 'reward_meter_std': 0.0003733669000212103, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9548450708389282, 'reward_total_composite_std': 0.11756306886672974, 'epoch': 0.05}

 35%|███▍      | 1146/3300 [1:02:27<3:41:38,  6.17s/it]INFO 04-11 23:39:18 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1147/3300 [1:02:33<3:39:41,  6.12s/it]2026-04-11 23:39:24,079 | INFO | train_grpo_train | metrics_logged mode=train step=1147

                                                       
{'loss': 0.0294, 'grad_norm': 15.72409725189209, 'learning_rate': 6.527272727272728e-06, 'num_tokens': 2585484.0, 'completions/mean_length': 32.625, 'completions/min_length': 30.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.89949631690979, 'rewards/meter/std': 0.18807001411914825, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.89949631690979, 'rewards/total_composite/std': 0.18807001411914825, 'reward': 0.89949631690979, 'reward_std': 0.18807001411914825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06001806631684303, 'sampling/sampling_logp_difference/max': 1.7198677062988281, 'sampling/importance_sampling_ratio/min': 0.29962798953056335, 'sampling/importance_sampling_ratio/mean': 1.0089141130447388, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21602893620729446, 'clip_ratio/low_mean': 0.01907169120386243, 'clip_ratio/low_min': 0.01907169120386243, 'clip_ratio/high_mean': 0.014285714365541935, 'clip_ratio/high_max': 0.014285714365541935, 'clip_ratio/region_mean': 0.033357405569404364, 'reward_total_mean': 0.89949631690979, 'reward_meter_mean': 0.89949631690979, 'reward_meter_std': 0.18807001411914825, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.89949631690979, 'reward_total_composite_std': 0.18807001411914825, 'epoch': 0.05}

 35%|███▍      | 1147/3300 [1:02:33<3:39:41,  6.12s/it]INFO 04-11 23:39:24 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1148/3300 [1:02:37<3:22:54,  5.66s/it]2026-04-11 23:39:28,652 | INFO | train_grpo_train | metrics_logged mode=train step=1148

                                                       
{'loss': 0.0136, 'grad_norm': 5.76182746887207, 'learning_rate': 6.524242424242425e-06, 'num_tokens': 2586905.0, 'completions/mean_length': 37.625, 'completions/min_length': 37.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9881135821342468, 'rewards/meter/std': 0.0235806442797184, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9881135821342468, 'rewards/total_composite/std': 0.0235806442797184, 'reward': 0.9881135821342468, 'reward_std': 0.02358064614236355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010363386012613773, 'sampling/sampling_logp_difference/max': 0.48454809188842773, 'sampling/importance_sampling_ratio/min': 0.6159754991531372, 'sampling/importance_sampling_ratio/mean': 1.0021705627441406, 'sampling/importance_sampling_ratio/max': 1.4336398839950562, 'entropy': 0.03891826095059514, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.00657894741743803, 'clip_ratio/high_max': 0.00657894741743803, 'clip_ratio/region_mean': 0.00657894741743803, 'reward_total_mean': 0.9881135821342468, 'reward_meter_mean': 0.9881135821342468, 'reward_meter_std': 0.0235806442797184, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9881135821342468, 'reward_total_composite_std': 0.0235806442797184, 'epoch': 0.05}

 35%|███▍      | 1148/3300 [1:02:37<3:22:54,  5.66s/it]INFO 04-11 23:39:28 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1149/3300 [1:02:42<3:17:37,  5.51s/it]2026-04-11 23:39:33,829 | INFO | train_grpo_train | metrics_logged mode=train step=1149

                                                       
{'loss': 0.0133, 'grad_norm': 2.983004331588745, 'learning_rate': 6.521212121212121e-06, 'num_tokens': 2588889.0, 'completions/mean_length': 77.0, 'completions/min_length': 71.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9948194026947021, 'rewards/meter/std': 0.007181616500020027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.9116114377975464, 'rewards/total_composite/std': 0.1519550234079361, 'reward': 0.9116114377975464, 'reward_std': 0.1519550234079361, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022046150639653206, 'sampling/sampling_logp_difference/max': 0.7371308207511902, 'sampling/importance_sampling_ratio/min': 0.4784848093986511, 'sampling/importance_sampling_ratio/mean': 1.003957748413086, 'sampling/importance_sampling_ratio/max': 1.373502492904663, 'entropy': 0.11253493744879961, 'clip_ratio/low_mean': 0.011157091706991196, 'clip_ratio/low_min': 0.011157091706991196, 'clip_ratio/high_mean': 0.021323838154785335, 'clip_ratio/high_max': 0.021323838154785335, 'clip_ratio/region_mean': 0.03248092986177653, 'reward_total_mean': 0.9116114377975464, 'reward_meter_mean': 0.9948194026947021, 'reward_meter_std': 0.007181616500020027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.9116114377975464, 'reward_total_composite_std': 0.1519550234079361, 'epoch': 0.05}

 35%|███▍      | 1149/3300 [1:02:42<3:17:37,  5.51s/it]INFO 04-11 23:39:34 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1150/3300 [1:02:49<3:29:00,  5.83s/it]2026-04-11 23:39:40,406 | INFO | train_grpo_train | metrics_logged mode=train step=1150

                                                       
{'loss': -0.0071, 'grad_norm': 1.0385583639144897, 'learning_rate': 6.5181818181818195e-06, 'num_tokens': 2592268.0, 'completions/mean_length': 213.375, 'completions/min_length': 212.0, 'completions/max_length': 218.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 213.375, 'completions/min_terminated_length': 212.0, 'completions/max_terminated_length': 218.0, 'rewards/meter/mean': 0.996587872505188, 'rewards/meter/std': 0.0006334602949209511, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6363636255264282, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5073539018630981, 'rewards/total_composite/std': 0.0003224806860089302, 'reward': 0.5073539018630981, 'reward_std': 0.00032248429488390684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0020132821518927813, 'sampling/sampling_logp_difference/max': 0.6653909683227539, 'sampling/importance_sampling_ratio/min': 0.5140725374221802, 'sampling/importance_sampling_ratio/mean': 1.000615119934082, 'sampling/importance_sampling_ratio/max': 1.7144917249679565, 'entropy': 0.006337960890959948, 'clip_ratio/low_mean': 0.000589622650295496, 'clip_ratio/low_min': 0.000589622650295496, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.000589622650295496, 'reward_total_mean': 0.5073539018630981, 'reward_meter_mean': 0.996587872505188, 'reward_meter_std': 0.0006334602949209511, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6363636255264282, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5073539018630981, 'reward_total_composite_std': 0.0003224806860089302, 'epoch': 0.05}

 35%|███▍      | 1150/3300 [1:02:49<3:29:00,  5.83s/it]INFO 04-11 23:39:40 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:24,  2.25s/it]

 23%|██▎       | 3/13 [00:08<00:29,  2.93s/it]

 31%|███       | 4/13 [00:13<00:35,  3.91s/it]

 38%|███▊      | 5/13 [00:19<00:34,  4.34s/it]

 46%|████▌     | 6/13 [00:25<00:34,  4.94s/it]

 54%|█████▍    | 7/13 [00:30<00:30,  5.09s/it]

 62%|██████▏   | 8/13 [00:36<00:26,  5.36s/it]

 69%|██████▉   | 9/13 [00:43<00:22,  5.74s/it]

 77%|███████▋  | 10/13 [00:48<00:17,  5.67s/it]

 85%|████████▍ | 11/13 [00:53<00:10,  5.39s/it]

 92%|█████████▏| 12/13 [00:59<00:05,  5.47s/it]

100%|██████████| 13/13 [01:04<00:00,  5.41s/it]2026-04-11 23:40:50,084 | INFO | train_grpo_train | metrics_logged mode=eval step=1150

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 69.6474, 'eval_samples_per_second': 1.493, 'eval_steps_per_second': 0.187, 'eval_num_tokens': 2592268.0, 'eval_completions/mean_length': 209.30769230769232, 'eval_completions/min_length': 63.53846153846154, 'eval_completions/max_length': 363.6923076923077, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 209.30769230769232, 'eval_completions/min_terminated_length': 63.53846153846154, 'eval_completions/max_terminated_length': 363.6923076923077, 'eval_rewards/meter/mean': 0.6754777202239404, 'eval_rewards/meter/std': 0.40071778343274045, 'eval_rewards/count_adherence/mean': 0.8692530210201557, 'eval_rewards/count_adherence/std': 0.11666750907897949, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.6637609280072726, 'eval_rewards/repeat_penalty/std': 0.22615024103568152, 'eval_rewards/total_composite/mean': 0.39963403802651626, 'eval_rewards/total_composite/std': 0.3056422472000122, 'eval_reward': 0.39963403802651626, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.007615503783409412, 'eval_sampling/sampling_logp_difference/max': 0.8880168061990005, 'eval_sampling/importance_sampling_ratio/min': 0.4256615604345615, 'eval_sampling/importance_sampling_ratio/mean': 1.001767112658574, 'eval_sampling/importance_sampling_ratio/max': 1.3572269219618578, 'eval_entropy': 0.06449572856609638, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.39963403802651626, 'eval_reward_meter_mean': 0.6754777202239404, 'eval_reward_meter_std': 0.40071778343274045, 'eval_reward_count_adherence_mean': 0.8692530210201557, 'eval_reward_count_adherence_std': 0.11666750907897949, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.6637609280072726, 'eval_reward_repeat_penalty_std': 0.22615024103568152, 'eval_reward_total_composite_mean': 0.39963403802651626, 'eval_reward_total_composite_std': 0.3056422472000122, 'epoch': 0.05}

 35%|███▍      | 1150/3300 [1:03:59<3:29:00,  5.83s/it]

100%|██████████| 13/13 [01:04<00:00,  5.41s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 35%|███▍      | 1151/3300 [1:04:07<16:28:06, 27.59s/it]2026-04-11 23:40:58,755 | INFO | train_grpo_train | metrics_logged mode=train step=1151

                                                        
{'loss': -0.0045, 'grad_norm': 0.24846500158309937, 'learning_rate': 6.515151515151516e-06, 'num_tokens': 2595104.0, 'completions/mean_length': 180.5, 'completions/min_length': 177.0, 'completions/max_length': 181.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 180.5, 'completions/min_terminated_length': 177.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.9975081086158752, 'rewards/meter/std': 0.000507785240188241, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4987540543079376, 'rewards/total_composite/std': 0.0002538926200941205, 'reward': 0.4987540543079376, 'reward_std': 0.0002538926200941205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0012201471254229546, 'sampling/sampling_logp_difference/max': 0.5224804878234863, 'sampling/importance_sampling_ratio/min': 0.90339595079422, 'sampling/importance_sampling_ratio/mean': 1.0008914470672607, 'sampling/importance_sampling_ratio/max': 1.6862051486968994, 'entropy': 0.0079670658451505, 'clip_ratio/low_mean': 0.0007062146905809641, 'clip_ratio/low_min': 0.0007062146905809641, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0007062146905809641, 'reward_total_mean': 0.4987540543079376, 'reward_meter_mean': 0.9975081086158752, 'reward_meter_std': 0.000507785240188241, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4987540543079376, 'reward_total_composite_std': 0.0002538926200941205, 'epoch': 0.05}

 35%|███▍      | 1151/3300 [1:04:07<16:28:06, 27.59s/it]INFO 04-11 23:40:58 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1152/3300 [1:04:12<12:22:30, 20.74s/it]2026-04-11 23:41:03,519 | INFO | train_grpo_train | metrics_logged mode=train step=1152

                                                        
{'loss': 0.0463, 'grad_norm': 17.063459396362305, 'learning_rate': 6.512121212121213e-06, 'num_tokens': 2596884.0, 'completions/mean_length': 66.5, 'completions/min_length': 57.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8263880014419556, 'rewards/meter/std': 0.2611541450023651, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8263880014419556, 'rewards/total_composite/std': 0.2611541450023651, 'reward': 0.8263880014419556, 'reward_std': 0.2611541450023651, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06748796254396439, 'sampling/sampling_logp_difference/max': 4.3918986320495605, 'sampling/importance_sampling_ratio/min': 0.012377207167446613, 'sampling/importance_sampling_ratio/mean': 1.0026739835739136, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3189387395977974, 'clip_ratio/low_mean': 0.020571128465235233, 'clip_ratio/low_min': 0.020571128465235233, 'clip_ratio/high_mean': 0.03683905629441142, 'clip_ratio/high_max': 0.03683905629441142, 'clip_ratio/region_mean': 0.057410184759646654, 'reward_total_mean': 0.8263880014419556, 'reward_meter_mean': 0.8263880014419556, 'reward_meter_std': 0.2611541450023651, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8263880014419556, 'reward_total_composite_std': 0.2611541450023651, 'epoch': 0.05}

 35%|███▍      | 1152/3300 [1:04:12<12:22:30, 20.74s/it]INFO 04-11 23:41:03 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1153/3300 [1:04:17<9:37:15, 16.13s/it] 2026-04-11 23:41:08,898 | INFO | train_grpo_train | metrics_logged mode=train step=1153

                                                       
{'loss': 0.0129, 'grad_norm': 3.4035446643829346, 'learning_rate': 6.5090909090909095e-06, 'num_tokens': 2599043.0, 'completions/mean_length': 114.875, 'completions/min_length': 112.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.875, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.9962472319602966, 'rewards/meter/std': 0.0029050002340227365, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8219736814498901, 'rewards/total_composite/std': 0.07157647609710693, 'reward': 0.8219736814498901, 'reward_std': 0.07157646119594574, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0268792062997818, 'sampling/sampling_logp_difference/max': 1.2601839303970337, 'sampling/importance_sampling_ratio/min': 0.283601850271225, 'sampling/importance_sampling_ratio/mean': 0.9999147057533264, 'sampling/importance_sampling_ratio/max': 1.9995946884155273, 'entropy': 0.1615551235154271, 'clip_ratio/low_mean': 0.018382577574811876, 'clip_ratio/low_min': 0.018382577574811876, 'clip_ratio/high_mean': 0.0021929824724793434, 'clip_ratio/high_max': 0.0021929824724793434, 'clip_ratio/region_mean': 0.02057556004729122, 'reward_total_mean': 0.8219736814498901, 'reward_meter_mean': 0.9962472319602966, 'reward_meter_std': 0.0029050002340227365, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8219736814498901, 'reward_total_composite_std': 0.07157647609710693, 'epoch': 0.05}

 35%|███▍      | 1153/3300 [1:04:17<9:37:15, 16.13s/it]INFO 04-11 23:41:09 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1154/3300 [1:04:22<7:29:08, 12.56s/it]2026-04-11 23:41:13,116 | INFO | train_grpo_train | metrics_logged mode=train step=1154

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.506060606060607e-06, 'num_tokens': 2600507.0, 'completions/mean_length': 37.0, 'completions/min_length': 37.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9963316321372986, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9963316321372986, 'rewards/total_composite/std': 0.0, 'reward': 0.9963316321372986, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0014280767645686865, 'sampling/sampling_logp_difference/max': 0.04976058006286621, 'sampling/importance_sampling_ratio/min': 0.9514572024345398, 'sampling/importance_sampling_ratio/mean': 1.0005978345870972, 'sampling/importance_sampling_ratio/max': 1.0367541313171387, 'entropy': 0.014172183815389872, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9963316321372986, 'reward_meter_mean': 0.9963316321372986, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9963316321372986, 'reward_total_composite_std': 0.0, 'epoch': 0.05}

 35%|███▍      | 1154/3300 [1:04:22<7:29:08, 12.56s/it]INFO 04-11 23:41:13 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1155/3300 [1:04:28<6:27:08, 10.83s/it]2026-04-11 23:41:19,913 | INFO | train_grpo_train | metrics_logged mode=train step=1155

                                                       
{'loss': -0.0352, 'grad_norm': 1.2776048183441162, 'learning_rate': 6.503030303030303e-06, 'num_tokens': 2603752.0, 'completions/mean_length': 222.625, 'completions/min_length': 213.0, 'completions/max_length': 241.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 222.625, 'completions/min_terminated_length': 213.0, 'completions/max_terminated_length': 241.0, 'rewards/meter/mean': 0.9973673820495605, 'rewards/meter/std': 0.000317126396112144, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.3440934121608734, 'rewards/repeat_penalty/std': 0.21156011521816254, 'rewards/total_composite/mean': 0.2860187888145447, 'rewards/total_composite/std': 0.17588701844215393, 'reward': 0.2860187888145447, 'reward_std': 0.17588701844215393, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012430640868842602, 'sampling/sampling_logp_difference/max': 2.1552317142486572, 'sampling/importance_sampling_ratio/min': 0.11587633937597275, 'sampling/importance_sampling_ratio/mean': 0.9986491203308105, 'sampling/importance_sampling_ratio/max': 1.5816864967346191, 'entropy': 0.04108358500525355, 'clip_ratio/low_mean': 0.005830406676977873, 'clip_ratio/low_min': 0.005830406676977873, 'clip_ratio/high_mean': 0.009366041573230177, 'clip_ratio/high_max': 0.009366041573230177, 'clip_ratio/region_mean': 0.01519644825020805, 'reward_total_mean': 0.2860187888145447, 'reward_meter_mean': 0.9973673820495605, 'reward_meter_std': 0.000317126396112144, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.3440934121608734, 'reward_repeat_penalty_std': 0.21156011521816254, 'reward_total_composite_mean': 0.2860187888145447, 'reward_total_composite_std': 0.17588701844215393, 'epoch': 0.05}

 35%|███▌      | 1155/3300 [1:04:28<6:27:08, 10.83s/it]INFO 04-11 23:41:20 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1156/3300 [1:04:33<5:18:23,  8.91s/it]2026-04-11 23:41:24,345 | INFO | train_grpo_train | metrics_logged mode=train step=1156

                                                       
{'loss': 0.0178, 'grad_norm': 7.644164562225342, 'learning_rate': 6.5000000000000004e-06, 'num_tokens': 2605292.0, 'completions/mean_length': 33.5, 'completions/min_length': 32.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9726088643074036, 'rewards/meter/std': 0.043839771300554276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9726088643074036, 'rewards/total_composite/std': 0.043839771300554276, 'reward': 0.9726088643074036, 'reward_std': 0.04383978247642517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041076283901929855, 'sampling/sampling_logp_difference/max': 1.6910080909729004, 'sampling/importance_sampling_ratio/min': 0.18433360755443573, 'sampling/importance_sampling_ratio/mean': 1.0058238506317139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16383972112089396, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.026069519110023975, 'clip_ratio/high_max': 0.026069519110023975, 'clip_ratio/region_mean': 0.02974598971195519, 'reward_total_mean': 0.9726088643074036, 'reward_meter_mean': 0.9726088643074036, 'reward_meter_std': 0.043839771300554276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9726088643074036, 'reward_total_composite_std': 0.043839771300554276, 'epoch': 0.05}

 35%|███▌      | 1156/3300 [1:04:33<5:18:23,  8.91s/it]INFO 04-11 23:41:24 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1157/3300 [1:04:37<4:31:45,  7.61s/it]2026-04-11 23:41:28,917 | INFO | train_grpo_train | metrics_logged mode=train step=1157

                                                       
{'loss': 0.0036, 'grad_norm': 5.127370357513428, 'learning_rate': 6.496969696969697e-06, 'num_tokens': 2607186.0, 'completions/mean_length': 62.75, 'completions/min_length': 61.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.75, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.4182388186454773, 'rewards/meter/std': 0.25260862708091736, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4182388186454773, 'rewards/total_composite/std': 0.25260862708091736, 'reward': 0.4182388186454773, 'reward_std': 0.25260859727859497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03417876362800598, 'sampling/sampling_logp_difference/max': 1.50307297706604, 'sampling/importance_sampling_ratio/min': 0.2224455326795578, 'sampling/importance_sampling_ratio/mean': 1.0079007148742676, 'sampling/importance_sampling_ratio/max': 1.7306538820266724, 'entropy': 0.1482592076063156, 'clip_ratio/low_mean': 0.03776956582441926, 'clip_ratio/low_min': 0.03776956582441926, 'clip_ratio/high_mean': 0.004000256070867181, 'clip_ratio/high_max': 0.004000256070867181, 'clip_ratio/region_mean': 0.04176982189528644, 'reward_total_mean': 0.4182388186454773, 'reward_meter_mean': 0.4182388186454773, 'reward_meter_std': 0.25260862708091736, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4182388186454773, 'reward_total_composite_std': 0.25260862708091736, 'epoch': 0.05}

 35%|███▌      | 1157/3300 [1:04:37<4:31:45,  7.61s/it]INFO 04-11 23:41:29 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1158/3300 [1:04:42<3:57:59,  6.67s/it]2026-04-11 23:41:33,383 | INFO | train_grpo_train | metrics_logged mode=train step=1158

                                                       
{'loss': 0.0042, 'grad_norm': 3.4404122829437256, 'learning_rate': 6.493939393939395e-06, 'num_tokens': 2608844.0, 'completions/mean_length': 55.25, 'completions/min_length': 54.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.1349601447582245, 'rewards/meter/std': 0.012392125092446804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7083333730697632, 'rewards/repeat_penalty/std': 0.11785111576318741, 'rewards/total_composite/mean': 0.09517530351877213, 'rewards/total_composite/std': 0.01430203951895237, 'reward': 0.09517530351877213, 'reward_std': 0.01430203951895237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016672344878315926, 'sampling/sampling_logp_difference/max': 1.2043662071228027, 'sampling/importance_sampling_ratio/min': 0.2998819947242737, 'sampling/importance_sampling_ratio/mean': 1.0016802549362183, 'sampling/importance_sampling_ratio/max': 1.3683043718338013, 'entropy': 0.07266200426965952, 'clip_ratio/low_mean': 0.008780332282185555, 'clip_ratio/low_min': 0.008780332282185555, 'clip_ratio/high_mean': 0.004545454401522875, 'clip_ratio/high_max': 0.004545454401522875, 'clip_ratio/region_mean': 0.01332578668370843, 'reward_total_mean': 0.09517530351877213, 'reward_meter_mean': 0.1349601447582245, 'reward_meter_std': 0.012392125092446804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7083333730697632, 'reward_repeat_penalty_std': 0.11785111576318741, 'reward_total_composite_mean': 0.09517530351877213, 'reward_total_composite_std': 0.01430203951895237, 'epoch': 0.05}

 35%|███▌      | 1158/3300 [1:04:42<3:57:59,  6.67s/it]INFO 04-11 23:41:33 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1159/3300 [1:04:46<3:31:42,  5.93s/it]2026-04-11 23:41:37,605 | INFO | train_grpo_train | metrics_logged mode=train step=1159

                                                       
{'loss': 0.0034, 'grad_norm': 10.866904258728027, 'learning_rate': 6.490909090909091e-06, 'num_tokens': 2610338.0, 'completions/mean_length': 32.75, 'completions/min_length': 32.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9911776781082153, 'rewards/meter/std': 0.0027705347165465355, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9911776781082153, 'rewards/total_composite/std': 0.0027705347165465355, 'reward': 0.9911776781082153, 'reward_std': 0.002770546358078718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017460720613598824, 'sampling/sampling_logp_difference/max': 0.969951868057251, 'sampling/importance_sampling_ratio/min': 0.3791012763977051, 'sampling/importance_sampling_ratio/mean': 1.002776861190796, 'sampling/importance_sampling_ratio/max': 1.708295226097107, 'entropy': 0.0744382287375629, 'clip_ratio/low_mean': 0.011482007801532745, 'clip_ratio/low_min': 0.011482007801532745, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.015269886702299118, 'reward_total_mean': 0.9911776781082153, 'reward_meter_mean': 0.9911776781082153, 'reward_meter_std': 0.0027705347165465355, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9911776781082153, 'reward_total_composite_std': 0.0027705347165465355, 'epoch': 0.05}

 35%|███▌      | 1159/3300 [1:04:46<3:31:42,  5.93s/it]INFO 04-11 23:41:37 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1160/3300 [1:04:50<3:14:34,  5.46s/it]2026-04-11 23:41:41,949 | INFO | train_grpo_train | metrics_logged mode=train step=1160

                                                       
{'loss': 0.0007, 'grad_norm': 5.433602809906006, 'learning_rate': 6.487878787878789e-06, 'num_tokens': 2612174.0, 'completions/mean_length': 56.5, 'completions/min_length': 53.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.5, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.19516333937644958, 'rewards/meter/std': 0.10145802795886993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.19516333937644958, 'rewards/total_composite/std': 0.10145802795886993, 'reward': 0.19516333937644958, 'reward_std': 0.10145802050828934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030529655516147614, 'sampling/sampling_logp_difference/max': 4.638236999511719, 'sampling/importance_sampling_ratio/min': 0.009674739092588425, 'sampling/importance_sampling_ratio/mean': 1.0025193691253662, 'sampling/importance_sampling_ratio/max': 1.9637987613677979, 'entropy': 0.07584757776930928, 'clip_ratio/low_mean': 0.004551473073661327, 'clip_ratio/low_min': 0.004551473073661327, 'clip_ratio/high_mean': 0.0021929824724793434, 'clip_ratio/high_max': 0.0021929824724793434, 'clip_ratio/region_mean': 0.006744455546140671, 'reward_total_mean': 0.19516333937644958, 'reward_meter_mean': 0.19516333937644958, 'reward_meter_std': 0.10145802795886993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.19516333937644958, 'reward_total_composite_std': 0.10145802795886993, 'epoch': 0.05}

 35%|███▌      | 1160/3300 [1:04:50<3:14:34,  5.46s/it]INFO 04-11 23:41:42 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1161/3300 [1:04:55<3:01:17,  5.09s/it]2026-04-11 23:41:46,170 | INFO | train_grpo_train | metrics_logged mode=train step=1161

                                                       
{'loss': -0.0148, 'grad_norm': 12.443059921264648, 'learning_rate': 6.484848484848485e-06, 'num_tokens': 2613517.0, 'completions/mean_length': 31.875, 'completions/min_length': 31.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9917822480201721, 'rewards/meter/std': 0.003353093983605504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9917822480201721, 'rewards/total_composite/std': 0.003353093983605504, 'reward': 0.9917822480201721, 'reward_std': 0.003353102831169963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02346481755375862, 'sampling/sampling_logp_difference/max': 0.5687723755836487, 'sampling/importance_sampling_ratio/min': 0.6050869226455688, 'sampling/importance_sampling_ratio/mean': 1.0028786659240723, 'sampling/importance_sampling_ratio/max': 1.7660976648330688, 'entropy': 0.11351501289755106, 'clip_ratio/low_mean': 0.02016128972172737, 'clip_ratio/low_min': 0.02016128972172737, 'clip_ratio/high_mean': 0.011600378900766373, 'clip_ratio/high_max': 0.011600378900766373, 'clip_ratio/region_mean': 0.031761668622493744, 'reward_total_mean': 0.9917822480201721, 'reward_meter_mean': 0.9917822480201721, 'reward_meter_std': 0.003353093983605504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9917822480201721, 'reward_total_composite_std': 0.003353093983605504, 'epoch': 0.05}

 35%|███▌      | 1161/3300 [1:04:55<3:01:17,  5.09s/it]INFO 04-11 23:41:46 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1162/3300 [1:05:00<3:02:16,  5.12s/it]2026-04-11 23:41:51,353 | INFO | train_grpo_train | metrics_logged mode=train step=1162

                                                       
{'loss': 0.0034, 'grad_norm': 3.7559328079223633, 'learning_rate': 6.481818181818182e-06, 'num_tokens': 2615967.0, 'completions/mean_length': 116.25, 'completions/min_length': 114.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.25, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9908610582351685, 'rewards/meter/std': 0.010183922946453094, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.891869068145752, 'rewards/total_composite/std': 0.10717974603176117, 'reward': 0.891869068145752, 'reward_std': 0.10717973858118057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03219734504818916, 'sampling/sampling_logp_difference/max': 1.7524470090866089, 'sampling/importance_sampling_ratio/min': 0.17334923148155212, 'sampling/importance_sampling_ratio/mean': 0.9961613416671753, 'sampling/importance_sampling_ratio/max': 1.661534309387207, 'entropy': 0.16215351317077875, 'clip_ratio/low_mean': 0.005444317124783993, 'clip_ratio/low_min': 0.005444317124783993, 'clip_ratio/high_mean': 0.027328122640028596, 'clip_ratio/high_max': 0.027328122640028596, 'clip_ratio/region_mean': 0.03277243976481259, 'reward_total_mean': 0.891869068145752, 'reward_meter_mean': 0.9908610582351685, 'reward_meter_std': 0.010183922946453094, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.891869068145752, 'reward_total_composite_std': 0.10717974603176117, 'epoch': 0.05}

 35%|███▌      | 1162/3300 [1:05:00<3:02:16,  5.12s/it]INFO 04-11 23:41:51 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1163/3300 [1:05:06<3:16:34,  5.52s/it]2026-04-11 23:41:57,817 | INFO | train_grpo_train | metrics_logged mode=train step=1163

                                                       
{'loss': -0.0674, 'grad_norm': 2.036916971206665, 'learning_rate': 6.478787878787879e-06, 'num_tokens': 2619154.0, 'completions/mean_length': 211.375, 'completions/min_length': 152.0, 'completions/max_length': 232.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 211.375, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 232.0, 'rewards/meter/mean': 0.8547403812408447, 'rewards/meter/std': 0.34573426842689514, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4711538553237915, 'rewards/repeat_penalty/std': 0.2268439084291458, 'rewards/total_composite/mean': 0.31233876943588257, 'rewards/total_composite/std': 0.20915131270885468, 'reward': 0.31233876943588257, 'reward_std': 0.20915129780769348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010163981467485428, 'sampling/sampling_logp_difference/max': 0.8715519905090332, 'sampling/importance_sampling_ratio/min': 0.41830188035964966, 'sampling/importance_sampling_ratio/mean': 0.9999427795410156, 'sampling/importance_sampling_ratio/max': 1.551490068435669, 'entropy': 0.044531215680763125, 'clip_ratio/low_mean': 0.003117707441560924, 'clip_ratio/low_min': 0.003117707441560924, 'clip_ratio/high_mean': 0.003341847565025091, 'clip_ratio/high_max': 0.003341847565025091, 'clip_ratio/region_mean': 0.006459555006586015, 'reward_total_mean': 0.31233876943588257, 'reward_meter_mean': 0.8547403812408447, 'reward_meter_std': 0.34573426842689514, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4711538553237915, 'reward_repeat_penalty_std': 0.2268439084291458, 'reward_total_composite_mean': 0.31233876943588257, 'reward_total_composite_std': 0.20915131270885468, 'epoch': 0.05}

 35%|███▌      | 1163/3300 [1:05:06<3:16:34,  5.52s/it]INFO 04-11 23:41:58 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1164/3300 [1:05:11<3:06:35,  5.24s/it]2026-04-11 23:42:02,410 | INFO | train_grpo_train | metrics_logged mode=train step=1164

                                                       
{'loss': 0.0062, 'grad_norm': 4.857386112213135, 'learning_rate': 6.475757575757576e-06, 'num_tokens': 2621022.0, 'completions/mean_length': 75.5, 'completions/min_length': 75.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.905701756477356, 'rewards/meter/std': 0.22805918753147125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.905701756477356, 'rewards/total_composite/std': 0.22805918753147125, 'reward': 0.905701756477356, 'reward_std': 0.22805917263031006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014857825823128223, 'sampling/sampling_logp_difference/max': 0.9470729827880859, 'sampling/importance_sampling_ratio/min': 0.38787469267845154, 'sampling/importance_sampling_ratio/mean': 1.003965139389038, 'sampling/importance_sampling_ratio/max': 1.5772894620895386, 'entropy': 0.06199400080367923, 'clip_ratio/low_mean': 0.0016447368543595076, 'clip_ratio/low_min': 0.0016447368543595076, 'clip_ratio/high_mean': 0.014912280952557921, 'clip_ratio/high_max': 0.014912280952557921, 'clip_ratio/region_mean': 0.01655701780691743, 'reward_total_mean': 0.905701756477356, 'reward_meter_mean': 0.905701756477356, 'reward_meter_std': 0.22805918753147125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.905701756477356, 'reward_total_composite_std': 0.22805918753147125, 'epoch': 0.05}

 35%|███▌      | 1164/3300 [1:05:11<3:06:35,  5.24s/it]INFO 04-11 23:42:02 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1165/3300 [1:05:15<2:55:22,  4.93s/it]2026-04-11 23:42:06,610 | INFO | train_grpo_train | metrics_logged mode=train step=1165

                                                       
{'loss': 0.0288, 'grad_norm': 4.955336570739746, 'learning_rate': 6.472727272727272e-06, 'num_tokens': 2622496.0, 'completions/mean_length': 40.25, 'completions/min_length': 37.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9979323148727417, 'rewards/meter/std': 0.0005710619152523577, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979323148727417, 'rewards/total_composite/std': 0.0005710619152523577, 'reward': 0.9979323148727417, 'reward_std': 0.0005710580153390765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05128004774451256, 'sampling/sampling_logp_difference/max': 1.0725364685058594, 'sampling/importance_sampling_ratio/min': 0.3421395719051361, 'sampling/importance_sampling_ratio/mean': 0.9936442971229553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19564981944859028, 'clip_ratio/low_mean': 0.009149970952421427, 'clip_ratio/low_min': 0.009149970952421427, 'clip_ratio/high_mean': 0.02869208576157689, 'clip_ratio/high_max': 0.02869208576157689, 'clip_ratio/region_mean': 0.03784205671399832, 'reward_total_mean': 0.9979323148727417, 'reward_meter_mean': 0.9979323148727417, 'reward_meter_std': 0.0005710619152523577, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979323148727417, 'reward_total_composite_std': 0.0005710619152523577, 'epoch': 0.05}

 35%|███▌      | 1165/3300 [1:05:15<2:55:22,  4.93s/it]INFO 04-11 23:42:06 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1166/3300 [1:05:20<2:50:10,  4.78s/it]2026-04-11 23:42:11,056 | INFO | train_grpo_train | metrics_logged mode=train step=1166

                                                       
{'loss': -0.0054, 'grad_norm': 10.79326057434082, 'learning_rate': 6.4696969696969705e-06, 'num_tokens': 2624300.0, 'completions/mean_length': 66.5, 'completions/min_length': 63.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9930996894836426, 'rewards/meter/std': 0.0033798664808273315, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9930996894836426, 'rewards/total_composite/std': 0.0033798664808273315, 'reward': 0.9930996894836426, 'reward_std': 0.003379874862730503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02763393148779869, 'sampling/sampling_logp_difference/max': 2.4080264568328857, 'sampling/importance_sampling_ratio/min': 0.08999272435903549, 'sampling/importance_sampling_ratio/mean': 0.9978261590003967, 'sampling/importance_sampling_ratio/max': 1.7809786796569824, 'entropy': 0.0714933155104518, 'clip_ratio/low_mean': 0.007695082924328744, 'clip_ratio/low_min': 0.007695082924328744, 'clip_ratio/high_mean': 0.014788191299885511, 'clip_ratio/high_max': 0.014788191299885511, 'clip_ratio/region_mean': 0.022483274224214256, 'reward_total_mean': 0.9930996894836426, 'reward_meter_mean': 0.9930996894836426, 'reward_meter_std': 0.0033798664808273315, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9930996894836426, 'reward_total_composite_std': 0.0033798664808273315, 'epoch': 0.05}

 35%|███▌      | 1166/3300 [1:05:20<2:50:10,  4.78s/it]INFO 04-11 23:42:11 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1167/3300 [1:05:26<3:04:35,  5.19s/it]2026-04-11 23:42:17,201 | INFO | train_grpo_train | metrics_logged mode=train step=1167

                                                       
{'loss': 0.0106, 'grad_norm': 0.7076108455657959, 'learning_rate': 6.466666666666667e-06, 'num_tokens': 2627412.0, 'completions/mean_length': 203.0, 'completions/min_length': 197.0, 'completions/max_length': 210.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 203.0, 'completions/min_terminated_length': 197.0, 'completions/max_terminated_length': 210.0, 'rewards/meter/mean': 0.07068999111652374, 'rewards/meter/std': 0.12723270058631897, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6153846383094788, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.03625128045678139, 'rewards/total_composite/std': 0.0652475357055664, 'reward': 0.03625128045678139, 'reward_std': 0.06524752825498581, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005202981643378735, 'sampling/sampling_logp_difference/max': 2.0959839820861816, 'sampling/importance_sampling_ratio/min': 0.12294920533895493, 'sampling/importance_sampling_ratio/mean': 1.00026535987854, 'sampling/importance_sampling_ratio/max': 1.9165208339691162, 'entropy': 0.020221689250320196, 'clip_ratio/low_mean': 0.005486392183229327, 'clip_ratio/low_min': 0.005486392183229327, 'clip_ratio/high_mean': 0.0012690355069935322, 'clip_ratio/high_max': 0.0012690355069935322, 'clip_ratio/region_mean': 0.006755427690222859, 'reward_total_mean': 0.03625128045678139, 'reward_meter_mean': 0.07068999111652374, 'reward_meter_std': 0.12723270058631897, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6153846383094788, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.03625128045678139, 'reward_total_composite_std': 0.0652475357055664, 'epoch': 0.05}

 35%|███▌      | 1167/3300 [1:05:26<3:04:35,  5.19s/it]INFO 04-11 23:42:17 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1168/3300 [1:05:30<2:59:49,  5.06s/it]2026-04-11 23:42:21,958 | INFO | train_grpo_train | metrics_logged mode=train step=1168

                                                       
{'loss': 0.0086, 'grad_norm': 3.631923198699951, 'learning_rate': 6.463636363636364e-06, 'num_tokens': 2629391.0, 'completions/mean_length': 86.375, 'completions/min_length': 85.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.375, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.07566258311271667, 'rewards/meter/std': 0.10104013234376907, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.06269732117652893, 'rewards/total_composite/std': 0.08014771342277527, 'reward': 0.06269732117652893, 'reward_std': 0.08014770597219467, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009703797288239002, 'sampling/sampling_logp_difference/max': 1.0345752239227295, 'sampling/importance_sampling_ratio/min': 0.35537731647491455, 'sampling/importance_sampling_ratio/mean': 0.9974919557571411, 'sampling/importance_sampling_ratio/max': 1.388146162033081, 'entropy': 0.04172563157044351, 'clip_ratio/low_mean': 0.0043441514717414975, 'clip_ratio/low_min': 0.0043441514717414975, 'clip_ratio/high_mean': 0.00147058826405555, 'clip_ratio/high_max': 0.00147058826405555, 'clip_ratio/region_mean': 0.005814739735797048, 'reward_total_mean': 0.06269732117652893, 'reward_meter_mean': 0.07566258311271667, 'reward_meter_std': 0.10104013234376907, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.06269732117652893, 'reward_total_composite_std': 0.08014771342277527, 'epoch': 0.05}

 35%|███▌      | 1168/3300 [1:05:30<2:59:49,  5.06s/it]INFO 04-11 23:42:22 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1169/3300 [1:05:38<3:22:14,  5.69s/it]2026-04-11 23:42:29,127 | INFO | train_grpo_train | metrics_logged mode=train step=1169

                                                       
{'loss': -0.0418, 'grad_norm': 1.607576847076416, 'learning_rate': 6.460606060606061e-06, 'num_tokens': 2633025.0, 'completions/mean_length': 262.25, 'completions/min_length': 234.0, 'completions/max_length': 285.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 262.25, 'completions/min_terminated_length': 234.0, 'completions/max_terminated_length': 285.0, 'rewards/meter/mean': 0.9962061643600464, 'rewards/meter/std': 0.0021561330650001764, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6634615659713745, 'rewards/repeat_penalty/std': 0.05723259598016739, 'rewards/total_composite/mean': 0.5507833957672119, 'rewards/total_composite/std': 0.04750329256057739, 'reward': 0.5507833957672119, 'reward_std': 0.04750329256057739, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010979145765304565, 'sampling/sampling_logp_difference/max': 1.3150863647460938, 'sampling/importance_sampling_ratio/min': 0.26845112442970276, 'sampling/importance_sampling_ratio/mean': 1.0013923645019531, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.047941830940544605, 'clip_ratio/low_mean': 0.0029717457364313304, 'clip_ratio/low_min': 0.0029717457364313304, 'clip_ratio/high_mean': 0.00317868051934056, 'clip_ratio/high_max': 0.00317868051934056, 'clip_ratio/region_mean': 0.00615042625577189, 'reward_total_mean': 0.5507833957672119, 'reward_meter_mean': 0.9962061643600464, 'reward_meter_std': 0.0021561330650001764, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6634615659713745, 'reward_repeat_penalty_std': 0.05723259598016739, 'reward_total_composite_mean': 0.5507833957672119, 'reward_total_composite_std': 0.04750329256057739, 'epoch': 0.05}

 35%|███▌      | 1169/3300 [1:05:38<3:22:14,  5.69s/it]INFO 04-11 23:42:29 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1170/3300 [1:05:42<3:10:42,  5.37s/it]2026-04-11 23:42:33,750 | INFO | train_grpo_train | metrics_logged mode=train step=1170

                                                       
{'loss': -0.0059, 'grad_norm': 4.66847562789917, 'learning_rate': 6.457575757575758e-06, 'num_tokens': 2635041.0, 'completions/mean_length': 80.0, 'completions/min_length': 79.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9650785326957703, 'rewards/meter/std': 0.06653696298599243, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9650785326957703, 'rewards/total_composite/std': 0.06653696298599243, 'reward': 0.9650785326957703, 'reward_std': 0.06653697043657303, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03951912745833397, 'sampling/sampling_logp_difference/max': 1.5156645774841309, 'sampling/importance_sampling_ratio/min': 0.2196621596813202, 'sampling/importance_sampling_ratio/mean': 1.0022578239440918, 'sampling/importance_sampling_ratio/max': 1.8339917659759521, 'entropy': 0.24884118884801865, 'clip_ratio/low_mean': 0.0031645570416003466, 'clip_ratio/low_min': 0.0031645570416003466, 'clip_ratio/high_mean': 0.02818627143278718, 'clip_ratio/high_max': 0.02818627143278718, 'clip_ratio/region_mean': 0.031350828474387527, 'reward_total_mean': 0.9650785326957703, 'reward_meter_mean': 0.9650785326957703, 'reward_meter_std': 0.06653696298599243, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9650785326957703, 'reward_total_composite_std': 0.06653696298599243, 'epoch': 0.05}

 35%|███▌      | 1170/3300 [1:05:42<3:10:42,  5.37s/it]INFO 04-11 23:42:33 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1171/3300 [1:05:47<3:03:30,  5.17s/it]2026-04-11 23:42:38,453 | INFO | train_grpo_train | metrics_logged mode=train step=1171

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.454545454545456e-06, 'num_tokens': 2636953.0, 'completions/mean_length': 85.0, 'completions/min_length': 85.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.0, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.3254895508289337, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.2603916525840759, 'rewards/total_composite/std': 0.0, 'reward': 0.2603916525840759, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0026430224534124136, 'sampling/sampling_logp_difference/max': 0.10097821056842804, 'sampling/importance_sampling_ratio/min': 0.9174838662147522, 'sampling/importance_sampling_ratio/mean': 1.0015860795974731, 'sampling/importance_sampling_ratio/max': 1.1062525510787964, 'entropy': 0.023862487636506557, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.2603916525840759, 'reward_meter_mean': 0.3254895508289337, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.2603916525840759, 'reward_total_composite_std': 0.0, 'epoch': 0.05}

 35%|███▌      | 1171/3300 [1:05:47<3:03:30,  5.17s/it]INFO 04-11 23:42:38 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1172/3300 [1:05:52<3:02:59,  5.16s/it]2026-04-11 23:42:43,583 | INFO | train_grpo_train | metrics_logged mode=train step=1172

                                                       
{'loss': 0.008, 'grad_norm': 4.03551721572876, 'learning_rate': 6.451515151515152e-06, 'num_tokens': 2639474.0, 'completions/mean_length': 118.125, 'completions/min_length': 114.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.125, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9958237409591675, 'rewards/meter/std': 0.0014982123393565416, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.921150803565979, 'rewards/total_composite/std': 0.10321994870901108, 'reward': 0.921150803565979, 'reward_std': 0.10321993380784988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027086030691862106, 'sampling/sampling_logp_difference/max': 1.4236351251602173, 'sampling/importance_sampling_ratio/min': 0.2408369481563568, 'sampling/importance_sampling_ratio/mean': 1.001880407333374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13357019051909447, 'clip_ratio/low_mean': 0.010513697401620448, 'clip_ratio/low_min': 0.010513697401620448, 'clip_ratio/high_mean': 0.015921411802992225, 'clip_ratio/high_max': 0.015921411802992225, 'clip_ratio/region_mean': 0.026435109204612672, 'reward_total_mean': 0.921150803565979, 'reward_meter_mean': 0.9958237409591675, 'reward_meter_std': 0.0014982123393565416, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.921150803565979, 'reward_total_composite_std': 0.10321994870901108, 'epoch': 0.05}

 36%|███▌      | 1172/3300 [1:05:52<3:02:59,  5.16s/it]INFO 04-11 23:42:43 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1173/3300 [1:05:57<2:57:57,  5.02s/it]2026-04-11 23:42:48,277 | INFO | train_grpo_train | metrics_logged mode=train step=1173

                                                       
{'loss': 0.008, 'grad_norm': 3.1960055828094482, 'learning_rate': 6.4484848484848496e-06, 'num_tokens': 2641397.0, 'completions/mean_length': 76.375, 'completions/min_length': 74.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.988898754119873, 'rewards/meter/std': 0.0167799461632967, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.988898754119873, 'rewards/total_composite/std': 0.0167799461632967, 'reward': 0.988898754119873, 'reward_std': 0.016779951751232147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015661096200346947, 'sampling/sampling_logp_difference/max': 1.2024707794189453, 'sampling/importance_sampling_ratio/min': 0.3004509508609772, 'sampling/importance_sampling_ratio/mean': 1.0082311630249023, 'sampling/importance_sampling_ratio/max': 1.5719304084777832, 'entropy': 0.08165451744571328, 'clip_ratio/low_mean': 0.004870129749178886, 'clip_ratio/low_min': 0.004870129749178886, 'clip_ratio/high_mean': 0.008183791185729206, 'clip_ratio/high_max': 0.008183791185729206, 'clip_ratio/region_mean': 0.013053920934908092, 'reward_total_mean': 0.988898754119873, 'reward_meter_mean': 0.988898754119873, 'reward_meter_std': 0.0167799461632967, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.988898754119873, 'reward_total_composite_std': 0.0167799461632967, 'epoch': 0.05}

 36%|███▌      | 1173/3300 [1:05:57<2:57:57,  5.02s/it]INFO 04-11 23:42:48 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1174/3300 [1:06:01<2:54:15,  4.92s/it]2026-04-11 23:42:52,959 | INFO | train_grpo_train | metrics_logged mode=train step=1174

                                                       
{'loss': -0.0038, 'grad_norm': 3.3341424465179443, 'learning_rate': 6.445454545454546e-06, 'num_tokens': 2643116.0, 'completions/mean_length': 75.875, 'completions/min_length': 75.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9900845289230347, 'rewards/meter/std': 0.006058650091290474, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9900845289230347, 'rewards/total_composite/std': 0.006058650091290474, 'reward': 0.9900845289230347, 'reward_std': 0.006058644037693739, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017089035362005234, 'sampling/sampling_logp_difference/max': 0.9457058906555176, 'sampling/importance_sampling_ratio/min': 0.38840532302856445, 'sampling/importance_sampling_ratio/mean': 1.0023661851882935, 'sampling/importance_sampling_ratio/max': 1.7328146696090698, 'entropy': 0.08360559307038784, 'clip_ratio/low_mean': 0.0033333334140479565, 'clip_ratio/low_min': 0.0033333334140479565, 'clip_ratio/high_mean': 0.003289473708719015, 'clip_ratio/high_max': 0.003289473708719015, 'clip_ratio/region_mean': 0.006622807122766972, 'reward_total_mean': 0.9900845289230347, 'reward_meter_mean': 0.9900845289230347, 'reward_meter_std': 0.006058650091290474, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9900845289230347, 'reward_total_composite_std': 0.006058650091290474, 'epoch': 0.05}

 36%|███▌      | 1174/3300 [1:06:01<2:54:15,  4.92s/it]INFO 04-11 23:42:53 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1175/3300 [1:06:06<2:54:03,  4.91s/it]2026-04-11 23:42:57,864 | INFO | train_grpo_train | metrics_logged mode=train step=1175

                                                       
{'loss': 0.0211, 'grad_norm': 3.098088502883911, 'learning_rate': 6.442424242424243e-06, 'num_tokens': 2645197.0, 'completions/mean_length': 86.125, 'completions/min_length': 85.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.125, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.2395191788673401, 'rewards/meter/std': 0.14119157195091248, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.19161534309387207, 'rewards/total_composite/std': 0.11295326054096222, 'reward': 0.19161534309387207, 'reward_std': 0.11295326054096222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008148287422955036, 'sampling/sampling_logp_difference/max': 1.1111087799072266, 'sampling/importance_sampling_ratio/min': 0.32919374108314514, 'sampling/importance_sampling_ratio/mean': 1.0021647214889526, 'sampling/importance_sampling_ratio/max': 1.5297051668167114, 'entropy': 0.05733899096958339, 'clip_ratio/low_mean': 0.002794080995954573, 'clip_ratio/low_min': 0.002794080995954573, 'clip_ratio/high_mean': 0.004411764908581972, 'clip_ratio/high_max': 0.004411764908581972, 'clip_ratio/region_mean': 0.007205845904536545, 'reward_total_mean': 0.19161534309387207, 'reward_meter_mean': 0.2395191788673401, 'reward_meter_std': 0.14119157195091248, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.19161534309387207, 'reward_total_composite_std': 0.11295326054096222, 'epoch': 0.05}

 36%|███▌      | 1175/3300 [1:06:06<2:54:03,  4.91s/it]INFO 04-11 23:42:58 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1176/3300 [1:06:14<3:26:23,  5.83s/it]2026-04-11 23:43:05,834 | INFO | train_grpo_train | metrics_logged mode=train step=1176

                                                       
{'loss': -0.0027, 'grad_norm': 1.6563795804977417, 'learning_rate': 6.43939393939394e-06, 'num_tokens': 2649214.0, 'completions/mean_length': 277.125, 'completions/min_length': 268.0, 'completions/max_length': 292.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 277.125, 'completions/min_terminated_length': 268.0, 'completions/max_terminated_length': 292.0, 'rewards/meter/mean': 0.9914854764938354, 'rewards/meter/std': 0.0076880063861608505, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7403846383094788, 'rewards/repeat_penalty/std': 0.15350237488746643, 'rewards/total_composite/mean': 0.6114773750305176, 'rewards/total_composite/std': 0.12588950991630554, 'reward': 0.6114773750305176, 'reward_std': 0.12588950991630554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015588824637234211, 'sampling/sampling_logp_difference/max': 1.1217985153198242, 'sampling/importance_sampling_ratio/min': 0.3256934881210327, 'sampling/importance_sampling_ratio/mean': 1.0004220008850098, 'sampling/importance_sampling_ratio/max': 1.965925931930542, 'entropy': 0.09004434384405613, 'clip_ratio/low_mean': 0.0026997023087460548, 'clip_ratio/low_min': 0.0026997023087460548, 'clip_ratio/high_mean': 0.013525328016839921, 'clip_ratio/high_max': 0.013525328016839921, 'clip_ratio/region_mean': 0.016225030325585976, 'reward_total_mean': 0.6114773750305176, 'reward_meter_mean': 0.9914854764938354, 'reward_meter_std': 0.0076880063861608505, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7403846383094788, 'reward_repeat_penalty_std': 0.15350237488746643, 'reward_total_composite_mean': 0.6114773750305176, 'reward_total_composite_std': 0.12588950991630554, 'epoch': 0.05}

 36%|███▌      | 1176/3300 [1:06:14<3:26:23,  5.83s/it]INFO 04-11 23:43:06 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1177/3300 [1:06:19<3:12:13,  5.43s/it]2026-04-11 23:43:10,336 | INFO | train_grpo_train | metrics_logged mode=train step=1177

                                                       
{'loss': -0.0359, 'grad_norm': 4.175478935241699, 'learning_rate': 6.436363636363637e-06, 'num_tokens': 2650978.0, 'completions/mean_length': 58.5, 'completions/min_length': 57.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.36279141902923584, 'rewards/meter/std': 0.24407321214675903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.36279141902923584, 'rewards/total_composite/std': 0.24407321214675903, 'reward': 0.36279141902923584, 'reward_std': 0.24407321214675903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009452205151319504, 'sampling/sampling_logp_difference/max': 0.6868042945861816, 'sampling/importance_sampling_ratio/min': 0.503181517124176, 'sampling/importance_sampling_ratio/mean': 1.0017210245132446, 'sampling/importance_sampling_ratio/max': 1.3822486400604248, 'entropy': 0.06206447468139231, 'clip_ratio/low_mean': 0.006048386916518211, 'clip_ratio/low_min': 0.006048386916518211, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.009954636916518211, 'reward_total_mean': 0.36279141902923584, 'reward_meter_mean': 0.36279141902923584, 'reward_meter_std': 0.24407321214675903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.36279141902923584, 'reward_total_composite_std': 0.24407321214675903, 'epoch': 0.05}

 36%|███▌      | 1177/3300 [1:06:19<3:12:13,  5.43s/it]INFO 04-11 23:43:10 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1178/3300 [1:06:28<3:56:41,  6.69s/it]2026-04-11 23:43:19,967 | INFO | train_grpo_train | metrics_logged mode=train step=1178

                                                       
{'loss': -0.1566, 'grad_norm': 1.462255835533142, 'learning_rate': 6.433333333333333e-06, 'num_tokens': 2653979.0, 'completions/mean_length': 267.125, 'completions/min_length': 218.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 232.1428680419922, 'completions/min_terminated_length': 218.0, 'completions/max_terminated_length': 257.0, 'rewards/meter/mean': 0.4442686140537262, 'rewards/meter/std': 0.2729472219944, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.6192708611488342, 'rewards/repeat_penalty/std': 0.21683759987354279, 'rewards/total_composite/mean': 0.2283807098865509, 'rewards/total_composite/std': 0.17121197283267975, 'reward': 0.2283807098865509, 'reward_std': 0.17121195793151855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012943691574037075, 'sampling/sampling_logp_difference/max': 1.0156912803649902, 'sampling/importance_sampling_ratio/min': 0.44583451747894287, 'sampling/importance_sampling_ratio/mean': 1.0050114393234253, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07222116971388459, 'clip_ratio/low_mean': 0.0021399176912382245, 'clip_ratio/low_min': 0.0021399176912382245, 'clip_ratio/high_mean': 0.007645888428669423, 'clip_ratio/high_max': 0.007645888428669423, 'clip_ratio/region_mean': 0.009785806119907647, 'reward_total_mean': 0.2283807098865509, 'reward_meter_mean': 0.4442686140537262, 'reward_meter_std': 0.2729472219944, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.6192708611488342, 'reward_repeat_penalty_std': 0.21683759987354279, 'reward_total_composite_mean': 0.2283807098865509, 'reward_total_composite_std': 0.17121197283267975, 'epoch': 0.05}

 36%|███▌      | 1178/3300 [1:06:28<3:56:41,  6.69s/it]INFO 04-11 23:43:20 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1179/3300 [1:06:33<3:33:42,  6.05s/it]2026-04-11 23:43:24,504 | INFO | train_grpo_train | metrics_logged mode=train step=1179

                                                       
{'loss': 0.0133, 'grad_norm': 11.408062934875488, 'learning_rate': 6.430303030303031e-06, 'num_tokens': 2655696.0, 'completions/mean_length': 68.625, 'completions/min_length': 66.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9946950078010559, 'rewards/meter/std': 0.002590329386293888, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946950078010559, 'rewards/total_composite/std': 0.002590329386293888, 'reward': 0.9946950078010559, 'reward_std': 0.002590324031189084, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026325104758143425, 'sampling/sampling_logp_difference/max': 1.1628682613372803, 'sampling/importance_sampling_ratio/min': 0.31258830428123474, 'sampling/importance_sampling_ratio/mean': 1.0001428127288818, 'sampling/importance_sampling_ratio/max': 1.6397497653961182, 'entropy': 0.09837264008820057, 'clip_ratio/low_mean': 0.014470615307800472, 'clip_ratio/low_min': 0.014470615307800472, 'clip_ratio/high_mean': 0.016267166705802083, 'clip_ratio/high_max': 0.016267166705802083, 'clip_ratio/region_mean': 0.030737782013602555, 'reward_total_mean': 0.9946950078010559, 'reward_meter_mean': 0.9946950078010559, 'reward_meter_std': 0.002590329386293888, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946950078010559, 'reward_total_composite_std': 0.002590329386293888, 'epoch': 0.05}

 36%|███▌      | 1179/3300 [1:06:33<3:33:42,  6.05s/it]INFO 04-11 23:43:24 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1180/3300 [1:06:37<3:16:44,  5.57s/it]2026-04-11 23:43:28,958 | INFO | train_grpo_train | metrics_logged mode=train step=1180

                                                       
{'loss': 0.0028, 'grad_norm': 7.856446266174316, 'learning_rate': 6.427272727272728e-06, 'num_tokens': 2657383.0, 'completions/mean_length': 55.875, 'completions/min_length': 52.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.7185721397399902, 'rewards/meter/std': 0.25754180550575256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7185721397399902, 'rewards/total_composite/std': 0.25754180550575256, 'reward': 0.7185721397399902, 'reward_std': 0.25754180550575256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03376821056008339, 'sampling/sampling_logp_difference/max': 1.3033967018127441, 'sampling/importance_sampling_ratio/min': 0.2716076672077179, 'sampling/importance_sampling_ratio/mean': 0.9931161403656006, 'sampling/importance_sampling_ratio/max': 1.7777073383331299, 'entropy': 0.142734014429152, 'clip_ratio/low_mean': 0.011504121124744415, 'clip_ratio/low_min': 0.011504121124744415, 'clip_ratio/high_mean': 0.0242516309954226, 'clip_ratio/high_max': 0.0242516309954226, 'clip_ratio/region_mean': 0.03575575212016702, 'reward_total_mean': 0.7185721397399902, 'reward_meter_mean': 0.7185721397399902, 'reward_meter_std': 0.25754180550575256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7185721397399902, 'reward_total_composite_std': 0.25754180550575256, 'epoch': 0.05}

 36%|███▌      | 1180/3300 [1:06:37<3:16:44,  5.57s/it]INFO 04-11 23:43:29 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1181/3300 [1:06:46<3:46:28,  6.41s/it]2026-04-11 23:43:37,343 | INFO | train_grpo_train | metrics_logged mode=train step=1181

                                                       
{'loss': 0.0075, 'grad_norm': 1.3380672931671143, 'learning_rate': 6.424242424242425e-06, 'num_tokens': 2662002.0, 'completions/mean_length': 371.375, 'completions/min_length': 359.0, 'completions/max_length': 384.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 371.375, 'completions/min_terminated_length': 359.0, 'completions/max_terminated_length': 384.0, 'rewards/meter/mean': 0.9918642044067383, 'rewards/meter/std': 0.006125128827989101, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7426470518112183, 'rewards/repeat_penalty/std': 0.08281681686639786, 'rewards/total_composite/mean': 0.6446620225906372, 'rewards/total_composite/std': 0.07332957535982132, 'reward': 0.6446620225906372, 'reward_std': 0.07332959026098251, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013688583858311176, 'sampling/sampling_logp_difference/max': 1.4129068851470947, 'sampling/importance_sampling_ratio/min': 0.2434346228837967, 'sampling/importance_sampling_ratio/mean': 1.0013055801391602, 'sampling/importance_sampling_ratio/max': 1.6264017820358276, 'entropy': 0.0724606835283339, 'clip_ratio/low_mean': 0.004685206571593881, 'clip_ratio/low_min': 0.004685206571593881, 'clip_ratio/high_mean': 0.006440639495849609, 'clip_ratio/high_max': 0.006440639495849609, 'clip_ratio/region_mean': 0.01112584606744349, 'reward_total_mean': 0.6446620225906372, 'reward_meter_mean': 0.9918642044067383, 'reward_meter_std': 0.006125128827989101, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7426470518112183, 'reward_repeat_penalty_std': 0.08281681686639786, 'reward_total_composite_mean': 0.6446620225906372, 'reward_total_composite_std': 0.07332957535982132, 'epoch': 0.05}

 36%|███▌      | 1181/3300 [1:06:46<3:46:28,  6.41s/it]INFO 04-11 23:43:37 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1182/3300 [1:06:50<3:27:21,  5.87s/it]2026-04-11 23:43:41,959 | INFO | train_grpo_train | metrics_logged mode=train step=1182

                                                       
{'loss': -0.0062, 'grad_norm': 2.379884719848633, 'learning_rate': 6.4212121212121215e-06, 'num_tokens': 2663974.0, 'completions/mean_length': 77.5, 'completions/min_length': 76.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9949516654014587, 'rewards/meter/std': 0.0005235640564933419, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949516654014587, 'rewards/total_composite/std': 0.0005235640564933419, 'reward': 0.9949516654014587, 'reward_std': 0.0005235668504610658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010198801755905151, 'sampling/sampling_logp_difference/max': 0.6724367141723633, 'sampling/importance_sampling_ratio/min': 0.5104632377624512, 'sampling/importance_sampling_ratio/mean': 0.9996510148048401, 'sampling/importance_sampling_ratio/max': 1.353779911994934, 'entropy': 0.04025013535283506, 'clip_ratio/low_mean': 0.008097166079096496, 'clip_ratio/low_min': 0.008097166079096496, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.008097166079096496, 'reward_total_mean': 0.9949516654014587, 'reward_meter_mean': 0.9949516654014587, 'reward_meter_std': 0.0005235640564933419, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949516654014587, 'reward_total_composite_std': 0.0005235640564933419, 'epoch': 0.05}

 36%|███▌      | 1182/3300 [1:06:50<3:27:21,  5.87s/it]INFO 04-11 23:43:42 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1183/3300 [1:06:54<3:07:00,  5.30s/it]2026-04-11 23:43:45,921 | INFO | train_grpo_train | metrics_logged mode=train step=1183

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.418181818181819e-06, 'num_tokens': 2665278.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9889696836471558, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9889696836471558, 'rewards/total_composite/std': 0.0, 'reward': 0.9889696836471558, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.005411628168076277, 'sampling/sampling_logp_difference/max': 0.15753476321697235, 'sampling/importance_sampling_ratio/min': 0.8542470932006836, 'sampling/importance_sampling_ratio/mean': 0.9997879266738892, 'sampling/importance_sampling_ratio/max': 1.0598801374435425, 'entropy': 0.046004267409443855, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9889696836471558, 'reward_meter_mean': 0.9889696836471558, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9889696836471558, 'reward_total_composite_std': 0.0, 'epoch': 0.05}

 36%|███▌      | 1183/3300 [1:06:54<3:07:00,  5.30s/it]INFO 04-11 23:43:46 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1184/3300 [1:06:59<3:03:16,  5.20s/it]2026-04-11 23:43:50,877 | INFO | train_grpo_train | metrics_logged mode=train step=1184

                                                       
{'loss': 0.038, 'grad_norm': 9.193962097167969, 'learning_rate': 6.415151515151515e-06, 'num_tokens': 2667336.0, 'completions/mean_length': 97.25, 'completions/min_length': 93.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.25, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.892284095287323, 'rewards/meter/std': 0.20539914071559906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.724637508392334, 'rewards/total_composite/std': 0.1372780054807663, 'reward': 0.724637508392334, 'reward_std': 0.1372780054807663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04928457364439964, 'sampling/sampling_logp_difference/max': 1.7316360473632812, 'sampling/importance_sampling_ratio/min': 0.17699459195137024, 'sampling/importance_sampling_ratio/mean': 0.9922058582305908, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16915534622967243, 'clip_ratio/low_mean': 0.012163461884483695, 'clip_ratio/low_min': 0.012163461884483695, 'clip_ratio/high_mean': 0.020877854549326003, 'clip_ratio/high_max': 0.020877854549326003, 'clip_ratio/region_mean': 0.0330413164338097, 'reward_total_mean': 0.724637508392334, 'reward_meter_mean': 0.892284095287323, 'reward_meter_std': 0.20539914071559906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.724637508392334, 'reward_total_composite_std': 0.1372780054807663, 'epoch': 0.05}

 36%|███▌      | 1184/3300 [1:06:59<3:03:16,  5.20s/it]INFO 04-11 23:43:51 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1185/3300 [1:07:04<2:57:08,  5.03s/it]2026-04-11 23:43:55,500 | INFO | train_grpo_train | metrics_logged mode=train step=1185

                                                       
{'loss': -0.0083, 'grad_norm': 1.140117883682251, 'learning_rate': 6.412121212121213e-06, 'num_tokens': 2669230.0, 'completions/mean_length': 78.75, 'completions/min_length': 78.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.75, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9956543445587158, 'rewards/meter/std': 0.0006610880373045802, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9956543445587158, 'rewards/total_composite/std': 0.0006610880373045802, 'reward': 0.9956543445587158, 'reward_std': 0.0006611024145968258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00999254360795021, 'sampling/sampling_logp_difference/max': 0.8007946014404297, 'sampling/importance_sampling_ratio/min': 0.44897207617759705, 'sampling/importance_sampling_ratio/mean': 1.0024654865264893, 'sampling/importance_sampling_ratio/max': 1.5469458103179932, 'entropy': 0.040787032805383205, 'clip_ratio/low_mean': 0.004807692370377481, 'clip_ratio/low_min': 0.004807692370377481, 'clip_ratio/high_mean': 0.0031645570416003466, 'clip_ratio/high_max': 0.0031645570416003466, 'clip_ratio/region_mean': 0.007972249411977828, 'reward_total_mean': 0.9956543445587158, 'reward_meter_mean': 0.9956543445587158, 'reward_meter_std': 0.0006610880373045802, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9956543445587158, 'reward_total_composite_std': 0.0006610880373045802, 'epoch': 0.05}

 36%|███▌      | 1185/3300 [1:07:04<2:57:08,  5.03s/it]INFO 04-11 23:43:55 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1186/3300 [1:07:10<3:12:23,  5.46s/it]2026-04-11 23:44:01,980 | INFO | train_grpo_train | metrics_logged mode=train step=1186

                                                       
{'loss': 0.0014, 'grad_norm': 1.235166072845459, 'learning_rate': 6.40909090909091e-06, 'num_tokens': 2672128.0, 'completions/mean_length': 239.25, 'completions/min_length': 236.0, 'completions/max_length': 242.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 239.25, 'completions/min_terminated_length': 236.0, 'completions/max_terminated_length': 242.0, 'rewards/meter/mean': 0.9895985722541809, 'rewards/meter/std': 0.0064008040353655815, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5795454978942871, 'rewards/repeat_penalty/std': 0.19399181008338928, 'rewards/total_composite/mean': 0.4595944285392761, 'rewards/total_composite/std': 0.15526776015758514, 'reward': 0.4595944285392761, 'reward_std': 0.15526776015758514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007288929540663958, 'sampling/sampling_logp_difference/max': 1.047019124031067, 'sampling/importance_sampling_ratio/min': 0.3509824275970459, 'sampling/importance_sampling_ratio/mean': 1.0002281665802002, 'sampling/importance_sampling_ratio/max': 1.5414388179779053, 'entropy': 0.029086438240483403, 'clip_ratio/low_mean': 0.0010416667209938169, 'clip_ratio/low_min': 0.0010416667209938169, 'clip_ratio/high_mean': 0.006786415702663362, 'clip_ratio/high_max': 0.006786415702663362, 'clip_ratio/region_mean': 0.007828082423657179, 'reward_total_mean': 0.4595944285392761, 'reward_meter_mean': 0.9895985722541809, 'reward_meter_std': 0.0064008040353655815, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5795454978942871, 'reward_repeat_penalty_std': 0.19399181008338928, 'reward_total_composite_mean': 0.4595944285392761, 'reward_total_composite_std': 0.15526776015758514, 'epoch': 0.05}

 36%|███▌      | 1186/3300 [1:07:10<3:12:23,  5.46s/it]INFO 04-11 23:44:02 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1187/3300 [1:07:19<3:42:18,  6.31s/it]2026-04-11 23:44:10,278 | INFO | train_grpo_train | metrics_logged mode=train step=1187

                                                       
{'loss': 0.0002, 'grad_norm': 1.61380934715271, 'learning_rate': 6.406060606060607e-06, 'num_tokens': 2676850.0, 'completions/mean_length': 363.25, 'completions/min_length': 348.0, 'completions/max_length': 377.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 363.25, 'completions/min_terminated_length': 348.0, 'completions/max_terminated_length': 377.0, 'rewards/meter/mean': 0.9924304485321045, 'rewards/meter/std': 0.012585594318807125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6617647409439087, 'rewards/repeat_penalty/std': 0.027230001986026764, 'rewards/total_composite/mean': 0.6568292379379272, 'rewards/total_composite/std': 0.0302126407623291, 'reward': 0.6568292379379272, 'reward_std': 0.030212653800845146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012508058920502663, 'sampling/sampling_logp_difference/max': 1.4292956590652466, 'sampling/importance_sampling_ratio/min': 0.2921198010444641, 'sampling/importance_sampling_ratio/mean': 1.0015922784805298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06546153873205185, 'clip_ratio/low_mean': 0.006156876712338999, 'clip_ratio/low_min': 0.006156876712338999, 'clip_ratio/high_mean': 0.0030575180426239967, 'clip_ratio/high_max': 0.0030575180426239967, 'clip_ratio/region_mean': 0.009214394754962996, 'reward_total_mean': 0.6568292379379272, 'reward_meter_mean': 0.9924304485321045, 'reward_meter_std': 0.012585594318807125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6617647409439087, 'reward_repeat_penalty_std': 0.027230001986026764, 'reward_total_composite_mean': 0.6568292379379272, 'reward_total_composite_std': 0.0302126407623291, 'epoch': 0.05}

 36%|███▌      | 1187/3300 [1:07:19<3:42:18,  6.31s/it]INFO 04-11 23:44:10 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1188/3300 [1:07:27<4:00:31,  6.83s/it]2026-04-11 23:44:18,325 | INFO | train_grpo_train | metrics_logged mode=train step=1188

                                                       
{'loss': 0.0066, 'grad_norm': 1.0509389638900757, 'learning_rate': 6.403030303030303e-06, 'num_tokens': 2681346.0, 'completions/mean_length': 364.0, 'completions/min_length': 360.0, 'completions/max_length': 368.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 364.0, 'completions/min_terminated_length': 360.0, 'completions/max_terminated_length': 368.0, 'rewards/meter/mean': 0.9938761591911316, 'rewards/meter/std': 0.0003768211172427982, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6029411554336548, 'rewards/repeat_penalty/std': 0.027230001986026764, 'rewards/total_composite/mean': 0.5992562770843506, 'rewards/total_composite/std': 0.027259474620223045, 'reward': 0.5992562770843506, 'reward_std': 0.0272594653069973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003385041607543826, 'sampling/sampling_logp_difference/max': 0.9387906789779663, 'sampling/importance_sampling_ratio/min': 0.3911004960536957, 'sampling/importance_sampling_ratio/mean': 1.0006787776947021, 'sampling/importance_sampling_ratio/max': 1.6698790788650513, 'entropy': 0.01585481537040323, 'clip_ratio/low_mean': 0.00206068845000118, 'clip_ratio/low_min': 0.00206068845000118, 'clip_ratio/high_mean': 0.0013888889516238123, 'clip_ratio/high_max': 0.0013888889516238123, 'clip_ratio/region_mean': 0.0034495774016249925, 'reward_total_mean': 0.5992562770843506, 'reward_meter_mean': 0.9938761591911316, 'reward_meter_std': 0.0003768211172427982, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6029411554336548, 'reward_repeat_penalty_std': 0.027230001986026764, 'reward_total_composite_mean': 0.5992562770843506, 'reward_total_composite_std': 0.027259474620223045, 'epoch': 0.05}

 36%|███▌      | 1188/3300 [1:07:27<4:00:31,  6.83s/it]INFO 04-11 23:44:18 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1189/3300 [1:07:31<3:34:49,  6.11s/it]2026-04-11 23:44:22,737 | INFO | train_grpo_train | metrics_logged mode=train step=1189

                                                       
{'loss': 0.0367, 'grad_norm': 5.781359672546387, 'learning_rate': 6.4000000000000006e-06, 'num_tokens': 2683179.0, 'completions/mean_length': 58.125, 'completions/min_length': 56.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9804081916809082, 'rewards/meter/std': 0.011680345050990582, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9804081916809082, 'rewards/total_composite/std': 0.011680345050990582, 'reward': 0.9804081916809082, 'reward_std': 0.011680349707603455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05228231102228165, 'sampling/sampling_logp_difference/max': 2.627631187438965, 'sampling/importance_sampling_ratio/min': 0.0722494050860405, 'sampling/importance_sampling_ratio/mean': 0.9941091537475586, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15911316499114037, 'clip_ratio/low_mean': 0.012431694194674492, 'clip_ratio/low_min': 0.012431694194674492, 'clip_ratio/high_mean': 0.052866541780531406, 'clip_ratio/high_max': 0.052866541780531406, 'clip_ratio/region_mean': 0.0652982359752059, 'reward_total_mean': 0.9804081916809082, 'reward_meter_mean': 0.9804081916809082, 'reward_meter_std': 0.011680345050990582, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9804081916809082, 'reward_total_composite_std': 0.011680345050990582, 'epoch': 0.05}

 36%|███▌      | 1189/3300 [1:07:31<3:34:49,  6.11s/it]INFO 04-11 23:44:22 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1190/3300 [1:07:36<3:23:11,  5.78s/it]2026-04-11 23:44:27,747 | INFO | train_grpo_train | metrics_logged mode=train step=1190

                                                       
{'loss': -0.0135, 'grad_norm': 6.697084426879883, 'learning_rate': 6.396969696969697e-06, 'num_tokens': 2685520.0, 'completions/mean_length': 113.625, 'completions/min_length': 109.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.625, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9840307831764221, 'rewards/meter/std': 0.004937502555549145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8035714626312256, 'rewards/repeat_penalty/std': 0.10628911107778549, 'rewards/total_composite/mean': 0.7905905246734619, 'rewards/total_composite/std': 0.10349281877279282, 'reward': 0.7905905246734619, 'reward_std': 0.10349280387163162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023248611018061638, 'sampling/sampling_logp_difference/max': 1.2992486953735352, 'sampling/importance_sampling_ratio/min': 0.3153161406517029, 'sampling/importance_sampling_ratio/mean': 1.004918098449707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09165592398494482, 'clip_ratio/low_mean': 0.010279399110004306, 'clip_ratio/low_min': 0.010279399110004306, 'clip_ratio/high_mean': 0.008562042959965765, 'clip_ratio/high_max': 0.008562042959965765, 'clip_ratio/region_mean': 0.01884144206997007, 'reward_total_mean': 0.7905905246734619, 'reward_meter_mean': 0.9840307831764221, 'reward_meter_std': 0.004937502555549145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8035714626312256, 'reward_repeat_penalty_std': 0.10628911107778549, 'reward_total_composite_mean': 0.7905905246734619, 'reward_total_composite_std': 0.10349281877279282, 'epoch': 0.05}

 36%|███▌      | 1190/3300 [1:07:36<3:23:11,  5.78s/it]INFO 04-11 23:44:27 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1191/3300 [1:07:41<3:15:42,  5.57s/it]2026-04-11 23:44:32,824 | INFO | train_grpo_train | metrics_logged mode=train step=1191

                                                       
{'loss': -0.0018, 'grad_norm': 3.869032859802246, 'learning_rate': 6.393939393939394e-06, 'num_tokens': 2687846.0, 'completions/mean_length': 117.75, 'completions/min_length': 113.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.75, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9962882399559021, 'rewards/meter/std': 0.002757209585979581, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8966987133026123, 'rewards/total_composite/std': 0.10691322386264801, 'reward': 0.8966987133026123, 'reward_std': 0.10691321641206741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03010808303952217, 'sampling/sampling_logp_difference/max': 1.3976101875305176, 'sampling/importance_sampling_ratio/min': 0.24718700349330902, 'sampling/importance_sampling_ratio/mean': 1.0035390853881836, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16953753679990768, 'clip_ratio/low_mean': 0.013930454850196838, 'clip_ratio/low_min': 0.013930454850196838, 'clip_ratio/high_mean': 0.01574227074161172, 'clip_ratio/high_max': 0.01574227074161172, 'clip_ratio/region_mean': 0.029672725591808558, 'reward_total_mean': 0.8966987133026123, 'reward_meter_mean': 0.9962882399559021, 'reward_meter_std': 0.002757209585979581, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8966987133026123, 'reward_total_composite_std': 0.10691322386264801, 'epoch': 0.05}

 36%|███▌      | 1191/3300 [1:07:41<3:15:42,  5.57s/it]INFO 04-11 23:44:33 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1192/3300 [1:07:46<3:04:30,  5.25s/it]2026-04-11 23:44:37,340 | INFO | train_grpo_train | metrics_logged mode=train step=1192

                                                       
{'loss': -0.0028, 'grad_norm': 5.973384380340576, 'learning_rate': 6.390909090909091e-06, 'num_tokens': 2689734.0, 'completions/mean_length': 79.0, 'completions/min_length': 79.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.0, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.994601309299469, 'rewards/meter/std': 0.003018211107701063, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.994601309299469, 'rewards/total_composite/std': 0.003018211107701063, 'reward': 0.994601309299469, 'reward_std': 0.0030182143673300743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01802201196551323, 'sampling/sampling_logp_difference/max': 1.7237977981567383, 'sampling/importance_sampling_ratio/min': 0.3428289294242859, 'sampling/importance_sampling_ratio/mean': 1.0011334419250488, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.033365981886163354, 'clip_ratio/low_mean': 0.0015822785208001733, 'clip_ratio/low_min': 0.0015822785208001733, 'clip_ratio/high_mean': 0.007911392254754901, 'clip_ratio/high_max': 0.007911392254754901, 'clip_ratio/region_mean': 0.009493670775555074, 'reward_total_mean': 0.994601309299469, 'reward_meter_mean': 0.994601309299469, 'reward_meter_std': 0.003018211107701063, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.994601309299469, 'reward_total_composite_std': 0.003018211107701063, 'epoch': 0.05}

 36%|███▌      | 1192/3300 [1:07:46<3:04:30,  5.25s/it]INFO 04-11 23:44:37 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1193/3300 [1:07:50<2:56:36,  5.03s/it]2026-04-11 23:44:41,852 | INFO | train_grpo_train | metrics_logged mode=train step=1193

                                                       
{'loss': -0.0592, 'grad_norm': 7.629351615905762, 'learning_rate': 6.387878787878789e-06, 'num_tokens': 2691559.0, 'completions/mean_length': 61.125, 'completions/min_length': 49.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8711988925933838, 'rewards/meter/std': 0.3447757959365845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8711988925933838, 'rewards/total_composite/std': 0.3447757959365845, 'reward': 0.8711988925933838, 'reward_std': 0.3447757661342621, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03702111169695854, 'sampling/sampling_logp_difference/max': 0.9225618839263916, 'sampling/importance_sampling_ratio/min': 0.3974994122982025, 'sampling/importance_sampling_ratio/mean': 1.0055021047592163, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16935504972934723, 'clip_ratio/low_mean': 0.015306122601032257, 'clip_ratio/low_min': 0.015306122601032257, 'clip_ratio/high_mean': 0.027820809744298458, 'clip_ratio/high_max': 0.027820809744298458, 'clip_ratio/region_mean': 0.043126932345330715, 'reward_total_mean': 0.8711988925933838, 'reward_meter_mean': 0.8711988925933838, 'reward_meter_std': 0.3447757959365845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8711988925933838, 'reward_total_composite_std': 0.3447757959365845, 'epoch': 0.05}

 36%|███▌      | 1193/3300 [1:07:50<2:56:36,  5.03s/it]INFO 04-11 23:44:42 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1194/3300 [1:07:55<2:53:12,  4.93s/it]2026-04-11 23:44:46,565 | INFO | train_grpo_train | metrics_logged mode=train step=1194

                                                       
{'loss': 0.0146, 'grad_norm': 4.607259750366211, 'learning_rate': 6.384848484848485e-06, 'num_tokens': 2693468.0, 'completions/mean_length': 83.625, 'completions/min_length': 81.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.625, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.9963547587394714, 'rewards/meter/std': 0.0017178525449708104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9963547587394714, 'rewards/total_composite/std': 0.0017178525449708104, 'reward': 0.9963547587394714, 'reward_std': 0.0017178438138216734, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0324261412024498, 'sampling/sampling_logp_difference/max': 1.778012990951538, 'sampling/importance_sampling_ratio/min': 0.16897356510162354, 'sampling/importance_sampling_ratio/mean': 1.0033197402954102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12350557837635279, 'clip_ratio/low_mean': 0.0029069767333567142, 'clip_ratio/low_min': 0.0029069767333567142, 'clip_ratio/high_mean': 0.02205510064959526, 'clip_ratio/high_max': 0.02205510064959526, 'clip_ratio/region_mean': 0.024962077382951975, 'reward_total_mean': 0.9963547587394714, 'reward_meter_mean': 0.9963547587394714, 'reward_meter_std': 0.0017178525449708104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9963547587394714, 'reward_total_composite_std': 0.0017178525449708104, 'epoch': 0.05}

 36%|███▌      | 1194/3300 [1:07:55<2:53:12,  4.93s/it]INFO 04-11 23:44:46 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1195/3300 [1:07:59<2:46:36,  4.75s/it]2026-04-11 23:44:50,879 | INFO | train_grpo_train | metrics_logged mode=train step=1195

                                                       
{'loss': -0.0016, 'grad_norm': 6.79469108581543, 'learning_rate': 6.381818181818182e-06, 'num_tokens': 2695188.0, 'completions/mean_length': 56.0, 'completions/min_length': 55.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.08778245747089386, 'rewards/meter/std': 0.11504703015089035, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.17817415297031403, 'rewards/total_composite/mean': 0.07255840301513672, 'rewards/total_composite/std': 0.08352638781070709, 'reward': 0.07255840301513672, 'reward_std': 0.0835263803601265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022929562255740166, 'sampling/sampling_logp_difference/max': 0.9827042818069458, 'sampling/importance_sampling_ratio/min': 0.37429752945899963, 'sampling/importance_sampling_ratio/mean': 1.0030452013015747, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11227265931665897, 'clip_ratio/low_mean': 0.006657268386334181, 'clip_ratio/low_min': 0.006657268386334181, 'clip_ratio/high_mean': 0.011043233331292868, 'clip_ratio/high_max': 0.011043233331292868, 'clip_ratio/region_mean': 0.01770050171762705, 'reward_total_mean': 0.07255840301513672, 'reward_meter_mean': 0.08778245747089386, 'reward_meter_std': 0.11504703015089035, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.17817415297031403, 'reward_total_composite_mean': 0.07255840301513672, 'reward_total_composite_std': 0.08352638781070709, 'epoch': 0.05}

 36%|███▌      | 1195/3300 [1:07:59<2:46:36,  4.75s/it]INFO 04-11 23:44:51 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1196/3300 [1:08:04<2:47:06,  4.77s/it]2026-04-11 23:44:55,682 | INFO | train_grpo_train | metrics_logged mode=train step=1196

                                                       
{'loss': 0.0143, 'grad_norm': 5.267541408538818, 'learning_rate': 6.37878787878788e-06, 'num_tokens': 2697251.0, 'completions/mean_length': 92.875, 'completions/min_length': 91.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.875, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9628203511238098, 'rewards/meter/std': 0.07420913130044937, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8394116759300232, 'rewards/total_composite/std': 0.0950227826833725, 'reward': 0.8394116759300232, 'reward_std': 0.0950227826833725, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.046500932425260544, 'sampling/sampling_logp_difference/max': 1.5942132472991943, 'sampling/importance_sampling_ratio/min': 0.20306822657585144, 'sampling/importance_sampling_ratio/mean': 0.9947278499603271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14395452290773392, 'clip_ratio/low_mean': 0.023972635506652296, 'clip_ratio/low_min': 0.023972635506652296, 'clip_ratio/high_mean': 0.012287983670830727, 'clip_ratio/high_max': 0.012287983670830727, 'clip_ratio/region_mean': 0.03626061917748302, 'reward_total_mean': 0.8394116759300232, 'reward_meter_mean': 0.9628203511238098, 'reward_meter_std': 0.07420913130044937, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8394116759300232, 'reward_total_composite_std': 0.0950227826833725, 'epoch': 0.05}

 36%|███▌      | 1196/3300 [1:08:04<2:47:06,  4.77s/it]INFO 04-11 23:44:55 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1197/3300 [1:08:14<3:35:54,  6.16s/it]2026-04-11 23:45:05,099 | INFO | train_grpo_train | metrics_logged mode=train step=1197

                                                       
{'loss': 0.0431, 'grad_norm': 0.9943876266479492, 'learning_rate': 6.375757575757576e-06, 'num_tokens': 2702278.0, 'completions/mean_length': 419.375, 'completions/min_length': 402.0, 'completions/max_length': 474.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 419.375, 'completions/min_terminated_length': 402.0, 'completions/max_terminated_length': 474.0, 'rewards/meter/mean': 0.8696904182434082, 'rewards/meter/std': 0.3263057768344879, 'rewards/count_adherence/mean': 0.6333333253860474, 'rewards/count_adherence/std': 0.0942809134721756, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6529605388641357, 'rewards/repeat_penalty/std': 0.05881762132048607, 'rewards/total_composite/mean': 0.3697792887687683, 'rewards/total_composite/std': 0.14553934335708618, 'reward': 0.3697792887687683, 'reward_std': 0.14553934335708618, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009604532271623611, 'sampling/sampling_logp_difference/max': 1.1023553609848022, 'sampling/importance_sampling_ratio/min': 0.3320879638195038, 'sampling/importance_sampling_ratio/mean': 1.0019516944885254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04937975201755762, 'clip_ratio/low_mean': 0.0029802137287333608, 'clip_ratio/low_min': 0.0029802137287333608, 'clip_ratio/high_mean': 0.004789536877069622, 'clip_ratio/high_max': 0.004789536877069622, 'clip_ratio/region_mean': 0.007769750605802983, 'reward_total_mean': 0.3697792887687683, 'reward_meter_mean': 0.8696904182434082, 'reward_meter_std': 0.3263057768344879, 'reward_count_adherence_mean': 0.6333333253860474, 'reward_count_adherence_std': 0.0942809134721756, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6529605388641357, 'reward_repeat_penalty_std': 0.05881762132048607, 'reward_total_composite_mean': 0.3697792887687683, 'reward_total_composite_std': 0.14553934335708618, 'epoch': 0.05}

 36%|███▋      | 1197/3300 [1:08:14<3:35:54,  6.16s/it]INFO 04-11 23:45:05 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1198/3300 [1:08:18<3:16:41,  5.61s/it]2026-04-11 23:45:09,436 | INFO | train_grpo_train | metrics_logged mode=train step=1198

                                                       
{'loss': 0.0032, 'grad_norm': 6.456137657165527, 'learning_rate': 6.372727272727274e-06, 'num_tokens': 2703954.0, 'completions/mean_length': 61.5, 'completions/min_length': 59.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.3377665877342224, 'rewards/meter/std': 0.33047032356262207, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.3377665877342224, 'rewards/total_composite/std': 0.33047032356262207, 'reward': 0.3377665877342224, 'reward_std': 0.33047032356262207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0402526780962944, 'sampling/sampling_logp_difference/max': 1.6591825485229492, 'sampling/importance_sampling_ratio/min': 0.19029447436332703, 'sampling/importance_sampling_ratio/mean': 1.0035641193389893, 'sampling/importance_sampling_ratio/max': 1.5698946714401245, 'entropy': 0.31682545877993107, 'clip_ratio/low_mean': 0.014116051141172647, 'clip_ratio/low_min': 0.014116051141172647, 'clip_ratio/high_mean': 0.006082487525418401, 'clip_ratio/high_max': 0.006082487525418401, 'clip_ratio/region_mean': 0.020198538666591048, 'reward_total_mean': 0.3377665877342224, 'reward_meter_mean': 0.3377665877342224, 'reward_meter_std': 0.33047032356262207, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.3377665877342224, 'reward_total_composite_std': 0.33047032356262207, 'epoch': 0.05}

 36%|███▋      | 1198/3300 [1:08:18<3:16:41,  5.61s/it]INFO 04-11 23:45:09 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1199/3300 [1:08:23<3:06:21,  5.32s/it]2026-04-11 23:45:14,079 | INFO | train_grpo_train | metrics_logged mode=train step=1199

                                                       
{'loss': 0.0199, 'grad_norm': 4.539976596832275, 'learning_rate': 6.3696969696969706e-06, 'num_tokens': 2705923.0, 'completions/mean_length': 81.125, 'completions/min_length': 79.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.125, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.5489683747291565, 'rewards/meter/std': 0.4779263734817505, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5489683747291565, 'rewards/total_composite/std': 0.4779263734817505, 'reward': 0.5489683747291565, 'reward_std': 0.4779263436794281, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010573667474091053, 'sampling/sampling_logp_difference/max': 0.9852040410041809, 'sampling/importance_sampling_ratio/min': 0.37336301803588867, 'sampling/importance_sampling_ratio/mean': 1.0014384984970093, 'sampling/importance_sampling_ratio/max': 1.4673889875411987, 'entropy': 0.04470323724672198, 'clip_ratio/low_mean': 0.004537329194135964, 'clip_ratio/low_min': 0.004537329194135964, 'clip_ratio/high_mean': 0.0031447785440832376, 'clip_ratio/high_max': 0.0031447785440832376, 'clip_ratio/region_mean': 0.0076821077382192016, 'reward_total_mean': 0.5489683747291565, 'reward_meter_mean': 0.5489683747291565, 'reward_meter_std': 0.4779263734817505, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5489683747291565, 'reward_total_composite_std': 0.4779263734817505, 'epoch': 0.05}

 36%|███▋      | 1199/3300 [1:08:23<3:06:21,  5.32s/it]INFO 04-11 23:45:14 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1200/3300 [1:08:27<2:56:15,  5.04s/it]2026-04-11 23:45:18,446 | INFO | train_grpo_train | metrics_logged mode=train step=1200

                                                       
{'loss': 0.018, 'grad_norm': 13.826656341552734, 'learning_rate': 6.366666666666668e-06, 'num_tokens': 2707810.0, 'completions/mean_length': 59.875, 'completions/min_length': 56.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.875, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8469290137290955, 'rewards/meter/std': 0.20266500115394592, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8469290137290955, 'rewards/total_composite/std': 0.20266500115394592, 'reward': 0.8469290137290955, 'reward_std': 0.20266498625278473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036583464592695236, 'sampling/sampling_logp_difference/max': 1.854140043258667, 'sampling/importance_sampling_ratio/min': 0.15658754110336304, 'sampling/importance_sampling_ratio/mean': 1.0088112354278564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09853810910135508, 'clip_ratio/low_mean': 0.0147115015424788, 'clip_ratio/low_min': 0.0147115015424788, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.01676068175584078, 'reward_total_mean': 0.8469290137290955, 'reward_meter_mean': 0.8469290137290955, 'reward_meter_std': 0.20266500115394592, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8469290137290955, 'reward_total_composite_std': 0.20266500115394592, 'epoch': 0.05}

 36%|███▋      | 1200/3300 [1:08:27<2:56:15,  5.04s/it]INFO 04-11 23:45:18 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:25,  2.30s/it]

 23%|██▎       | 3/13 [00:08<00:28,  2.88s/it]

 31%|███       | 4/13 [00:14<00:35,  3.95s/it]

 38%|███▊      | 5/13 [00:19<00:35,  4.41s/it]

 46%|████▌     | 6/13 [00:25<00:34,  4.99s/it]

 54%|█████▍    | 7/13 [00:30<00:29,  4.96s/it]

 62%|██████▏   | 8/13 [00:36<00:26,  5.37s/it]

 69%|██████▉   | 9/13 [00:43<00:23,  5.84s/it]

 77%|███████▋  | 10/13 [00:48<00:16,  5.66s/it]

 85%|████████▍ | 11/13 [00:53<00:10,  5.39s/it]

 92%|█████████▏| 12/13 [00:59<00:05,  5.56s/it]

100%|██████████| 13/13 [01:04<00:00,  5.37s/it]2026-04-11 23:46:28,030 | INFO | train_grpo_train | metrics_logged mode=eval step=1200

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 69.5055, 'eval_samples_per_second': 1.496, 'eval_steps_per_second': 0.187, 'eval_num_tokens': 2707810.0, 'eval_completions/mean_length': 210.4903846153846, 'eval_completions/min_length': 63.69230769230769, 'eval_completions/max_length': 365.9230769230769, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 210.4903846153846, 'eval_completions/min_terminated_length': 63.69230769230769, 'eval_completions/max_terminated_length': 365.9230769230769, 'eval_rewards/meter/mean': 0.675024688243866, 'eval_rewards/meter/std': 0.4085804086465102, 'eval_rewards/count_adherence/mean': 0.8891764970926138, 'eval_rewards/count_adherence/std': 0.11591204485067955, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.7145149661944463, 'eval_rewards/repeat_penalty/std': 0.17899919931705183, 'eval_rewards/total_composite/mean': 0.4327001617505, 'eval_rewards/total_composite/std': 0.3292730886202592, 'eval_reward': 0.4327001617505, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.009530584721897658, 'eval_sampling/sampling_logp_difference/max': 0.8752486155583308, 'eval_sampling/importance_sampling_ratio/min': 0.4270020562868852, 'eval_sampling/importance_sampling_ratio/mean': 1.0017014145851135, 'eval_sampling/importance_sampling_ratio/max': 1.3788938980836134, 'eval_entropy': 0.07712389929936482, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.4327001617505, 'eval_reward_meter_mean': 0.675024688243866, 'eval_reward_meter_std': 0.4085804086465102, 'eval_reward_count_adherence_mean': 0.8891764970926138, 'eval_reward_count_adherence_std': 0.11591204485067955, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.7145149661944463, 'eval_reward_repeat_penalty_std': 0.17899919931705183, 'eval_reward_total_composite_mean': 0.4327001617505, 'eval_reward_total_composite_std': 0.3292730886202592, 'epoch': 0.05}

 36%|███▋      | 1200/3300 [1:09:37<2:56:15,  5.04s/it]

100%|██████████| 13/13 [01:04<00:00,  5.37s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 36%|███▋      | 1201/3300 [1:09:44<15:32:54, 26.67s/it]2026-04-11 23:46:35,713 | INFO | train_grpo_train | metrics_logged mode=train step=1201

                                                        
{'loss': 0.0014, 'grad_norm': 1.0187959671020508, 'learning_rate': 6.363636363636364e-06, 'num_tokens': 2709651.0, 'completions/mean_length': 79.125, 'completions/min_length': 79.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9956833124160767, 'rewards/meter/std': 4.2210078390780836e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9956833124160767, 'rewards/total_composite/std': 4.2210078390780836e-05, 'reward': 0.9956833124160767, 'reward_std': 4.2219100578222424e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0033457186073064804, 'sampling/sampling_logp_difference/max': 0.5514485836029053, 'sampling/importance_sampling_ratio/min': 0.5761146545410156, 'sampling/importance_sampling_ratio/mean': 1.0009052753448486, 'sampling/importance_sampling_ratio/max': 1.1379395723342896, 'entropy': 0.022257187170907855, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0015822785208001733, 'clip_ratio/high_max': 0.0015822785208001733, 'clip_ratio/region_mean': 0.0015822785208001733, 'reward_total_mean': 0.9956833124160767, 'reward_meter_mean': 0.9956833124160767, 'reward_meter_std': 4.2210078390780836e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9956833124160767, 'reward_total_composite_std': 4.2210078390780836e-05, 'epoch': 0.05}

 36%|███▋      | 1201/3300 [1:09:44<15:32:54, 26.67s/it]INFO 04-11 23:46:35 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1202/3300 [1:09:49<11:43:35, 20.12s/it]2026-04-11 23:46:40,467 | INFO | train_grpo_train | metrics_logged mode=train step=1202

                                                        
{'loss': -0.0406, 'grad_norm': 8.3601655960083, 'learning_rate': 6.3606060606060615e-06, 'num_tokens': 2711138.0, 'completions/mean_length': 29.875, 'completions/min_length': 27.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.875, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.644133985042572, 'rewards/meter/std': 0.4584461748600006, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.644133985042572, 'rewards/total_composite/std': 0.4584461748600006, 'reward': 0.644133985042572, 'reward_std': 0.4584461748600006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034943368285894394, 'sampling/sampling_logp_difference/max': 1.0919749736785889, 'sampling/importance_sampling_ratio/min': 0.3355531394481659, 'sampling/importance_sampling_ratio/mean': 1.0082812309265137, 'sampling/importance_sampling_ratio/max': 1.6411713361740112, 'entropy': 0.21921741589903831, 'clip_ratio/low_mean': 0.03160349419340491, 'clip_ratio/low_min': 0.03160349419340491, 'clip_ratio/high_mean': 0.012096773833036423, 'clip_ratio/high_max': 0.012096773833036423, 'clip_ratio/region_mean': 0.043700268026441336, 'reward_total_mean': 0.644133985042572, 'reward_meter_mean': 0.644133985042572, 'reward_meter_std': 0.4584461748600006, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.644133985042572, 'reward_total_composite_std': 0.4584461748600006, 'epoch': 0.05}

 36%|███▋      | 1202/3300 [1:09:49<11:43:35, 20.12s/it]INFO 04-11 23:46:40 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1203/3300 [1:09:54<9:04:52, 15.59s/it] 2026-04-11 23:46:45,480 | INFO | train_grpo_train | metrics_logged mode=train step=1203

                                                       
{'loss': 0.0002, 'grad_norm': 4.518101215362549, 'learning_rate': 6.357575757575758e-06, 'num_tokens': 2712910.0, 'completions/mean_length': 60.5, 'completions/min_length': 59.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.203798308968544, 'rewards/meter/std': 0.2958202362060547, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.203798308968544, 'rewards/total_composite/std': 0.2958202362060547, 'reward': 0.203798308968544, 'reward_std': 0.2958202362060547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05345018953084946, 'sampling/sampling_logp_difference/max': 1.703664779663086, 'sampling/importance_sampling_ratio/min': 0.18201525509357452, 'sampling/importance_sampling_ratio/mean': 1.004960060119629, 'sampling/importance_sampling_ratio/max': 1.886362910270691, 'entropy': 0.37029214575886726, 'clip_ratio/low_mean': 0.04342324007302523, 'clip_ratio/low_min': 0.04342324007302523, 'clip_ratio/high_mean': 0.008405112195760012, 'clip_ratio/high_max': 0.008405112195760012, 'clip_ratio/region_mean': 0.05182835226878524, 'reward_total_mean': 0.203798308968544, 'reward_meter_mean': 0.203798308968544, 'reward_meter_std': 0.2958202362060547, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.203798308968544, 'reward_total_composite_std': 0.2958202362060547, 'epoch': 0.05}

 36%|███▋      | 1203/3300 [1:09:54<9:04:52, 15.59s/it]INFO 04-11 23:46:45 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1204/3300 [1:09:59<7:16:17, 12.49s/it]2026-04-11 23:46:50,778 | INFO | train_grpo_train | metrics_logged mode=train step=1204

                                                       
{'loss': 0.004, 'grad_norm': 3.396805763244629, 'learning_rate': 6.354545454545455e-06, 'num_tokens': 2715103.0, 'completions/mean_length': 94.125, 'completions/min_length': 91.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.125, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9817042350769043, 'rewards/meter/std': 0.002063595922663808, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8099536895751953, 'rewards/total_composite/std': 0.070185586810112, 'reward': 0.8099536895751953, 'reward_std': 0.070185586810112, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022159146144986153, 'sampling/sampling_logp_difference/max': 1.2238504886627197, 'sampling/importance_sampling_ratio/min': 0.29409554600715637, 'sampling/importance_sampling_ratio/mean': 1.0024614334106445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06619423814117908, 'clip_ratio/low_mean': 0.01712075702380389, 'clip_ratio/low_min': 0.01712075702380389, 'clip_ratio/high_mean': 0.003989361692219973, 'clip_ratio/high_max': 0.003989361692219973, 'clip_ratio/region_mean': 0.021110118716023862, 'reward_total_mean': 0.8099536895751953, 'reward_meter_mean': 0.9817042350769043, 'reward_meter_std': 0.002063595922663808, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8099536895751953, 'reward_total_composite_std': 0.070185586810112, 'epoch': 0.05}

 36%|███▋      | 1204/3300 [1:09:59<7:16:17, 12.49s/it]INFO 04-11 23:46:51 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1205/3300 [1:10:04<5:58:02, 10.25s/it]2026-04-11 23:46:55,747 | INFO | train_grpo_train | metrics_logged mode=train step=1205

                                                       
{'loss': 0.0175, 'grad_norm': 4.248809814453125, 'learning_rate': 6.3515151515151516e-06, 'num_tokens': 2717019.0, 'completions/mean_length': 78.5, 'completions/min_length': 75.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9937853813171387, 'rewards/meter/std': 0.0054353163577616215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9937853813171387, 'rewards/total_composite/std': 0.0054353163577616215, 'reward': 0.9937853813171387, 'reward_std': 0.0054353224113583565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03144395351409912, 'sampling/sampling_logp_difference/max': 1.310187816619873, 'sampling/importance_sampling_ratio/min': 0.2697693705558777, 'sampling/importance_sampling_ratio/mean': 1.0043216943740845, 'sampling/importance_sampling_ratio/max': 1.6284278631210327, 'entropy': 0.18105985596776009, 'clip_ratio/low_mean': 0.004518072120845318, 'clip_ratio/low_min': 0.004518072120845318, 'clip_ratio/high_mean': 0.014675587648525834, 'clip_ratio/high_max': 0.014675587648525834, 'clip_ratio/region_mean': 0.019193659769371152, 'reward_total_mean': 0.9937853813171387, 'reward_meter_mean': 0.9937853813171387, 'reward_meter_std': 0.0054353163577616215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9937853813171387, 'reward_total_composite_std': 0.0054353163577616215, 'epoch': 0.05}

 37%|███▋      | 1205/3300 [1:10:04<5:58:02, 10.25s/it]INFO 04-11 23:46:55 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1206/3300 [1:10:09<5:02:37,  8.67s/it]2026-04-11 23:47:00,725 | INFO | train_grpo_train | metrics_logged mode=train step=1206

                                                       
{'loss': -0.0049, 'grad_norm': 2.2625913619995117, 'learning_rate': 6.34848484848485e-06, 'num_tokens': 2718833.0, 'completions/mean_length': 77.75, 'completions/min_length': 76.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.75, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9951738119125366, 'rewards/meter/std': 0.00032568458118475974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9951738119125366, 'rewards/total_composite/std': 0.00032568458118475974, 'reward': 0.9951738119125366, 'reward_std': 0.00032568458118475974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004397572483867407, 'sampling/sampling_logp_difference/max': 0.4512134790420532, 'sampling/importance_sampling_ratio/min': 0.636854887008667, 'sampling/importance_sampling_ratio/mean': 1.0006308555603027, 'sampling/importance_sampling_ratio/max': 1.3814133405685425, 'entropy': 0.022122491151094437, 'clip_ratio/low_mean': 0.001623376621864736, 'clip_ratio/low_min': 0.001623376621864736, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.001623376621864736, 'reward_total_mean': 0.9951738119125366, 'reward_meter_mean': 0.9951738119125366, 'reward_meter_std': 0.00032568458118475974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9951738119125366, 'reward_total_composite_std': 0.00032568458118475974, 'epoch': 0.05}

 37%|███▋      | 1206/3300 [1:10:09<5:02:37,  8.67s/it]INFO 04-11 23:47:00 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1207/3300 [1:10:15<4:28:31,  7.70s/it]2026-04-11 23:47:06,149 | INFO | train_grpo_train | metrics_logged mode=train step=1207

                                                       
{'loss': -0.005, 'grad_norm': 7.367405891418457, 'learning_rate': 6.345454545454546e-06, 'num_tokens': 2721333.0, 'completions/mean_length': 124.5, 'completions/min_length': 114.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.5, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.8035296201705933, 'rewards/meter/std': 0.35723742842674255, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.5791441202163696, 'rewards/total_composite/std': 0.24694091081619263, 'reward': 0.5791441202163696, 'reward_std': 0.24694089591503143, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022883908823132515, 'sampling/sampling_logp_difference/max': 1.5613296031951904, 'sampling/importance_sampling_ratio/min': 0.20985685288906097, 'sampling/importance_sampling_ratio/mean': 1.0008379220962524, 'sampling/importance_sampling_ratio/max': 1.7591142654418945, 'entropy': 0.1076383925974369, 'clip_ratio/low_mean': 0.005122669972479343, 'clip_ratio/low_min': 0.005122669972479343, 'clip_ratio/high_mean': 0.013700352283194661, 'clip_ratio/high_max': 0.013700352283194661, 'clip_ratio/region_mean': 0.018823022255674005, 'reward_total_mean': 0.5791441202163696, 'reward_meter_mean': 0.8035296201705933, 'reward_meter_std': 0.35723742842674255, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.5791441202163696, 'reward_total_composite_std': 0.24694091081619263, 'epoch': 0.05}

 37%|███▋      | 1207/3300 [1:10:15<4:28:31,  7.70s/it]INFO 04-11 23:47:06 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1208/3300 [1:10:19<3:57:30,  6.81s/it]2026-04-11 23:47:10,896 | INFO | train_grpo_train | metrics_logged mode=train step=1208

                                                       
{'loss': -0.0147, 'grad_norm': 5.575094699859619, 'learning_rate': 6.342424242424243e-06, 'num_tokens': 2723045.0, 'completions/mean_length': 63.0, 'completions/min_length': 55.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.3075833320617676, 'rewards/meter/std': 0.3925836980342865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.3075833320617676, 'rewards/total_composite/std': 0.3925836980342865, 'reward': 0.3075833320617676, 'reward_std': 0.3925836980342865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07028155028820038, 'sampling/sampling_logp_difference/max': 1.5979199409484863, 'sampling/importance_sampling_ratio/min': 0.20231691002845764, 'sampling/importance_sampling_ratio/mean': 0.9958197474479675, 'sampling/importance_sampling_ratio/max': 1.9152469635009766, 'entropy': 0.3939699362963438, 'clip_ratio/low_mean': 0.03671912616118789, 'clip_ratio/low_min': 0.03671912616118789, 'clip_ratio/high_mean': 0.016762672923505306, 'clip_ratio/high_max': 0.016762672923505306, 'clip_ratio/region_mean': 0.05348179908469319, 'reward_total_mean': 0.3075833320617676, 'reward_meter_mean': 0.3075833320617676, 'reward_meter_std': 0.3925836980342865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.3075833320617676, 'reward_total_composite_std': 0.3925836980342865, 'epoch': 0.05}

 37%|███▋      | 1208/3300 [1:10:19<3:57:30,  6.81s/it]INFO 04-11 23:47:11 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1209/3300 [1:10:27<4:01:36,  6.93s/it]2026-04-11 23:47:18,111 | INFO | train_grpo_train | metrics_logged mode=train step=1209

                                                       
{'loss': -0.0039, 'grad_norm': 2.477656364440918, 'learning_rate': 6.33939393939394e-06, 'num_tokens': 2726311.0, 'completions/mean_length': 230.25, 'completions/min_length': 220.0, 'completions/max_length': 246.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 230.25, 'completions/min_terminated_length': 220.0, 'completions/max_terminated_length': 246.0, 'rewards/meter/mean': 0.9955352544784546, 'rewards/meter/std': 0.0017380027566105127, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6704545617103577, 'rewards/repeat_penalty/std': 0.13690368831157684, 'rewards/total_composite/mean': 0.5340035557746887, 'rewards/total_composite/std': 0.10914173722267151, 'reward': 0.5340035557746887, 'reward_std': 0.10914173722267151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020786819979548454, 'sampling/sampling_logp_difference/max': 1.8532099723815918, 'sampling/importance_sampling_ratio/min': 0.15673324465751648, 'sampling/importance_sampling_ratio/mean': 1.0021883249282837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10937195271253586, 'clip_ratio/low_mean': 0.004963513347320259, 'clip_ratio/low_min': 0.004963513347320259, 'clip_ratio/high_mean': 0.010197082534432411, 'clip_ratio/high_max': 0.010197082534432411, 'clip_ratio/region_mean': 0.01516059588175267, 'reward_total_mean': 0.5340035557746887, 'reward_meter_mean': 0.9955352544784546, 'reward_meter_std': 0.0017380027566105127, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6704545617103577, 'reward_repeat_penalty_std': 0.13690368831157684, 'reward_total_composite_mean': 0.5340035557746887, 'reward_total_composite_std': 0.10914173722267151, 'epoch': 0.05}

 37%|███▋      | 1209/3300 [1:10:27<4:01:36,  6.93s/it]INFO 04-11 23:47:18 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1210/3300 [1:10:33<4:00:53,  6.92s/it]2026-04-11 23:47:24,988 | INFO | train_grpo_train | metrics_logged mode=train step=1210

                                                       
{'loss': 0.0075, 'grad_norm': 0.6899715662002563, 'learning_rate': 6.336363636363637e-06, 'num_tokens': 2729552.0, 'completions/mean_length': 232.125, 'completions/min_length': 228.0, 'completions/max_length': 235.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 232.125, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.9961996078491211, 'rewards/meter/std': 0.0006288865115493536, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4545454680919647, 'rewards/repeat_penalty/std': 0.11902794241905212, 'rewards/total_composite/mean': 0.36228621006011963, 'rewards/total_composite/std': 0.09505681693553925, 'reward': 0.36228621006011963, 'reward_std': 0.09505681693553925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005651684943586588, 'sampling/sampling_logp_difference/max': 0.983978271484375, 'sampling/importance_sampling_ratio/min': 0.3738209903240204, 'sampling/importance_sampling_ratio/mean': 1.0020835399627686, 'sampling/importance_sampling_ratio/max': 1.5110461711883545, 'entropy': 0.03385857096873224, 'clip_ratio/low_mean': 0.0016003104392439127, 'clip_ratio/low_min': 0.0016003104392439127, 'clip_ratio/high_mean': 0.001087038719560951, 'clip_ratio/high_max': 0.001087038719560951, 'clip_ratio/region_mean': 0.0026873491588048637, 'reward_total_mean': 0.36228621006011963, 'reward_meter_mean': 0.9961996078491211, 'reward_meter_std': 0.0006288865115493536, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4545454680919647, 'reward_repeat_penalty_std': 0.11902794241905212, 'reward_total_composite_mean': 0.36228621006011963, 'reward_total_composite_std': 0.09505681693553925, 'epoch': 0.05}

 37%|███▋      | 1210/3300 [1:10:33<4:00:53,  6.92s/it]INFO 04-11 23:47:25 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1211/3300 [1:10:41<4:10:15,  7.19s/it]2026-04-11 23:47:32,808 | INFO | train_grpo_train | metrics_logged mode=train step=1211

                                                       
{'loss': 0.0165, 'grad_norm': 2.939035177230835, 'learning_rate': 6.333333333333333e-06, 'num_tokens': 2733597.0, 'completions/mean_length': 305.625, 'completions/min_length': 289.0, 'completions/max_length': 320.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 305.625, 'completions/min_terminated_length': 289.0, 'completions/max_terminated_length': 320.0, 'rewards/meter/mean': 0.9956304430961609, 'rewards/meter/std': 0.0010864537907764316, 'rewards/count_adherence/mean': 0.9444444179534912, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7098382711410522, 'rewards/repeat_penalty/std': 0.09350526332855225, 'rewards/total_composite/mean': 0.6673710942268372, 'rewards/total_composite/std': 0.0976538211107254, 'reward': 0.6673710942268372, 'reward_std': 0.0976538211107254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01374336052685976, 'sampling/sampling_logp_difference/max': 2.966305732727051, 'sampling/importance_sampling_ratio/min': 0.06939523667097092, 'sampling/importance_sampling_ratio/mean': 1.001654028892517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04799028439447284, 'clip_ratio/low_mean': 0.00556039166986011, 'clip_ratio/low_min': 0.00556039166986011, 'clip_ratio/high_mean': 0.005265707382932305, 'clip_ratio/high_max': 0.005265707382932305, 'clip_ratio/region_mean': 0.010826099052792415, 'reward_total_mean': 0.6673710942268372, 'reward_meter_mean': 0.9956304430961609, 'reward_meter_std': 0.0010864537907764316, 'reward_count_adherence_mean': 0.9444444179534912, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7098382711410522, 'reward_repeat_penalty_std': 0.09350526332855225, 'reward_total_composite_mean': 0.6673710942268372, 'reward_total_composite_std': 0.0976538211107254, 'epoch': 0.05}

 37%|███▋      | 1211/3300 [1:10:41<4:10:15,  7.19s/it]INFO 04-11 23:47:33 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1212/3300 [1:10:47<3:51:22,  6.65s/it]2026-04-11 23:47:38,201 | INFO | train_grpo_train | metrics_logged mode=train step=1212

                                                       
{'loss': 0.0002, 'grad_norm': 1.3120684623718262, 'learning_rate': 6.330303030303031e-06, 'num_tokens': 2735462.0, 'completions/mean_length': 62.125, 'completions/min_length': 61.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.125, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.993018627166748, 'rewards/meter/std': 0.0011156242107972503, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.993018627166748, 'rewards/total_composite/std': 0.0011156242107972503, 'reward': 0.993018627166748, 'reward_std': 0.0011156280525028706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016475966200232506, 'sampling/sampling_logp_difference/max': 1.2096130847930908, 'sampling/importance_sampling_ratio/min': 0.4961382746696472, 'sampling/importance_sampling_ratio/mean': 1.0012403726577759, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05741153517737985, 'clip_ratio/low_mean': 0.010017690248787403, 'clip_ratio/low_min': 0.010017690248787403, 'clip_ratio/high_mean': 0.004032257944345474, 'clip_ratio/high_max': 0.004032257944345474, 'clip_ratio/region_mean': 0.014049948193132877, 'reward_total_mean': 0.993018627166748, 'reward_meter_mean': 0.993018627166748, 'reward_meter_std': 0.0011156242107972503, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.993018627166748, 'reward_total_composite_std': 0.0011156242107972503, 'epoch': 0.05}

 37%|███▋      | 1212/3300 [1:10:47<3:51:22,  6.65s/it]INFO 04-11 23:47:38 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1213/3300 [1:10:52<3:38:47,  6.29s/it]2026-04-11 23:47:43,654 | INFO | train_grpo_train | metrics_logged mode=train step=1213

                                                       
{'loss': -0.0134, 'grad_norm': 5.671288967132568, 'learning_rate': 6.327272727272727e-06, 'num_tokens': 2737205.0, 'completions/mean_length': 61.875, 'completions/min_length': 59.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.875, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9939371347427368, 'rewards/meter/std': 0.0013982808450236917, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9939371347427368, 'rewards/total_composite/std': 0.0013982808450236917, 'reward': 0.9939371347427368, 'reward_std': 0.0013982723467051983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02799602597951889, 'sampling/sampling_logp_difference/max': 1.1926932334899902, 'sampling/importance_sampling_ratio/min': 0.3034030497074127, 'sampling/importance_sampling_ratio/mean': 0.9959902763366699, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08121698070317507, 'clip_ratio/low_mean': 0.012202381622046232, 'clip_ratio/low_min': 0.012202381622046232, 'clip_ratio/high_mean': 0.007692307815887034, 'clip_ratio/high_max': 0.007692307815887034, 'clip_ratio/region_mean': 0.019894689437933266, 'reward_total_mean': 0.9939371347427368, 'reward_meter_mean': 0.9939371347427368, 'reward_meter_std': 0.0013982808450236917, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9939371347427368, 'reward_total_composite_std': 0.0013982808450236917, 'epoch': 0.05}

 37%|███▋      | 1213/3300 [1:10:52<3:38:47,  6.29s/it]INFO 04-11 23:47:43 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1214/3300 [1:11:01<4:09:41,  7.18s/it]2026-04-11 23:47:52,918 | INFO | train_grpo_train | metrics_logged mode=train step=1214

                                                       
{'loss': -0.0101, 'grad_norm': 1.0214275121688843, 'learning_rate': 6.324242424242425e-06, 'num_tokens': 2741692.0, 'completions/mean_length': 323.875, 'completions/min_length': 316.0, 'completions/max_length': 362.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 323.875, 'completions/min_terminated_length': 316.0, 'completions/max_terminated_length': 362.0, 'rewards/meter/mean': 0.9967888593673706, 'rewards/meter/std': 0.0008384960819967091, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.035355325788259506, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6199448704719543, 'rewards/repeat_penalty/std': 0.08540944010019302, 'rewards/total_composite/mean': 0.5024136304855347, 'rewards/total_composite/std': 0.07446450740098953, 'reward': 0.5024136304855347, 'reward_std': 0.07446449995040894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007462342269718647, 'sampling/sampling_logp_difference/max': 2.224062919616699, 'sampling/importance_sampling_ratio/min': 0.10816873610019684, 'sampling/importance_sampling_ratio/mean': 1.0000638961791992, 'sampling/importance_sampling_ratio/max': 1.6564884185791016, 'entropy': 0.0330718404147774, 'clip_ratio/low_mean': 0.0023547938617412, 'clip_ratio/low_min': 0.0023547938617412, 'clip_ratio/high_mean': 0.002272886282298714, 'clip_ratio/high_max': 0.002272886282298714, 'clip_ratio/region_mean': 0.004627680144039914, 'reward_total_mean': 0.5024136304855347, 'reward_meter_mean': 0.9967888593673706, 'reward_meter_std': 0.0008384960819967091, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.035355325788259506, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6199448704719543, 'reward_repeat_penalty_std': 0.08540944010019302, 'reward_total_composite_mean': 0.5024136304855347, 'reward_total_composite_std': 0.07446450740098953, 'epoch': 0.05}

 37%|███▋      | 1214/3300 [1:11:01<4:09:41,  7.18s/it]INFO 04-11 23:47:53 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1215/3300 [1:11:09<4:13:30,  7.30s/it]2026-04-11 23:48:00,478 | INFO | train_grpo_train | metrics_logged mode=train step=1215

                                                       
{'loss': 0.0064, 'grad_norm': 2.1483280658721924, 'learning_rate': 6.3212121212121216e-06, 'num_tokens': 2745177.0, 'completions/mean_length': 241.625, 'completions/min_length': 235.0, 'completions/max_length': 246.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 241.625, 'completions/min_terminated_length': 235.0, 'completions/max_terminated_length': 246.0, 'rewards/meter/mean': 0.9970427751541138, 'rewards/meter/std': 0.0011380411451682448, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.0739355981349945, 'rewards/total_composite/mean': 0.7299137115478516, 'rewards/total_composite/std': 0.07295885682106018, 'reward': 0.7299137115478516, 'reward_std': 0.07295886427164078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016915393993258476, 'sampling/sampling_logp_difference/max': 1.4178450107574463, 'sampling/importance_sampling_ratio/min': 0.2422354817390442, 'sampling/importance_sampling_ratio/mean': 1.000618815422058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.061672994401305914, 'clip_ratio/low_mean': 0.0077110049314796925, 'clip_ratio/low_min': 0.0077110049314796925, 'clip_ratio/high_mean': 0.006238285917788744, 'clip_ratio/high_max': 0.006238285917788744, 'clip_ratio/region_mean': 0.013949290849268436, 'reward_total_mean': 0.7299137115478516, 'reward_meter_mean': 0.9970427751541138, 'reward_meter_std': 0.0011380411451682448, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.0739355981349945, 'reward_total_composite_mean': 0.7299137115478516, 'reward_total_composite_std': 0.07295885682106018, 'epoch': 0.05}

 37%|███▋      | 1215/3300 [1:11:09<4:13:30,  7.30s/it]INFO 04-11 23:48:00 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1216/3300 [1:11:15<4:02:54,  6.99s/it]2026-04-11 23:48:06,768 | INFO | train_grpo_train | metrics_logged mode=train step=1216

                                                       
{'loss': -0.0023, 'grad_norm': 1.560193657875061, 'learning_rate': 6.318181818181819e-06, 'num_tokens': 2747538.0, 'completions/mean_length': 124.125, 'completions/min_length': 122.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.125, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.99362713098526, 'rewards/meter/std': 0.0015980940079316497, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7678571343421936, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.7628991007804871, 'rewards/total_composite/std': 0.07266794145107269, 'reward': 0.7628991007804871, 'reward_std': 0.07266796380281448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010110512375831604, 'sampling/sampling_logp_difference/max': 0.9531111717224121, 'sampling/importance_sampling_ratio/min': 0.3855396807193756, 'sampling/importance_sampling_ratio/mean': 0.9995880722999573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03512992733158171, 'clip_ratio/low_mean': 0.003024590201675892, 'clip_ratio/low_min': 0.003024590201675892, 'clip_ratio/high_mean': 0.006018434185534716, 'clip_ratio/high_max': 0.006018434185534716, 'clip_ratio/region_mean': 0.009043024387210608, 'reward_total_mean': 0.7628991007804871, 'reward_meter_mean': 0.99362713098526, 'reward_meter_std': 0.0015980940079316497, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7678571343421936, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.7628991007804871, 'reward_total_composite_std': 0.07266794145107269, 'epoch': 0.05}

 37%|███▋      | 1216/3300 [1:11:15<4:02:54,  6.99s/it]INFO 04-11 23:48:07 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1217/3300 [1:11:21<3:52:07,  6.69s/it]2026-04-11 23:48:12,760 | INFO | train_grpo_train | metrics_logged mode=train step=1217

                                                       
{'loss': 0.0084, 'grad_norm': 4.285740852355957, 'learning_rate': 6.315151515151515e-06, 'num_tokens': 2749384.0, 'completions/mean_length': 75.75, 'completions/min_length': 71.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.41745519638061523, 'rewards/meter/std': 0.3040141463279724, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.3990457057952881, 'rewards/total_composite/std': 0.3067740499973297, 'reward': 0.3990457057952881, 'reward_std': 0.3067740499973297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03294277563691139, 'sampling/sampling_logp_difference/max': 1.4083929061889648, 'sampling/importance_sampling_ratio/min': 0.24453596770763397, 'sampling/importance_sampling_ratio/mean': 1.002768635749817, 'sampling/importance_sampling_ratio/max': 1.500490427017212, 'entropy': 0.20184578280895948, 'clip_ratio/low_mean': 0.009840250364504755, 'clip_ratio/low_min': 0.009840250364504755, 'clip_ratio/high_mean': 0.00835704104974866, 'clip_ratio/high_max': 0.00835704104974866, 'clip_ratio/region_mean': 0.018197291414253414, 'reward_total_mean': 0.3990457057952881, 'reward_meter_mean': 0.41745519638061523, 'reward_meter_std': 0.3040141463279724, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.3990457057952881, 'reward_total_composite_std': 0.3067740499973297, 'epoch': 0.05}

 37%|███▋      | 1217/3300 [1:11:21<3:52:07,  6.69s/it]INFO 04-11 23:48:13 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1218/3300 [1:11:27<3:47:58,  6.57s/it]2026-04-11 23:48:19,051 | INFO | train_grpo_train | metrics_logged mode=train step=1218

                                                       
{'loss': -0.0029, 'grad_norm': 4.59775972366333, 'learning_rate': 6.3121212121212125e-06, 'num_tokens': 2751115.0, 'completions/mean_length': 64.375, 'completions/min_length': 63.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9942485094070435, 'rewards/meter/std': 0.003570063039660454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9942485094070435, 'rewards/total_composite/std': 0.003570063039660454, 'reward': 0.9942485094070435, 'reward_std': 0.0035700735170394182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014056126587092876, 'sampling/sampling_logp_difference/max': 1.6094756126403809, 'sampling/importance_sampling_ratio/min': 0.19999246299266815, 'sampling/importance_sampling_ratio/mean': 1.001212239265442, 'sampling/importance_sampling_ratio/max': 1.7857288122177124, 'entropy': 0.048549097031354904, 'clip_ratio/low_mean': 0.003937252098694444, 'clip_ratio/low_min': 0.003937252098694444, 'clip_ratio/high_mean': 0.003968254197388887, 'clip_ratio/high_max': 0.003968254197388887, 'clip_ratio/region_mean': 0.007905506296083331, 'reward_total_mean': 0.9942485094070435, 'reward_meter_mean': 0.9942485094070435, 'reward_meter_std': 0.003570063039660454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9942485094070435, 'reward_total_composite_std': 0.003570063039660454, 'epoch': 0.05}

 37%|███▋      | 1218/3300 [1:11:28<3:47:58,  6.57s/it]INFO 04-11 23:48:19 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1219/3300 [1:11:34<3:42:17,  6.41s/it]2026-04-11 23:48:25,068 | INFO | train_grpo_train | metrics_logged mode=train step=1219

                                                       
{'loss': -0.0049, 'grad_norm': 4.931517124176025, 'learning_rate': 6.309090909090909e-06, 'num_tokens': 2753014.0, 'completions/mean_length': 78.375, 'completions/min_length': 76.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.375, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9358141422271729, 'rewards/meter/std': 0.024736735969781876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8663658499717712, 'rewards/total_composite/std': 0.10636419802904129, 'reward': 0.8663658499717712, 'reward_std': 0.1063641905784607, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03355337679386139, 'sampling/sampling_logp_difference/max': 2.9294357299804688, 'sampling/importance_sampling_ratio/min': 0.05342717468738556, 'sampling/importance_sampling_ratio/mean': 0.9951846599578857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09925029519945383, 'clip_ratio/low_mean': 0.008117978577502072, 'clip_ratio/low_min': 0.008117978577502072, 'clip_ratio/high_mean': 0.023660059785470366, 'clip_ratio/high_max': 0.023660059785470366, 'clip_ratio/region_mean': 0.03177803836297244, 'reward_total_mean': 0.8663658499717712, 'reward_meter_mean': 0.9358141422271729, 'reward_meter_std': 0.024736735969781876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8663658499717712, 'reward_total_composite_std': 0.10636419802904129, 'epoch': 0.05}

 37%|███▋      | 1219/3300 [1:11:34<3:42:17,  6.41s/it]INFO 04-11 23:48:25 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1220/3300 [1:11:38<3:20:47,  5.79s/it]2026-04-11 23:48:29,421 | INFO | train_grpo_train | metrics_logged mode=train step=1220

                                                       
{'loss': 0.0002, 'grad_norm': 0.8351545929908752, 'learning_rate': 6.306060606060607e-06, 'num_tokens': 2754558.0, 'completions/mean_length': 37.0, 'completions/min_length': 37.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9974673390388489, 'rewards/meter/std': 1.9203745978302322e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974673390388489, 'rewards/total_composite/std': 1.9203745978302322e-05, 'reward': 0.9974673390388489, 'reward_std': 1.9203745978302322e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010439498350024223, 'sampling/sampling_logp_difference/max': 0.9242507219314575, 'sampling/importance_sampling_ratio/min': 0.39682865142822266, 'sampling/importance_sampling_ratio/mean': 1.0021648406982422, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03344483021646738, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.0033783784601837397, 'clip_ratio/high_max': 0.0033783784601837397, 'clip_ratio/region_mean': 0.006756756920367479, 'reward_total_mean': 0.9974673390388489, 'reward_meter_mean': 0.9974673390388489, 'reward_meter_std': 1.9203745978302322e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974673390388489, 'reward_total_composite_std': 1.9203745978302322e-05, 'epoch': 0.05}

 37%|███▋      | 1220/3300 [1:11:38<3:20:47,  5.79s/it]INFO 04-11 23:48:29 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1221/3300 [1:11:45<3:38:08,  6.30s/it]2026-04-11 23:48:36,892 | INFO | train_grpo_train | metrics_logged mode=train step=1221

                                                       
{'loss': 0.0928, 'grad_norm': 3.4342517852783203, 'learning_rate': 6.303030303030303e-06, 'num_tokens': 2757419.0, 'completions/mean_length': 174.625, 'completions/min_length': 162.0, 'completions/max_length': 219.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 174.625, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 219.0, 'rewards/meter/mean': 0.9972714185714722, 'rewards/meter/std': 0.0009564639185555279, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8042929172515869, 'rewards/repeat_penalty/std': 0.05763205140829086, 'rewards/total_composite/mean': 0.7635921239852905, 'rewards/total_composite/std': 0.1048756018280983, 'reward': 0.7635921239852905, 'reward_std': 0.1048755943775177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025953466072678566, 'sampling/sampling_logp_difference/max': 4.84648323059082, 'sampling/importance_sampling_ratio/min': 0.007855957373976707, 'sampling/importance_sampling_ratio/mean': 0.9992591738700867, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06981830345466733, 'clip_ratio/low_mean': 0.0034246575087308884, 'clip_ratio/low_min': 0.0034246575087308884, 'clip_ratio/high_mean': 0.00996978604234755, 'clip_ratio/high_max': 0.00996978604234755, 'clip_ratio/region_mean': 0.013394443551078439, 'reward_total_mean': 0.7635921239852905, 'reward_meter_mean': 0.9972714185714722, 'reward_meter_std': 0.0009564639185555279, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8042929172515869, 'reward_repeat_penalty_std': 0.05763205140829086, 'reward_total_composite_mean': 0.7635921239852905, 'reward_total_composite_std': 0.1048756018280983, 'epoch': 0.05}

 37%|███▋      | 1221/3300 [1:11:45<3:38:08,  6.30s/it]INFO 04-11 23:48:37 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1222/3300 [1:11:53<3:52:35,  6.72s/it]2026-04-11 23:48:44,589 | INFO | train_grpo_train | metrics_logged mode=train step=1222

                                                       
{'loss': -0.0077, 'grad_norm': 1.311091661453247, 'learning_rate': 6.300000000000001e-06, 'num_tokens': 2761573.0, 'completions/mean_length': 276.25, 'completions/min_length': 270.0, 'completions/max_length': 286.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 276.25, 'completions/min_terminated_length': 270.0, 'completions/max_terminated_length': 286.0, 'rewards/meter/mean': 0.9943939447402954, 'rewards/meter/std': 0.001025979407131672, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5, 'rewards/repeat_penalty/std': 0.15936382114887238, 'rewards/total_composite/mean': 0.4972769021987915, 'rewards/total_composite/std': 0.158652201294899, 'reward': 0.4972769021987915, 'reward_std': 0.1586521863937378, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013553553260862827, 'sampling/sampling_logp_difference/max': 3.245562791824341, 'sampling/importance_sampling_ratio/min': 0.03894663602113724, 'sampling/importance_sampling_ratio/mean': 0.9976973533630371, 'sampling/importance_sampling_ratio/max': 1.5534347295761108, 'entropy': 0.04638162930496037, 'clip_ratio/low_mean': 0.0027206970553379506, 'clip_ratio/low_min': 0.0027206970553379506, 'clip_ratio/high_mean': 0.006255614600377157, 'clip_ratio/high_max': 0.006255614600377157, 'clip_ratio/region_mean': 0.008976311655715108, 'reward_total_mean': 0.4972769021987915, 'reward_meter_mean': 0.9943939447402954, 'reward_meter_std': 0.001025979407131672, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5, 'reward_repeat_penalty_std': 0.15936382114887238, 'reward_total_composite_mean': 0.4972769021987915, 'reward_total_composite_std': 0.158652201294899, 'epoch': 0.05}

 37%|███▋      | 1222/3300 [1:11:53<3:52:35,  6.72s/it]INFO 04-11 23:48:44 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1223/3300 [1:11:58<3:30:40,  6.09s/it]2026-04-11 23:48:49,203 | INFO | train_grpo_train | metrics_logged mode=train step=1223

                                                       
{'loss': -0.003, 'grad_norm': 2.368471145629883, 'learning_rate': 6.296969696969697e-06, 'num_tokens': 2763119.0, 'completions/mean_length': 34.25, 'completions/min_length': 34.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9942933320999146, 'rewards/meter/std': 8.478895324515179e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9942933320999146, 'rewards/total_composite/std': 8.478895324515179e-05, 'reward': 0.9942933320999146, 'reward_std': 8.477975643472746e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009194603189826012, 'sampling/sampling_logp_difference/max': 0.555586576461792, 'sampling/importance_sampling_ratio/min': 0.5737356543540955, 'sampling/importance_sampling_ratio/mean': 1.0032249689102173, 'sampling/importance_sampling_ratio/max': 1.5068020820617676, 'entropy': 0.043639433570206165, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.0035714285913854837, 'clip_ratio/high_max': 0.0035714285913854837, 'clip_ratio/region_mean': 0.007247899193316698, 'reward_total_mean': 0.9942933320999146, 'reward_meter_mean': 0.9942933320999146, 'reward_meter_std': 8.478895324515179e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9942933320999146, 'reward_total_composite_std': 8.478895324515179e-05, 'epoch': 0.05}

 37%|███▋      | 1223/3300 [1:11:58<3:30:40,  6.09s/it]INFO 04-11 23:48:49 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1224/3300 [1:12:03<3:19:09,  5.76s/it]2026-04-11 23:48:54,188 | INFO | train_grpo_train | metrics_logged mode=train step=1224

                                                       
{'loss': 0.0036, 'grad_norm': 2.613860607147217, 'learning_rate': 6.293939393939394e-06, 'num_tokens': 2764882.0, 'completions/mean_length': 75.375, 'completions/min_length': 72.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9953185319900513, 'rewards/meter/std': 0.0019400938181206584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.8708760142326355, 'rewards/total_composite/std': 0.17157325148582458, 'reward': 0.8708760142326355, 'reward_std': 0.1715732365846634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028062716126441956, 'sampling/sampling_logp_difference/max': 2.739466667175293, 'sampling/importance_sampling_ratio/min': 0.06460478901863098, 'sampling/importance_sampling_ratio/mean': 1.003928542137146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.141799321398139, 'clip_ratio/low_mean': 0.006672008661553264, 'clip_ratio/low_min': 0.006672008661553264, 'clip_ratio/high_mean': 0.016601469949819148, 'clip_ratio/high_max': 0.016601469949819148, 'clip_ratio/region_mean': 0.02327347861137241, 'reward_total_mean': 0.8708760142326355, 'reward_meter_mean': 0.9953185319900513, 'reward_meter_std': 0.0019400938181206584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.8708760142326355, 'reward_total_composite_std': 0.17157325148582458, 'epoch': 0.05}

 37%|███▋      | 1224/3300 [1:12:03<3:19:09,  5.76s/it]INFO 04-11 23:48:54 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1225/3300 [1:12:09<3:26:56,  5.98s/it]2026-04-11 23:49:00,703 | INFO | train_grpo_train | metrics_logged mode=train step=1225

                                                       
{'loss': -0.0013, 'grad_norm': 0.7861045002937317, 'learning_rate': 6.290909090909092e-06, 'num_tokens': 2767323.0, 'completions/mean_length': 122.125, 'completions/min_length': 122.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.125, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9720443487167358, 'rewards/meter/std': 0.0028714225627481937, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.2857142984867096, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.277726948261261, 'rewards/total_composite/std': 0.000820409506559372, 'reward': 0.277726948261261, 'reward_std': 0.0008204064215533435, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009138394147157669, 'sampling/sampling_logp_difference/max': 2.7687785625457764, 'sampling/importance_sampling_ratio/min': 0.12622418999671936, 'sampling/importance_sampling_ratio/mean': 0.9990015625953674, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.018890328239649534, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0030487803742289543, 'clip_ratio/high_max': 0.0030487803742289543, 'clip_ratio/region_mean': 0.0030487803742289543, 'reward_total_mean': 0.277726948261261, 'reward_meter_mean': 0.9720443487167358, 'reward_meter_std': 0.0028714225627481937, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.2857142984867096, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.277726948261261, 'reward_total_composite_std': 0.000820409506559372, 'epoch': 0.05}

 37%|███▋      | 1225/3300 [1:12:09<3:26:56,  5.98s/it]INFO 04-11 23:49:00 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1226/3300 [1:12:14<3:15:09,  5.65s/it]2026-04-11 23:49:05,568 | INFO | train_grpo_train | metrics_logged mode=train step=1226

                                                       
{'loss': -0.0158, 'grad_norm': 7.526054382324219, 'learning_rate': 6.287878787878788e-06, 'num_tokens': 2768989.0, 'completions/mean_length': 62.25, 'completions/min_length': 58.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.25, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8702743053436279, 'rewards/meter/std': 0.2847423553466797, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.6277604103088379, 'rewards/total_composite/std': 0.21942587196826935, 'reward': 0.6277604103088379, 'reward_std': 0.21942584216594696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027990855276584625, 'sampling/sampling_logp_difference/max': 1.2952322959899902, 'sampling/importance_sampling_ratio/min': 0.273834228515625, 'sampling/importance_sampling_ratio/mean': 1.007242202758789, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10299600008875132, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.020103482995182276, 'clip_ratio/high_max': 0.020103482995182276, 'clip_ratio/region_mean': 0.024413827806711197, 'reward_total_mean': 0.6277604103088379, 'reward_meter_mean': 0.8702743053436279, 'reward_meter_std': 0.2847423553466797, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.6277604103088379, 'reward_total_composite_std': 0.21942587196826935, 'epoch': 0.05}

 37%|███▋      | 1226/3300 [1:12:14<3:15:09,  5.65s/it]INFO 04-11 23:49:05 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1227/3300 [1:12:18<3:01:59,  5.27s/it]2026-04-11 23:49:09,947 | INFO | train_grpo_train | metrics_logged mode=train step=1227

                                                       
{'loss': -0.0048, 'grad_norm': 1.5801409482955933, 'learning_rate': 6.284848484848486e-06, 'num_tokens': 2770476.0, 'completions/mean_length': 32.875, 'completions/min_length': 32.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9896509051322937, 'rewards/meter/std': 0.00014116229431238025, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9896509051322937, 'rewards/total_composite/std': 0.00014116229431238025, 'reward': 0.9896509051322937, 'reward_std': 0.00014115417434368283, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016333717852830887, 'sampling/sampling_logp_difference/max': 1.630277156829834, 'sampling/importance_sampling_ratio/min': 0.19587527215480804, 'sampling/importance_sampling_ratio/mean': 0.9984188675880432, 'sampling/importance_sampling_ratio/max': 1.3669874668121338, 'entropy': 0.03647553990595043, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.9896509051322937, 'reward_meter_mean': 0.9896509051322937, 'reward_meter_std': 0.00014116229431238025, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9896509051322937, 'reward_total_composite_std': 0.00014116229431238025, 'epoch': 0.05}

 37%|███▋      | 1227/3300 [1:12:18<3:01:59,  5.27s/it]INFO 04-11 23:49:10 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1228/3300 [1:12:24<3:05:44,  5.38s/it]2026-04-11 23:49:15,584 | INFO | train_grpo_train | metrics_logged mode=train step=1228

                                                       
{'loss': 0.0107, 'grad_norm': 3.883788824081421, 'learning_rate': 6.2818181818181825e-06, 'num_tokens': 2772825.0, 'completions/mean_length': 131.625, 'completions/min_length': 126.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.625, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.9974247217178345, 'rewards/meter/std': 0.001207518856972456, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8035714626312256, 'rewards/repeat_penalty/std': 0.10628911107778549, 'rewards/total_composite/mean': 0.8014565706253052, 'rewards/total_composite/std': 0.10568944364786148, 'reward': 0.8014565706253052, 'reward_std': 0.10568942874670029, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01869257725775242, 'sampling/sampling_logp_difference/max': 3.0793256759643555, 'sampling/importance_sampling_ratio/min': 0.04599026218056679, 'sampling/importance_sampling_ratio/mean': 1.0004632472991943, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0591668093111366, 'clip_ratio/low_mean': 0.004665242100600153, 'clip_ratio/low_min': 0.004665242100600153, 'clip_ratio/high_mean': 0.009555137949064374, 'clip_ratio/high_max': 0.009555137949064374, 'clip_ratio/region_mean': 0.014220380049664527, 'reward_total_mean': 0.8014565706253052, 'reward_meter_mean': 0.9974247217178345, 'reward_meter_std': 0.001207518856972456, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8035714626312256, 'reward_repeat_penalty_std': 0.10628911107778549, 'reward_total_composite_mean': 0.8014565706253052, 'reward_total_composite_std': 0.10568944364786148, 'epoch': 0.05}

 37%|███▋      | 1228/3300 [1:12:24<3:05:44,  5.38s/it]INFO 04-11 23:49:15 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1229/3300 [1:12:29<2:59:03,  5.19s/it]2026-04-11 23:49:20,327 | INFO | train_grpo_train | metrics_logged mode=train step=1229

                                                       
{'loss': 0.0111, 'grad_norm': 6.317329406738281, 'learning_rate': 6.27878787878788e-06, 'num_tokens': 2774629.0, 'completions/mean_length': 57.5, 'completions/min_length': 56.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.11471378803253174, 'rewards/meter/std': 0.10604586452245712, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.11471378803253174, 'rewards/total_composite/std': 0.10604586452245712, 'reward': 0.11471378803253174, 'reward_std': 0.10604586452245712, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033396147191524506, 'sampling/sampling_logp_difference/max': 1.2361993789672852, 'sampling/importance_sampling_ratio/min': 0.2904861271381378, 'sampling/importance_sampling_ratio/mean': 1.000756025314331, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15489636361598969, 'clip_ratio/low_mean': 0.006359649356454611, 'clip_ratio/low_min': 0.006359649356454611, 'clip_ratio/high_mean': 0.008696309756487608, 'clip_ratio/high_max': 0.008696309756487608, 'clip_ratio/region_mean': 0.015055959112942219, 'reward_total_mean': 0.11471378803253174, 'reward_meter_mean': 0.11471378803253174, 'reward_meter_std': 0.10604586452245712, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.11471378803253174, 'reward_total_composite_std': 0.10604586452245712, 'epoch': 0.05}

 37%|███▋      | 1229/3300 [1:12:29<2:59:03,  5.19s/it]INFO 04-11 23:49:20 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1230/3300 [1:12:33<2:52:07,  4.99s/it]2026-04-11 23:49:24,853 | INFO | train_grpo_train | metrics_logged mode=train step=1230

                                                       
{'loss': 0.0035, 'grad_norm': 3.3790931701660156, 'learning_rate': 6.275757575757576e-06, 'num_tokens': 2776041.0, 'completions/mean_length': 33.5, 'completions/min_length': 33.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9702644944190979, 'rewards/meter/std': 0.0008763103978708386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9702644944190979, 'rewards/total_composite/std': 0.0008763103978708386, 'reward': 0.9702644944190979, 'reward_std': 0.0008763210498727858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022185776382684708, 'sampling/sampling_logp_difference/max': 1.2655248641967773, 'sampling/importance_sampling_ratio/min': 0.3646947145462036, 'sampling/importance_sampling_ratio/mean': 1.0119966268539429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07495713606476784, 'clip_ratio/low_mean': 0.007464349502697587, 'clip_ratio/low_min': 0.007464349502697587, 'clip_ratio/high_mean': 0.0036764706019312143, 'clip_ratio/high_max': 0.0036764706019312143, 'clip_ratio/region_mean': 0.011140820104628801, 'reward_total_mean': 0.9702644944190979, 'reward_meter_mean': 0.9702644944190979, 'reward_meter_std': 0.0008763103978708386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9702644944190979, 'reward_total_composite_std': 0.0008763103978708386, 'epoch': 0.05}

 37%|███▋      | 1230/3300 [1:12:33<2:52:07,  4.99s/it]INFO 04-11 23:49:25 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1231/3300 [1:12:38<2:50:40,  4.95s/it]2026-04-11 23:49:29,709 | INFO | train_grpo_train | metrics_logged mode=train step=1231

                                                       
{'loss': -0.0146, 'grad_norm': 6.076712608337402, 'learning_rate': 6.2727272727272734e-06, 'num_tokens': 2777725.0, 'completions/mean_length': 58.5, 'completions/min_length': 52.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.5, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6401726603507996, 'rewards/meter/std': 0.34182804822921753, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6401726603507996, 'rewards/total_composite/std': 0.34182804822921753, 'reward': 0.6401726603507996, 'reward_std': 0.34182801842689514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06882074475288391, 'sampling/sampling_logp_difference/max': 1.4318437576293945, 'sampling/importance_sampling_ratio/min': 0.24655072391033173, 'sampling/importance_sampling_ratio/mean': 1.0154814720153809, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5035125408321619, 'clip_ratio/low_mean': 0.028201664797961712, 'clip_ratio/low_min': 0.028201664797961712, 'clip_ratio/high_mean': 0.03711274731904268, 'clip_ratio/high_max': 0.03711274731904268, 'clip_ratio/region_mean': 0.0653144121170044, 'reward_total_mean': 0.6401726603507996, 'reward_meter_mean': 0.6401726603507996, 'reward_meter_std': 0.34182804822921753, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6401726603507996, 'reward_total_composite_std': 0.34182804822921753, 'epoch': 0.05}

 37%|███▋      | 1231/3300 [1:12:38<2:50:40,  4.95s/it]INFO 04-11 23:49:29 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1232/3300 [1:12:43<2:51:11,  4.97s/it]2026-04-11 23:49:34,717 | INFO | train_grpo_train | metrics_logged mode=train step=1232

                                                       
{'loss': -0.0004, 'grad_norm': 4.830690860748291, 'learning_rate': 6.26969696969697e-06, 'num_tokens': 2779575.0, 'completions/mean_length': 76.25, 'completions/min_length': 74.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8763631582260132, 'rewards/meter/std': 0.33702343702316284, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8763631582260132, 'rewards/total_composite/std': 0.33702343702316284, 'reward': 0.8763631582260132, 'reward_std': 0.33702343702316284, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03446757048368454, 'sampling/sampling_logp_difference/max': 3.854651927947998, 'sampling/importance_sampling_ratio/min': 0.021180974319577217, 'sampling/importance_sampling_ratio/mean': 0.9978901743888855, 'sampling/importance_sampling_ratio/max': 1.7376562356948853, 'entropy': 0.1252383915707469, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.01965810963883996, 'clip_ratio/high_max': 0.01965810963883996, 'clip_ratio/region_mean': 0.01965810963883996, 'reward_total_mean': 0.8763631582260132, 'reward_meter_mean': 0.8763631582260132, 'reward_meter_std': 0.33702343702316284, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8763631582260132, 'reward_total_composite_std': 0.33702343702316284, 'epoch': 0.05}

 37%|███▋      | 1232/3300 [1:12:43<2:51:11,  4.97s/it]INFO 04-11 23:49:35 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1233/3300 [1:12:50<3:10:15,  5.52s/it]2026-04-11 23:49:41,537 | INFO | train_grpo_train | metrics_logged mode=train step=1233

                                                       
{'loss': 0.0154, 'grad_norm': 2.373234748840332, 'learning_rate': 6.266666666666668e-06, 'num_tokens': 2782216.0, 'completions/mean_length': 150.125, 'completions/min_length': 140.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.125, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9954185485839844, 'rewards/meter/std': 0.0016214650822803378, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6785714626312256, 'rewards/repeat_penalty/std': 0.14787118136882782, 'rewards/total_composite/mean': 0.6755450963973999, 'rewards/total_composite/std': 0.1476629376411438, 'reward': 0.6755450963973999, 'reward_std': 0.1476629078388214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013783158734440804, 'sampling/sampling_logp_difference/max': 1.0632364749908447, 'sampling/importance_sampling_ratio/min': 0.34533634781837463, 'sampling/importance_sampling_ratio/mean': 1.002487301826477, 'sampling/importance_sampling_ratio/max': 1.8110910654067993, 'entropy': 0.08290160167962313, 'clip_ratio/low_mean': 0.004945102846249938, 'clip_ratio/low_min': 0.004945102846249938, 'clip_ratio/high_mean': 0.006812343490310013, 'clip_ratio/high_max': 0.006812343490310013, 'clip_ratio/region_mean': 0.011757446336559951, 'reward_total_mean': 0.6755450963973999, 'reward_meter_mean': 0.9954185485839844, 'reward_meter_std': 0.0016214650822803378, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6785714626312256, 'reward_repeat_penalty_std': 0.14787118136882782, 'reward_total_composite_mean': 0.6755450963973999, 'reward_total_composite_std': 0.1476629376411438, 'epoch': 0.05}

 37%|███▋      | 1233/3300 [1:12:50<3:10:15,  5.52s/it]INFO 04-11 23:49:41 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1234/3300 [1:12:57<3:26:51,  6.01s/it]2026-04-11 23:49:48,675 | INFO | train_grpo_train | metrics_logged mode=train step=1234

                                                       
{'loss': -0.002, 'grad_norm': 1.7159820795059204, 'learning_rate': 6.263636363636364e-06, 'num_tokens': 2785815.0, 'completions/mean_length': 242.875, 'completions/min_length': 227.0, 'completions/max_length': 260.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 242.875, 'completions/min_terminated_length': 227.0, 'completions/max_terminated_length': 260.0, 'rewards/meter/mean': 0.9932608008384705, 'rewards/meter/std': 0.003682313719764352, 'rewards/count_adherence/mean': 0.737500011920929, 'rewards/count_adherence/std': 0.05175492912530899, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6418956518173218, 'rewards/repeat_penalty/std': 0.06913584470748901, 'rewards/total_composite/mean': 0.4686053991317749, 'rewards/total_composite/std': 0.04152140021324158, 'reward': 0.4686053991317749, 'reward_std': 0.04152139648795128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022554513067007065, 'sampling/sampling_logp_difference/max': 3.497123956680298, 'sampling/importance_sampling_ratio/min': 0.030284356325864792, 'sampling/importance_sampling_ratio/mean': 1.0014477968215942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09457159182056785, 'clip_ratio/low_mean': 0.005789120797999203, 'clip_ratio/low_min': 0.005789120797999203, 'clip_ratio/high_mean': 0.00862484163371846, 'clip_ratio/high_max': 0.00862484163371846, 'clip_ratio/region_mean': 0.014413962431717664, 'reward_total_mean': 0.4686053991317749, 'reward_meter_mean': 0.9932608008384705, 'reward_meter_std': 0.003682313719764352, 'reward_count_adherence_mean': 0.737500011920929, 'reward_count_adherence_std': 0.05175492912530899, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6418956518173218, 'reward_repeat_penalty_std': 0.06913584470748901, 'reward_total_composite_mean': 0.4686053991317749, 'reward_total_composite_std': 0.04152140021324158, 'epoch': 0.05}

 37%|███▋      | 1234/3300 [1:12:57<3:26:51,  6.01s/it]INFO 04-11 23:49:48 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1235/3300 [1:13:02<3:15:29,  5.68s/it]2026-04-11 23:49:53,590 | INFO | train_grpo_train | metrics_logged mode=train step=1235

                                                       
{'loss': 0.011, 'grad_norm': 15.415648460388184, 'learning_rate': 6.260606060606062e-06, 'num_tokens': 2787651.0, 'completions/mean_length': 80.5, 'completions/min_length': 79.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.5, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9555257558822632, 'rewards/meter/std': 0.013856525532901287, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9315690398216248, 'rewards/total_composite/std': 0.06806813180446625, 'reward': 0.9315690398216248, 'reward_std': 0.06806813925504684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04006582126021385, 'sampling/sampling_logp_difference/max': 5.584328651428223, 'sampling/importance_sampling_ratio/min': 0.003756270743906498, 'sampling/importance_sampling_ratio/mean': 0.9921613335609436, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10135750938206911, 'clip_ratio/low_mean': 0.006172839552164078, 'clip_ratio/low_min': 0.006172839552164078, 'clip_ratio/high_mean': 0.031061405315995216, 'clip_ratio/high_max': 0.031061405315995216, 'clip_ratio/region_mean': 0.037234244868159294, 'reward_total_mean': 0.9315690398216248, 'reward_meter_mean': 0.9555257558822632, 'reward_meter_std': 0.013856525532901287, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9315690398216248, 'reward_total_composite_std': 0.06806813180446625, 'epoch': 0.05}

 37%|███▋      | 1235/3300 [1:13:02<3:15:29,  5.68s/it]INFO 04-11 23:49:53 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1236/3300 [1:13:08<3:21:08,  5.85s/it]2026-04-11 23:49:59,829 | INFO | train_grpo_train | metrics_logged mode=train step=1236

                                                       
{'loss': 0.0094, 'grad_norm': 1.9874340295791626, 'learning_rate': 6.257575757575758e-06, 'num_tokens': 2790696.0, 'completions/mean_length': 181.625, 'completions/min_length': 173.0, 'completions/max_length': 186.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 181.625, 'completions/min_terminated_length': 173.0, 'completions/max_terminated_length': 186.0, 'rewards/meter/mean': 0.9979325532913208, 'rewards/meter/std': 0.00011955534137086943, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.737500011920929, 'rewards/repeat_penalty/std': 0.07440238445997238, 'rewards/total_composite/mean': 0.5256986618041992, 'rewards/total_composite/std': 0.053062621504068375, 'reward': 0.5256986618041992, 'reward_std': 0.05306261032819748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012886044569313526, 'sampling/sampling_logp_difference/max': 1.4943835735321045, 'sampling/importance_sampling_ratio/min': 0.22438688576221466, 'sampling/importance_sampling_ratio/mean': 1.0016508102416992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.050744547275826335, 'clip_ratio/low_mean': 0.002725058700889349, 'clip_ratio/low_min': 0.002725058700889349, 'clip_ratio/high_mean': 0.006200521253049374, 'clip_ratio/high_max': 0.006200521253049374, 'clip_ratio/region_mean': 0.008925579953938723, 'reward_total_mean': 0.5256986618041992, 'reward_meter_mean': 0.9979325532913208, 'reward_meter_std': 0.00011955534137086943, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.737500011920929, 'reward_repeat_penalty_std': 0.07440238445997238, 'reward_total_composite_mean': 0.5256986618041992, 'reward_total_composite_std': 0.053062621504068375, 'epoch': 0.05}

 37%|███▋      | 1236/3300 [1:13:08<3:21:08,  5.85s/it]INFO 04-11 23:50:00 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1237/3300 [1:13:13<3:09:43,  5.52s/it]2026-04-11 23:50:04,580 | INFO | train_grpo_train | metrics_logged mode=train step=1237

                                                       
{'loss': 0.0102, 'grad_norm': 6.145448684692383, 'learning_rate': 6.254545454545455e-06, 'num_tokens': 2792675.0, 'completions/mean_length': 67.375, 'completions/min_length': 64.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.35231250524520874, 'rewards/meter/std': 0.3081033229827881, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.35225385427474976, 'rewards/total_composite/std': 0.30817967653274536, 'reward': 0.35225385427474976, 'reward_std': 0.30817967653274536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0802878737449646, 'sampling/sampling_logp_difference/max': 3.271935224533081, 'sampling/importance_sampling_ratio/min': 0.037932947278022766, 'sampling/importance_sampling_ratio/mean': 1.002850890159607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3711269237101078, 'clip_ratio/low_mean': 0.03385214158333838, 'clip_ratio/low_min': 0.03385214158333838, 'clip_ratio/high_mean': 0.028940699994564056, 'clip_ratio/high_max': 0.028940699994564056, 'clip_ratio/region_mean': 0.06279284157790244, 'reward_total_mean': 0.35225385427474976, 'reward_meter_mean': 0.35231250524520874, 'reward_meter_std': 0.3081033229827881, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.35225385427474976, 'reward_total_composite_std': 0.30817967653274536, 'epoch': 0.05}

 37%|███▋      | 1237/3300 [1:13:13<3:09:43,  5.52s/it]INFO 04-11 23:50:04 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1238/3300 [1:13:18<3:06:39,  5.43s/it]2026-04-11 23:50:09,809 | INFO | train_grpo_train | metrics_logged mode=train step=1238

                                                       
{'loss': -0.0256, 'grad_norm': 6.303664684295654, 'learning_rate': 6.251515151515152e-06, 'num_tokens': 2794768.0, 'completions/mean_length': 85.625, 'completions/min_length': 78.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.625, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.6144058704376221, 'rewards/meter/std': 0.3429763913154602, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.452033668756485, 'rewards/total_composite/std': 0.2605564594268799, 'reward': 0.452033668756485, 'reward_std': 0.2605564594268799, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05436992645263672, 'sampling/sampling_logp_difference/max': 2.363379955291748, 'sampling/importance_sampling_ratio/min': 0.09410162270069122, 'sampling/importance_sampling_ratio/mean': 1.0040868520736694, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1898734886199236, 'clip_ratio/low_mean': 0.009122495306655765, 'clip_ratio/low_min': 0.009122495306655765, 'clip_ratio/high_mean': 0.017562699620611966, 'clip_ratio/high_max': 0.017562699620611966, 'clip_ratio/region_mean': 0.02668519492726773, 'reward_total_mean': 0.452033668756485, 'reward_meter_mean': 0.6144058704376221, 'reward_meter_std': 0.3429763913154602, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.452033668756485, 'reward_total_composite_std': 0.2605564594268799, 'epoch': 0.05}

 38%|███▊      | 1238/3300 [1:13:18<3:06:39,  5.43s/it]INFO 04-11 23:50:10 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1239/3300 [1:13:23<2:56:49,  5.15s/it]2026-04-11 23:50:14,294 | INFO | train_grpo_train | metrics_logged mode=train step=1239

                                                       
{'loss': 0.0011, 'grad_norm': 11.653518676757812, 'learning_rate': 6.248484848484849e-06, 'num_tokens': 2796176.0, 'completions/mean_length': 34.0, 'completions/min_length': 33.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9623700380325317, 'rewards/meter/std': 0.017837192863225937, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9623700380325317, 'rewards/total_composite/std': 0.017837192863225937, 'reward': 0.9623700380325317, 'reward_std': 0.017837194725871086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017896274104714394, 'sampling/sampling_logp_difference/max': 1.5471727848052979, 'sampling/importance_sampling_ratio/min': 0.21284890174865723, 'sampling/importance_sampling_ratio/mean': 0.9969296455383301, 'sampling/importance_sampling_ratio/max': 1.2581915855407715, 'entropy': 0.04970884299837053, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.007464349502697587, 'reward_total_mean': 0.9623700380325317, 'reward_meter_mean': 0.9623700380325317, 'reward_meter_std': 0.017837192863225937, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9623700380325317, 'reward_total_composite_std': 0.017837192863225937, 'epoch': 0.05}

 38%|███▊      | 1239/3300 [1:13:23<2:56:49,  5.15s/it]INFO 04-11 23:50:14 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1240/3300 [1:13:28<2:53:02,  5.04s/it]2026-04-11 23:50:19,083 | INFO | train_grpo_train | metrics_logged mode=train step=1240

                                                       
{'loss': -0.0239, 'grad_norm': 18.420658111572266, 'learning_rate': 6.245454545454545e-06, 'num_tokens': 2797992.0, 'completions/mean_length': 65.0, 'completions/min_length': 58.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.750508189201355, 'rewards/meter/std': 0.30968019366264343, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.750508189201355, 'rewards/total_composite/std': 0.30968019366264343, 'reward': 0.750508189201355, 'reward_std': 0.3096802234649658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.052898943424224854, 'sampling/sampling_logp_difference/max': 2.5542707443237305, 'sampling/importance_sampling_ratio/min': 0.07774890959262848, 'sampling/importance_sampling_ratio/mean': 0.9967520833015442, 'sampling/importance_sampling_ratio/max': 1.6765276193618774, 'entropy': 0.17542924359440804, 'clip_ratio/low_mean': 0.023755351547151804, 'clip_ratio/low_min': 0.023755351547151804, 'clip_ratio/high_mean': 0.037521140300668776, 'clip_ratio/high_max': 0.037521140300668776, 'clip_ratio/region_mean': 0.06127649184782058, 'reward_total_mean': 0.750508189201355, 'reward_meter_mean': 0.750508189201355, 'reward_meter_std': 0.30968019366264343, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.750508189201355, 'reward_total_composite_std': 0.30968019366264343, 'epoch': 0.05}

 38%|███▊      | 1240/3300 [1:13:28<2:53:02,  5.04s/it]INFO 04-11 23:50:19 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1241/3300 [1:13:33<2:59:46,  5.24s/it]2026-04-11 23:50:24,783 | INFO | train_grpo_train | metrics_logged mode=train step=1241

                                                       
{'loss': 0.0106, 'grad_norm': 5.958988189697266, 'learning_rate': 6.2424242424242434e-06, 'num_tokens': 2799886.0, 'completions/mean_length': 69.75, 'completions/min_length': 67.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9957868456840515, 'rewards/meter/std': 0.0033736079931259155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957868456840515, 'rewards/total_composite/std': 0.0033736079931259155, 'reward': 0.9957868456840515, 'reward_std': 0.0033736105542629957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0355050228536129, 'sampling/sampling_logp_difference/max': 1.2876784801483154, 'sampling/importance_sampling_ratio/min': 0.2759105861186981, 'sampling/importance_sampling_ratio/mean': 0.9955186247825623, 'sampling/importance_sampling_ratio/max': 1.7876514196395874, 'entropy': 0.134712896309793, 'clip_ratio/low_mean': 0.007147361640818417, 'clip_ratio/low_min': 0.007147361640818417, 'clip_ratio/high_mean': 0.030546388239599764, 'clip_ratio/high_max': 0.030546388239599764, 'clip_ratio/region_mean': 0.03769374988041818, 'reward_total_mean': 0.9957868456840515, 'reward_meter_mean': 0.9957868456840515, 'reward_meter_std': 0.0033736079931259155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957868456840515, 'reward_total_composite_std': 0.0033736079931259155, 'epoch': 0.05}

 38%|███▊      | 1241/3300 [1:13:33<2:59:46,  5.24s/it]INFO 04-11 23:50:25 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1242/3300 [1:13:38<2:55:57,  5.13s/it]2026-04-11 23:50:29,660 | INFO | train_grpo_train | metrics_logged mode=train step=1242

                                                       
{'loss': 0.0265, 'grad_norm': 16.907014846801758, 'learning_rate': 6.23939393939394e-06, 'num_tokens': 2801684.0, 'completions/mean_length': 70.75, 'completions/min_length': 68.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.75, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9978107810020447, 'rewards/meter/std': 0.0009332987247034907, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978107810020447, 'rewards/total_composite/std': 0.0009332987247034907, 'reward': 0.9978107810020447, 'reward_std': 0.0009333047200925648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03522227331995964, 'sampling/sampling_logp_difference/max': 2.570969581604004, 'sampling/importance_sampling_ratio/min': 0.07646137475967407, 'sampling/importance_sampling_ratio/mean': 1.003254771232605, 'sampling/importance_sampling_ratio/max': 1.5690217018127441, 'entropy': 0.14225120842456818, 'clip_ratio/low_mean': 0.006905802641995251, 'clip_ratio/low_min': 0.006905802641995251, 'clip_ratio/high_mean': 0.024632065324112773, 'clip_ratio/high_max': 0.024632065324112773, 'clip_ratio/region_mean': 0.031537867966108024, 'reward_total_mean': 0.9978107810020447, 'reward_meter_mean': 0.9978107810020447, 'reward_meter_std': 0.0009332987247034907, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978107810020447, 'reward_total_composite_std': 0.0009332987247034907, 'epoch': 0.05}

 38%|███▊      | 1242/3300 [1:13:38<2:55:57,  5.13s/it]INFO 04-11 23:50:29 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1243/3300 [1:13:43<2:53:08,  5.05s/it]2026-04-11 23:50:34,531 | INFO | train_grpo_train | metrics_logged mode=train step=1243

                                                       
{'loss': 0.0058, 'grad_norm': 8.487194061279297, 'learning_rate': 6.236363636363637e-06, 'num_tokens': 2803443.0, 'completions/mean_length': 63.875, 'completions/min_length': 63.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9128004312515259, 'rewards/meter/std': 0.03738430514931679, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9128004312515259, 'rewards/total_composite/std': 0.03738430514931679, 'reward': 0.9128004312515259, 'reward_std': 0.037384290248155594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04461809620261192, 'sampling/sampling_logp_difference/max': 5.818354606628418, 'sampling/importance_sampling_ratio/min': 0.002972492016851902, 'sampling/importance_sampling_ratio/mean': 1.0069886445999146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10157017782330513, 'clip_ratio/low_mean': 0.021577381063252687, 'clip_ratio/low_min': 0.021577381063252687, 'clip_ratio/high_mean': 0.017581941094249487, 'clip_ratio/high_max': 0.017581941094249487, 'clip_ratio/region_mean': 0.039159322157502174, 'reward_total_mean': 0.9128004312515259, 'reward_meter_mean': 0.9128004312515259, 'reward_meter_std': 0.03738430514931679, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9128004312515259, 'reward_total_composite_std': 0.03738430514931679, 'epoch': 0.05}

 38%|███▊      | 1243/3300 [1:13:43<2:53:08,  5.05s/it]INFO 04-11 23:50:34 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1244/3300 [1:13:49<3:00:16,  5.26s/it]2026-04-11 23:50:40,279 | INFO | train_grpo_train | metrics_logged mode=train step=1244

                                                       
{'loss': -0.0386, 'grad_norm': 2.8616859912872314, 'learning_rate': 6.2333333333333335e-06, 'num_tokens': 2805756.0, 'completions/mean_length': 128.125, 'completions/min_length': 110.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.125, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9177215695381165, 'rewards/meter/std': 0.08298756927251816, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8363094925880432, 'rewards/repeat_penalty/std': 0.09127599745988846, 'rewards/total_composite/mean': 0.7401801347732544, 'rewards/total_composite/std': 0.10087206214666367, 'reward': 0.7401801347732544, 'reward_std': 0.10087206959724426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021584779024124146, 'sampling/sampling_logp_difference/max': 0.7692317962646484, 'sampling/importance_sampling_ratio/min': 0.46336889266967773, 'sampling/importance_sampling_ratio/mean': 1.0052604675292969, 'sampling/importance_sampling_ratio/max': 1.7540781497955322, 'entropy': 0.1307503404095769, 'clip_ratio/low_mean': 0.009928334911819547, 'clip_ratio/low_min': 0.009928334911819547, 'clip_ratio/high_mean': 0.009529390663374215, 'clip_ratio/high_max': 0.009529390663374215, 'clip_ratio/region_mean': 0.019457725575193763, 'reward_total_mean': 0.7401801347732544, 'reward_meter_mean': 0.9177215695381165, 'reward_meter_std': 0.08298756927251816, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8363094925880432, 'reward_repeat_penalty_std': 0.09127599745988846, 'reward_total_composite_mean': 0.7401801347732544, 'reward_total_composite_std': 0.10087206214666367, 'epoch': 0.05}

 38%|███▊      | 1244/3300 [1:13:49<3:00:16,  5.26s/it]INFO 04-11 23:50:40 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1245/3300 [1:13:56<3:17:33,  5.77s/it]2026-04-11 23:50:47,230 | INFO | train_grpo_train | metrics_logged mode=train step=1245

                                                       
{'loss': 0.0098, 'grad_norm': 2.1829257011413574, 'learning_rate': 6.230303030303031e-06, 'num_tokens': 2809292.0, 'completions/mean_length': 230.0, 'completions/min_length': 218.0, 'completions/max_length': 239.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 230.0, 'completions/min_terminated_length': 218.0, 'completions/max_terminated_length': 239.0, 'rewards/meter/mean': 0.9974883198738098, 'rewards/meter/std': 0.0011670741951093078, 'rewards/count_adherence/mean': 0.7638888955116272, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7868589758872986, 'rewards/repeat_penalty/std': 0.07841575890779495, 'rewards/total_composite/mean': 0.5989092588424683, 'rewards/total_composite/std': 0.061690803617239, 'reward': 0.5989092588424683, 'reward_std': 0.061690811067819595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01827998273074627, 'sampling/sampling_logp_difference/max': 2.269944667816162, 'sampling/importance_sampling_ratio/min': 0.10331789404153824, 'sampling/importance_sampling_ratio/mean': 0.9981163740158081, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06460519693791866, 'clip_ratio/low_mean': 0.009747638367116451, 'clip_ratio/low_min': 0.009747638367116451, 'clip_ratio/high_mean': 0.004898735671304166, 'clip_ratio/high_max': 0.004898735671304166, 'clip_ratio/region_mean': 0.014646374038420618, 'reward_total_mean': 0.5989092588424683, 'reward_meter_mean': 0.9974883198738098, 'reward_meter_std': 0.0011670741951093078, 'reward_count_adherence_mean': 0.7638888955116272, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7868589758872986, 'reward_repeat_penalty_std': 0.07841575890779495, 'reward_total_composite_mean': 0.5989092588424683, 'reward_total_composite_std': 0.061690803617239, 'epoch': 0.05}

 38%|███▊      | 1245/3300 [1:13:56<3:17:33,  5.77s/it]INFO 04-11 23:50:47 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1246/3300 [1:14:01<3:09:11,  5.53s/it]2026-04-11 23:50:52,193 | INFO | train_grpo_train | metrics_logged mode=train step=1246

                                                       
{'loss': -0.0011, 'grad_norm': 5.752688407897949, 'learning_rate': 6.227272727272727e-06, 'num_tokens': 2811283.0, 'completions/mean_length': 68.875, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7574069499969482, 'rewards/meter/std': 0.31513679027557373, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7574069499969482, 'rewards/total_composite/std': 0.31513679027557373, 'reward': 0.7574069499969482, 'reward_std': 0.31513676047325134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017493808642029762, 'sampling/sampling_logp_difference/max': 0.742142915725708, 'sampling/importance_sampling_ratio/min': 0.4760926067829132, 'sampling/importance_sampling_ratio/mean': 0.9997113347053528, 'sampling/importance_sampling_ratio/max': 1.6327474117279053, 'entropy': 0.08360549993813038, 'clip_ratio/low_mean': 0.007273018010891974, 'clip_ratio/low_min': 0.007273018010891974, 'clip_ratio/high_mean': 0.007221258128993213, 'clip_ratio/high_max': 0.007221258128993213, 'clip_ratio/region_mean': 0.014494276139885187, 'reward_total_mean': 0.7574069499969482, 'reward_meter_mean': 0.7574069499969482, 'reward_meter_std': 0.31513679027557373, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7574069499969482, 'reward_total_composite_std': 0.31513679027557373, 'epoch': 0.05}

 38%|███▊      | 1246/3300 [1:14:01<3:09:11,  5.53s/it]INFO 04-11 23:50:52 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1247/3300 [1:14:06<3:06:21,  5.45s/it]2026-04-11 23:50:57,453 | INFO | train_grpo_train | metrics_logged mode=train step=1247

                                                       
{'loss': -0.0245, 'grad_norm': 3.6555347442626953, 'learning_rate': 6.224242424242425e-06, 'num_tokens': 2813271.0, 'completions/mean_length': 95.5, 'completions/min_length': 82.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.5, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.95280921459198, 'rewards/meter/std': 0.017720038071274757, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8177590370178223, 'rewards/total_composite/std': 0.12049128115177155, 'reward': 0.8177590370178223, 'reward_std': 0.12049128115177155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020115824416279793, 'sampling/sampling_logp_difference/max': 0.5828499794006348, 'sampling/importance_sampling_ratio/min': 0.5583049654960632, 'sampling/importance_sampling_ratio/mean': 1.0025370121002197, 'sampling/importance_sampling_ratio/max': 1.4265241622924805, 'entropy': 0.13955960143357515, 'clip_ratio/low_mean': 0.011977351736277342, 'clip_ratio/low_min': 0.011977351736277342, 'clip_ratio/high_mean': 0.007731958641670644, 'clip_ratio/high_max': 0.007731958641670644, 'clip_ratio/region_mean': 0.019709310377947986, 'reward_total_mean': 0.8177590370178223, 'reward_meter_mean': 0.95280921459198, 'reward_meter_std': 0.017720038071274757, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8177590370178223, 'reward_total_composite_std': 0.12049128115177155, 'epoch': 0.05}

 38%|███▊      | 1247/3300 [1:14:06<3:06:21,  5.45s/it]INFO 04-11 23:50:57 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1248/3300 [1:14:11<3:01:50,  5.32s/it]2026-04-11 23:51:02,467 | INFO | train_grpo_train | metrics_logged mode=train step=1248

                                                       
{'loss': 0.0024, 'grad_norm': 3.7371089458465576, 'learning_rate': 6.221212121212121e-06, 'num_tokens': 2815189.0, 'completions/mean_length': 76.75, 'completions/min_length': 75.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.75, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9958518147468567, 'rewards/meter/std': 0.0027126488275825977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9958518147468567, 'rewards/total_composite/std': 0.0027126488275825977, 'reward': 0.9958518147468567, 'reward_std': 0.0027126302011311054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03340546786785126, 'sampling/sampling_logp_difference/max': 3.952702045440674, 'sampling/importance_sampling_ratio/min': 0.019202744588255882, 'sampling/importance_sampling_ratio/mean': 0.999036431312561, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14584199711680412, 'clip_ratio/low_mean': 0.004895833437331021, 'clip_ratio/low_min': 0.004895833437331021, 'clip_ratio/high_mean': 0.009827935369685292, 'clip_ratio/high_max': 0.009827935369685292, 'clip_ratio/region_mean': 0.014723768807016313, 'reward_total_mean': 0.9958518147468567, 'reward_meter_mean': 0.9958518147468567, 'reward_meter_std': 0.0027126488275825977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9958518147468567, 'reward_total_composite_std': 0.0027126488275825977, 'epoch': 0.05}

 38%|███▊      | 1248/3300 [1:14:11<3:01:50,  5.32s/it]INFO 04-11 23:51:02 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1249/3300 [1:14:16<3:00:26,  5.28s/it]2026-04-11 23:51:07,657 | INFO | train_grpo_train | metrics_logged mode=train step=1249

                                                       
{'loss': -0.0042, 'grad_norm': 2.240663528442383, 'learning_rate': 6.218181818181819e-06, 'num_tokens': 2817451.0, 'completions/mean_length': 100.75, 'completions/min_length': 99.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.75, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9981772899627686, 'rewards/meter/std': 0.0003791229974012822, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.898347020149231, 'rewards/total_composite/std': 0.1065896674990654, 'reward': 0.898347020149231, 'reward_std': 0.1065896600484848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012546716257929802, 'sampling/sampling_logp_difference/max': 0.8134353160858154, 'sampling/importance_sampling_ratio/min': 0.516797661781311, 'sampling/importance_sampling_ratio/mean': 1.003831148147583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06938632484525442, 'clip_ratio/low_mean': 0.007488123839721084, 'clip_ratio/low_min': 0.007488123839721084, 'clip_ratio/high_mean': 0.0036526747280731797, 'clip_ratio/high_max': 0.0036526747280731797, 'clip_ratio/region_mean': 0.011140798567794263, 'reward_total_mean': 0.898347020149231, 'reward_meter_mean': 0.9981772899627686, 'reward_meter_std': 0.0003791229974012822, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.898347020149231, 'reward_total_composite_std': 0.1065896674990654, 'epoch': 0.05}

 38%|███▊      | 1249/3300 [1:14:16<3:00:26,  5.28s/it]INFO 04-11 23:51:07 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1250/3300 [1:14:21<2:57:00,  5.18s/it]2026-04-11 23:51:12,607 | INFO | train_grpo_train | metrics_logged mode=train step=1250

                                                       
{'loss': -0.0122, 'grad_norm': 2.518706798553467, 'learning_rate': 6.215151515151515e-06, 'num_tokens': 2819322.0, 'completions/mean_length': 79.875, 'completions/min_length': 77.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9909250736236572, 'rewards/meter/std': 0.0028839793521910906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9494280815124512, 'rewards/total_composite/std': 0.11536786705255508, 'reward': 0.9494280815124512, 'reward_std': 0.11536785960197449, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0175126064568758, 'sampling/sampling_logp_difference/max': 2.6809237003326416, 'sampling/importance_sampling_ratio/min': 0.0684998482465744, 'sampling/importance_sampling_ratio/mean': 1.0048216581344604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09262991230934858, 'clip_ratio/low_mean': 0.004870129749178886, 'clip_ratio/low_min': 0.004870129749178886, 'clip_ratio/high_mean': 0.01242569915484637, 'clip_ratio/high_max': 0.01242569915484637, 'clip_ratio/region_mean': 0.017295828904025257, 'reward_total_mean': 0.9494280815124512, 'reward_meter_mean': 0.9909250736236572, 'reward_meter_std': 0.0028839793521910906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9494280815124512, 'reward_total_composite_std': 0.11536786705255508, 'epoch': 0.05}

 38%|███▊      | 1250/3300 [1:14:21<2:57:00,  5.18s/it]INFO 04-11 23:51:12 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:26,  2.37s/it]

 23%|██▎       | 3/13 [00:07<00:24,  2.50s/it]

 31%|███       | 4/13 [00:11<00:28,  3.12s/it]

 38%|███▊      | 5/13 [00:15<00:27,  3.48s/it]

 46%|████▌     | 6/13 [00:20<00:27,  3.98s/it]

 54%|█████▍    | 7/13 [00:24<00:24,  4.00s/it]

 62%|██████▏   | 8/13 [00:29<00:21,  4.35s/it]

 69%|██████▉   | 9/13 [00:35<00:18,  4.64s/it]

 77%|███████▋  | 10/13 [00:39<00:13,  4.47s/it]

 85%|████████▍ | 11/13 [00:43<00:08,  4.28s/it]

 92%|█████████▏| 12/13 [00:47<00:04,  4.32s/it]

100%|██████████| 13/13 [00:51<00:00,  4.27s/it]2026-04-11 23:52:09,004 | INFO | train_grpo_train | metrics_logged mode=eval step=1250

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 56.3598, 'eval_samples_per_second': 1.845, 'eval_steps_per_second': 0.231, 'eval_num_tokens': 2819322.0, 'eval_completions/mean_length': 172.58653846153845, 'eval_completions/min_length': 65.3076923076923, 'eval_completions/max_length': 286.84615384615387, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 172.58653846153845, 'eval_completions/min_terminated_length': 65.3076923076923, 'eval_completions/max_terminated_length': 286.84615384615387, 'eval_rewards/meter/mean': 0.689526264484112, 'eval_rewards/meter/std': 0.4130670749224149, 'eval_rewards/count_adherence/mean': 0.8135509353417617, 'eval_rewards/count_adherence/std': 0.17685549706220627, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8221937005336468, 'eval_rewards/repeat_penalty/std': 0.12831746328335542, 'eval_rewards/total_composite/mean': 0.47181554253284747, 'eval_rewards/total_composite/std': 0.3225762511675174, 'eval_reward': 0.47181554253284747, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.012653420428530527, 'eval_sampling/sampling_logp_difference/max': 0.8327025633591872, 'eval_sampling/importance_sampling_ratio/min': 0.4489339819321266, 'eval_sampling/importance_sampling_ratio/mean': 1.002467971581679, 'eval_sampling/importance_sampling_ratio/max': 1.3588372377248912, 'eval_entropy': 0.10610828204796864, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.47181554253284747, 'eval_reward_meter_mean': 0.689526264484112, 'eval_reward_meter_std': 0.4130670749224149, 'eval_reward_count_adherence_mean': 0.8135509353417617, 'eval_reward_count_adherence_std': 0.17685549706220627, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8221937005336468, 'eval_reward_repeat_penalty_std': 0.12831746328335542, 'eval_reward_total_composite_mean': 0.47181554253284747, 'eval_reward_total_composite_std': 0.3225762511675174, 'epoch': 0.05}

 38%|███▊      | 1250/3300 [1:15:18<2:57:00,  5.18s/it]

100%|██████████| 13/13 [00:51<00:00,  4.27s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 38%|███▊      | 1251/3300 [1:15:26<13:06:17, 23.02s/it]2026-04-11 23:52:17,273 | INFO | train_grpo_train | metrics_logged mode=train step=1251

                                                        
{'loss': 0.0013, 'grad_norm': 6.843512058258057, 'learning_rate': 6.212121212121213e-06, 'num_tokens': 2821356.0, 'completions/mean_length': 99.25, 'completions/min_length': 96.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.25, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9614334106445312, 'rewards/meter/std': 0.015211360529065132, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8895110487937927, 'rewards/total_composite/std': 0.10235293209552765, 'reward': 0.8895110487937927, 'reward_std': 0.10235292464494705, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039652157574892044, 'sampling/sampling_logp_difference/max': 3.9673361778259277, 'sampling/importance_sampling_ratio/min': 0.018923774361610413, 'sampling/importance_sampling_ratio/mean': 0.9989413022994995, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09751791087910533, 'clip_ratio/low_mean': 0.007605619612149894, 'clip_ratio/low_min': 0.007605619612149894, 'clip_ratio/high_mean': 0.01761134958360344, 'clip_ratio/high_max': 0.01761134958360344, 'clip_ratio/region_mean': 0.025216969195753336, 'reward_total_mean': 0.8895110487937927, 'reward_meter_mean': 0.9614334106445312, 'reward_meter_std': 0.015211360529065132, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8895110487937927, 'reward_total_composite_std': 0.10235293209552765, 'epoch': 0.05}

 38%|███▊      | 1251/3300 [1:15:26<13:06:17, 23.02s/it]INFO 04-11 23:52:17 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1252/3300 [1:15:32<10:16:54, 18.07s/it]2026-04-11 23:52:23,795 | INFO | train_grpo_train | metrics_logged mode=train step=1252

                                                        
{'loss': 0.0046, 'grad_norm': 1.9272445440292358, 'learning_rate': 6.209090909090909e-06, 'num_tokens': 2824023.0, 'completions/mean_length': 147.375, 'completions/min_length': 143.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.375, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9962889552116394, 'rewards/meter/std': 0.0008069492760114372, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.7294289469718933, 'rewards/total_composite/std': 0.050382453948259354, 'reward': 0.7294289469718933, 'reward_std': 0.050382472574710846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014175360091030598, 'sampling/sampling_logp_difference/max': 1.445974588394165, 'sampling/importance_sampling_ratio/min': 0.23551644384860992, 'sampling/importance_sampling_ratio/mean': 1.0037672519683838, 'sampling/importance_sampling_ratio/max': 1.5003139972686768, 'entropy': 0.09728506673127413, 'clip_ratio/low_mean': 0.01109143253415823, 'clip_ratio/low_min': 0.01109143253415823, 'clip_ratio/high_mean': 0.0025510203558951616, 'clip_ratio/high_max': 0.0025510203558951616, 'clip_ratio/region_mean': 0.013642452890053391, 'reward_total_mean': 0.7294289469718933, 'reward_meter_mean': 0.9962889552116394, 'reward_meter_std': 0.0008069492760114372, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.7294289469718933, 'reward_total_composite_std': 0.050382453948259354, 'epoch': 0.05}

 38%|███▊      | 1252/3300 [1:15:32<10:16:54, 18.07s/it]INFO 04-11 23:52:24 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1253/3300 [1:15:40<8:32:01, 15.01s/it] 2026-04-11 23:52:31,647 | INFO | train_grpo_train | metrics_logged mode=train step=1253

                                                       
{'loss': -0.0071, 'grad_norm': 2.051161050796509, 'learning_rate': 6.206060606060606e-06, 'num_tokens': 2828134.0, 'completions/mean_length': 290.875, 'completions/min_length': 269.0, 'completions/max_length': 313.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 290.875, 'completions/min_terminated_length': 269.0, 'completions/max_terminated_length': 313.0, 'rewards/meter/mean': 0.9930562376976013, 'rewards/meter/std': 0.01025536097586155, 'rewards/count_adherence/mean': 0.5576923489570618, 'rewards/count_adherence/std': 0.03560846298933029, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6976648569107056, 'rewards/repeat_penalty/std': 0.03868091478943825, 'rewards/total_composite/mean': 0.3865600824356079, 'rewards/total_composite/std': 0.0352591797709465, 'reward': 0.3865600824356079, 'reward_std': 0.0352591797709465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02420906163752079, 'sampling/sampling_logp_difference/max': 2.5028674602508545, 'sampling/importance_sampling_ratio/min': 0.11250852793455124, 'sampling/importance_sampling_ratio/mean': 0.9969112277030945, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09894186165183783, 'clip_ratio/low_mean': 0.005023152916692197, 'clip_ratio/low_min': 0.005023152916692197, 'clip_ratio/high_mean': 0.008442942867986858, 'clip_ratio/high_max': 0.008442942867986858, 'clip_ratio/region_mean': 0.013466095784679055, 'reward_total_mean': 0.3865600824356079, 'reward_meter_mean': 0.9930562376976013, 'reward_meter_std': 0.01025536097586155, 'reward_count_adherence_mean': 0.5576923489570618, 'reward_count_adherence_std': 0.03560846298933029, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6976648569107056, 'reward_repeat_penalty_std': 0.03868091478943825, 'reward_total_composite_mean': 0.3865600824356079, 'reward_total_composite_std': 0.0352591797709465, 'epoch': 0.05}

 38%|███▊      | 1253/3300 [1:15:40<8:32:01, 15.01s/it]INFO 04-11 23:52:31 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1254/3300 [1:15:45<6:49:49, 12.02s/it]2026-04-11 23:52:36,691 | INFO | train_grpo_train | metrics_logged mode=train step=1254

                                                       
{'loss': -0.0038, 'grad_norm': 4.324112892150879, 'learning_rate': 6.203030303030304e-06, 'num_tokens': 2829964.0, 'completions/mean_length': 67.75, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9599640965461731, 'rewards/meter/std': 0.03301013261079788, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9599640965461731, 'rewards/total_composite/std': 0.03301013261079788, 'reward': 0.9599640965461731, 'reward_std': 0.03301013633608818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016246721148490906, 'sampling/sampling_logp_difference/max': 2.111745834350586, 'sampling/importance_sampling_ratio/min': 0.12102648615837097, 'sampling/importance_sampling_ratio/mean': 1.0017342567443848, 'sampling/importance_sampling_ratio/max': 1.6136887073516846, 'entropy': 0.052088204305619, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.00916612590663135, 'clip_ratio/high_max': 0.00916612590663135, 'clip_ratio/region_mean': 0.011031797504983842, 'reward_total_mean': 0.9599640965461731, 'reward_meter_mean': 0.9599640965461731, 'reward_meter_std': 0.03301013261079788, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9599640965461731, 'reward_total_composite_std': 0.03301013261079788, 'epoch': 0.05}

 38%|███▊      | 1254/3300 [1:15:45<6:49:49, 12.02s/it]INFO 04-11 23:52:36 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1255/3300 [1:15:50<5:41:05, 10.01s/it]2026-04-11 23:52:42,005 | INFO | train_grpo_train | metrics_logged mode=train step=1255

                                                       
{'loss': -0.0, 'grad_norm': 4.520086765289307, 'learning_rate': 6.200000000000001e-06, 'num_tokens': 2832143.0, 'completions/mean_length': 95.375, 'completions/min_length': 92.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.375, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.6753093600273132, 'rewards/meter/std': 0.18275485932826996, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.5924432277679443, 'rewards/total_composite/std': 0.18791036307811737, 'reward': 0.5924432277679443, 'reward_std': 0.18791036307811737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036074861884117126, 'sampling/sampling_logp_difference/max': 2.096856117248535, 'sampling/importance_sampling_ratio/min': 0.12284202128648758, 'sampling/importance_sampling_ratio/mean': 1.0073623657226562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12418541312217712, 'clip_ratio/low_mean': 0.015763227711431682, 'clip_ratio/low_min': 0.015763227711431682, 'clip_ratio/high_mean': 0.015558094717562199, 'clip_ratio/high_max': 0.015558094717562199, 'clip_ratio/region_mean': 0.03132132242899388, 'reward_total_mean': 0.5924432277679443, 'reward_meter_mean': 0.6753093600273132, 'reward_meter_std': 0.18275485932826996, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.5924432277679443, 'reward_total_composite_std': 0.18791036307811737, 'epoch': 0.05}

 38%|███▊      | 1255/3300 [1:15:51<5:41:05, 10.01s/it]INFO 04-11 23:52:42 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1256/3300 [1:15:55<4:48:23,  8.47s/it]2026-04-11 23:52:46,873 | INFO | train_grpo_train | metrics_logged mode=train step=1256

                                                       
{'loss': 0.016, 'grad_norm': 4.7573018074035645, 'learning_rate': 6.196969696969698e-06, 'num_tokens': 2833894.0, 'completions/mean_length': 64.875, 'completions/min_length': 63.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.34324562549591064, 'rewards/meter/std': 0.38487473130226135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.34324562549591064, 'rewards/total_composite/std': 0.38487473130226135, 'reward': 0.34324562549591064, 'reward_std': 0.38487470149993896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06118984892964363, 'sampling/sampling_logp_difference/max': 2.886340618133545, 'sampling/importance_sampling_ratio/min': 0.05577996000647545, 'sampling/importance_sampling_ratio/mean': 1.0009076595306396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3142077159136534, 'clip_ratio/low_mean': 0.024938606191426516, 'clip_ratio/low_min': 0.024938606191426516, 'clip_ratio/high_mean': 0.015424644108861685, 'clip_ratio/high_max': 0.015424644108861685, 'clip_ratio/region_mean': 0.0403632503002882, 'reward_total_mean': 0.34324562549591064, 'reward_meter_mean': 0.34324562549591064, 'reward_meter_std': 0.38487473130226135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.34324562549591064, 'reward_total_composite_std': 0.38487473130226135, 'epoch': 0.05}

 38%|███▊      | 1256/3300 [1:15:55<4:48:23,  8.47s/it]INFO 04-11 23:52:47 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1257/3300 [1:16:00<4:08:23,  7.29s/it]2026-04-11 23:52:51,437 | INFO | train_grpo_train | metrics_logged mode=train step=1257

                                                       
{'loss': 0.0292, 'grad_norm': 11.342290878295898, 'learning_rate': 6.1939393939393944e-06, 'num_tokens': 2835400.0, 'completions/mean_length': 33.25, 'completions/min_length': 32.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.5455847978591919, 'rewards/meter/std': 0.46300008893013, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5455847978591919, 'rewards/total_composite/std': 0.46300008893013, 'reward': 0.5455847978591919, 'reward_std': 0.4630000591278076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05044575408101082, 'sampling/sampling_logp_difference/max': 0.9388256072998047, 'sampling/importance_sampling_ratio/min': 0.3910868465900421, 'sampling/importance_sampling_ratio/mean': 1.0081783533096313, 'sampling/importance_sampling_ratio/max': 1.4548543691635132, 'entropy': 0.2989681512117386, 'clip_ratio/low_mean': 0.03709893091581762, 'clip_ratio/low_min': 0.03709893091581762, 'clip_ratio/high_mean': 0.011482007801532745, 'clip_ratio/high_max': 0.011482007801532745, 'clip_ratio/region_mean': 0.048580938717350364, 'reward_total_mean': 0.5455847978591919, 'reward_meter_mean': 0.5455847978591919, 'reward_meter_std': 0.46300008893013, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5455847978591919, 'reward_total_composite_std': 0.46300008893013, 'epoch': 0.05}

 38%|███▊      | 1257/3300 [1:16:00<4:08:23,  7.29s/it]INFO 04-11 23:52:51 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1258/3300 [1:16:07<4:08:32,  7.30s/it]2026-04-11 23:52:58,758 | INFO | train_grpo_train | metrics_logged mode=train step=1258

                                                       
{'loss': 0.003, 'grad_norm': 1.4305702447891235, 'learning_rate': 6.190909090909092e-06, 'num_tokens': 2839026.0, 'completions/mean_length': 265.25, 'completions/min_length': 259.0, 'completions/max_length': 270.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 265.25, 'completions/min_terminated_length': 259.0, 'completions/max_terminated_length': 270.0, 'rewards/meter/mean': 0.9982133507728577, 'rewards/meter/std': 0.0003504555206745863, 'rewards/count_adherence/mean': 0.6363636255264282, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.701923131942749, 'rewards/repeat_penalty/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.4458819031715393, 'rewards/total_composite/std': 0.017329318448901176, 'reward': 0.4458819031715393, 'reward_std': 0.017329316586256027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011669746600091457, 'sampling/sampling_logp_difference/max': 1.233236312866211, 'sampling/importance_sampling_ratio/min': 0.2913481593132019, 'sampling/importance_sampling_ratio/mean': 1.0004147291183472, 'sampling/importance_sampling_ratio/max': 1.5016297101974487, 'entropy': 0.06276637688279152, 'clip_ratio/low_mean': 0.006089399073971435, 'clip_ratio/low_min': 0.006089399073971435, 'clip_ratio/high_mean': 0.0009433962404727936, 'clip_ratio/high_max': 0.0009433962404727936, 'clip_ratio/region_mean': 0.007032795314444229, 'reward_total_mean': 0.4458819031715393, 'reward_meter_mean': 0.9982133507728577, 'reward_meter_std': 0.0003504555206745863, 'reward_count_adherence_mean': 0.6363636255264282, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.701923131942749, 'reward_repeat_penalty_std': 0.027196412906050682, 'reward_total_composite_mean': 0.4458819031715393, 'reward_total_composite_std': 0.017329318448901176, 'epoch': 0.05}

 38%|███▊      | 1258/3300 [1:16:07<4:08:32,  7.30s/it]INFO 04-11 23:52:58 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1259/3300 [1:16:12<3:46:49,  6.67s/it]2026-04-11 23:53:03,944 | INFO | train_grpo_train | metrics_logged mode=train step=1259

                                                       
{'loss': 0.013, 'grad_norm': 3.326974630355835, 'learning_rate': 6.187878787878788e-06, 'num_tokens': 2841255.0, 'completions/mean_length': 107.625, 'completions/min_length': 102.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.625, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.5407921671867371, 'rewards/meter/std': 0.3177524209022522, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.925000011920929, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.4992244839668274, 'rewards/total_composite/std': 0.31581348180770874, 'reward': 0.4992244839668274, 'reward_std': 0.31581348180770874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04232693091034889, 'sampling/sampling_logp_difference/max': 1.4409921169281006, 'sampling/importance_sampling_ratio/min': 0.24779127538204193, 'sampling/importance_sampling_ratio/mean': 1.0075656175613403, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18995575420558453, 'clip_ratio/low_mean': 0.01721830479800701, 'clip_ratio/low_min': 0.01721830479800701, 'clip_ratio/high_mean': 0.015308146830648184, 'clip_ratio/high_max': 0.015308146830648184, 'clip_ratio/region_mean': 0.032526451628655195, 'reward_total_mean': 0.4992244839668274, 'reward_meter_mean': 0.5407921671867371, 'reward_meter_std': 0.3177524209022522, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.925000011920929, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.4992244839668274, 'reward_total_composite_std': 0.31581348180770874, 'epoch': 0.05}

 38%|███▊      | 1259/3300 [1:16:12<3:46:49,  6.67s/it]INFO 04-11 23:53:04 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1260/3300 [1:16:17<3:23:08,  5.97s/it]2026-04-11 23:53:08,301 | INFO | train_grpo_train | metrics_logged mode=train step=1260

                                                       
{'loss': 0.0127, 'grad_norm': 2.4279935359954834, 'learning_rate': 6.184848484848485e-06, 'num_tokens': 2842846.0, 'completions/mean_length': 40.875, 'completions/min_length': 38.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9946430921554565, 'rewards/meter/std': 0.006536941975355148, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946430921554565, 'rewards/total_composite/std': 0.006536941975355148, 'reward': 0.9946430921554565, 'reward_std': 0.00653693825006485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03152947872877121, 'sampling/sampling_logp_difference/max': 1.162613868713379, 'sampling/importance_sampling_ratio/min': 0.3126678168773651, 'sampling/importance_sampling_ratio/mean': 0.9979214072227478, 'sampling/importance_sampling_ratio/max': 1.3470659255981445, 'entropy': 0.174057693220675, 'clip_ratio/low_mean': 0.0029761905316263437, 'clip_ratio/low_min': 0.0029761905316263437, 'clip_ratio/high_mean': 0.028182906797155738, 'clip_ratio/high_max': 0.028182906797155738, 'clip_ratio/region_mean': 0.03115909732878208, 'reward_total_mean': 0.9946430921554565, 'reward_meter_mean': 0.9946430921554565, 'reward_meter_std': 0.006536941975355148, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946430921554565, 'reward_total_composite_std': 0.006536941975355148, 'epoch': 0.05}

 38%|███▊      | 1260/3300 [1:16:17<3:23:08,  5.97s/it]INFO 04-11 23:53:08 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1261/3300 [1:16:22<3:20:02,  5.89s/it]2026-04-11 23:53:13,980 | INFO | train_grpo_train | metrics_logged mode=train step=1261

                                                       
{'loss': 0.0029, 'grad_norm': 5.053354263305664, 'learning_rate': 6.181818181818182e-06, 'num_tokens': 2844913.0, 'completions/mean_length': 93.375, 'completions/min_length': 91.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.375, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.39358216524124146, 'rewards/meter/std': 0.3734220564365387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.3705918788909912, 'rewards/total_composite/std': 0.34063175320625305, 'reward': 0.3705918788909912, 'reward_std': 0.34063172340393066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.052313853055238724, 'sampling/sampling_logp_difference/max': 1.3460525274276733, 'sampling/importance_sampling_ratio/min': 0.26026561856269836, 'sampling/importance_sampling_ratio/mean': 1.0092734098434448, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34710839204490185, 'clip_ratio/low_mean': 0.024214978329837322, 'clip_ratio/low_min': 0.024214978329837322, 'clip_ratio/high_mean': 0.027944298926740885, 'clip_ratio/high_max': 0.027944298926740885, 'clip_ratio/region_mean': 0.05215927725657821, 'reward_total_mean': 0.3705918788909912, 'reward_meter_mean': 0.39358216524124146, 'reward_meter_std': 0.3734220564365387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.3705918788909912, 'reward_total_composite_std': 0.34063175320625305, 'epoch': 0.05}

 38%|███▊      | 1261/3300 [1:16:22<3:20:02,  5.89s/it]INFO 04-11 23:53:14 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1262/3300 [1:16:28<3:19:33,  5.87s/it]2026-04-11 23:53:19,832 | INFO | train_grpo_train | metrics_logged mode=train step=1262

                                                       
{'loss': 0.0561, 'grad_norm': 2.785076856613159, 'learning_rate': 6.17878787878788e-06, 'num_tokens': 2847577.0, 'completions/mean_length': 158.0, 'completions/min_length': 143.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.0, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.504827618598938, 'rewards/meter/std': 0.3096640408039093, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7850378751754761, 'rewards/repeat_penalty/std': 0.07399878650903702, 'rewards/total_composite/mean': 0.32526856660842896, 'rewards/total_composite/std': 0.17344774305820465, 'reward': 0.32526856660842896, 'reward_std': 0.17344774305820465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028311705216765404, 'sampling/sampling_logp_difference/max': 1.1957664489746094, 'sampling/importance_sampling_ratio/min': 0.3024720251560211, 'sampling/importance_sampling_ratio/mean': 0.9964918494224548, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0898361736908555, 'clip_ratio/low_mean': 0.010547005338594317, 'clip_ratio/low_min': 0.010547005338594317, 'clip_ratio/high_mean': 0.018799963407218456, 'clip_ratio/high_max': 0.018799963407218456, 'clip_ratio/region_mean': 0.029346968745812774, 'reward_total_mean': 0.32526856660842896, 'reward_meter_mean': 0.504827618598938, 'reward_meter_std': 0.3096640408039093, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7850378751754761, 'reward_repeat_penalty_std': 0.07399878650903702, 'reward_total_composite_mean': 0.32526856660842896, 'reward_total_composite_std': 0.17344774305820465, 'epoch': 0.05}

 38%|███▊      | 1262/3300 [1:16:28<3:19:33,  5.87s/it]INFO 04-11 23:53:20 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1263/3300 [1:16:36<3:34:11,  6.31s/it]2026-04-11 23:53:27,151 | INFO | train_grpo_train | metrics_logged mode=train step=1263

                                                       
{'loss': 0.0009, 'grad_norm': 1.8219730854034424, 'learning_rate': 6.175757575757576e-06, 'num_tokens': 2851398.0, 'completions/mean_length': 272.625, 'completions/min_length': 262.0, 'completions/max_length': 285.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 272.625, 'completions/min_terminated_length': 262.0, 'completions/max_terminated_length': 285.0, 'rewards/meter/mean': 0.9851720929145813, 'rewards/meter/std': 0.025058675557374954, 'rewards/count_adherence/mean': 0.6363636255264282, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7403846383094788, 'rewards/repeat_penalty/std': 0.07047118246555328, 'rewards/total_composite/mean': 0.4640815854072571, 'rewards/total_composite/std': 0.0445028655230999, 'reward': 0.4640815854072571, 'reward_std': 0.044502876698970795, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023550348356366158, 'sampling/sampling_logp_difference/max': 1.8511247634887695, 'sampling/importance_sampling_ratio/min': 0.1570604145526886, 'sampling/importance_sampling_ratio/mean': 0.9996957182884216, 'sampling/importance_sampling_ratio/max': 1.9002686738967896, 'entropy': 0.11504366714507341, 'clip_ratio/low_mean': 0.007349682127824053, 'clip_ratio/low_min': 0.007349682127824053, 'clip_ratio/high_mean': 0.006790850544348359, 'clip_ratio/high_max': 0.006790850544348359, 'clip_ratio/region_mean': 0.014140532672172412, 'reward_total_mean': 0.4640815854072571, 'reward_meter_mean': 0.9851720929145813, 'reward_meter_std': 0.025058675557374954, 'reward_count_adherence_mean': 0.6363636255264282, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7403846383094788, 'reward_repeat_penalty_std': 0.07047118246555328, 'reward_total_composite_mean': 0.4640815854072571, 'reward_total_composite_std': 0.0445028655230999, 'epoch': 0.05}

 38%|███▊      | 1263/3300 [1:16:36<3:34:11,  6.31s/it]INFO 04-11 23:53:27 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1264/3300 [1:16:40<3:16:27,  5.79s/it]2026-04-11 23:53:31,728 | INFO | train_grpo_train | metrics_logged mode=train step=1264

                                                       
{'loss': 0.0151, 'grad_norm': 6.026924133300781, 'learning_rate': 6.1727272727272735e-06, 'num_tokens': 2853090.0, 'completions/mean_length': 58.5, 'completions/min_length': 56.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.5, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.2589147090911865, 'rewards/meter/std': 0.21604736149311066, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.2589147090911865, 'rewards/total_composite/std': 0.21604736149311066, 'reward': 0.2589147090911865, 'reward_std': 0.21604733169078827, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039983682334423065, 'sampling/sampling_logp_difference/max': 0.9148893356323242, 'sampling/importance_sampling_ratio/min': 0.4005609452724457, 'sampling/importance_sampling_ratio/mean': 1.0070089101791382, 'sampling/importance_sampling_ratio/max': 1.932633399963379, 'entropy': 0.23071717843413353, 'clip_ratio/low_mean': 0.021268213167786598, 'clip_ratio/low_min': 0.021268213167786598, 'clip_ratio/high_mean': 0.00657894741743803, 'clip_ratio/high_max': 0.00657894741743803, 'clip_ratio/region_mean': 0.02784716058522463, 'reward_total_mean': 0.2589147090911865, 'reward_meter_mean': 0.2589147090911865, 'reward_meter_std': 0.21604736149311066, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.2589147090911865, 'reward_total_composite_std': 0.21604736149311066, 'epoch': 0.05}

 38%|███▊      | 1264/3300 [1:16:40<3:16:27,  5.79s/it]INFO 04-11 23:53:31 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1265/3300 [1:16:45<3:01:41,  5.36s/it]2026-04-11 23:53:36,075 | INFO | train_grpo_train | metrics_logged mode=train step=1265

                                                       
{'loss': 0.0293, 'grad_norm': 16.352861404418945, 'learning_rate': 6.16969696969697e-06, 'num_tokens': 2854666.0, 'completions/mean_length': 36.0, 'completions/min_length': 34.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8565147519111633, 'rewards/meter/std': 0.3440394103527069, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8565147519111633, 'rewards/total_composite/std': 0.3440394103527069, 'reward': 0.8565147519111633, 'reward_std': 0.3440393805503845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.057592298835515976, 'sampling/sampling_logp_difference/max': 1.1175026893615723, 'sampling/importance_sampling_ratio/min': 0.327095627784729, 'sampling/importance_sampling_ratio/mean': 0.9966188073158264, 'sampling/importance_sampling_ratio/max': 1.7478100061416626, 'entropy': 0.4582943990826607, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.05259183724410832, 'clip_ratio/high_max': 0.05259183724410832, 'clip_ratio/region_mean': 0.05597021570429206, 'reward_total_mean': 0.8565147519111633, 'reward_meter_mean': 0.8565147519111633, 'reward_meter_std': 0.3440394103527069, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8565147519111633, 'reward_total_composite_std': 0.3440394103527069, 'epoch': 0.05}

 38%|███▊      | 1265/3300 [1:16:45<3:01:41,  5.36s/it]INFO 04-11 23:53:36 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1266/3300 [1:16:49<2:56:06,  5.20s/it]2026-04-11 23:53:40,894 | INFO | train_grpo_train | metrics_logged mode=train step=1266

                                                       
{'loss': 0.0058, 'grad_norm': 3.7630727291107178, 'learning_rate': 6.166666666666667e-06, 'num_tokens': 2856589.0, 'completions/mean_length': 77.375, 'completions/min_length': 75.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.375, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9948470592498779, 'rewards/meter/std': 0.0029209780041128397, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948470592498779, 'rewards/total_composite/std': 0.0029209780041128397, 'reward': 0.9948470592498779, 'reward_std': 0.0029209787026047707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04110397398471832, 'sampling/sampling_logp_difference/max': 1.2458505630493164, 'sampling/importance_sampling_ratio/min': 0.28769612312316895, 'sampling/importance_sampling_ratio/mean': 1.0063776969909668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2823401764035225, 'clip_ratio/low_mean': 0.025986650260165334, 'clip_ratio/low_min': 0.025986650260165334, 'clip_ratio/high_mean': 0.016067266347818077, 'clip_ratio/high_max': 0.016067266347818077, 'clip_ratio/region_mean': 0.04205391660798341, 'reward_total_mean': 0.9948470592498779, 'reward_meter_mean': 0.9948470592498779, 'reward_meter_std': 0.0029209780041128397, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948470592498779, 'reward_total_composite_std': 0.0029209780041128397, 'epoch': 0.05}

 38%|███▊      | 1266/3300 [1:16:49<2:56:06,  5.20s/it]INFO 04-11 23:53:41 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1267/3300 [1:16:54<2:50:57,  5.05s/it]2026-04-11 23:53:45,590 | INFO | train_grpo_train | metrics_logged mode=train step=1267

                                                       
{'loss': 0.0264, 'grad_norm': 7.6448774337768555, 'learning_rate': 6.163636363636364e-06, 'num_tokens': 2858420.0, 'completions/mean_length': 63.875, 'completions/min_length': 58.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.6281806230545044, 'rewards/meter/std': 0.4045161306858063, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6281806230545044, 'rewards/total_composite/std': 0.4045161306858063, 'reward': 0.6281806230545044, 'reward_std': 0.4045161306858063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08960650116205215, 'sampling/sampling_logp_difference/max': 1.5403270721435547, 'sampling/importance_sampling_ratio/min': 0.21431100368499756, 'sampling/importance_sampling_ratio/mean': 1.0073941946029663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5581622421741486, 'clip_ratio/low_mean': 0.02490918291732669, 'clip_ratio/low_min': 0.02490918291732669, 'clip_ratio/high_mean': 0.035607312340289354, 'clip_ratio/high_max': 0.035607312340289354, 'clip_ratio/region_mean': 0.06051649525761604, 'reward_total_mean': 0.6281806230545044, 'reward_meter_mean': 0.6281806230545044, 'reward_meter_std': 0.4045161306858063, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6281806230545044, 'reward_total_composite_std': 0.4045161306858063, 'epoch': 0.05}

 38%|███▊      | 1267/3300 [1:16:54<2:50:57,  5.05s/it]INFO 04-11 23:53:45 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1268/3300 [1:16:59<2:48:35,  4.98s/it]2026-04-11 23:53:50,410 | INFO | train_grpo_train | metrics_logged mode=train step=1268

                                                       
{'loss': -0.008, 'grad_norm': 6.307591438293457, 'learning_rate': 6.160606060606062e-06, 'num_tokens': 2860190.0, 'completions/mean_length': 76.25, 'completions/min_length': 73.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9219943881034851, 'rewards/meter/std': 0.2038099318742752, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9219943881034851, 'rewards/total_composite/std': 0.2038099318742752, 'reward': 0.9219943881034851, 'reward_std': 0.2038099467754364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03656025603413582, 'sampling/sampling_logp_difference/max': 1.101435661315918, 'sampling/importance_sampling_ratio/min': 0.3323935270309448, 'sampling/importance_sampling_ratio/mean': 0.99973064661026, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1600959748029709, 'clip_ratio/low_mean': 0.0034246575087308884, 'clip_ratio/low_min': 0.0034246575087308884, 'clip_ratio/high_mean': 0.028985073906369507, 'clip_ratio/high_max': 0.028985073906369507, 'clip_ratio/region_mean': 0.032409731415100396, 'reward_total_mean': 0.9219943881034851, 'reward_meter_mean': 0.9219943881034851, 'reward_meter_std': 0.2038099318742752, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9219943881034851, 'reward_total_composite_std': 0.2038099318742752, 'epoch': 0.05}

 38%|███▊      | 1268/3300 [1:16:59<2:48:35,  4.98s/it]INFO 04-11 23:53:50 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1269/3300 [1:17:04<2:46:49,  4.93s/it]2026-04-11 23:53:55,223 | INFO | train_grpo_train | metrics_logged mode=train step=1269

                                                       
{'loss': 0.0401, 'grad_norm': 16.08408546447754, 'learning_rate': 6.157575757575758e-06, 'num_tokens': 2862222.0, 'completions/mean_length': 69.0, 'completions/min_length': 66.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.8981978893280029, 'rewards/meter/std': 0.22402629256248474, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8981978893280029, 'rewards/total_composite/std': 0.22402629256248474, 'reward': 0.8981978893280029, 'reward_std': 0.22402627766132355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03928679972887039, 'sampling/sampling_logp_difference/max': 2.344562530517578, 'sampling/importance_sampling_ratio/min': 0.0958891436457634, 'sampling/importance_sampling_ratio/mean': 0.9943932890892029, 'sampling/importance_sampling_ratio/max': 1.8017491102218628, 'entropy': 0.09085107641294599, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.034388539381325245, 'clip_ratio/high_max': 0.034388539381325245, 'clip_ratio/region_mean': 0.037766917841508985, 'reward_total_mean': 0.8981978893280029, 'reward_meter_mean': 0.8981978893280029, 'reward_meter_std': 0.22402629256248474, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8981978893280029, 'reward_total_composite_std': 0.22402629256248474, 'epoch': 0.05}

 38%|███▊      | 1269/3300 [1:17:04<2:46:49,  4.93s/it]INFO 04-11 23:53:55 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1270/3300 [1:17:12<3:25:16,  6.07s/it]2026-04-11 23:54:03,948 | INFO | train_grpo_train | metrics_logged mode=train step=1270

                                                       
{'loss': -0.0005, 'grad_norm': 1.1973329782485962, 'learning_rate': 6.154545454545455e-06, 'num_tokens': 2866664.0, 'completions/mean_length': 375.25, 'completions/min_length': 359.0, 'completions/max_length': 388.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 375.25, 'completions/min_terminated_length': 359.0, 'completions/max_terminated_length': 388.0, 'rewards/meter/mean': 0.9982415437698364, 'rewards/meter/std': 0.00024763745022937655, 'rewards/count_adherence/mean': 0.5263158082962036, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6842105388641357, 'rewards/repeat_penalty/std': 0.07443230599164963, 'rewards/total_composite/mean': 0.35947954654693604, 'rewards/total_composite/std': 0.039129648357629776, 'reward': 0.35947954654693604, 'reward_std': 0.039129648357629776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01408437080681324, 'sampling/sampling_logp_difference/max': 1.9516000747680664, 'sampling/importance_sampling_ratio/min': 0.14204660058021545, 'sampling/importance_sampling_ratio/mean': 1.0016210079193115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07593974284827709, 'clip_ratio/low_mean': 0.004985210340237245, 'clip_ratio/low_min': 0.004985210340237245, 'clip_ratio/high_mean': 0.006335443118587136, 'clip_ratio/high_max': 0.006335443118587136, 'clip_ratio/region_mean': 0.011320653458824381, 'reward_total_mean': 0.35947954654693604, 'reward_meter_mean': 0.9982415437698364, 'reward_meter_std': 0.00024763745022937655, 'reward_count_adherence_mean': 0.5263158082962036, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6842105388641357, 'reward_repeat_penalty_std': 0.07443230599164963, 'reward_total_composite_mean': 0.35947954654693604, 'reward_total_composite_std': 0.039129648357629776, 'epoch': 0.05}

 38%|███▊      | 1270/3300 [1:17:12<3:25:16,  6.07s/it]INFO 04-11 23:54:04 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1271/3300 [1:17:17<3:12:44,  5.70s/it]2026-04-11 23:54:08,790 | INFO | train_grpo_train | metrics_logged mode=train step=1271

                                                       
{'loss': 0.0134, 'grad_norm': 3.943135976791382, 'learning_rate': 6.151515151515152e-06, 'num_tokens': 2868449.0, 'completions/mean_length': 76.125, 'completions/min_length': 75.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.125, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9946237802505493, 'rewards/meter/std': 0.0009093984263017774, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946237802505493, 'rewards/total_composite/std': 0.0009093984263017774, 'reward': 0.9946237802505493, 'reward_std': 0.0009094090783037245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02133559249341488, 'sampling/sampling_logp_difference/max': 0.8648512959480286, 'sampling/importance_sampling_ratio/min': 0.4211141765117645, 'sampling/importance_sampling_ratio/mean': 1.0047322511672974, 'sampling/importance_sampling_ratio/max': 1.8516945838928223, 'entropy': 0.13653591368347406, 'clip_ratio/low_mean': 0.0032056551426649094, 'clip_ratio/low_min': 0.0032056551426649094, 'clip_ratio/high_mean': 0.00831140368245542, 'clip_ratio/high_max': 0.00831140368245542, 'clip_ratio/region_mean': 0.01151705882512033, 'reward_total_mean': 0.9946237802505493, 'reward_meter_mean': 0.9946237802505493, 'reward_meter_std': 0.0009093984263017774, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946237802505493, 'reward_total_composite_std': 0.0009093984263017774, 'epoch': 0.05}

 39%|███▊      | 1271/3300 [1:17:17<3:12:44,  5.70s/it]INFO 04-11 23:54:09 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1272/3300 [1:17:22<3:02:49,  5.41s/it]2026-04-11 23:54:13,520 | INFO | train_grpo_train | metrics_logged mode=train step=1272

                                                       
{'loss': 0.0295, 'grad_norm': 5.4271039962768555, 'learning_rate': 6.148484848484849e-06, 'num_tokens': 2870337.0, 'completions/mean_length': 67.0, 'completions/min_length': 63.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7235432863235474, 'rewards/meter/std': 0.3634967803955078, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7235432863235474, 'rewards/total_composite/std': 0.3634967803955078, 'reward': 0.7235432863235474, 'reward_std': 0.3634967505931854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06481360644102097, 'sampling/sampling_logp_difference/max': 3.7088615894317627, 'sampling/importance_sampling_ratio/min': 0.024505402892827988, 'sampling/importance_sampling_ratio/mean': 0.9991426467895508, 'sampling/importance_sampling_ratio/max': 1.7601451873779297, 'entropy': 0.26362900249660015, 'clip_ratio/low_mean': 0.010900080436840653, 'clip_ratio/low_min': 0.010900080436840653, 'clip_ratio/high_mean': 0.036419710610061884, 'clip_ratio/high_max': 0.036419710610061884, 'clip_ratio/region_mean': 0.04731979104690254, 'reward_total_mean': 0.7235432863235474, 'reward_meter_mean': 0.7235432863235474, 'reward_meter_std': 0.3634967803955078, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7235432863235474, 'reward_total_composite_std': 0.3634967803955078, 'epoch': 0.05}

 39%|███▊      | 1272/3300 [1:17:22<3:02:49,  5.41s/it]INFO 04-11 23:54:13 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1273/3300 [1:17:28<3:05:34,  5.49s/it]2026-04-11 23:54:19,210 | INFO | train_grpo_train | metrics_logged mode=train step=1273

                                                       
{'loss': 0.0059, 'grad_norm': 3.146674394607544, 'learning_rate': 6.1454545454545454e-06, 'num_tokens': 2872992.0, 'completions/mean_length': 152.875, 'completions/min_length': 149.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 152.875, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.9979577660560608, 'rewards/meter/std': 0.0006660494254902005, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.855392336845398, 'rewards/total_composite/std': 0.0005708994576707482, 'reward': 0.855392336845398, 'reward_std': 0.0005708927637897432, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01988840475678444, 'sampling/sampling_logp_difference/max': 1.3840652704238892, 'sampling/importance_sampling_ratio/min': 0.25055789947509766, 'sampling/importance_sampling_ratio/mean': 0.9994745850563049, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10434958059340715, 'clip_ratio/low_mean': 0.004043337190523744, 'clip_ratio/low_min': 0.004043337190523744, 'clip_ratio/high_mean': 0.00823544233571738, 'clip_ratio/high_max': 0.00823544233571738, 'clip_ratio/region_mean': 0.012278779526241124, 'reward_total_mean': 0.855392336845398, 'reward_meter_mean': 0.9979577660560608, 'reward_meter_std': 0.0006660494254902005, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.855392336845398, 'reward_total_composite_std': 0.0005708994576707482, 'epoch': 0.05}

 39%|███▊      | 1273/3300 [1:17:28<3:05:34,  5.49s/it]INFO 04-11 23:54:19 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1274/3300 [1:17:32<2:54:10,  5.16s/it]2026-04-11 23:54:23,586 | INFO | train_grpo_train | metrics_logged mode=train step=1274

                                                       
{'loss': -0.0028, 'grad_norm': 9.43035888671875, 'learning_rate': 6.142424242424243e-06, 'num_tokens': 2874585.0, 'completions/mean_length': 42.125, 'completions/min_length': 42.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9962718486785889, 'rewards/meter/std': 0.0009969644015654922, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9962718486785889, 'rewards/total_composite/std': 0.0009969644015654922, 'reward': 0.9962718486785889, 'reward_std': 0.0009969675447791815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02699565887451172, 'sampling/sampling_logp_difference/max': 0.7795038223266602, 'sampling/importance_sampling_ratio/min': 0.547152042388916, 'sampling/importance_sampling_ratio/mean': 1.0059658288955688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1509600691497326, 'clip_ratio/low_mean': 0.0029761905316263437, 'clip_ratio/low_min': 0.0029761905316263437, 'clip_ratio/high_mean': 0.008859357796609402, 'clip_ratio/high_max': 0.008859357796609402, 'clip_ratio/region_mean': 0.011835548328235745, 'reward_total_mean': 0.9962718486785889, 'reward_meter_mean': 0.9962718486785889, 'reward_meter_std': 0.0009969644015654922, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9962718486785889, 'reward_total_composite_std': 0.0009969644015654922, 'epoch': 0.05}

 39%|███▊      | 1274/3300 [1:17:32<2:54:10,  5.16s/it]INFO 04-11 23:54:23 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1275/3300 [1:17:38<2:57:41,  5.26s/it]2026-04-11 23:54:29,099 | INFO | train_grpo_train | metrics_logged mode=train step=1275

                                                       
{'loss': -0.0071, 'grad_norm': 3.5774126052856445, 'learning_rate': 6.139393939393939e-06, 'num_tokens': 2877026.0, 'completions/mean_length': 128.125, 'completions/min_length': 121.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.125, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9509094953536987, 'rewards/meter/std': 0.036813125014305115, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8314633369445801, 'rewards/total_composite/std': 0.04581277072429657, 'reward': 0.8314633369445801, 'reward_std': 0.04581277072429657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035852231085300446, 'sampling/sampling_logp_difference/max': 1.1192617416381836, 'sampling/importance_sampling_ratio/min': 0.32652077078819275, 'sampling/importance_sampling_ratio/mean': 1.0062079429626465, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20628871954977512, 'clip_ratio/low_mean': 0.007767904899083078, 'clip_ratio/low_min': 0.007767904899083078, 'clip_ratio/high_mean': 0.012963371758814901, 'clip_ratio/high_max': 0.012963371758814901, 'clip_ratio/region_mean': 0.02073127665789798, 'reward_total_mean': 0.8314633369445801, 'reward_meter_mean': 0.9509094953536987, 'reward_meter_std': 0.036813125014305115, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8314633369445801, 'reward_total_composite_std': 0.04581277072429657, 'epoch': 0.05}

 39%|███▊      | 1275/3300 [1:17:38<2:57:41,  5.26s/it]INFO 04-11 23:54:29 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1276/3300 [1:17:43<2:55:13,  5.19s/it]2026-04-11 23:54:34,130 | INFO | train_grpo_train | metrics_logged mode=train step=1276

                                                       
{'loss': 0.0392, 'grad_norm': 7.507855415344238, 'learning_rate': 6.136363636363637e-06, 'num_tokens': 2879100.0, 'completions/mean_length': 89.25, 'completions/min_length': 84.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.25, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.414013534784317, 'rewards/meter/std': 0.36701732873916626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.3967258036136627, 'rewards/total_composite/std': 0.3643990457057953, 'reward': 0.3967258036136627, 'reward_std': 0.3643990457057953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09024675190448761, 'sampling/sampling_logp_difference/max': 1.623225212097168, 'sampling/importance_sampling_ratio/min': 0.19726145267486572, 'sampling/importance_sampling_ratio/mean': 1.0259032249450684, 'sampling/importance_sampling_ratio/max': 1.9257968664169312, 'entropy': 0.8716921396553516, 'clip_ratio/low_mean': 0.0324383438564837, 'clip_ratio/low_min': 0.0324383438564837, 'clip_ratio/high_mean': 0.028633129317313433, 'clip_ratio/high_max': 0.028633129317313433, 'clip_ratio/region_mean': 0.06107147317379713, 'reward_total_mean': 0.3967258036136627, 'reward_meter_mean': 0.414013534784317, 'reward_meter_std': 0.36701732873916626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.3967258036136627, 'reward_total_composite_std': 0.3643990457057953, 'epoch': 0.05}

 39%|███▊      | 1276/3300 [1:17:43<2:55:13,  5.19s/it]INFO 04-11 23:54:34 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1277/3300 [1:17:49<3:04:15,  5.46s/it]2026-04-11 23:54:40,227 | INFO | train_grpo_train | metrics_logged mode=train step=1277

                                                       
{'loss': 0.0264, 'grad_norm': 5.259714126586914, 'learning_rate': 6.133333333333334e-06, 'num_tokens': 2881970.0, 'completions/mean_length': 177.75, 'completions/min_length': 170.0, 'completions/max_length': 184.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 177.75, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 184.0, 'rewards/meter/mean': 0.8953356742858887, 'rewards/meter/std': 0.2748320698738098, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.05143444612622261, 'rewards/total_composite/mean': 0.6403263807296753, 'rewards/total_composite/std': 0.19905686378479004, 'reward': 0.6403263807296753, 'reward_std': 0.19905687868595123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04297341778874397, 'sampling/sampling_logp_difference/max': 1.2142763137817383, 'sampling/importance_sampling_ratio/min': 0.2969248294830322, 'sampling/importance_sampling_ratio/mean': 1.0098189115524292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3461093604564667, 'clip_ratio/low_mean': 0.012976306956261396, 'clip_ratio/low_min': 0.012976306956261396, 'clip_ratio/high_mean': 0.02701578661799431, 'clip_ratio/high_max': 0.02701578661799431, 'clip_ratio/region_mean': 0.039992093574255705, 'reward_total_mean': 0.6403263807296753, 'reward_meter_mean': 0.8953356742858887, 'reward_meter_std': 0.2748320698738098, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.05143444612622261, 'reward_total_composite_mean': 0.6403263807296753, 'reward_total_composite_std': 0.19905686378479004, 'epoch': 0.05}

 39%|███▊      | 1277/3300 [1:17:49<3:04:15,  5.46s/it]INFO 04-11 23:54:40 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1278/3300 [1:17:53<2:55:05,  5.20s/it]2026-04-11 23:54:44,794 | INFO | train_grpo_train | metrics_logged mode=train step=1278

                                                       
{'loss': 0.0461, 'grad_norm': 10.583206176757812, 'learning_rate': 6.130303030303031e-06, 'num_tokens': 2883696.0, 'completions/mean_length': 48.75, 'completions/min_length': 42.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8016507625579834, 'rewards/meter/std': 0.21309679746627808, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8016507625579834, 'rewards/total_composite/std': 0.21309679746627808, 'reward': 0.8016507625579834, 'reward_std': 0.21309679746627808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07192375510931015, 'sampling/sampling_logp_difference/max': 1.8451597690582275, 'sampling/importance_sampling_ratio/min': 0.15800006687641144, 'sampling/importance_sampling_ratio/mean': 0.992801308631897, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2942120973020792, 'clip_ratio/low_mean': 0.012557290028780699, 'clip_ratio/low_min': 0.012557290028780699, 'clip_ratio/high_mean': 0.054509096313267946, 'clip_ratio/high_max': 0.054509096313267946, 'clip_ratio/region_mean': 0.06706638634204865, 'reward_total_mean': 0.8016507625579834, 'reward_meter_mean': 0.8016507625579834, 'reward_meter_std': 0.21309679746627808, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8016507625579834, 'reward_total_composite_std': 0.21309679746627808, 'epoch': 0.05}

 39%|███▊      | 1278/3300 [1:17:53<2:55:05,  5.20s/it]INFO 04-11 23:54:45 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1279/3300 [1:17:58<2:53:34,  5.15s/it]2026-04-11 23:54:49,847 | INFO | train_grpo_train | metrics_logged mode=train step=1279

                                                       
{'loss': -0.028, 'grad_norm': 4.000077247619629, 'learning_rate': 6.127272727272727e-06, 'num_tokens': 2885855.0, 'completions/mean_length': 86.875, 'completions/min_length': 79.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.875, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.7521673440933228, 'rewards/meter/std': 0.3259941041469574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.1511857807636261, 'rewards/total_composite/mean': 0.625595211982727, 'rewards/total_composite/std': 0.32554712891578674, 'reward': 0.625595211982727, 'reward_std': 0.32554712891578674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03764592856168747, 'sampling/sampling_logp_difference/max': 1.77433443069458, 'sampling/importance_sampling_ratio/min': 0.45753392577171326, 'sampling/importance_sampling_ratio/mean': 1.005535364151001, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20182008855044842, 'clip_ratio/low_mean': 0.010369318537414074, 'clip_ratio/low_min': 0.010369318537414074, 'clip_ratio/high_mean': 0.0231478811474517, 'clip_ratio/high_max': 0.0231478811474517, 'clip_ratio/region_mean': 0.03351719968486577, 'reward_total_mean': 0.625595211982727, 'reward_meter_mean': 0.7521673440933228, 'reward_meter_std': 0.3259941041469574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.1511857807636261, 'reward_total_composite_mean': 0.625595211982727, 'reward_total_composite_std': 0.32554712891578674, 'epoch': 0.05}

 39%|███▉      | 1279/3300 [1:17:58<2:53:34,  5.15s/it]INFO 04-11 23:54:50 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1280/3300 [1:18:03<2:53:48,  5.16s/it]2026-04-11 23:54:55,035 | INFO | train_grpo_train | metrics_logged mode=train step=1280

                                                       
{'loss': 0.0056, 'grad_norm': 5.375721454620361, 'learning_rate': 6.1242424242424245e-06, 'num_tokens': 2888100.0, 'completions/mean_length': 110.625, 'completions/min_length': 107.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.625, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.32401683926582336, 'rewards/meter/std': 0.4113844931125641, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.3231015205383301, 'rewards/total_composite/std': 0.41212281584739685, 'reward': 0.3231015205383301, 'reward_std': 0.41212278604507446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07019690424203873, 'sampling/sampling_logp_difference/max': 4.393073081970215, 'sampling/importance_sampling_ratio/min': 0.012362679466605186, 'sampling/importance_sampling_ratio/mean': 1.0015076398849487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3459767811000347, 'clip_ratio/low_mean': 0.035150347743183374, 'clip_ratio/low_min': 0.035150347743183374, 'clip_ratio/high_mean': 0.020130750257521868, 'clip_ratio/high_max': 0.020130750257521868, 'clip_ratio/region_mean': 0.05528109800070524, 'reward_total_mean': 0.3231015205383301, 'reward_meter_mean': 0.32401683926582336, 'reward_meter_std': 0.4113844931125641, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.3231015205383301, 'reward_total_composite_std': 0.41212281584739685, 'epoch': 0.05}

 39%|███▉      | 1280/3300 [1:18:04<2:53:48,  5.16s/it]INFO 04-11 23:54:55 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1281/3300 [1:18:08<2:49:36,  5.04s/it]2026-04-11 23:54:59,791 | INFO | train_grpo_train | metrics_logged mode=train step=1281

                                                       
{'loss': 0.0157, 'grad_norm': 7.6968994140625, 'learning_rate': 6.121212121212121e-06, 'num_tokens': 2889991.0, 'completions/mean_length': 70.375, 'completions/min_length': 69.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9908488988876343, 'rewards/meter/std': 0.0105263227596879, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9908488988876343, 'rewards/total_composite/std': 0.0105263227596879, 'reward': 0.9908488988876343, 'reward_std': 0.010526325553655624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044087208807468414, 'sampling/sampling_logp_difference/max': 1.6382884979248047, 'sampling/importance_sampling_ratio/min': 0.19431231915950775, 'sampling/importance_sampling_ratio/mean': 0.9948356747627258, 'sampling/importance_sampling_ratio/max': 1.7191059589385986, 'entropy': 0.2260036300867796, 'clip_ratio/low_mean': 0.005307539715431631, 'clip_ratio/low_min': 0.005307539715431631, 'clip_ratio/high_mean': 0.03385740250814706, 'clip_ratio/high_max': 0.03385740250814706, 'clip_ratio/region_mean': 0.03916494222357869, 'reward_total_mean': 0.9908488988876343, 'reward_meter_mean': 0.9908488988876343, 'reward_meter_std': 0.0105263227596879, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9908488988876343, 'reward_total_composite_std': 0.0105263227596879, 'epoch': 0.05}

 39%|███▉      | 1281/3300 [1:18:08<2:49:36,  5.04s/it]INFO 04-11 23:55:00 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1282/3300 [1:18:16<3:16:08,  5.83s/it]2026-04-11 23:55:07,468 | INFO | train_grpo_train | metrics_logged mode=train step=1282

                                                       
{'loss': 0.0068, 'grad_norm': 2.044459581375122, 'learning_rate': 6.118181818181819e-06, 'num_tokens': 2894526.0, 'completions/mean_length': 303.875, 'completions/min_length': 290.0, 'completions/max_length': 311.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 303.875, 'completions/min_terminated_length': 290.0, 'completions/max_terminated_length': 311.0, 'rewards/meter/mean': 0.9781317114830017, 'rewards/meter/std': 0.027103744447231293, 'rewards/count_adherence/mean': 0.5714285969734192, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7749999761581421, 'rewards/repeat_penalty/std': 0.11233452707529068, 'rewards/total_composite/mean': 0.43306678533554077, 'rewards/total_composite/std': 0.0634385421872139, 'reward': 0.43306678533554077, 'reward_std': 0.0634385421872139, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03700267896056175, 'sampling/sampling_logp_difference/max': 3.257213592529297, 'sampling/importance_sampling_ratio/min': 0.0384955108165741, 'sampling/importance_sampling_ratio/mean': 1.0035651922225952, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20566194783896208, 'clip_ratio/low_mean': 0.010263397532980889, 'clip_ratio/low_min': 0.010263397532980889, 'clip_ratio/high_mean': 0.013975600013509393, 'clip_ratio/high_max': 0.013975600013509393, 'clip_ratio/region_mean': 0.024238997546490282, 'reward_total_mean': 0.43306678533554077, 'reward_meter_mean': 0.9781317114830017, 'reward_meter_std': 0.027103744447231293, 'reward_count_adherence_mean': 0.5714285969734192, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7749999761581421, 'reward_repeat_penalty_std': 0.11233452707529068, 'reward_total_composite_mean': 0.43306678533554077, 'reward_total_composite_std': 0.0634385421872139, 'epoch': 0.05}

 39%|███▉      | 1282/3300 [1:18:16<3:16:08,  5.83s/it]INFO 04-11 23:55:07 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1283/3300 [1:18:21<3:06:37,  5.55s/it]2026-04-11 23:55:12,363 | INFO | train_grpo_train | metrics_logged mode=train step=1283

                                                       
{'loss': 0.0097, 'grad_norm': 5.159550666809082, 'learning_rate': 6.115151515151516e-06, 'num_tokens': 2896475.0, 'completions/mean_length': 76.625, 'completions/min_length': 72.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9953988790512085, 'rewards/meter/std': 0.0025027033407241106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953988790512085, 'rewards/total_composite/std': 0.0025027033407241106, 'reward': 0.9953988790512085, 'reward_std': 0.0025027082301676273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.053737834095954895, 'sampling/sampling_logp_difference/max': 1.5215466022491455, 'sampling/importance_sampling_ratio/min': 0.21837389469146729, 'sampling/importance_sampling_ratio/mean': 1.0098578929901123, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43742145597934723, 'clip_ratio/low_mean': 0.009698634734377265, 'clip_ratio/low_min': 0.009698634734377265, 'clip_ratio/high_mean': 0.030571716954000294, 'clip_ratio/high_max': 0.030571716954000294, 'clip_ratio/region_mean': 0.04027035168837756, 'reward_total_mean': 0.9953988790512085, 'reward_meter_mean': 0.9953988790512085, 'reward_meter_std': 0.0025027033407241106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953988790512085, 'reward_total_composite_std': 0.0025027033407241106, 'epoch': 0.05}

 39%|███▉      | 1283/3300 [1:18:21<3:06:37,  5.55s/it]INFO 04-11 23:55:12 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1284/3300 [1:18:25<2:55:02,  5.21s/it]2026-04-11 23:55:16,775 | INFO | train_grpo_train | metrics_logged mode=train step=1284

                                                       
{'loss': -0.0204, 'grad_norm': 8.960368156433105, 'learning_rate': 6.112121212121213e-06, 'num_tokens': 2897859.0, 'completions/mean_length': 36.0, 'completions/min_length': 34.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7910915613174438, 'rewards/meter/std': 0.36750105023384094, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7910915613174438, 'rewards/total_composite/std': 0.36750105023384094, 'reward': 0.7910915613174438, 'reward_std': 0.36750105023384094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06175791099667549, 'sampling/sampling_logp_difference/max': 1.3975987434387207, 'sampling/importance_sampling_ratio/min': 0.24718981981277466, 'sampling/importance_sampling_ratio/mean': 0.996502161026001, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31284985691308975, 'clip_ratio/low_mean': 0.03277311008423567, 'clip_ratio/low_min': 0.03277311008423567, 'clip_ratio/high_mean': 0.020760516403242946, 'clip_ratio/high_max': 0.020760516403242946, 'clip_ratio/region_mean': 0.053533626487478614, 'reward_total_mean': 0.7910915613174438, 'reward_meter_mean': 0.7910915613174438, 'reward_meter_std': 0.36750105023384094, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7910915613174438, 'reward_total_composite_std': 0.36750105023384094, 'epoch': 0.05}

 39%|███▉      | 1284/3300 [1:18:25<2:55:02,  5.21s/it]INFO 04-11 23:55:17 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1285/3300 [1:18:31<2:58:12,  5.31s/it]2026-04-11 23:55:22,307 | INFO | train_grpo_train | metrics_logged mode=train step=1285

                                                       
{'loss': -0.0018, 'grad_norm': 4.468361854553223, 'learning_rate': 6.10909090909091e-06, 'num_tokens': 2900285.0, 'completions/mean_length': 133.25, 'completions/min_length': 125.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.25, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.7679275274276733, 'rewards/meter/std': 0.3219233751296997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.7135751247406006, 'rewards/total_composite/std': 0.3173206150531769, 'reward': 0.7135751247406006, 'reward_std': 0.31732064485549927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04117295891046524, 'sampling/sampling_logp_difference/max': 1.2831401824951172, 'sampling/importance_sampling_ratio/min': 0.27716556191444397, 'sampling/importance_sampling_ratio/mean': 1.006422996520996, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.272802772000432, 'clip_ratio/low_mean': 0.013498181011527777, 'clip_ratio/low_min': 0.013498181011527777, 'clip_ratio/high_mean': 0.01694088790100068, 'clip_ratio/high_max': 0.01694088790100068, 'clip_ratio/region_mean': 0.030439068912528455, 'reward_total_mean': 0.7135751247406006, 'reward_meter_mean': 0.7679275274276733, 'reward_meter_std': 0.3219233751296997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.7135751247406006, 'reward_total_composite_std': 0.3173206150531769, 'epoch': 0.05}

 39%|███▉      | 1285/3300 [1:18:31<2:58:12,  5.31s/it]INFO 04-11 23:55:22 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1286/3300 [1:18:38<3:12:40,  5.74s/it]2026-04-11 23:55:29,062 | INFO | train_grpo_train | metrics_logged mode=train step=1286

                                                       
{'loss': 0.0163, 'grad_norm': 2.972513437271118, 'learning_rate': 6.106060606060606e-06, 'num_tokens': 2903604.0, 'completions/mean_length': 219.875, 'completions/min_length': 204.0, 'completions/max_length': 233.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 219.875, 'completions/min_terminated_length': 204.0, 'completions/max_terminated_length': 233.0, 'rewards/meter/mean': 0.8396387100219727, 'rewards/meter/std': 0.2860766649246216, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8863636255264282, 'rewards/repeat_penalty/std': 0.06428244709968567, 'rewards/total_composite/mean': 0.6425139904022217, 'rewards/total_composite/std': 0.22964133322238922, 'reward': 0.6425139904022217, 'reward_std': 0.22964133322238922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0455651730298996, 'sampling/sampling_logp_difference/max': 1.4966444969177246, 'sampling/importance_sampling_ratio/min': 0.26713767647743225, 'sampling/importance_sampling_ratio/mean': 1.0046546459197998, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3109611961990595, 'clip_ratio/low_mean': 0.007508100010454655, 'clip_ratio/low_min': 0.007508100010454655, 'clip_ratio/high_mean': 0.027195949805900455, 'clip_ratio/high_max': 0.027195949805900455, 'clip_ratio/region_mean': 0.03470404981635511, 'reward_total_mean': 0.6425139904022217, 'reward_meter_mean': 0.8396387100219727, 'reward_meter_std': 0.2860766649246216, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8863636255264282, 'reward_repeat_penalty_std': 0.06428244709968567, 'reward_total_composite_mean': 0.6425139904022217, 'reward_total_composite_std': 0.22964133322238922, 'epoch': 0.05}

 39%|███▉      | 1286/3300 [1:18:38<3:12:40,  5.74s/it]INFO 04-11 23:55:29 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1287/3300 [1:18:44<3:18:32,  5.92s/it]2026-04-11 23:55:35,396 | INFO | train_grpo_train | metrics_logged mode=train step=1287

                                                       
{'loss': -0.0036, 'grad_norm': 3.1675779819488525, 'learning_rate': 6.103030303030304e-06, 'num_tokens': 2906858.0, 'completions/mean_length': 189.75, 'completions/min_length': 182.0, 'completions/max_length': 199.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 189.75, 'completions/min_terminated_length': 182.0, 'completions/max_terminated_length': 199.0, 'rewards/meter/mean': 0.9928793907165527, 'rewards/meter/std': 0.01269504800438881, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.09848947077989578, 'rewards/total_composite/mean': 0.7119789123535156, 'rewards/total_composite/std': 0.07652789354324341, 'reward': 0.7119789123535156, 'reward_std': 0.0765279084444046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0419112965464592, 'sampling/sampling_logp_difference/max': 3.135072708129883, 'sampling/importance_sampling_ratio/min': 0.04349659010767937, 'sampling/importance_sampling_ratio/mean': 1.0034816265106201, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22143690288066864, 'clip_ratio/low_mean': 0.007304175465833396, 'clip_ratio/low_min': 0.007304175465833396, 'clip_ratio/high_mean': 0.020334920845925808, 'clip_ratio/high_max': 0.020334920845925808, 'clip_ratio/region_mean': 0.027639096311759204, 'reward_total_mean': 0.7119789123535156, 'reward_meter_mean': 0.9928793907165527, 'reward_meter_std': 0.01269504800438881, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.09848947077989578, 'reward_total_composite_mean': 0.7119789123535156, 'reward_total_composite_std': 0.07652789354324341, 'epoch': 0.05}

 39%|███▉      | 1287/3300 [1:18:44<3:18:32,  5.92s/it]INFO 04-11 23:55:35 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1288/3300 [1:18:51<3:31:38,  6.31s/it]2026-04-11 23:55:42,622 | INFO | train_grpo_train | metrics_logged mode=train step=1288

                                                       
{'loss': 0.0056, 'grad_norm': 1.9611791372299194, 'learning_rate': 6.1e-06, 'num_tokens': 2910730.0, 'completions/mean_length': 272.0, 'completions/min_length': 269.0, 'completions/max_length': 275.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 272.0, 'completions/min_terminated_length': 269.0, 'completions/max_terminated_length': 275.0, 'rewards/meter/mean': 0.9904348254203796, 'rewards/meter/std': 0.015252761542797089, 'rewards/count_adherence/mean': 0.7777777910232544, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8365384340286255, 'rewards/repeat_penalty/std': 0.08661473542451859, 'rewards/total_composite/mean': 0.6442359685897827, 'rewards/total_composite/std': 0.06582249701023102, 'reward': 0.6442359685897827, 'reward_std': 0.06582249701023102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025431334972381592, 'sampling/sampling_logp_difference/max': 2.2604598999023438, 'sampling/importance_sampling_ratio/min': 0.10430250316858292, 'sampling/importance_sampling_ratio/mean': 1.0058023929595947, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.139354033395648, 'clip_ratio/low_mean': 0.008273915416793898, 'clip_ratio/low_min': 0.008273915416793898, 'clip_ratio/high_mean': 0.00875260157044977, 'clip_ratio/high_max': 0.00875260157044977, 'clip_ratio/region_mean': 0.017026516987243667, 'reward_total_mean': 0.6442359685897827, 'reward_meter_mean': 0.9904348254203796, 'reward_meter_std': 0.015252761542797089, 'reward_count_adherence_mean': 0.7777777910232544, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8365384340286255, 'reward_repeat_penalty_std': 0.08661473542451859, 'reward_total_composite_mean': 0.6442359685897827, 'reward_total_composite_std': 0.06582249701023102, 'epoch': 0.05}

 39%|███▉      | 1288/3300 [1:18:51<3:31:38,  6.31s/it]INFO 04-11 23:55:42 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1289/3300 [1:18:56<3:16:42,  5.87s/it]2026-04-11 23:55:47,458 | INFO | train_grpo_train | metrics_logged mode=train step=1289

                                                       
{'loss': 0.0464, 'grad_norm': 5.8379316329956055, 'learning_rate': 6.096969696969698e-06, 'num_tokens': 2912411.0, 'completions/mean_length': 68.125, 'completions/min_length': 62.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8378640413284302, 'rewards/meter/std': 0.22375915944576263, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8378640413284302, 'rewards/total_composite/std': 0.22375915944576263, 'reward': 0.8378640413284302, 'reward_std': 0.22375912964344025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051838938146829605, 'sampling/sampling_logp_difference/max': 1.0195064544677734, 'sampling/importance_sampling_ratio/min': 0.360772967338562, 'sampling/importance_sampling_ratio/mean': 1.0112100839614868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32122630439698696, 'clip_ratio/low_mean': 0.01717152213677764, 'clip_ratio/low_min': 0.01717152213677764, 'clip_ratio/high_mean': 0.0282916008727625, 'clip_ratio/high_max': 0.0282916008727625, 'clip_ratio/region_mean': 0.04546312300954014, 'reward_total_mean': 0.8378640413284302, 'reward_meter_mean': 0.8378640413284302, 'reward_meter_std': 0.22375915944576263, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8378640413284302, 'reward_total_composite_std': 0.22375915944576263, 'epoch': 0.05}

 39%|███▉      | 1289/3300 [1:18:56<3:16:42,  5.87s/it]INFO 04-11 23:55:47 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1290/3300 [1:19:03<3:29:09,  6.24s/it]2026-04-11 23:55:54,578 | INFO | train_grpo_train | metrics_logged mode=train step=1290

                                                       
{'loss': 0.0012, 'grad_norm': 2.28285813331604, 'learning_rate': 6.0939393939393946e-06, 'num_tokens': 2915945.0, 'completions/mean_length': 251.75, 'completions/min_length': 245.0, 'completions/max_length': 264.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 251.75, 'completions/min_terminated_length': 245.0, 'completions/max_terminated_length': 264.0, 'rewards/meter/mean': 0.017678312957286835, 'rewards/meter/std': 0.021944953128695488, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8269230723381042, 'rewards/repeat_penalty/std': 0.09859537333250046, 'rewards/total_composite/mean': 0.012334112077951431, 'rewards/total_composite/std': 0.01488783210515976, 'reward': 0.012334112077951431, 'reward_std': 0.01488783210515976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041879359632730484, 'sampling/sampling_logp_difference/max': 1.4119462966918945, 'sampling/importance_sampling_ratio/min': 0.2436685711145401, 'sampling/importance_sampling_ratio/mean': 1.0024718046188354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28565115854144096, 'clip_ratio/low_mean': 0.01538592274300754, 'clip_ratio/low_min': 0.01538592274300754, 'clip_ratio/high_mean': 0.01394439465366304, 'clip_ratio/high_max': 0.01394439465366304, 'clip_ratio/region_mean': 0.02933031739667058, 'reward_total_mean': 0.012334112077951431, 'reward_meter_mean': 0.017678312957286835, 'reward_meter_std': 0.021944953128695488, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8269230723381042, 'reward_repeat_penalty_std': 0.09859537333250046, 'reward_total_composite_mean': 0.012334112077951431, 'reward_total_composite_std': 0.01488783210515976, 'epoch': 0.05}

 39%|███▉      | 1290/3300 [1:19:03<3:29:09,  6.24s/it]INFO 04-11 23:55:54 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1291/3300 [1:19:08<3:18:04,  5.92s/it]2026-04-11 23:55:59,727 | INFO | train_grpo_train | metrics_logged mode=train step=1291

                                                       
{'loss': -0.0041, 'grad_norm': 3.836810350418091, 'learning_rate': 6.090909090909092e-06, 'num_tokens': 2918069.0, 'completions/mean_length': 102.5, 'completions/min_length': 100.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.5, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9957513809204102, 'rewards/meter/std': 0.0030371833126991987, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9708477258682251, 'rewards/total_composite/std': 0.07034352421760559, 'reward': 0.9708477258682251, 'reward_std': 0.07034352421760559, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03093818947672844, 'sampling/sampling_logp_difference/max': 0.7555546760559082, 'sampling/importance_sampling_ratio/min': 0.4889427423477173, 'sampling/importance_sampling_ratio/mean': 1.0092651844024658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21113021485507488, 'clip_ratio/low_mean': 0.0062500000931322575, 'clip_ratio/low_min': 0.0062500000931322575, 'clip_ratio/high_mean': 0.024315623799338937, 'clip_ratio/high_max': 0.024315623799338937, 'clip_ratio/region_mean': 0.030565623892471194, 'reward_total_mean': 0.9708477258682251, 'reward_meter_mean': 0.9957513809204102, 'reward_meter_std': 0.0030371833126991987, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9708477258682251, 'reward_total_composite_std': 0.07034352421760559, 'epoch': 0.05}

 39%|███▉      | 1291/3300 [1:19:08<3:18:04,  5.92s/it]INFO 04-11 23:55:59 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1292/3300 [1:19:13<3:07:18,  5.60s/it]2026-04-11 23:56:04,580 | INFO | train_grpo_train | metrics_logged mode=train step=1292

                                                       
{'loss': 0.0073, 'grad_norm': 4.021605968475342, 'learning_rate': 6.087878787878788e-06, 'num_tokens': 2919873.0, 'completions/mean_length': 73.5, 'completions/min_length': 71.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9728573560714722, 'rewards/meter/std': 0.0436580516397953, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9728573560714722, 'rewards/total_composite/std': 0.0436580516397953, 'reward': 0.9728573560714722, 'reward_std': 0.0436580590903759, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06107901781797409, 'sampling/sampling_logp_difference/max': 1.4575915336608887, 'sampling/importance_sampling_ratio/min': 0.23279628157615662, 'sampling/importance_sampling_ratio/mean': 1.0169715881347656, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4877397455275059, 'clip_ratio/low_mean': 0.006756756920367479, 'clip_ratio/low_min': 0.006756756920367479, 'clip_ratio/high_mean': 0.03230128774885088, 'clip_ratio/high_max': 0.03230128774885088, 'clip_ratio/region_mean': 0.03905804466921836, 'reward_total_mean': 0.9728573560714722, 'reward_meter_mean': 0.9728573560714722, 'reward_meter_std': 0.0436580516397953, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9728573560714722, 'reward_total_composite_std': 0.0436580516397953, 'epoch': 0.05}

 39%|███▉      | 1292/3300 [1:19:13<3:07:18,  5.60s/it]INFO 04-11 23:56:04 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1293/3300 [1:19:18<3:00:33,  5.40s/it]2026-04-11 23:56:09,513 | INFO | train_grpo_train | metrics_logged mode=train step=1293

                                                       
{'loss': 0.0401, 'grad_norm': 6.868655681610107, 'learning_rate': 6.0848484848484855e-06, 'num_tokens': 2921458.0, 'completions/mean_length': 37.125, 'completions/min_length': 36.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9967739582061768, 'rewards/meter/std': 0.002069387584924698, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9967739582061768, 'rewards/total_composite/std': 0.002069387584924698, 'reward': 0.9967739582061768, 'reward_std': 0.00206937943585217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04344024881720543, 'sampling/sampling_logp_difference/max': 2.1784615516662598, 'sampling/importance_sampling_ratio/min': 0.11321556568145752, 'sampling/importance_sampling_ratio/mean': 0.9899824857711792, 'sampling/importance_sampling_ratio/max': 1.8224308490753174, 'entropy': 0.13870495092123747, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.027407341869547963, 'clip_ratio/high_max': 0.027407341869547963, 'clip_ratio/region_mean': 0.027407341869547963, 'reward_total_mean': 0.9967739582061768, 'reward_meter_mean': 0.9967739582061768, 'reward_meter_std': 0.002069387584924698, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9967739582061768, 'reward_total_composite_std': 0.002069387584924698, 'epoch': 0.05}

 39%|███▉      | 1293/3300 [1:19:18<3:00:33,  5.40s/it]INFO 04-11 23:56:09 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1294/3300 [1:19:23<2:57:45,  5.32s/it]2026-04-11 23:56:14,641 | INFO | train_grpo_train | metrics_logged mode=train step=1294

                                                       
{'loss': 0.0191, 'grad_norm': 4.397799491882324, 'learning_rate': 6.081818181818182e-06, 'num_tokens': 2923230.0, 'completions/mean_length': 67.5, 'completions/min_length': 64.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.938705325126648, 'rewards/meter/std': 0.06307699531316757, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.938705325126648, 'rewards/total_composite/std': 0.06307699531316757, 'reward': 0.938705325126648, 'reward_std': 0.06307698041200638, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05176489055156708, 'sampling/sampling_logp_difference/max': 2.194624423980713, 'sampling/importance_sampling_ratio/min': 0.11140039563179016, 'sampling/importance_sampling_ratio/mean': 1.0048556327819824, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30966538563370705, 'clip_ratio/low_mean': 0.005599473137408495, 'clip_ratio/low_min': 0.005599473137408495, 'clip_ratio/high_mean': 0.031401457847096026, 'clip_ratio/high_max': 0.031401457847096026, 'clip_ratio/region_mean': 0.03700093098450452, 'reward_total_mean': 0.938705325126648, 'reward_meter_mean': 0.938705325126648, 'reward_meter_std': 0.06307699531316757, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.938705325126648, 'reward_total_composite_std': 0.06307699531316757, 'epoch': 0.05}

 39%|███▉      | 1294/3300 [1:19:23<2:57:45,  5.32s/it]INFO 04-11 23:56:14 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1295/3300 [1:19:28<2:52:31,  5.16s/it]2026-04-11 23:56:19,444 | INFO | train_grpo_train | metrics_logged mode=train step=1295

                                                       
{'loss': 0.0019, 'grad_norm': 7.243402481079102, 'learning_rate': 6.07878787878788e-06, 'num_tokens': 2924989.0, 'completions/mean_length': 66.875, 'completions/min_length': 62.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9284621477127075, 'rewards/meter/std': 0.13208235800266266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9284621477127075, 'rewards/total_composite/std': 0.13208235800266266, 'reward': 0.9284621477127075, 'reward_std': 0.13208235800266266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05147900432348251, 'sampling/sampling_logp_difference/max': 1.085993766784668, 'sampling/importance_sampling_ratio/min': 0.33756616711616516, 'sampling/importance_sampling_ratio/mean': 1.0104353427886963, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3127963822335005, 'clip_ratio/low_mean': 0.011194029822945595, 'clip_ratio/low_min': 0.011194029822945595, 'clip_ratio/high_mean': 0.032499466673471034, 'clip_ratio/high_max': 0.032499466673471034, 'clip_ratio/region_mean': 0.04369349649641663, 'reward_total_mean': 0.9284621477127075, 'reward_meter_mean': 0.9284621477127075, 'reward_meter_std': 0.13208235800266266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9284621477127075, 'reward_total_composite_std': 0.13208235800266266, 'epoch': 0.05}

 39%|███▉      | 1295/3300 [1:19:28<2:52:31,  5.16s/it]INFO 04-11 23:56:19 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1296/3300 [1:19:33<2:53:04,  5.18s/it]2026-04-11 23:56:24,672 | INFO | train_grpo_train | metrics_logged mode=train step=1296

                                                       
{'loss': 0.0071, 'grad_norm': 5.327840328216553, 'learning_rate': 6.0757575757575755e-06, 'num_tokens': 2927167.0, 'completions/mean_length': 108.25, 'completions/min_length': 104.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.25, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.8149079084396362, 'rewards/meter/std': 0.24847427010536194, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8149079084396362, 'rewards/total_composite/std': 0.24847427010536194, 'reward': 0.8149079084396362, 'reward_std': 0.24847425520420074, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06685814261436462, 'sampling/sampling_logp_difference/max': 1.5687246322631836, 'sampling/importance_sampling_ratio/min': 0.20831067860126495, 'sampling/importance_sampling_ratio/mean': 1.0199620723724365, 'sampling/importance_sampling_ratio/max': 1.9841198921203613, 'entropy': 0.6056771166622639, 'clip_ratio/low_mean': 0.01175213698297739, 'clip_ratio/low_min': 0.01175213698297739, 'clip_ratio/high_mean': 0.04434093181043863, 'clip_ratio/high_max': 0.04434093181043863, 'clip_ratio/region_mean': 0.05609306879341602, 'reward_total_mean': 0.8149079084396362, 'reward_meter_mean': 0.8149079084396362, 'reward_meter_std': 0.24847427010536194, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8149079084396362, 'reward_total_composite_std': 0.24847427010536194, 'epoch': 0.05}

 39%|███▉      | 1296/3300 [1:19:33<2:53:04,  5.18s/it]INFO 04-11 23:56:24 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1297/3300 [1:19:38<2:49:33,  5.08s/it]2026-04-11 23:56:29,515 | INFO | train_grpo_train | metrics_logged mode=train step=1297

                                                       
{'loss': 0.019, 'grad_norm': 4.677238941192627, 'learning_rate': 6.072727272727274e-06, 'num_tokens': 2929000.0, 'completions/mean_length': 73.125, 'completions/min_length': 71.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9951352477073669, 'rewards/meter/std': 0.0018285271944478154, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9951352477073669, 'rewards/total_composite/std': 0.0018285271944478154, 'reward': 0.9951352477073669, 'reward_std': 0.0018285303376615047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0533045269548893, 'sampling/sampling_logp_difference/max': 2.0506200790405273, 'sampling/importance_sampling_ratio/min': 0.1286551058292389, 'sampling/importance_sampling_ratio/mean': 1.010929822921753, 'sampling/importance_sampling_ratio/max': 1.4575989246368408, 'entropy': 0.3857220932841301, 'clip_ratio/low_mean': 0.015135752153582871, 'clip_ratio/low_min': 0.015135752153582871, 'clip_ratio/high_mean': 0.02069655992090702, 'clip_ratio/high_max': 0.02069655992090702, 'clip_ratio/region_mean': 0.03583231207448989, 'reward_total_mean': 0.9951352477073669, 'reward_meter_mean': 0.9951352477073669, 'reward_meter_std': 0.0018285271944478154, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9951352477073669, 'reward_total_composite_std': 0.0018285271944478154, 'epoch': 0.05}

 39%|███▉      | 1297/3300 [1:19:38<2:49:33,  5.08s/it]INFO 04-11 23:56:29 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1298/3300 [1:19:42<2:42:17,  4.86s/it]2026-04-11 23:56:33,873 | INFO | train_grpo_train | metrics_logged mode=train step=1298

                                                       
{'loss': 0.0077, 'grad_norm': 7.133460521697998, 'learning_rate': 6.06969696969697e-06, 'num_tokens': 2930480.0, 'completions/mean_length': 34.0, 'completions/min_length': 33.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9618321657180786, 'rewards/meter/std': 0.022519638761878014, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9618321657180786, 'rewards/total_composite/std': 0.022519638761878014, 'reward': 0.9618321657180786, 'reward_std': 0.022519640624523163, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04984576627612114, 'sampling/sampling_logp_difference/max': 0.7256646156311035, 'sampling/importance_sampling_ratio/min': 0.4840027689933777, 'sampling/importance_sampling_ratio/mean': 1.023314118385315, 'sampling/importance_sampling_ratio/max': 1.5166521072387695, 'entropy': 0.4743592441082001, 'clip_ratio/low_mean': 0.014928699005395174, 'clip_ratio/low_min': 0.014928699005395174, 'clip_ratio/high_mean': 0.02205882384441793, 'clip_ratio/high_max': 0.02205882384441793, 'clip_ratio/region_mean': 0.036987522849813104, 'reward_total_mean': 0.9618321657180786, 'reward_meter_mean': 0.9618321657180786, 'reward_meter_std': 0.022519638761878014, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9618321657180786, 'reward_total_composite_std': 0.022519638761878014, 'epoch': 0.05}

 39%|███▉      | 1298/3300 [1:19:42<2:42:17,  4.86s/it]INFO 04-11 23:56:34 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1299/3300 [1:19:47<2:43:46,  4.91s/it]2026-04-11 23:56:38,894 | INFO | train_grpo_train | metrics_logged mode=train step=1299

                                                       
{'loss': -0.0198, 'grad_norm': 6.881636142730713, 'learning_rate': 6.066666666666667e-06, 'num_tokens': 2932528.0, 'completions/mean_length': 91.0, 'completions/min_length': 86.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.0, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.37875261902809143, 'rewards/meter/std': 0.31527000665664673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.36679619550704956, 'rewards/total_composite/std': 0.3127610683441162, 'reward': 0.36679619550704956, 'reward_std': 0.3127610683441162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0896112248301506, 'sampling/sampling_logp_difference/max': 1.3641595840454102, 'sampling/importance_sampling_ratio/min': 0.2555953860282898, 'sampling/importance_sampling_ratio/mean': 1.0225510597229004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7791374623775482, 'clip_ratio/low_mean': 0.04831210756674409, 'clip_ratio/low_min': 0.04831210756674409, 'clip_ratio/high_mean': 0.02266476070508361, 'clip_ratio/high_max': 0.02266476070508361, 'clip_ratio/region_mean': 0.0709768682718277, 'reward_total_mean': 0.36679619550704956, 'reward_meter_mean': 0.37875261902809143, 'reward_meter_std': 0.31527000665664673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.36679619550704956, 'reward_total_composite_std': 0.3127610683441162, 'epoch': 0.05}

 39%|███▉      | 1299/3300 [1:19:47<2:43:46,  4.91s/it]INFO 04-11 23:56:39 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1300/3300 [1:19:52<2:38:56,  4.77s/it]2026-04-11 23:56:43,328 | INFO | train_grpo_train | metrics_logged mode=train step=1300

                                                       
{'loss': 0.0288, 'grad_norm': 10.192344665527344, 'learning_rate': 6.063636363636364e-06, 'num_tokens': 2933984.0, 'completions/mean_length': 38.0, 'completions/min_length': 37.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.6407955884933472, 'rewards/meter/std': 0.46960803866386414, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6407955884933472, 'rewards/total_composite/std': 0.46960803866386414, 'reward': 0.6407955884933472, 'reward_std': 0.46960803866386414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07035049796104431, 'sampling/sampling_logp_difference/max': 1.5472084283828735, 'sampling/importance_sampling_ratio/min': 0.21284130215644836, 'sampling/importance_sampling_ratio/mean': 1.007232427597046, 'sampling/importance_sampling_ratio/max': 1.5791682004928589, 'entropy': 0.4784490168094635, 'clip_ratio/low_mean': 0.016562293516471982, 'clip_ratio/low_min': 0.016562293516471982, 'clip_ratio/high_mean': 0.04649715404957533, 'clip_ratio/high_max': 0.04649715404957533, 'clip_ratio/region_mean': 0.06305944756604731, 'reward_total_mean': 0.6407955884933472, 'reward_meter_mean': 0.6407955884933472, 'reward_meter_std': 0.46960803866386414, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6407955884933472, 'reward_total_composite_std': 0.46960803866386414, 'epoch': 0.05}

 39%|███▉      | 1300/3300 [1:19:52<2:38:56,  4.77s/it]INFO 04-11 23:56:43 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:24,  2.22s/it]

 23%|██▎       | 3/13 [00:07<00:25,  2.50s/it]

 31%|███       | 4/13 [00:11<00:29,  3.29s/it]

 38%|███▊      | 5/13 [00:16<00:28,  3.60s/it]

 46%|████▌     | 6/13 [00:20<00:27,  3.94s/it]

 54%|█████▍    | 7/13 [00:24<00:24,  4.01s/it]

 62%|██████▏   | 8/13 [00:29<00:21,  4.26s/it]

 69%|██████▉   | 9/13 [00:34<00:18,  4.54s/it]

 77%|███████▋  | 10/13 [00:39<00:13,  4.44s/it]

 85%|████████▍ | 11/13 [00:43<00:08,  4.29s/it]

 92%|█████████▏| 12/13 [00:47<00:04,  4.26s/it]

100%|██████████| 13/13 [00:51<00:00,  4.18s/it]2026-04-11 23:57:39,339 | INFO | train_grpo_train | metrics_logged mode=eval step=1300

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 55.9753, 'eval_samples_per_second': 1.858, 'eval_steps_per_second': 0.232, 'eval_num_tokens': 2933984.0, 'eval_completions/mean_length': 170.44230769230768, 'eval_completions/min_length': 62.07692307692308, 'eval_completions/max_length': 285.9230769230769, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 170.44230769230768, 'eval_completions/min_terminated_length': 62.07692307692308, 'eval_completions/max_terminated_length': 285.9230769230769, 'eval_rewards/meter/mean': 0.5045671004515427, 'eval_rewards/meter/std': 0.4269758417056157, 'eval_rewards/count_adherence/mean': 0.8285529888593234, 'eval_rewards/count_adherence/std': 0.17163905444053504, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.8525567375696622, 'eval_rewards/repeat_penalty/std': 0.13138744378319153, 'eval_rewards/total_composite/mean': 0.36767612856168014, 'eval_rewards/total_composite/std': 0.36054489933527434, 'eval_reward': 0.36767612856168014, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.02532697569292325, 'eval_sampling/sampling_logp_difference/max': 1.0623591863192046, 'eval_sampling/importance_sampling_ratio/min': 0.35662598334825957, 'eval_sampling/importance_sampling_ratio/mean': 1.0066082569269033, 'eval_sampling/importance_sampling_ratio/max': 1.4160877099403968, 'eval_entropy': 0.24410619873266953, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.36767612856168014, 'eval_reward_meter_mean': 0.5045671004515427, 'eval_reward_meter_std': 0.4269758417056157, 'eval_reward_count_adherence_mean': 0.8285529888593234, 'eval_reward_count_adherence_std': 0.17163905444053504, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.8525567375696622, 'eval_reward_repeat_penalty_std': 0.13138744378319153, 'eval_reward_total_composite_mean': 0.36767612856168014, 'eval_reward_total_composite_std': 0.36054489933527434, 'epoch': 0.05}

 39%|███▉      | 1300/3300 [1:20:48<2:38:56,  4.77s/it]

100%|██████████| 13/13 [00:51<00:00,  4.18s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 39%|███▉      | 1301/3300 [1:20:56<12:29:36, 22.50s/it]2026-04-11 23:57:47,205 | INFO | train_grpo_train | metrics_logged mode=train step=1301

                                                        
{'loss': 0.0021, 'grad_norm': 5.742159843444824, 'learning_rate': 6.060606060606061e-06, 'num_tokens': 2935839.0, 'completions/mean_length': 71.875, 'completions/min_length': 67.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.5558804869651794, 'rewards/meter/std': 0.43805626034736633, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5558804869651794, 'rewards/total_composite/std': 0.43805626034736633, 'reward': 0.5558804869651794, 'reward_std': 0.43805623054504395, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.061652250587940216, 'sampling/sampling_logp_difference/max': 1.8877463340759277, 'sampling/importance_sampling_ratio/min': 0.15141265094280243, 'sampling/importance_sampling_ratio/mean': 1.0061253309249878, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47279578633606434, 'clip_ratio/low_mean': 0.010581943904981017, 'clip_ratio/low_min': 0.010581943904981017, 'clip_ratio/high_mean': 0.03823378193192184, 'clip_ratio/high_max': 0.03823378193192184, 'clip_ratio/region_mean': 0.04881572583690286, 'reward_total_mean': 0.5558804869651794, 'reward_meter_mean': 0.5558804869651794, 'reward_meter_std': 0.43805626034736633, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5558804869651794, 'reward_total_composite_std': 0.43805626034736633, 'epoch': 0.05}

 39%|███▉      | 1301/3300 [1:20:56<12:29:36, 22.50s/it]INFO 04-11 23:57:47 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1302/3300 [1:21:04<10:06:10, 18.20s/it]2026-04-11 23:57:55,387 | INFO | train_grpo_train | metrics_logged mode=train step=1302

                                                        
{'loss': 0.0079, 'grad_norm': 1.432262897491455, 'learning_rate': 6.057575757575757e-06, 'num_tokens': 2939778.0, 'completions/mean_length': 322.375, 'completions/min_length': 317.0, 'completions/max_length': 328.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 322.375, 'completions/min_terminated_length': 317.0, 'completions/max_terminated_length': 328.0, 'rewards/meter/mean': 0.8052475452423096, 'rewards/meter/std': 0.0374782457947731, 'rewards/count_adherence/mean': 0.5333333611488342, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7666666507720947, 'rewards/repeat_penalty/std': 0.06172133609652519, 'rewards/total_composite/mean': 0.32962822914123535, 'rewards/total_composite/std': 0.0352957658469677, 'reward': 0.32962822914123535, 'reward_std': 0.0352957658469677, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024004999548196793, 'sampling/sampling_logp_difference/max': 1.421916127204895, 'sampling/importance_sampling_ratio/min': 0.24125131964683533, 'sampling/importance_sampling_ratio/mean': 1.0048707723617554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1677823355421424, 'clip_ratio/low_mean': 0.013534625410102308, 'clip_ratio/low_min': 0.013534625410102308, 'clip_ratio/high_mean': 0.005509414477273822, 'clip_ratio/high_max': 0.005509414477273822, 'clip_ratio/region_mean': 0.01904403988737613, 'reward_total_mean': 0.32962822914123535, 'reward_meter_mean': 0.8052475452423096, 'reward_meter_std': 0.0374782457947731, 'reward_count_adherence_mean': 0.5333333611488342, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7666666507720947, 'reward_repeat_penalty_std': 0.06172133609652519, 'reward_total_composite_mean': 0.32962822914123535, 'reward_total_composite_std': 0.0352957658469677, 'epoch': 0.05}

 39%|███▉      | 1302/3300 [1:21:04<10:06:10, 18.20s/it]INFO 04-11 23:57:55 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1303/3300 [1:21:09<8:00:23, 14.43s/it] 2026-04-11 23:58:01,018 | INFO | train_grpo_train | metrics_logged mode=train step=1303

                                                       
{'loss': 0.0168, 'grad_norm': 3.591283082962036, 'learning_rate': 6.0545454545454555e-06, 'num_tokens': 2942114.0, 'completions/mean_length': 119.0, 'completions/min_length': 113.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.0, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.8989412188529968, 'rewards/meter/std': 0.138026162981987, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7678571343421936, 'rewards/repeat_penalty/std': 0.15152288973331451, 'rewards/total_composite/mean': 0.6865410804748535, 'rewards/total_composite/std': 0.16533121466636658, 'reward': 0.6865410804748535, 'reward_std': 0.16533119976520538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027758914977312088, 'sampling/sampling_logp_difference/max': 1.322387456893921, 'sampling/importance_sampling_ratio/min': 0.26649829745292664, 'sampling/importance_sampling_ratio/mean': 1.0041345357894897, 'sampling/importance_sampling_ratio/max': 1.522925853729248, 'entropy': 0.17016714438796043, 'clip_ratio/low_mean': 0.010266162920743227, 'clip_ratio/low_min': 0.010266162920743227, 'clip_ratio/high_mean': 0.018164411187171936, 'clip_ratio/high_max': 0.018164411187171936, 'clip_ratio/region_mean': 0.028430574107915163, 'reward_total_mean': 0.6865410804748535, 'reward_meter_mean': 0.8989412188529968, 'reward_meter_std': 0.138026162981987, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7678571343421936, 'reward_repeat_penalty_std': 0.15152288973331451, 'reward_total_composite_mean': 0.6865410804748535, 'reward_total_composite_std': 0.16533121466636658, 'epoch': 0.05}

 39%|███▉      | 1303/3300 [1:21:10<8:00:23, 14.43s/it]INFO 04-11 23:58:01 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1304/3300 [1:21:17<6:49:33, 12.31s/it]2026-04-11 23:58:08,377 | INFO | train_grpo_train | metrics_logged mode=train step=1304

                                                       
{'loss': -0.0515, 'grad_norm': 3.418581962585449, 'learning_rate': 6.051515151515152e-06, 'num_tokens': 2945734.0, 'completions/mean_length': 247.5, 'completions/min_length': 232.0, 'completions/max_length': 277.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 247.5, 'completions/min_terminated_length': 232.0, 'completions/max_terminated_length': 277.0, 'rewards/meter/mean': 0.9368978142738342, 'rewards/meter/std': 0.09892511367797852, 'rewards/count_adherence/mean': 0.796875, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7849650382995605, 'rewards/repeat_penalty/std': 0.07891790568828583, 'rewards/total_composite/mean': 0.5886770486831665, 'rewards/total_composite/std': 0.11287946254014969, 'reward': 0.5886770486831665, 'reward_std': 0.11287945508956909, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024336956441402435, 'sampling/sampling_logp_difference/max': 1.6146456003189087, 'sampling/importance_sampling_ratio/min': 0.19896115362644196, 'sampling/importance_sampling_ratio/mean': 1.0041065216064453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1491470541805029, 'clip_ratio/low_mean': 0.00956214708276093, 'clip_ratio/low_min': 0.00956214708276093, 'clip_ratio/high_mean': 0.011577953351661563, 'clip_ratio/high_max': 0.011577953351661563, 'clip_ratio/region_mean': 0.021140100434422493, 'reward_total_mean': 0.5886770486831665, 'reward_meter_mean': 0.9368978142738342, 'reward_meter_std': 0.09892511367797852, 'reward_count_adherence_mean': 0.796875, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7849650382995605, 'reward_repeat_penalty_std': 0.07891790568828583, 'reward_total_composite_mean': 0.5886770486831665, 'reward_total_composite_std': 0.11287946254014969, 'epoch': 0.05}

 40%|███▉      | 1304/3300 [1:21:17<6:49:33, 12.31s/it]INFO 04-11 23:58:08 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1305/3300 [1:21:23<5:46:27, 10.42s/it]2026-04-11 23:58:14,401 | INFO | train_grpo_train | metrics_logged mode=train step=1305

                                                       
{'loss': 0.0351, 'grad_norm': 3.999876022338867, 'learning_rate': 6.048484848484849e-06, 'num_tokens': 2948213.0, 'completions/mean_length': 132.875, 'completions/min_length': 125.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.875, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.45057880878448486, 'rewards/meter/std': 0.39947640895843506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.4297538995742798, 'rewards/total_composite/std': 0.3954527974128723, 'reward': 0.4297538995742798, 'reward_std': 0.3954527974128723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03056526370346546, 'sampling/sampling_logp_difference/max': 1.2996227741241455, 'sampling/importance_sampling_ratio/min': 0.2726346254348755, 'sampling/importance_sampling_ratio/mean': 1.0062963962554932, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15318176615983248, 'clip_ratio/low_mean': 0.006469378364272416, 'clip_ratio/low_min': 0.006469378364272416, 'clip_ratio/high_mean': 0.02036819839850068, 'clip_ratio/high_max': 0.02036819839850068, 'clip_ratio/region_mean': 0.026837576762773097, 'reward_total_mean': 0.4297538995742798, 'reward_meter_mean': 0.45057880878448486, 'reward_meter_std': 0.39947640895843506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.4297538995742798, 'reward_total_composite_std': 0.3954527974128723, 'epoch': 0.05}

 40%|███▉      | 1305/3300 [1:21:23<5:46:27, 10.42s/it]INFO 04-11 23:58:14 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1306/3300 [1:21:28<4:51:41,  8.78s/it]2026-04-11 23:58:19,329 | INFO | train_grpo_train | metrics_logged mode=train step=1306

                                                       
{'loss': 0.0182, 'grad_norm': 5.305455207824707, 'learning_rate': 6.0454545454545456e-06, 'num_tokens': 2949699.0, 'completions/mean_length': 38.75, 'completions/min_length': 33.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9870239496231079, 'rewards/meter/std': 0.016622211784124374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9870239496231079, 'rewards/total_composite/std': 0.016622211784124374, 'reward': 0.9870239496231079, 'reward_std': 0.016622209921479225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05584869906306267, 'sampling/sampling_logp_difference/max': 1.2581281661987305, 'sampling/importance_sampling_ratio/min': 0.2841854691505432, 'sampling/importance_sampling_ratio/mean': 1.015398621559143, 'sampling/importance_sampling_ratio/max': 1.6217622756958008, 'entropy': 0.4362948350608349, 'clip_ratio/low_mean': 0.0031250000465661287, 'clip_ratio/low_min': 0.0031250000465661287, 'clip_ratio/high_mean': 0.031160252634435892, 'clip_ratio/high_max': 0.031160252634435892, 'clip_ratio/region_mean': 0.03428525268100202, 'reward_total_mean': 0.9870239496231079, 'reward_meter_mean': 0.9870239496231079, 'reward_meter_std': 0.016622211784124374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9870239496231079, 'reward_total_composite_std': 0.016622211784124374, 'epoch': 0.05}

 40%|███▉      | 1306/3300 [1:21:28<4:51:41,  8.78s/it]INFO 04-11 23:58:19 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1307/3300 [1:21:33<4:14:10,  7.65s/it]2026-04-11 23:58:24,354 | INFO | train_grpo_train | metrics_logged mode=train step=1307

                                                       
{'loss': 0.0252, 'grad_norm': 4.432555198669434, 'learning_rate': 6.042424242424243e-06, 'num_tokens': 2951929.0, 'completions/mean_length': 96.75, 'completions/min_length': 86.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.75, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.6255627274513245, 'rewards/meter/std': 0.33046483993530273, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.14880475401878357, 'rewards/total_composite/mean': 0.5558610558509827, 'rewards/total_composite/std': 0.2833639085292816, 'reward': 0.5558610558509827, 'reward_std': 0.28336387872695923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048519354313611984, 'sampling/sampling_logp_difference/max': 1.356567621231079, 'sampling/importance_sampling_ratio/min': 0.25754326581954956, 'sampling/importance_sampling_ratio/mean': 1.008318543434143, 'sampling/importance_sampling_ratio/max': 1.8040260076522827, 'entropy': 0.36078083515167236, 'clip_ratio/low_mean': 0.012753612943924963, 'clip_ratio/low_min': 0.012753612943924963, 'clip_ratio/high_mean': 0.023235379019752145, 'clip_ratio/high_max': 0.023235379019752145, 'clip_ratio/region_mean': 0.03598899196367711, 'reward_total_mean': 0.5558610558509827, 'reward_meter_mean': 0.6255627274513245, 'reward_meter_std': 0.33046483993530273, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.14880475401878357, 'reward_total_composite_mean': 0.5558610558509827, 'reward_total_composite_std': 0.2833639085292816, 'epoch': 0.05}

 40%|███▉      | 1307/3300 [1:21:33<4:14:10,  7.65s/it]INFO 04-11 23:58:24 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1308/3300 [1:21:37<3:42:38,  6.71s/it]2026-04-11 23:58:28,852 | INFO | train_grpo_train | metrics_logged mode=train step=1308

                                                       
{'loss': 0.0072, 'grad_norm': 6.026689529418945, 'learning_rate': 6.039393939393939e-06, 'num_tokens': 2953756.0, 'completions/mean_length': 64.375, 'completions/min_length': 63.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9559545516967773, 'rewards/meter/std': 0.06238928064703941, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9559545516967773, 'rewards/total_composite/std': 0.06238928064703941, 'reward': 0.9559545516967773, 'reward_std': 0.062389273196458817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031952135264873505, 'sampling/sampling_logp_difference/max': 2.3071088790893555, 'sampling/importance_sampling_ratio/min': 0.09954863786697388, 'sampling/importance_sampling_ratio/mean': 1.000700831413269, 'sampling/importance_sampling_ratio/max': 1.5961639881134033, 'entropy': 0.11791138723492622, 'clip_ratio/low_mean': 0.003846153849735856, 'clip_ratio/low_min': 0.003846153849735856, 'clip_ratio/high_mean': 0.021340470295399427, 'clip_ratio/high_max': 0.021340470295399427, 'clip_ratio/region_mean': 0.025186624145135283, 'reward_total_mean': 0.9559545516967773, 'reward_meter_mean': 0.9559545516967773, 'reward_meter_std': 0.06238928064703941, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9559545516967773, 'reward_total_composite_std': 0.06238928064703941, 'epoch': 0.05}

 40%|███▉      | 1308/3300 [1:21:37<3:42:38,  6.71s/it]INFO 04-11 23:58:29 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1309/3300 [1:21:41<3:16:15,  5.91s/it]2026-04-11 23:58:32,921 | INFO | train_grpo_train | metrics_logged mode=train step=1309

                                                       
{'loss': 0.0141, 'grad_norm': 6.288991928100586, 'learning_rate': 6.0363636363636365e-06, 'num_tokens': 2955303.0, 'completions/mean_length': 34.375, 'completions/min_length': 32.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.8419111371040344, 'rewards/meter/std': 0.3378627300262451, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8419111371040344, 'rewards/total_composite/std': 0.3378627300262451, 'reward': 0.8419111371040344, 'reward_std': 0.33786270022392273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03979434072971344, 'sampling/sampling_logp_difference/max': 0.6886262893676758, 'sampling/importance_sampling_ratio/min': 0.5022655725479126, 'sampling/importance_sampling_ratio/mean': 1.0118752717971802, 'sampling/importance_sampling_ratio/max': 1.8161826133728027, 'entropy': 0.3678742181509733, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.030007666442543268, 'clip_ratio/high_max': 0.030007666442543268, 'clip_ratio/region_mean': 0.03357909503392875, 'reward_total_mean': 0.8419111371040344, 'reward_meter_mean': 0.8419111371040344, 'reward_meter_std': 0.3378627300262451, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8419111371040344, 'reward_total_composite_std': 0.3378627300262451, 'epoch': 0.05}

 40%|███▉      | 1309/3300 [1:21:41<3:16:15,  5.91s/it]INFO 04-11 23:58:33 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1310/3300 [1:21:46<3:01:49,  5.48s/it]2026-04-11 23:58:37,393 | INFO | train_grpo_train | metrics_logged mode=train step=1310

                                                       
{'loss': -0.0014, 'grad_norm': 4.350625514984131, 'learning_rate': 6.033333333333335e-06, 'num_tokens': 2957280.0, 'completions/mean_length': 70.125, 'completions/min_length': 69.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9980206489562988, 'rewards/meter/std': 0.00018548252410255373, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980206489562988, 'rewards/total_composite/std': 0.00018548252410255373, 'reward': 0.9980206489562988, 'reward_std': 0.00018549479136709124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022730743512511253, 'sampling/sampling_logp_difference/max': 1.1889805793762207, 'sampling/importance_sampling_ratio/min': 0.3045315444469452, 'sampling/importance_sampling_ratio/mean': 1.0046977996826172, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10299927368760109, 'clip_ratio/low_mean': 0.010817805537953973, 'clip_ratio/low_min': 0.010817805537953973, 'clip_ratio/high_mean': 0.0035462776431813836, 'clip_ratio/high_max': 0.0035462776431813836, 'clip_ratio/region_mean': 0.014364083181135356, 'reward_total_mean': 0.9980206489562988, 'reward_meter_mean': 0.9980206489562988, 'reward_meter_std': 0.00018548252410255373, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980206489562988, 'reward_total_composite_std': 0.00018548252410255373, 'epoch': 0.05}

 40%|███▉      | 1310/3300 [1:21:46<3:01:49,  5.48s/it]INFO 04-11 23:58:37 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1311/3300 [1:21:51<2:59:32,  5.42s/it]2026-04-11 23:58:42,655 | INFO | train_grpo_train | metrics_logged mode=train step=1311

                                                       
{'loss': 0.0065, 'grad_norm': 2.434211015701294, 'learning_rate': 6.030303030303031e-06, 'num_tokens': 2959801.0, 'completions/mean_length': 140.125, 'completions/min_length': 139.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.125, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.9983865022659302, 'rewards/meter/std': 2.5610979719203897e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.09155284613370895, 'rewards/total_composite/mean': 0.7309621572494507, 'rewards/total_composite/std': 0.09140855073928833, 'reward': 0.7309621572494507, 'reward_std': 0.09140852838754654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021905627101659775, 'sampling/sampling_logp_difference/max': 1.7227790355682373, 'sampling/importance_sampling_ratio/min': 0.17856921255588531, 'sampling/importance_sampling_ratio/mean': 1.0040990114212036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11066047195345163, 'clip_ratio/low_mean': 0.01160750730196014, 'clip_ratio/low_min': 0.01160750730196014, 'clip_ratio/high_mean': 0.0035714286495931447, 'clip_ratio/high_max': 0.0035714286495931447, 'clip_ratio/region_mean': 0.015178935951553285, 'reward_total_mean': 0.7309621572494507, 'reward_meter_mean': 0.9983865022659302, 'reward_meter_std': 2.5610979719203897e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.09155284613370895, 'reward_total_composite_mean': 0.7309621572494507, 'reward_total_composite_std': 0.09140855073928833, 'epoch': 0.05}

 40%|███▉      | 1311/3300 [1:21:51<2:59:32,  5.42s/it]INFO 04-11 23:58:42 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1312/3300 [1:21:55<2:48:51,  5.10s/it]2026-04-11 23:58:47,005 | INFO | train_grpo_train | metrics_logged mode=train step=1312

                                                       
{'loss': -0.0003, 'grad_norm': 5.163435459136963, 'learning_rate': 6.027272727272728e-06, 'num_tokens': 2961576.0, 'completions/mean_length': 59.875, 'completions/min_length': 59.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.875, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8866318464279175, 'rewards/meter/std': 0.18304355442523956, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8866318464279175, 'rewards/total_composite/std': 0.18304355442523956, 'reward': 0.8866318464279175, 'reward_std': 0.18304355442523956, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02904520370066166, 'sampling/sampling_logp_difference/max': 0.9866929054260254, 'sampling/importance_sampling_ratio/min': 0.3728075623512268, 'sampling/importance_sampling_ratio/mean': 1.0048235654830933, 'sampling/importance_sampling_ratio/max': 1.9061437845230103, 'entropy': 0.16338308528065681, 'clip_ratio/low_mean': 0.006250000325962901, 'clip_ratio/low_min': 0.006250000325962901, 'clip_ratio/high_mean': 0.016737288795411587, 'clip_ratio/high_max': 0.016737288795411587, 'clip_ratio/region_mean': 0.022987289121374488, 'reward_total_mean': 0.8866318464279175, 'reward_meter_mean': 0.8866318464279175, 'reward_meter_std': 0.18304355442523956, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8866318464279175, 'reward_total_composite_std': 0.18304355442523956, 'epoch': 0.05}

 40%|███▉      | 1312/3300 [1:21:56<2:48:51,  5.10s/it]INFO 04-11 23:58:47 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1313/3300 [1:22:01<2:57:25,  5.36s/it]2026-04-11 23:58:52,975 | INFO | train_grpo_train | metrics_logged mode=train step=1313

                                                       
{'loss': 0.0116, 'grad_norm': 3.3765552043914795, 'learning_rate': 6.024242424242425e-06, 'num_tokens': 2964564.0, 'completions/mean_length': 172.5, 'completions/min_length': 168.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 172.5, 'completions/min_terminated_length': 168.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.7292158007621765, 'rewards/meter/std': 0.2301097810268402, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.819444477558136, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.49107223749160767, 'rewards/total_composite/std': 0.1431042104959488, 'reward': 0.49107223749160767, 'reward_std': 0.1431041955947876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040278006345033646, 'sampling/sampling_logp_difference/max': 1.5737600326538086, 'sampling/importance_sampling_ratio/min': 0.20726439356803894, 'sampling/importance_sampling_ratio/mean': 1.005394697189331, 'sampling/importance_sampling_ratio/max': 1.8441625833511353, 'entropy': 0.2585892491042614, 'clip_ratio/low_mean': 0.010145591222681105, 'clip_ratio/low_min': 0.010145591222681105, 'clip_ratio/high_mean': 0.01937754324171692, 'clip_ratio/high_max': 0.01937754324171692, 'clip_ratio/region_mean': 0.029523134464398026, 'reward_total_mean': 0.49107223749160767, 'reward_meter_mean': 0.7292158007621765, 'reward_meter_std': 0.2301097810268402, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.819444477558136, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.49107223749160767, 'reward_total_composite_std': 0.1431042104959488, 'epoch': 0.05}

 40%|███▉      | 1313/3300 [1:22:01<2:57:25,  5.36s/it]INFO 04-11 23:58:53 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1314/3300 [1:22:06<2:51:59,  5.20s/it]2026-04-11 23:58:57,792 | INFO | train_grpo_train | metrics_logged mode=train step=1314

                                                       
{'loss': -0.0091, 'grad_norm': 5.761319637298584, 'learning_rate': 6.021212121212122e-06, 'num_tokens': 2966484.0, 'completions/mean_length': 74.0, 'completions/min_length': 70.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.7603874802589417, 'rewards/meter/std': 0.4266778230667114, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7603874802589417, 'rewards/total_composite/std': 0.4266778230667114, 'reward': 0.7603874802589417, 'reward_std': 0.4266778230667114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042963117361068726, 'sampling/sampling_logp_difference/max': 1.180001974105835, 'sampling/importance_sampling_ratio/min': 0.3072781264781952, 'sampling/importance_sampling_ratio/mean': 0.9964911341667175, 'sampling/importance_sampling_ratio/max': 1.84821355342865, 'entropy': 0.24265113845467567, 'clip_ratio/low_mean': 0.005260617821477354, 'clip_ratio/low_min': 0.005260617821477354, 'clip_ratio/high_mean': 0.043520811945199966, 'clip_ratio/high_max': 0.043520811945199966, 'clip_ratio/region_mean': 0.04878142976667732, 'reward_total_mean': 0.7603874802589417, 'reward_meter_mean': 0.7603874802589417, 'reward_meter_std': 0.4266778230667114, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7603874802589417, 'reward_total_composite_std': 0.4266778230667114, 'epoch': 0.05}

 40%|███▉      | 1314/3300 [1:22:06<2:51:59,  5.20s/it]INFO 04-11 23:58:58 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1315/3300 [1:22:11<2:51:15,  5.18s/it]2026-04-11 23:59:02,925 | INFO | train_grpo_train | metrics_logged mode=train step=1315

                                                       
{'loss': 0.0157, 'grad_norm': 12.346835136413574, 'learning_rate': 6.018181818181818e-06, 'num_tokens': 2968868.0, 'completions/mean_length': 109.0, 'completions/min_length': 106.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.0, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.9957406520843506, 'rewards/meter/std': 0.0016969876596704125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957406520843506, 'rewards/total_composite/std': 0.0016969876596704125, 'reward': 0.9957406520843506, 'reward_std': 0.0016970024444162846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.050970617681741714, 'sampling/sampling_logp_difference/max': 1.3137094974517822, 'sampling/importance_sampling_ratio/min': 0.26882100105285645, 'sampling/importance_sampling_ratio/mean': 1.0082180500030518, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3228848744183779, 'clip_ratio/low_mean': 0.013646618113853037, 'clip_ratio/low_min': 0.013646618113853037, 'clip_ratio/high_mean': 0.018337647430598736, 'clip_ratio/high_max': 0.018337647430598736, 'clip_ratio/region_mean': 0.03198426554445177, 'reward_total_mean': 0.9957406520843506, 'reward_meter_mean': 0.9957406520843506, 'reward_meter_std': 0.0016969876596704125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957406520843506, 'reward_total_composite_std': 0.0016969876596704125, 'epoch': 0.05}

 40%|███▉      | 1315/3300 [1:22:11<2:51:15,  5.18s/it]INFO 04-11 23:59:03 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1316/3300 [1:22:16<2:46:19,  5.03s/it]2026-04-11 23:59:07,611 | INFO | train_grpo_train | metrics_logged mode=train step=1316

                                                       
{'loss': 0.0078, 'grad_norm': 5.835085868835449, 'learning_rate': 6.015151515151516e-06, 'num_tokens': 2970696.0, 'completions/mean_length': 63.5, 'completions/min_length': 62.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.5, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5161525011062622, 'rewards/meter/std': 0.3657090961933136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4984230399131775, 'rewards/total_composite/std': 0.3891359865665436, 'reward': 0.4984230399131775, 'reward_std': 0.3891359865665436, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03714925795793533, 'sampling/sampling_logp_difference/max': 2.0932350158691406, 'sampling/importance_sampling_ratio/min': 0.12328764796257019, 'sampling/importance_sampling_ratio/mean': 1.006494402885437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21594705432653427, 'clip_ratio/low_mean': 0.025765649508684874, 'clip_ratio/low_min': 0.025765649508684874, 'clip_ratio/high_mean': 0.01568986615166068, 'clip_ratio/high_max': 0.01568986615166068, 'clip_ratio/region_mean': 0.041455515660345554, 'reward_total_mean': 0.4984230399131775, 'reward_meter_mean': 0.5161525011062622, 'reward_meter_std': 0.3657090961933136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4984230399131775, 'reward_total_composite_std': 0.3891359865665436, 'epoch': 0.05}

 40%|███▉      | 1316/3300 [1:22:16<2:46:19,  5.03s/it]INFO 04-11 23:59:07 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1317/3300 [1:22:21<2:43:05,  4.93s/it]2026-04-11 23:59:12,325 | INFO | train_grpo_train | metrics_logged mode=train step=1317

                                                       
{'loss': 0.0052, 'grad_norm': 3.077263832092285, 'learning_rate': 6.012121212121213e-06, 'num_tokens': 2972590.0, 'completions/mean_length': 75.75, 'completions/min_length': 74.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.2975860834121704, 'rewards/meter/std': 0.04687010124325752, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.26845037937164307, 'rewards/total_composite/std': 0.04051537066698074, 'reward': 0.26845037937164307, 'reward_std': 0.040515363216400146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02295822836458683, 'sampling/sampling_logp_difference/max': 0.9676837921142578, 'sampling/importance_sampling_ratio/min': 0.37996208667755127, 'sampling/importance_sampling_ratio/mean': 0.999930739402771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09577750787138939, 'clip_ratio/low_mean': 0.008204055950045586, 'clip_ratio/low_min': 0.008204055950045586, 'clip_ratio/high_mean': 0.006602564244531095, 'clip_ratio/high_max': 0.006602564244531095, 'clip_ratio/region_mean': 0.01480662019457668, 'reward_total_mean': 0.26845037937164307, 'reward_meter_mean': 0.2975860834121704, 'reward_meter_std': 0.04687010124325752, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.26845037937164307, 'reward_total_composite_std': 0.04051537066698074, 'epoch': 0.05}

 40%|███▉      | 1317/3300 [1:22:21<2:43:05,  4.93s/it]INFO 04-11 23:59:12 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1318/3300 [1:22:25<2:37:22,  4.76s/it]2026-04-11 23:59:16,689 | INFO | train_grpo_train | metrics_logged mode=train step=1318

                                                       
{'loss': 0.0333, 'grad_norm': 6.030852317810059, 'learning_rate': 6.00909090909091e-06, 'num_tokens': 2974432.0, 'completions/mean_length': 51.25, 'completions/min_length': 48.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.813197135925293, 'rewards/meter/std': 0.24947111308574677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.813197135925293, 'rewards/total_composite/std': 0.24947111308574677, 'reward': 0.813197135925293, 'reward_std': 0.24947108328342438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047421231865882874, 'sampling/sampling_logp_difference/max': 1.33819580078125, 'sampling/importance_sampling_ratio/min': 0.26231852173805237, 'sampling/importance_sampling_ratio/mean': 1.0154815912246704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.268137663602829, 'clip_ratio/low_mean': 0.009302935097366571, 'clip_ratio/low_min': 0.009302935097366571, 'clip_ratio/high_mean': 0.012667326722294092, 'clip_ratio/high_max': 0.012667326722294092, 'clip_ratio/region_mean': 0.021970261819660664, 'reward_total_mean': 0.813197135925293, 'reward_meter_mean': 0.813197135925293, 'reward_meter_std': 0.24947111308574677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.813197135925293, 'reward_total_composite_std': 0.24947111308574677, 'epoch': 0.05}

 40%|███▉      | 1318/3300 [1:22:25<2:37:22,  4.76s/it]INFO 04-11 23:59:16 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1319/3300 [1:22:32<2:59:05,  5.42s/it]2026-04-11 23:59:23,653 | INFO | train_grpo_train | metrics_logged mode=train step=1319

                                                       
{'loss': 0.0064, 'grad_norm': 2.1162424087524414, 'learning_rate': 6.0060606060606065e-06, 'num_tokens': 2978169.0, 'completions/mean_length': 256.125, 'completions/min_length': 232.0, 'completions/max_length': 267.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 256.125, 'completions/min_terminated_length': 232.0, 'completions/max_terminated_length': 267.0, 'rewards/meter/mean': 0.9216792583465576, 'rewards/meter/std': 0.17728787660598755, 'rewards/count_adherence/mean': 0.859375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7475961446762085, 'rewards/repeat_penalty/std': 0.0792488381266594, 'rewards/total_composite/mean': 0.586740255355835, 'rewards/total_composite/std': 0.11694769561290741, 'reward': 0.586740255355835, 'reward_std': 0.11694768071174622, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026936307549476624, 'sampling/sampling_logp_difference/max': 2.076201915740967, 'sampling/importance_sampling_ratio/min': 0.12540560960769653, 'sampling/importance_sampling_ratio/mean': 1.0010830163955688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14165427815169096, 'clip_ratio/low_mean': 0.006019041407853365, 'clip_ratio/low_min': 0.006019041407853365, 'clip_ratio/high_mean': 0.012993455864489079, 'clip_ratio/high_max': 0.012993455864489079, 'clip_ratio/region_mean': 0.019012497272342443, 'reward_total_mean': 0.586740255355835, 'reward_meter_mean': 0.9216792583465576, 'reward_meter_std': 0.17728787660598755, 'reward_count_adherence_mean': 0.859375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7475961446762085, 'reward_repeat_penalty_std': 0.0792488381266594, 'reward_total_composite_mean': 0.586740255355835, 'reward_total_composite_std': 0.11694769561290741, 'epoch': 0.05}

 40%|███▉      | 1319/3300 [1:22:32<2:59:05,  5.42s/it]INFO 04-11 23:59:23 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1320/3300 [1:22:37<2:50:08,  5.16s/it]2026-04-11 23:59:28,182 | INFO | train_grpo_train | metrics_logged mode=train step=1320

                                                       
{'loss': 0.0292, 'grad_norm': 6.174464225769043, 'learning_rate': 6.003030303030304e-06, 'num_tokens': 2980004.0, 'completions/mean_length': 63.375, 'completions/min_length': 59.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.375, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6976158022880554, 'rewards/meter/std': 0.3385031819343567, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6976158022880554, 'rewards/total_composite/std': 0.3385031819343567, 'reward': 0.6976158022880554, 'reward_std': 0.3385031521320343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06314612179994583, 'sampling/sampling_logp_difference/max': 1.4771289825439453, 'sampling/importance_sampling_ratio/min': 0.22829218208789825, 'sampling/importance_sampling_ratio/mean': 1.0181591510772705, 'sampling/importance_sampling_ratio/max': 1.8571363687515259, 'entropy': 0.4745039828121662, 'clip_ratio/low_mean': 0.007752403849735856, 'clip_ratio/low_min': 0.007752403849735856, 'clip_ratio/high_mean': 0.03416440007276833, 'clip_ratio/high_max': 0.03416440007276833, 'clip_ratio/region_mean': 0.04191680392250419, 'reward_total_mean': 0.6976158022880554, 'reward_meter_mean': 0.6976158022880554, 'reward_meter_std': 0.3385031819343567, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6976158022880554, 'reward_total_composite_std': 0.3385031819343567, 'epoch': 0.05}

 40%|████      | 1320/3300 [1:22:37<2:50:08,  5.16s/it]INFO 04-11 23:59:28 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1321/3300 [1:22:41<2:44:39,  4.99s/it]2026-04-11 23:59:32,793 | INFO | train_grpo_train | metrics_logged mode=train step=1321

                                                       
{'loss': 0.0145, 'grad_norm': 4.180057525634766, 'learning_rate': 6e-06, 'num_tokens': 2981843.0, 'completions/mean_length': 72.875, 'completions/min_length': 68.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.8855767846107483, 'rewards/meter/std': 0.09589327871799469, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8855767846107483, 'rewards/total_composite/std': 0.09589327871799469, 'reward': 0.8855767846107483, 'reward_std': 0.09589327871799469, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03412603959441185, 'sampling/sampling_logp_difference/max': 1.985456943511963, 'sampling/importance_sampling_ratio/min': 0.13731785118579865, 'sampling/importance_sampling_ratio/mean': 0.9989095330238342, 'sampling/importance_sampling_ratio/max': 1.6174578666687012, 'entropy': 0.1369091346859932, 'clip_ratio/low_mean': 0.008787594037130475, 'clip_ratio/low_min': 0.008787594037130475, 'clip_ratio/high_mean': 0.026141698006540537, 'clip_ratio/high_max': 0.026141698006540537, 'clip_ratio/region_mean': 0.03492929204367101, 'reward_total_mean': 0.8855767846107483, 'reward_meter_mean': 0.8855767846107483, 'reward_meter_std': 0.09589327871799469, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8855767846107483, 'reward_total_composite_std': 0.09589327871799469, 'epoch': 0.05}

 40%|████      | 1321/3300 [1:22:41<2:44:39,  4.99s/it]INFO 04-11 23:59:33 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1322/3300 [1:22:45<2:37:03,  4.76s/it]2026-04-11 23:59:37,024 | INFO | train_grpo_train | metrics_logged mode=train step=1322

                                                       
{'loss': -0.0035, 'grad_norm': 3.336163282394409, 'learning_rate': 5.996969696969697e-06, 'num_tokens': 2983446.0, 'completions/mean_length': 41.375, 'completions/min_length': 40.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9977743029594421, 'rewards/meter/std': 0.0008235168061219156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977743029594421, 'rewards/total_composite/std': 0.0008235168061219156, 'reward': 0.9977743029594421, 'reward_std': 0.0008235144196078181, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033282674849033356, 'sampling/sampling_logp_difference/max': 1.2528910636901855, 'sampling/importance_sampling_ratio/min': 0.2856776714324951, 'sampling/importance_sampling_ratio/mean': 1.0045777559280396, 'sampling/importance_sampling_ratio/max': 1.5313419103622437, 'entropy': 0.23666546866297722, 'clip_ratio/low_mean': 0.01510935788974166, 'clip_ratio/low_min': 0.01510935788974166, 'clip_ratio/high_mean': 0.009073751280084252, 'clip_ratio/high_max': 0.009073751280084252, 'clip_ratio/region_mean': 0.02418310916982591, 'reward_total_mean': 0.9977743029594421, 'reward_meter_mean': 0.9977743029594421, 'reward_meter_std': 0.0008235168061219156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977743029594421, 'reward_total_composite_std': 0.0008235168061219156, 'epoch': 0.05}

 40%|████      | 1322/3300 [1:22:46<2:37:03,  4.76s/it]INFO 04-11 23:59:37 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1323/3300 [1:22:50<2:35:07,  4.71s/it]2026-04-11 23:59:41,602 | INFO | train_grpo_train | metrics_logged mode=train step=1323

                                                       
{'loss': 0.0011, 'grad_norm': 7.640705108642578, 'learning_rate': 5.993939393939394e-06, 'num_tokens': 2985219.0, 'completions/mean_length': 62.625, 'completions/min_length': 52.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.4877135455608368, 'rewards/meter/std': 0.3409081995487213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4877135455608368, 'rewards/total_composite/std': 0.3409081995487213, 'reward': 0.4877135455608368, 'reward_std': 0.3409081995487213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08788689225912094, 'sampling/sampling_logp_difference/max': 1.4542689323425293, 'sampling/importance_sampling_ratio/min': 0.23357106745243073, 'sampling/importance_sampling_ratio/mean': 1.0144366025924683, 'sampling/importance_sampling_ratio/max': 1.7955073118209839, 'entropy': 0.6359868273139, 'clip_ratio/low_mean': 0.025560760172083974, 'clip_ratio/low_min': 0.025560760172083974, 'clip_ratio/high_mean': 0.023659866768866777, 'clip_ratio/high_max': 0.023659866768866777, 'clip_ratio/region_mean': 0.04922062694095075, 'reward_total_mean': 0.4877135455608368, 'reward_meter_mean': 0.4877135455608368, 'reward_meter_std': 0.3409081995487213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4877135455608368, 'reward_total_composite_std': 0.3409081995487213, 'epoch': 0.05}

 40%|████      | 1323/3300 [1:22:50<2:35:07,  4.71s/it]INFO 04-11 23:59:41 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1324/3300 [1:22:54<2:30:04,  4.56s/it]2026-04-11 23:59:45,810 | INFO | train_grpo_train | metrics_logged mode=train step=1324

                                                       
{'loss': 0.0016, 'grad_norm': 6.885532855987549, 'learning_rate': 5.990909090909092e-06, 'num_tokens': 2986901.0, 'completions/mean_length': 41.25, 'completions/min_length': 39.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8787177801132202, 'rewards/meter/std': 0.3274337947368622, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8787177801132202, 'rewards/total_composite/std': 0.3274337947368622, 'reward': 0.8787177801132202, 'reward_std': 0.3274337649345398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03787970542907715, 'sampling/sampling_logp_difference/max': 0.6567374467849731, 'sampling/importance_sampling_ratio/min': 0.5185403823852539, 'sampling/importance_sampling_ratio/mean': 1.0028668642044067, 'sampling/importance_sampling_ratio/max': 1.5247844457626343, 'entropy': 0.2660053391009569, 'clip_ratio/low_mean': 0.006097560748457909, 'clip_ratio/low_min': 0.006097560748457909, 'clip_ratio/high_mean': 0.023976000724360347, 'clip_ratio/high_max': 0.023976000724360347, 'clip_ratio/region_mean': 0.030073561472818255, 'reward_total_mean': 0.8787177801132202, 'reward_meter_mean': 0.8787177801132202, 'reward_meter_std': 0.3274337947368622, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8787177801132202, 'reward_total_composite_std': 0.3274337947368622, 'epoch': 0.05}

 40%|████      | 1324/3300 [1:22:54<2:30:04,  4.56s/it]INFO 04-11 23:59:46 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1325/3300 [1:22:59<2:36:04,  4.74s/it]2026-04-11 23:59:50,980 | INFO | train_grpo_train | metrics_logged mode=train step=1325

                                                       
{'loss': -0.0087, 'grad_norm': 3.35534405708313, 'learning_rate': 5.987878787878788e-06, 'num_tokens': 2988985.0, 'completions/mean_length': 108.5, 'completions/min_length': 104.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.5, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.7801897525787354, 'rewards/meter/std': 0.25841277837753296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.824999988079071, 'rewards/repeat_penalty/std': 0.12817399203777313, 'rewards/total_composite/mean': 0.6235833168029785, 'rewards/total_composite/std': 0.18524423241615295, 'reward': 0.6235833168029785, 'reward_std': 0.18524421751499176, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03167083114385605, 'sampling/sampling_logp_difference/max': 1.2086710929870605, 'sampling/importance_sampling_ratio/min': 0.29859381914138794, 'sampling/importance_sampling_ratio/mean': 1.0031850337982178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15250306809321046, 'clip_ratio/low_mean': 0.014129945659078658, 'clip_ratio/low_min': 0.014129945659078658, 'clip_ratio/high_mean': 0.017529349657706916, 'clip_ratio/high_max': 0.017529349657706916, 'clip_ratio/region_mean': 0.031659295316785574, 'reward_total_mean': 0.6235833168029785, 'reward_meter_mean': 0.7801897525787354, 'reward_meter_std': 0.25841277837753296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.824999988079071, 'reward_repeat_penalty_std': 0.12817399203777313, 'reward_total_composite_mean': 0.6235833168029785, 'reward_total_composite_std': 0.18524423241615295, 'epoch': 0.05}

 40%|████      | 1325/3300 [1:22:59<2:36:04,  4.74s/it]INFO 04-11 23:59:51 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1326/3300 [1:23:07<2:58:52,  5.44s/it]2026-04-11 23:59:58,042 | INFO | train_grpo_train | metrics_logged mode=train step=1326

                                                       
{'loss': 0.0087, 'grad_norm': 2.242483139038086, 'learning_rate': 5.984848484848486e-06, 'num_tokens': 2992787.0, 'completions/mean_length': 258.25, 'completions/min_length': 245.0, 'completions/max_length': 268.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 258.25, 'completions/min_terminated_length': 245.0, 'completions/max_terminated_length': 268.0, 'rewards/meter/mean': 0.9577704668045044, 'rewards/meter/std': 0.11232133209705353, 'rewards/count_adherence/mean': 0.5833333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.807692289352417, 'rewards/repeat_penalty/std': 0.082234226167202, 'rewards/total_composite/mean': 0.4503657817840576, 'rewards/total_composite/std': 0.06605512648820877, 'reward': 0.4503657817840576, 'reward_std': 0.06605512648820877, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02666037157177925, 'sampling/sampling_logp_difference/max': 1.849881649017334, 'sampling/importance_sampling_ratio/min': 0.15725576877593994, 'sampling/importance_sampling_ratio/mean': 1.0053751468658447, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.151836016215384, 'clip_ratio/low_mean': 0.01110323509783484, 'clip_ratio/low_min': 0.01110323509783484, 'clip_ratio/high_mean': 0.014489957015030086, 'clip_ratio/high_max': 0.014489957015030086, 'clip_ratio/region_mean': 0.025593192112864926, 'reward_total_mean': 0.4503657817840576, 'reward_meter_mean': 0.9577704668045044, 'reward_meter_std': 0.11232133209705353, 'reward_count_adherence_mean': 0.5833333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.807692289352417, 'reward_repeat_penalty_std': 0.082234226167202, 'reward_total_composite_mean': 0.4503657817840576, 'reward_total_composite_std': 0.06605512648820877, 'epoch': 0.05}

 40%|████      | 1326/3300 [1:23:07<2:58:52,  5.44s/it]INFO 04-11 23:59:58 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1327/3300 [1:23:12<2:58:23,  5.43s/it]2026-04-12 00:00:03,437 | INFO | train_grpo_train | metrics_logged mode=train step=1327

                                                       
{'loss': -0.0023, 'grad_norm': 2.6014902591705322, 'learning_rate': 5.981818181818182e-06, 'num_tokens': 2995231.0, 'completions/mean_length': 139.5, 'completions/min_length': 133.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 139.5, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9958108067512512, 'rewards/meter/std': 0.0007901726639829576, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571343421936, 'rewards/repeat_penalty/std': 0.10101525485515594, 'rewards/total_composite/mean': 0.8890672922134399, 'rewards/total_composite/std': 0.1000855341553688, 'reward': 0.8890672922134399, 'reward_std': 0.10008554905653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03180097043514252, 'sampling/sampling_logp_difference/max': 1.4053449630737305, 'sampling/importance_sampling_ratio/min': 0.24528244137763977, 'sampling/importance_sampling_ratio/mean': 1.0085175037384033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22298285737633705, 'clip_ratio/low_mean': 0.008893042104318738, 'clip_ratio/low_min': 0.008893042104318738, 'clip_ratio/high_mean': 0.002672330185305327, 'clip_ratio/high_max': 0.002672330185305327, 'clip_ratio/region_mean': 0.011565372289624065, 'reward_total_mean': 0.8890672922134399, 'reward_meter_mean': 0.9958108067512512, 'reward_meter_std': 0.0007901726639829576, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571343421936, 'reward_repeat_penalty_std': 0.10101525485515594, 'reward_total_composite_mean': 0.8890672922134399, 'reward_total_composite_std': 0.1000855341553688, 'epoch': 0.05}

 40%|████      | 1327/3300 [1:23:12<2:58:23,  5.43s/it]INFO 04-12 00:00:03 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1328/3300 [1:23:17<2:50:24,  5.18s/it]2026-04-12 00:00:08,062 | INFO | train_grpo_train | metrics_logged mode=train step=1328

                                                       
{'loss': 0.0114, 'grad_norm': 9.022930145263672, 'learning_rate': 5.978787878787879e-06, 'num_tokens': 2997079.0, 'completions/mean_length': 61.0, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8425968885421753, 'rewards/meter/std': 0.31753626465797424, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8425968885421753, 'rewards/total_composite/std': 0.31753626465797424, 'reward': 0.8425968885421753, 'reward_std': 0.31753626465797424, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04912829399108887, 'sampling/sampling_logp_difference/max': 1.3530278205871582, 'sampling/importance_sampling_ratio/min': 0.2584564983844757, 'sampling/importance_sampling_ratio/mean': 1.0045350790023804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17815150693058968, 'clip_ratio/low_mean': 0.01024590153247118, 'clip_ratio/low_min': 0.01024590153247118, 'clip_ratio/high_mean': 0.024630822706967592, 'clip_ratio/high_max': 0.024630822706967592, 'clip_ratio/region_mean': 0.03487672423943877, 'reward_total_mean': 0.8425968885421753, 'reward_meter_mean': 0.8425968885421753, 'reward_meter_std': 0.31753626465797424, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8425968885421753, 'reward_total_composite_std': 0.31753626465797424, 'epoch': 0.05}

 40%|████      | 1328/3300 [1:23:17<2:50:24,  5.18s/it]INFO 04-12 00:00:08 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1329/3300 [1:23:22<2:48:35,  5.13s/it]2026-04-12 00:00:13,070 | INFO | train_grpo_train | metrics_logged mode=train step=1329

                                                       
{'loss': 0.0153, 'grad_norm': 3.7103993892669678, 'learning_rate': 5.975757575757576e-06, 'num_tokens': 2999253.0, 'completions/mean_length': 89.75, 'completions/min_length': 87.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.75, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.7116844058036804, 'rewards/meter/std': 0.3912390172481537, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.12817397713661194, 'rewards/total_composite/mean': 0.5941954255104065, 'rewards/total_composite/std': 0.35485145449638367, 'reward': 0.5941954255104065, 'reward_std': 0.35485145449638367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026401184499263763, 'sampling/sampling_logp_difference/max': 0.8573803901672363, 'sampling/importance_sampling_ratio/min': 0.4242720603942871, 'sampling/importance_sampling_ratio/mean': 1.0010446310043335, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1314184283837676, 'clip_ratio/low_mean': 0.006886049755848944, 'clip_ratio/low_min': 0.006886049755848944, 'clip_ratio/high_mean': 0.02516206307336688, 'clip_ratio/high_max': 0.02516206307336688, 'clip_ratio/region_mean': 0.032048112829215825, 'reward_total_mean': 0.5941954255104065, 'reward_meter_mean': 0.7116844058036804, 'reward_meter_std': 0.3912390172481537, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.12817397713661194, 'reward_total_composite_mean': 0.5941954255104065, 'reward_total_composite_std': 0.35485145449638367, 'epoch': 0.05}

 40%|████      | 1329/3300 [1:23:22<2:48:35,  5.13s/it]INFO 04-12 00:00:13 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1330/3300 [1:23:26<2:45:23,  5.04s/it]2026-04-12 00:00:17,888 | INFO | train_grpo_train | metrics_logged mode=train step=1330

                                                       
{'loss': 0.0134, 'grad_norm': 6.578309535980225, 'learning_rate': 5.972727272727274e-06, 'num_tokens': 3001055.0, 'completions/mean_length': 70.25, 'completions/min_length': 68.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8781044483184814, 'rewards/meter/std': 0.32534006237983704, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8781044483184814, 'rewards/total_composite/std': 0.32534006237983704, 'reward': 0.8781044483184814, 'reward_std': 0.32534006237983704, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04965127259492874, 'sampling/sampling_logp_difference/max': 1.5530381202697754, 'sampling/importance_sampling_ratio/min': 0.21160411834716797, 'sampling/importance_sampling_ratio/mean': 1.002303957939148, 'sampling/importance_sampling_ratio/max': 1.5147264003753662, 'entropy': 0.29124583303928375, 'clip_ratio/low_mean': 0.0052083334885537624, 'clip_ratio/low_min': 0.0052083334885537624, 'clip_ratio/high_mean': 0.025107774534262717, 'clip_ratio/high_max': 0.025107774534262717, 'clip_ratio/region_mean': 0.03031610802281648, 'reward_total_mean': 0.8781044483184814, 'reward_meter_mean': 0.8781044483184814, 'reward_meter_std': 0.32534006237983704, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8781044483184814, 'reward_total_composite_std': 0.32534006237983704, 'epoch': 0.05}

 40%|████      | 1330/3300 [1:23:26<2:45:23,  5.04s/it]INFO 04-12 00:00:18 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1331/3300 [1:23:32<2:51:32,  5.23s/it]2026-04-12 00:00:23,584 | INFO | train_grpo_train | metrics_logged mode=train step=1331

                                                       
{'loss': -0.0004, 'grad_norm': 2.5885565280914307, 'learning_rate': 5.96969696969697e-06, 'num_tokens': 3003703.0, 'completions/mean_length': 142.0, 'completions/min_length': 135.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.0, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.8050373792648315, 'rewards/meter/std': 0.2827489972114563, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.08399210125207901, 'rewards/total_composite/mean': 0.5267435908317566, 'rewards/total_composite/std': 0.19810138642787933, 'reward': 0.5267435908317566, 'reward_std': 0.19810135662555695, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032939184457063675, 'sampling/sampling_logp_difference/max': 1.6217889785766602, 'sampling/importance_sampling_ratio/min': 0.19754497706890106, 'sampling/importance_sampling_ratio/mean': 0.9992785453796387, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1445506028831005, 'clip_ratio/low_mean': 0.006243940908461809, 'clip_ratio/low_min': 0.006243940908461809, 'clip_ratio/high_mean': 0.028128837468102574, 'clip_ratio/high_max': 0.028128837468102574, 'clip_ratio/region_mean': 0.034372778376564384, 'reward_total_mean': 0.5267435908317566, 'reward_meter_mean': 0.8050373792648315, 'reward_meter_std': 0.2827489972114563, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.08399210125207901, 'reward_total_composite_mean': 0.5267435908317566, 'reward_total_composite_std': 0.19810138642787933, 'epoch': 0.05}

 40%|████      | 1331/3300 [1:23:32<2:51:32,  5.23s/it]INFO 04-12 00:00:23 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1332/3300 [1:23:38<2:58:12,  5.43s/it]2026-04-12 00:00:29,472 | INFO | train_grpo_train | metrics_logged mode=train step=1332

                                                       
{'loss': 0.0143, 'grad_norm': 5.604155540466309, 'learning_rate': 5.966666666666667e-06, 'num_tokens': 3006083.0, 'completions/mean_length': 134.5, 'completions/min_length': 129.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.5, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.48692476749420166, 'rewards/meter/std': 0.34781885147094727, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8214285373687744, 'rewards/repeat_penalty/std': 0.10101525485515594, 'rewards/total_composite/mean': 0.39111724495887756, 'rewards/total_composite/std': 0.2647561728954315, 'reward': 0.39111724495887756, 'reward_std': 0.2647561728954315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04181791469454765, 'sampling/sampling_logp_difference/max': 0.9087643623352051, 'sampling/importance_sampling_ratio/min': 0.403021901845932, 'sampling/importance_sampling_ratio/mean': 1.007440209388733, 'sampling/importance_sampling_ratio/max': 1.928771734237671, 'entropy': 0.2525683268904686, 'clip_ratio/low_mean': 0.015480266651138663, 'clip_ratio/low_min': 0.015480266651138663, 'clip_ratio/high_mean': 0.01978508895263076, 'clip_ratio/high_max': 0.01978508895263076, 'clip_ratio/region_mean': 0.03526535560376942, 'reward_total_mean': 0.39111724495887756, 'reward_meter_mean': 0.48692476749420166, 'reward_meter_std': 0.34781885147094727, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8214285373687744, 'reward_repeat_penalty_std': 0.10101525485515594, 'reward_total_composite_mean': 0.39111724495887756, 'reward_total_composite_std': 0.2647561728954315, 'epoch': 0.05}

 40%|████      | 1332/3300 [1:23:38<2:58:12,  5.43s/it]INFO 04-12 00:00:29 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1333/3300 [1:23:43<2:51:06,  5.22s/it]2026-04-12 00:00:34,217 | INFO | train_grpo_train | metrics_logged mode=train step=1333

                                                       
{'loss': 0.0109, 'grad_norm': 8.868410110473633, 'learning_rate': 5.963636363636364e-06, 'num_tokens': 3007848.0, 'completions/mean_length': 59.625, 'completions/min_length': 54.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.2418718934059143, 'rewards/meter/std': 0.20975346863269806, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.22981885075569153, 'rewards/total_composite/std': 0.2094106674194336, 'reward': 0.22981885075569153, 'reward_std': 0.2094106674194336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06638707220554352, 'sampling/sampling_logp_difference/max': 1.021674633026123, 'sampling/importance_sampling_ratio/min': 0.3599916100502014, 'sampling/importance_sampling_ratio/mean': 1.0142555236816406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5658834353089333, 'clip_ratio/low_mean': 0.041454336838796735, 'clip_ratio/low_min': 0.041454336838796735, 'clip_ratio/high_mean': 0.027125590480864048, 'clip_ratio/high_max': 0.027125590480864048, 'clip_ratio/region_mean': 0.06857992731966078, 'reward_total_mean': 0.22981885075569153, 'reward_meter_mean': 0.2418718934059143, 'reward_meter_std': 0.20975346863269806, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.22981885075569153, 'reward_total_composite_std': 0.2094106674194336, 'epoch': 0.05}

 40%|████      | 1333/3300 [1:23:43<2:51:06,  5.22s/it]INFO 04-12 00:00:34 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1334/3300 [1:23:48<2:53:35,  5.30s/it]2026-04-12 00:00:39,673 | INFO | train_grpo_train | metrics_logged mode=train step=1334

                                                       
{'loss': -0.0233, 'grad_norm': 4.963626384735107, 'learning_rate': 5.960606060606061e-06, 'num_tokens': 3010034.0, 'completions/mean_length': 98.25, 'completions/min_length': 83.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.25, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.5051434636116028, 'rewards/meter/std': 0.3091416358947754, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.1414213478565216, 'rewards/total_composite/mean': 0.40488189458847046, 'rewards/total_composite/std': 0.2789016366004944, 'reward': 0.40488189458847046, 'reward_std': 0.278901606798172, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04572216421365738, 'sampling/sampling_logp_difference/max': 2.452214241027832, 'sampling/importance_sampling_ratio/min': 0.08610272407531738, 'sampling/importance_sampling_ratio/mean': 1.0055177211761475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29656792618334293, 'clip_ratio/low_mean': 0.01672628615051508, 'clip_ratio/low_min': 0.01672628615051508, 'clip_ratio/high_mean': 0.020469399401918054, 'clip_ratio/high_max': 0.020469399401918054, 'clip_ratio/region_mean': 0.03719568555243313, 'reward_total_mean': 0.40488189458847046, 'reward_meter_mean': 0.5051434636116028, 'reward_meter_std': 0.3091416358947754, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.1414213478565216, 'reward_total_composite_mean': 0.40488189458847046, 'reward_total_composite_std': 0.2789016366004944, 'epoch': 0.05}

 40%|████      | 1334/3300 [1:23:48<2:53:35,  5.30s/it]INFO 04-12 00:00:39 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1335/3300 [1:23:54<2:59:08,  5.47s/it]2026-04-12 00:00:45,545 | INFO | train_grpo_train | metrics_logged mode=train step=1335

                                                       
{'loss': 0.02, 'grad_norm': 4.286172866821289, 'learning_rate': 5.9575757575757575e-06, 'num_tokens': 3012715.0, 'completions/mean_length': 141.125, 'completions/min_length': 139.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.125, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.891250491142273, 'rewards/meter/std': 0.16874396800994873, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7142857313156128, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6366074681282043, 'rewards/total_composite/std': 0.12053140252828598, 'reward': 0.6366074681282043, 'reward_std': 0.12053138017654419, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013281167484819889, 'sampling/sampling_logp_difference/max': 1.791276454925537, 'sampling/importance_sampling_ratio/min': 0.16674718260765076, 'sampling/importance_sampling_ratio/mean': 1.0008882284164429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0390528803691268, 'clip_ratio/low_mean': 0.0008389261784031987, 'clip_ratio/low_min': 0.0008389261784031987, 'clip_ratio/high_mean': 0.007086224795784801, 'clip_ratio/high_max': 0.007086224795784801, 'clip_ratio/region_mean': 0.007925150974188, 'reward_total_mean': 0.6366074681282043, 'reward_meter_mean': 0.891250491142273, 'reward_meter_std': 0.16874396800994873, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7142857313156128, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6366074681282043, 'reward_total_composite_std': 0.12053140252828598, 'epoch': 0.05}

 40%|████      | 1335/3300 [1:23:54<2:59:08,  5.47s/it]INFO 04-12 00:00:45 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1336/3300 [1:24:01<3:18:22,  6.06s/it]2026-04-12 00:00:52,981 | INFO | train_grpo_train | metrics_logged mode=train step=1336

                                                       
{'loss': -0.0184, 'grad_norm': 3.4981236457824707, 'learning_rate': 5.954545454545455e-06, 'num_tokens': 3016037.0, 'completions/mean_length': 219.25, 'completions/min_length': 202.0, 'completions/max_length': 235.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 219.25, 'completions/min_terminated_length': 202.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.9961979389190674, 'rewards/meter/std': 0.0030242139473557472, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8533653616905212, 'rewards/repeat_penalty/std': 0.06368338316679001, 'rewards/total_composite/mean': 0.7166441082954407, 'rewards/total_composite/std': 0.06638932228088379, 'reward': 0.7166441082954407, 'reward_std': 0.06638932973146439, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03909594938158989, 'sampling/sampling_logp_difference/max': 1.3188912868499756, 'sampling/importance_sampling_ratio/min': 0.2674316465854645, 'sampling/importance_sampling_ratio/mean': 1.0017598867416382, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20540771167725325, 'clip_ratio/low_mean': 0.014849004452116787, 'clip_ratio/low_min': 0.014849004452116787, 'clip_ratio/high_mean': 0.022171670105308294, 'clip_ratio/high_max': 0.022171670105308294, 'clip_ratio/region_mean': 0.03702067455742508, 'reward_total_mean': 0.7166441082954407, 'reward_meter_mean': 0.9961979389190674, 'reward_meter_std': 0.0030242139473557472, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8533653616905212, 'reward_repeat_penalty_std': 0.06368338316679001, 'reward_total_composite_mean': 0.7166441082954407, 'reward_total_composite_std': 0.06638932228088379, 'epoch': 0.05}

 40%|████      | 1336/3300 [1:24:01<3:18:22,  6.06s/it]INFO 04-12 00:00:53 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1337/3300 [1:24:09<3:35:42,  6.59s/it]2026-04-12 00:01:00,820 | INFO | train_grpo_train | metrics_logged mode=train step=1337

                                                       
{'loss': -0.0154, 'grad_norm': 1.7854704856872559, 'learning_rate': 5.951515151515151e-06, 'num_tokens': 3020248.0, 'completions/mean_length': 292.375, 'completions/min_length': 277.0, 'completions/max_length': 322.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 292.375, 'completions/min_terminated_length': 277.0, 'completions/max_terminated_length': 322.0, 'rewards/meter/mean': 0.9976522922515869, 'rewards/meter/std': 0.0010374977719038725, 'rewards/count_adherence/mean': 0.5500000715255737, 'rewards/count_adherence/std': 0.030860668048262596, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7024509906768799, 'rewards/repeat_penalty/std': 0.09470956027507782, 'rewards/total_composite/mean': 0.3849773406982422, 'rewards/total_composite/std': 0.05099942535161972, 'reward': 0.3849773406982422, 'reward_std': 0.05099942535161972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015756171196699142, 'sampling/sampling_logp_difference/max': 1.7869529724121094, 'sampling/importance_sampling_ratio/min': 0.16746968030929565, 'sampling/importance_sampling_ratio/mean': 1.0004962682724, 'sampling/importance_sampling_ratio/max': 1.787358045578003, 'entropy': 0.10823496524244547, 'clip_ratio/low_mean': 0.004836487962165847, 'clip_ratio/low_min': 0.004836487962165847, 'clip_ratio/high_mean': 0.007067535363603383, 'clip_ratio/high_max': 0.007067535363603383, 'clip_ratio/region_mean': 0.01190402332576923, 'reward_total_mean': 0.3849773406982422, 'reward_meter_mean': 0.9976522922515869, 'reward_meter_std': 0.0010374977719038725, 'reward_count_adherence_mean': 0.5500000715255737, 'reward_count_adherence_std': 0.030860668048262596, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7024509906768799, 'reward_repeat_penalty_std': 0.09470956027507782, 'reward_total_composite_mean': 0.3849773406982422, 'reward_total_composite_std': 0.05099942535161972, 'epoch': 0.05}

 41%|████      | 1337/3300 [1:24:09<3:35:42,  6.59s/it]INFO 04-12 00:01:01 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1338/3300 [1:24:14<3:15:45,  5.99s/it]2026-04-12 00:01:05,390 | INFO | train_grpo_train | metrics_logged mode=train step=1338

                                                       
{'loss': 0.018, 'grad_norm': 10.472454071044922, 'learning_rate': 5.948484848484849e-06, 'num_tokens': 3021936.0, 'completions/mean_length': 50.0, 'completions/min_length': 47.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6975492238998413, 'rewards/meter/std': 0.3380868434906006, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.15430334210395813, 'rewards/total_composite/mean': 0.6181282997131348, 'rewards/total_composite/std': 0.2992437779903412, 'reward': 0.6181282997131348, 'reward_std': 0.2992437779903412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0779750868678093, 'sampling/sampling_logp_difference/max': 6.2996506690979, 'sampling/importance_sampling_ratio/min': 0.0018369464669376612, 'sampling/importance_sampling_ratio/mean': 0.9909482598304749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21370963286608458, 'clip_ratio/low_mean': 0.03290023095905781, 'clip_ratio/low_min': 0.03290023095905781, 'clip_ratio/high_mean': 0.024901960510760546, 'clip_ratio/high_max': 0.024901960510760546, 'clip_ratio/region_mean': 0.057802191469818354, 'reward_total_mean': 0.6181282997131348, 'reward_meter_mean': 0.6975492238998413, 'reward_meter_std': 0.3380868434906006, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.15430334210395813, 'reward_total_composite_mean': 0.6181282997131348, 'reward_total_composite_std': 0.2992437779903412, 'epoch': 0.05}

 41%|████      | 1338/3300 [1:24:14<3:15:45,  5.99s/it]INFO 04-12 00:01:05 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1339/3300 [1:24:19<3:08:43,  5.77s/it]2026-04-12 00:01:10,669 | INFO | train_grpo_train | metrics_logged mode=train step=1339

                                                       
{'loss': 0.0, 'grad_norm': 2.900134325027466, 'learning_rate': 5.9454545454545465e-06, 'num_tokens': 3024310.0, 'completions/mean_length': 115.75, 'completions/min_length': 106.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.75, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.981162428855896, 'rewards/meter/std': 0.015516383573412895, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7857142686843872, 'rewards/repeat_penalty/std': 0.1322600096464157, 'rewards/total_composite/mean': 0.7711630463600159, 'rewards/total_composite/std': 0.1326371133327484, 'reward': 0.7711630463600159, 'reward_std': 0.1326371133327484, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030475087463855743, 'sampling/sampling_logp_difference/max': 1.8554028272628784, 'sampling/importance_sampling_ratio/min': 0.15638993680477142, 'sampling/importance_sampling_ratio/mean': 0.9987893104553223, 'sampling/importance_sampling_ratio/max': 1.8718475103378296, 'entropy': 0.13537064753472805, 'clip_ratio/low_mean': 0.01771387830376625, 'clip_ratio/low_min': 0.01771387830376625, 'clip_ratio/high_mean': 0.019089363981038332, 'clip_ratio/high_max': 0.019089363981038332, 'clip_ratio/region_mean': 0.03680324228480458, 'reward_total_mean': 0.7711630463600159, 'reward_meter_mean': 0.981162428855896, 'reward_meter_std': 0.015516383573412895, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7857142686843872, 'reward_repeat_penalty_std': 0.1322600096464157, 'reward_total_composite_mean': 0.7711630463600159, 'reward_total_composite_std': 0.1326371133327484, 'epoch': 0.05}

 41%|████      | 1339/3300 [1:24:19<3:08:43,  5.77s/it]INFO 04-12 00:01:10 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1340/3300 [1:24:24<3:00:57,  5.54s/it]2026-04-12 00:01:15,660 | INFO | train_grpo_train | metrics_logged mode=train step=1340

                                                       
{'loss': 0.0012, 'grad_norm': 2.7907607555389404, 'learning_rate': 5.942424242424243e-06, 'num_tokens': 3026369.0, 'completions/mean_length': 90.375, 'completions/min_length': 87.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.375, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9822173118591309, 'rewards/meter/std': 0.01175969373434782, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7857738733291626, 'rewards/total_composite/std': 0.009407748468220234, 'reward': 0.7857738733291626, 'reward_std': 0.009407754056155682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031274572014808655, 'sampling/sampling_logp_difference/max': 1.289837121963501, 'sampling/importance_sampling_ratio/min': 0.27531561255455017, 'sampling/importance_sampling_ratio/mean': 1.0027801990509033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16755217872560024, 'clip_ratio/low_mean': 0.01111179729923606, 'clip_ratio/low_min': 0.01111179729923606, 'clip_ratio/high_mean': 0.020700859487988055, 'clip_ratio/high_max': 0.020700859487988055, 'clip_ratio/region_mean': 0.031812656787224114, 'reward_total_mean': 0.7857738733291626, 'reward_meter_mean': 0.9822173118591309, 'reward_meter_std': 0.01175969373434782, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7857738733291626, 'reward_total_composite_std': 0.009407748468220234, 'epoch': 0.05}

 41%|████      | 1340/3300 [1:24:24<3:00:57,  5.54s/it]INFO 04-12 00:01:15 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1341/3300 [1:24:29<2:52:25,  5.28s/it]2026-04-12 00:01:20,368 | INFO | train_grpo_train | metrics_logged mode=train step=1341

                                                       
{'loss': 0.0155, 'grad_norm': 7.174100399017334, 'learning_rate': 5.93939393939394e-06, 'num_tokens': 3028190.0, 'completions/mean_length': 60.625, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.985583484172821, 'rewards/meter/std': 0.008517028763890266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.985583484172821, 'rewards/total_composite/std': 0.008517028763890266, 'reward': 0.985583484172821, 'reward_std': 0.008517015725374222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030380286276340485, 'sampling/sampling_logp_difference/max': 1.27167809009552, 'sampling/importance_sampling_ratio/min': 0.28036075830459595, 'sampling/importance_sampling_ratio/mean': 1.005619764328003, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17691593896597624, 'clip_ratio/low_mean': 0.010416667209938169, 'clip_ratio/low_min': 0.010416667209938169, 'clip_ratio/high_mean': 0.01639674766920507, 'clip_ratio/high_max': 0.01639674766920507, 'clip_ratio/region_mean': 0.026813414879143238, 'reward_total_mean': 0.985583484172821, 'reward_meter_mean': 0.985583484172821, 'reward_meter_std': 0.008517028763890266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.985583484172821, 'reward_total_composite_std': 0.008517028763890266, 'epoch': 0.05}

 41%|████      | 1341/3300 [1:24:29<2:52:25,  5.28s/it]INFO 04-12 00:01:20 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1342/3300 [1:24:34<2:55:41,  5.38s/it]2026-04-12 00:01:25,962 | INFO | train_grpo_train | metrics_logged mode=train step=1342

                                                       
{'loss': 0.0171, 'grad_norm': 4.300682544708252, 'learning_rate': 5.936363636363637e-06, 'num_tokens': 3030079.0, 'completions/mean_length': 69.125, 'completions/min_length': 67.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.963999330997467, 'rewards/meter/std': 0.005599203985184431, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.963999330997467, 'rewards/total_composite/std': 0.005599203985184431, 'reward': 0.963999330997467, 'reward_std': 0.00559920072555542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011351129971444607, 'sampling/sampling_logp_difference/max': 0.8801417350769043, 'sampling/importance_sampling_ratio/min': 0.4147241413593292, 'sampling/importance_sampling_ratio/mean': 1.0003573894500732, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03910455713048577, 'clip_ratio/low_mean': 0.003597308532334864, 'clip_ratio/low_min': 0.003597308532334864, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.005462980130687356, 'reward_total_mean': 0.963999330997467, 'reward_meter_mean': 0.963999330997467, 'reward_meter_std': 0.005599203985184431, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.963999330997467, 'reward_total_composite_std': 0.005599203985184431, 'epoch': 0.05}

 41%|████      | 1342/3300 [1:24:34<2:55:41,  5.38s/it]INFO 04-12 00:01:26 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1343/3300 [1:24:39<2:50:12,  5.22s/it]2026-04-12 00:01:30,795 | INFO | train_grpo_train | metrics_logged mode=train step=1343

                                                       
{'loss': 0.029, 'grad_norm': 3.7703700065612793, 'learning_rate': 5.933333333333335e-06, 'num_tokens': 3031857.0, 'completions/mean_length': 66.25, 'completions/min_length': 63.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.4224083423614502, 'rewards/meter/std': 0.23921163380146027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4224083423614502, 'rewards/total_composite/std': 0.23921163380146027, 'reward': 0.4224083423614502, 'reward_std': 0.23921160399913788, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059910353273153305, 'sampling/sampling_logp_difference/max': 3.5546302795410156, 'sampling/importance_sampling_ratio/min': 0.0285919439047575, 'sampling/importance_sampling_ratio/mean': 0.9989436268806458, 'sampling/importance_sampling_ratio/max': 1.9895342588424683, 'entropy': 0.24532775953412056, 'clip_ratio/low_mean': 0.01829817472025752, 'clip_ratio/low_min': 0.01829817472025752, 'clip_ratio/high_mean': 0.013258611783385277, 'clip_ratio/high_max': 0.013258611783385277, 'clip_ratio/region_mean': 0.0315567865036428, 'reward_total_mean': 0.4224083423614502, 'reward_meter_mean': 0.4224083423614502, 'reward_meter_std': 0.23921163380146027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4224083423614502, 'reward_total_composite_std': 0.23921163380146027, 'epoch': 0.05}

 41%|████      | 1343/3300 [1:24:39<2:50:12,  5.22s/it]INFO 04-12 00:01:31 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1344/3300 [1:24:45<2:53:07,  5.31s/it]2026-04-12 00:01:36,322 | INFO | train_grpo_train | metrics_logged mode=train step=1344

                                                       
{'loss': 0.0095, 'grad_norm': 6.5239176750183105, 'learning_rate': 5.93030303030303e-06, 'num_tokens': 3034322.0, 'completions/mean_length': 140.125, 'completions/min_length': 136.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.125, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.900975227355957, 'rewards/meter/std': 0.25265270471572876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.84794020652771, 'rewards/total_composite/std': 0.24164395034313202, 'reward': 0.84794020652771, 'reward_std': 0.24164395034313202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054831597954034805, 'sampling/sampling_logp_difference/max': 1.7345802783966064, 'sampling/importance_sampling_ratio/min': 0.17647425830364227, 'sampling/importance_sampling_ratio/mean': 1.0002692937850952, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.352537851780653, 'clip_ratio/low_mean': 0.01140083302743733, 'clip_ratio/low_min': 0.01140083302743733, 'clip_ratio/high_mean': 0.033956656232476234, 'clip_ratio/high_max': 0.033956656232476234, 'clip_ratio/region_mean': 0.045357489259913564, 'reward_total_mean': 0.84794020652771, 'reward_meter_mean': 0.900975227355957, 'reward_meter_std': 0.25265270471572876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.84794020652771, 'reward_total_composite_std': 0.24164395034313202, 'epoch': 0.05}

 41%|████      | 1344/3300 [1:24:45<2:53:07,  5.31s/it]INFO 04-12 00:01:36 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1345/3300 [1:24:49<2:46:21,  5.11s/it]2026-04-12 00:01:40,951 | INFO | train_grpo_train | metrics_logged mode=train step=1345

                                                       
{'loss': -0.0074, 'grad_norm': 3.739021062850952, 'learning_rate': 5.927272727272728e-06, 'num_tokens': 3036050.0, 'completions/mean_length': 60.0, 'completions/min_length': 59.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.0, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9513925313949585, 'rewards/meter/std': 0.06354161351919174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9513925313949585, 'rewards/total_composite/std': 0.06354161351919174, 'reward': 0.9513925313949585, 'reward_std': 0.06354160606861115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040932025760412216, 'sampling/sampling_logp_difference/max': 1.3692903518676758, 'sampling/importance_sampling_ratio/min': 0.25428733229637146, 'sampling/importance_sampling_ratio/mean': 0.9930896162986755, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20175675675272942, 'clip_ratio/low_mean': 0.0190677959471941, 'clip_ratio/low_min': 0.0190677959471941, 'clip_ratio/high_mean': 0.01858155010268092, 'clip_ratio/high_max': 0.01858155010268092, 'clip_ratio/region_mean': 0.03764934604987502, 'reward_total_mean': 0.9513925313949585, 'reward_meter_mean': 0.9513925313949585, 'reward_meter_std': 0.06354161351919174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9513925313949585, 'reward_total_composite_std': 0.06354161351919174, 'epoch': 0.05}

 41%|████      | 1345/3300 [1:24:49<2:46:21,  5.11s/it]INFO 04-12 00:01:41 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1346/3300 [1:24:55<2:50:32,  5.24s/it]2026-04-12 00:01:46,495 | INFO | train_grpo_train | metrics_logged mode=train step=1346

                                                       
{'loss': 0.0129, 'grad_norm': 3.906574249267578, 'learning_rate': 5.924242424242425e-06, 'num_tokens': 3038243.0, 'completions/mean_length': 132.125, 'completions/min_length': 128.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.125, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.902446985244751, 'rewards/meter/std': 0.14302004873752594, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.6613233089447021, 'rewards/total_composite/std': 0.11659383773803711, 'reward': 0.6613233089447021, 'reward_std': 0.11659382283687592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013300405815243721, 'sampling/sampling_logp_difference/max': 1.1266558170318604, 'sampling/importance_sampling_ratio/min': 0.3424758315086365, 'sampling/importance_sampling_ratio/mean': 1.0032755136489868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04485098086297512, 'clip_ratio/low_mean': 0.0009328357991762459, 'clip_ratio/low_min': 0.0009328357991762459, 'clip_ratio/high_mean': 0.011403630953282118, 'clip_ratio/high_max': 0.011403630953282118, 'clip_ratio/region_mean': 0.012336466752458364, 'reward_total_mean': 0.6613233089447021, 'reward_meter_mean': 0.902446985244751, 'reward_meter_std': 0.14302004873752594, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.6613233089447021, 'reward_total_composite_std': 0.11659383773803711, 'epoch': 0.05}

 41%|████      | 1346/3300 [1:24:55<2:50:32,  5.24s/it]INFO 04-12 00:01:46 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1347/3300 [1:25:00<2:50:29,  5.24s/it]2026-04-12 00:01:51,730 | INFO | train_grpo_train | metrics_logged mode=train step=1347

                                                       
{'loss': 0.0011, 'grad_norm': 3.224806070327759, 'learning_rate': 5.921212121212122e-06, 'num_tokens': 3040443.0, 'completions/mean_length': 111.0, 'completions/min_length': 108.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.0, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9691954851150513, 'rewards/meter/std': 0.036640021950006485, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9197216033935547, 'rewards/total_composite/std': 0.08622133731842041, 'reward': 0.9197216033935547, 'reward_std': 0.08622132986783981, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03579000383615494, 'sampling/sampling_logp_difference/max': 1.3639929294586182, 'sampling/importance_sampling_ratio/min': 0.2556380033493042, 'sampling/importance_sampling_ratio/mean': 1.011066198348999, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.27726688235998154, 'clip_ratio/low_mean': 0.010106078116223216, 'clip_ratio/low_min': 0.010106078116223216, 'clip_ratio/high_mean': 0.024769189301878214, 'clip_ratio/high_max': 0.024769189301878214, 'clip_ratio/region_mean': 0.03487526741810143, 'reward_total_mean': 0.9197216033935547, 'reward_meter_mean': 0.9691954851150513, 'reward_meter_std': 0.036640021950006485, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9197216033935547, 'reward_total_composite_std': 0.08622133731842041, 'epoch': 0.05}

 41%|████      | 1347/3300 [1:25:00<2:50:29,  5.24s/it]INFO 04-12 00:01:51 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1348/3300 [1:25:05<2:44:39,  5.06s/it]2026-04-12 00:01:56,380 | INFO | train_grpo_train | metrics_logged mode=train step=1348

                                                       
{'loss': 0.011, 'grad_norm': 4.513327121734619, 'learning_rate': 5.9181818181818184e-06, 'num_tokens': 3042433.0, 'completions/mean_length': 71.75, 'completions/min_length': 67.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.99162358045578, 'rewards/meter/std': 0.011257669888436794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99162358045578, 'rewards/total_composite/std': 0.011257669888436794, 'reward': 0.99162358045578, 'reward_std': 0.011257672682404518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06130176782608032, 'sampling/sampling_logp_difference/max': 1.3990116119384766, 'sampling/importance_sampling_ratio/min': 0.24684081971645355, 'sampling/importance_sampling_ratio/mean': 1.0111957788467407, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.404640544205904, 'clip_ratio/low_mean': 0.006873097387142479, 'clip_ratio/low_min': 0.006873097387142479, 'clip_ratio/high_mean': 0.0280318089062348, 'clip_ratio/high_max': 0.0280318089062348, 'clip_ratio/region_mean': 0.03490490629337728, 'reward_total_mean': 0.99162358045578, 'reward_meter_mean': 0.99162358045578, 'reward_meter_std': 0.011257669888436794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99162358045578, 'reward_total_composite_std': 0.011257669888436794, 'epoch': 0.05}

 41%|████      | 1348/3300 [1:25:05<2:44:39,  5.06s/it]INFO 04-12 00:01:56 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1349/3300 [1:25:10<2:42:21,  4.99s/it]2026-04-12 00:02:01,213 | INFO | train_grpo_train | metrics_logged mode=train step=1349

                                                       
{'loss': -0.0009, 'grad_norm': 3.323253631591797, 'learning_rate': 5.915151515151516e-06, 'num_tokens': 3044432.0, 'completions/mean_length': 89.875, 'completions/min_length': 88.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.875, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9421913623809814, 'rewards/meter/std': 0.0664798840880394, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.753753125667572, 'rewards/total_composite/std': 0.0531839095056057, 'reward': 0.753753125667572, 'reward_std': 0.0531838983297348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014572631567716599, 'sampling/sampling_logp_difference/max': 0.9959890842437744, 'sampling/importance_sampling_ratio/min': 0.3693579435348511, 'sampling/importance_sampling_ratio/mean': 0.9989533424377441, 'sampling/importance_sampling_ratio/max': 1.4543837308883667, 'entropy': 0.10254091210663319, 'clip_ratio/low_mean': 0.004213483072817326, 'clip_ratio/low_min': 0.004213483072817326, 'clip_ratio/high_mean': 0.009848485118709505, 'clip_ratio/high_max': 0.009848485118709505, 'clip_ratio/region_mean': 0.01406196819152683, 'reward_total_mean': 0.753753125667572, 'reward_meter_mean': 0.9421913623809814, 'reward_meter_std': 0.0664798840880394, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.753753125667572, 'reward_total_composite_std': 0.0531839095056057, 'epoch': 0.05}

 41%|████      | 1349/3300 [1:25:10<2:42:21,  4.99s/it]INFO 04-12 00:02:01 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1350/3300 [1:25:14<2:38:33,  4.88s/it]2026-04-12 00:02:05,826 | INFO | train_grpo_train | metrics_logged mode=train step=1350

                                                       
{'loss': 0.0046, 'grad_norm': 4.666144847869873, 'learning_rate': 5.912121212121212e-06, 'num_tokens': 3046237.0, 'completions/mean_length': 63.625, 'completions/min_length': 61.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5202029347419739, 'rewards/meter/std': 0.32961180806159973, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5202029347419739, 'rewards/total_composite/std': 0.32961180806159973, 'reward': 0.5202029347419739, 'reward_std': 0.32961177825927734, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03747888654470444, 'sampling/sampling_logp_difference/max': 1.3192882537841797, 'sampling/importance_sampling_ratio/min': 0.2673254907131195, 'sampling/importance_sampling_ratio/mean': 0.9964258074760437, 'sampling/importance_sampling_ratio/max': 1.43716299533844, 'entropy': 0.18225187622010708, 'clip_ratio/low_mean': 0.015552184893749654, 'clip_ratio/low_min': 0.015552184893749654, 'clip_ratio/high_mean': 0.007785359863191843, 'clip_ratio/high_max': 0.007785359863191843, 'clip_ratio/region_mean': 0.023337544756941497, 'reward_total_mean': 0.5202029347419739, 'reward_meter_mean': 0.5202029347419739, 'reward_meter_std': 0.32961180806159973, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5202029347419739, 'reward_total_composite_std': 0.32961180806159973, 'epoch': 0.05}

 41%|████      | 1350/3300 [1:25:14<2:38:33,  4.88s/it]INFO 04-12 00:02:06 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:25,  2.29s/it]

 23%|██▎       | 3/13 [00:07<00:25,  2.53s/it]

 31%|███       | 4/13 [00:11<00:29,  3.23s/it]

 38%|███▊      | 5/13 [00:16<00:28,  3.57s/it]

 46%|████▌     | 6/13 [00:21<00:28,  4.09s/it]

 54%|█████▍    | 7/13 [00:25<00:24,  4.05s/it]

 62%|██████▏   | 8/13 [00:30<00:21,  4.39s/it]

 69%|██████▉   | 9/13 [00:36<00:19,  4.89s/it]

 77%|███████▋  | 10/13 [00:40<00:14,  4.72s/it]

 85%|████████▍ | 11/13 [00:45<00:09,  4.68s/it]

 92%|█████████▏| 12/13 [00:49<00:04,  4.70s/it]

100%|██████████| 13/13 [00:54<00:00,  4.52s/it]2026-04-12 00:03:04,968 | INFO | train_grpo_train | metrics_logged mode=eval step=1350

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 59.1029, 'eval_samples_per_second': 1.76, 'eval_steps_per_second': 0.22, 'eval_num_tokens': 3046237.0, 'eval_completions/mean_length': 172.09615384615384, 'eval_completions/min_length': 61.46153846153846, 'eval_completions/max_length': 307.0, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 172.09615384615384, 'eval_completions/min_terminated_length': 61.46153846153846, 'eval_completions/max_terminated_length': 307.0, 'eval_rewards/meter/mean': 0.6744922651694372, 'eval_rewards/meter/std': 0.41629832753768337, 'eval_rewards/count_adherence/mean': 0.8681816137754, 'eval_rewards/count_adherence/std': 0.14402351528406143, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.7968475085038406, 'eval_rewards/repeat_penalty/std': 0.16538272224939787, 'eval_rewards/total_composite/mean': 0.47231991015947783, 'eval_rewards/total_composite/std': 0.3407691912009166, 'eval_reward': 0.47231991015947783, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.016293376182707455, 'eval_sampling/sampling_logp_difference/max': 1.005841695345365, 'eval_sampling/importance_sampling_ratio/min': 0.37638011689369494, 'eval_sampling/importance_sampling_ratio/mean': 1.0026835478269136, 'eval_sampling/importance_sampling_ratio/max': 1.2998670798081617, 'eval_entropy': 0.14777196657199126, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.47231991015947783, 'eval_reward_meter_mean': 0.6744922651694372, 'eval_reward_meter_std': 0.41629832753768337, 'eval_reward_count_adherence_mean': 0.8681816137754, 'eval_reward_count_adherence_std': 0.14402351528406143, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.7968475085038406, 'eval_reward_repeat_penalty_std': 0.16538272224939787, 'eval_reward_total_composite_mean': 0.47231991015947783, 'eval_reward_total_composite_std': 0.3407691912009166, 'epoch': 0.05}

 41%|████      | 1350/3300 [1:26:13<2:38:33,  4.88s/it]

100%|██████████| 13/13 [00:54<00:00,  4.52s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 41%|████      | 1351/3300 [1:26:22<12:47:45, 23.64s/it]2026-04-12 00:03:13,228 | INFO | train_grpo_train | metrics_logged mode=train step=1351

                                                        
{'loss': 0.0176, 'grad_norm': 3.602323532104492, 'learning_rate': 5.90909090909091e-06, 'num_tokens': 3048749.0, 'completions/mean_length': 134.0, 'completions/min_length': 128.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.0, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.8950950503349304, 'rewards/meter/std': 0.18451449275016785, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.7975612878799438, 'rewards/total_composite/std': 0.16415363550186157, 'reward': 0.7975612878799438, 'reward_std': 0.16415365040302277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036450449377298355, 'sampling/sampling_logp_difference/max': 1.9212646484375, 'sampling/importance_sampling_ratio/min': 0.1464216709136963, 'sampling/importance_sampling_ratio/mean': 1.0027098655700684, 'sampling/importance_sampling_ratio/max': 1.8117905855178833, 'entropy': 0.21928934194147587, 'clip_ratio/low_mean': 0.008094369200989604, 'clip_ratio/low_min': 0.008094369200989604, 'clip_ratio/high_mean': 0.03014001634437591, 'clip_ratio/high_max': 0.03014001634437591, 'clip_ratio/region_mean': 0.03823438554536551, 'reward_total_mean': 0.7975612878799438, 'reward_meter_mean': 0.8950950503349304, 'reward_meter_std': 0.18451449275016785, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.7975612878799438, 'reward_total_composite_std': 0.16415363550186157, 'epoch': 0.05}

 41%|████      | 1351/3300 [1:26:22<12:47:45, 23.64s/it]INFO 04-12 00:03:13 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1352/3300 [1:26:26<9:38:14, 17.81s/it] 2026-04-12 00:03:17,448 | INFO | train_grpo_train | metrics_logged mode=train step=1352

                                                       
{'loss': 0.0112, 'grad_norm': 3.5144999027252197, 'learning_rate': 5.906060606060607e-06, 'num_tokens': 3050215.0, 'completions/mean_length': 32.25, 'completions/min_length': 31.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9494646191596985, 'rewards/meter/std': 0.08898033201694489, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9494646191596985, 'rewards/total_composite/std': 0.08898033201694489, 'reward': 0.9494646191596985, 'reward_std': 0.08898033201694489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03137016296386719, 'sampling/sampling_logp_difference/max': 1.034781575202942, 'sampling/importance_sampling_ratio/min': 0.35530397295951843, 'sampling/importance_sampling_ratio/mean': 0.9998236298561096, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14527534414082766, 'clip_ratio/low_mean': 0.011482007801532745, 'clip_ratio/low_min': 0.011482007801532745, 'clip_ratio/high_mean': 0.011844757944345474, 'clip_ratio/high_max': 0.011844757944345474, 'clip_ratio/region_mean': 0.02332676574587822, 'reward_total_mean': 0.9494646191596985, 'reward_meter_mean': 0.9494646191596985, 'reward_meter_std': 0.08898033201694489, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9494646191596985, 'reward_total_composite_std': 0.08898033201694489, 'epoch': 0.05}

 41%|████      | 1352/3300 [1:26:26<9:38:14, 17.81s/it]INFO 04-12 00:03:17 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1353/3300 [1:26:31<7:30:00, 13.87s/it]2026-04-12 00:03:22,117 | INFO | train_grpo_train | metrics_logged mode=train step=1353

                                                       
{'loss': 0.0173, 'grad_norm': 7.210882186889648, 'learning_rate': 5.903030303030304e-06, 'num_tokens': 3052065.0, 'completions/mean_length': 68.25, 'completions/min_length': 64.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9697257876396179, 'rewards/meter/std': 0.039108145982027054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9697257876396179, 'rewards/total_composite/std': 0.039108145982027054, 'reward': 0.9697257876396179, 'reward_std': 0.03910814970731735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03900406137108803, 'sampling/sampling_logp_difference/max': 1.7512474060058594, 'sampling/importance_sampling_ratio/min': 0.173557311296463, 'sampling/importance_sampling_ratio/mean': 1.003574252128601, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19455979391932487, 'clip_ratio/low_mean': 0.014525586506351829, 'clip_ratio/low_min': 0.014525586506351829, 'clip_ratio/high_mean': 0.031015241518616676, 'clip_ratio/high_max': 0.031015241518616676, 'clip_ratio/region_mean': 0.045540828024968505, 'reward_total_mean': 0.9697257876396179, 'reward_meter_mean': 0.9697257876396179, 'reward_meter_std': 0.039108145982027054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9697257876396179, 'reward_total_composite_std': 0.039108145982027054, 'epoch': 0.05}

 41%|████      | 1353/3300 [1:26:31<7:30:00, 13.87s/it]INFO 04-12 00:03:22 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1354/3300 [1:26:36<6:06:29, 11.30s/it]2026-04-12 00:03:27,418 | INFO | train_grpo_train | metrics_logged mode=train step=1354

                                                       
{'loss': 0.0013, 'grad_norm': 2.7349421977996826, 'learning_rate': 5.9e-06, 'num_tokens': 3054440.0, 'completions/mean_length': 128.875, 'completions/min_length': 127.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.875, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9711472392082214, 'rewards/meter/std': 0.022641118615865707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7678571343421936, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.7450422048568726, 'rewards/total_composite/std': 0.06550093740224838, 'reward': 0.7450422048568726, 'reward_std': 0.06550093740224838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01926073245704174, 'sampling/sampling_logp_difference/max': 1.9643745422363281, 'sampling/importance_sampling_ratio/min': 0.14024357497692108, 'sampling/importance_sampling_ratio/mean': 1.0016677379608154, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07222465332597494, 'clip_ratio/low_mean': 0.004853119375184178, 'clip_ratio/low_min': 0.004853119375184178, 'clip_ratio/high_mean': 0.009631440974771976, 'clip_ratio/high_max': 0.009631440974771976, 'clip_ratio/region_mean': 0.014484560349956155, 'reward_total_mean': 0.7450422048568726, 'reward_meter_mean': 0.9711472392082214, 'reward_meter_std': 0.022641118615865707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7678571343421936, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.7450422048568726, 'reward_total_composite_std': 0.06550093740224838, 'epoch': 0.05}

 41%|████      | 1354/3300 [1:26:36<6:06:29, 11.30s/it]INFO 04-12 00:03:27 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1355/3300 [1:26:40<4:58:13,  9.20s/it]2026-04-12 00:03:31,725 | INFO | train_grpo_train | metrics_logged mode=train step=1355

                                                       
{'loss': 0.017, 'grad_norm': 13.958351135253906, 'learning_rate': 5.8969696969696975e-06, 'num_tokens': 3055950.0, 'completions/mean_length': 33.75, 'completions/min_length': 32.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9982504844665527, 'rewards/meter/std': 0.0002638747973833233, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982504844665527, 'rewards/total_composite/std': 0.0002638747973833233, 'reward': 0.9982504844665527, 'reward_std': 0.00026387875550426543, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0569852739572525, 'sampling/sampling_logp_difference/max': 1.649427890777588, 'sampling/importance_sampling_ratio/min': 0.19215981662273407, 'sampling/importance_sampling_ratio/mean': 1.005803108215332, 'sampling/importance_sampling_ratio/max': 1.8974801301956177, 'entropy': 0.3209527414292097, 'clip_ratio/low_mean': 0.011488970601931214, 'clip_ratio/low_min': 0.011488970601931214, 'clip_ratio/high_mean': 0.04795599193312228, 'clip_ratio/high_max': 0.04795599193312228, 'clip_ratio/region_mean': 0.05944496253505349, 'reward_total_mean': 0.9982504844665527, 'reward_meter_mean': 0.9982504844665527, 'reward_meter_std': 0.0002638747973833233, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982504844665527, 'reward_total_composite_std': 0.0002638747973833233, 'epoch': 0.05}

 41%|████      | 1355/3300 [1:26:40<4:58:13,  9.20s/it]INFO 04-12 00:03:31 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1356/3300 [1:26:44<4:07:42,  7.65s/it]2026-04-12 00:03:35,740 | INFO | train_grpo_train | metrics_logged mode=train step=1356

                                                       
{'loss': 0.0063, 'grad_norm': 8.597617149353027, 'learning_rate': 5.893939393939394e-06, 'num_tokens': 3057379.0, 'completions/mean_length': 31.625, 'completions/min_length': 31.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9754250049591064, 'rewards/meter/std': 0.010440711863338947, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9754250049591064, 'rewards/total_composite/std': 0.010440711863338947, 'reward': 0.9754250049591064, 'reward_std': 0.010440702550113201, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02750094048678875, 'sampling/sampling_logp_difference/max': 1.307356834411621, 'sampling/importance_sampling_ratio/min': 0.2705341875553131, 'sampling/importance_sampling_ratio/mean': 0.9994874596595764, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0775892292149365, 'clip_ratio/low_mean': 0.023815523833036423, 'clip_ratio/low_min': 0.023815523833036423, 'clip_ratio/high_mean': 0.007938507944345474, 'clip_ratio/high_max': 0.007938507944345474, 'clip_ratio/region_mean': 0.0317540317773819, 'reward_total_mean': 0.9754250049591064, 'reward_meter_mean': 0.9754250049591064, 'reward_meter_std': 0.010440711863338947, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9754250049591064, 'reward_total_composite_std': 0.010440711863338947, 'epoch': 0.05}

 41%|████      | 1356/3300 [1:26:44<4:07:42,  7.65s/it]INFO 04-12 00:03:35 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1357/3300 [1:26:49<3:35:51,  6.67s/it]2026-04-12 00:03:40,117 | INFO | train_grpo_train | metrics_logged mode=train step=1357

                                                       
{'loss': -0.0049, 'grad_norm': 4.431914329528809, 'learning_rate': 5.890909090909091e-06, 'num_tokens': 3059262.0, 'completions/mean_length': 61.375, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.3558826744556427, 'rewards/meter/std': 0.09776563197374344, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.3558826744556427, 'rewards/total_composite/std': 0.09776563197374344, 'reward': 0.3558826744556427, 'reward_std': 0.09776563197374344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04409278184175491, 'sampling/sampling_logp_difference/max': 1.596221923828125, 'sampling/importance_sampling_ratio/min': 0.20266073942184448, 'sampling/importance_sampling_ratio/mean': 1.008901596069336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2319873459637165, 'clip_ratio/low_mean': 0.01653005462139845, 'clip_ratio/low_min': 0.01653005462139845, 'clip_ratio/high_mean': 0.028264103457331657, 'clip_ratio/high_max': 0.028264103457331657, 'clip_ratio/region_mean': 0.044794158078730106, 'reward_total_mean': 0.3558826744556427, 'reward_meter_mean': 0.3558826744556427, 'reward_meter_std': 0.09776563197374344, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.3558826744556427, 'reward_total_composite_std': 0.09776563197374344, 'epoch': 0.05}

 41%|████      | 1357/3300 [1:26:49<3:35:51,  6.67s/it]INFO 04-12 00:03:40 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1358/3300 [1:26:53<3:16:00,  6.06s/it]2026-04-12 00:03:44,754 | INFO | train_grpo_train | metrics_logged mode=train step=1358

                                                       
{'loss': -0.003, 'grad_norm': 6.204784393310547, 'learning_rate': 5.887878787878788e-06, 'num_tokens': 3061057.0, 'completions/mean_length': 76.375, 'completions/min_length': 74.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9842313528060913, 'rewards/meter/std': 0.035786159336566925, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9842313528060913, 'rewards/total_composite/std': 0.035786159336566925, 'reward': 0.9842313528060913, 'reward_std': 0.03578615561127663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042138196527957916, 'sampling/sampling_logp_difference/max': 1.1789560317993164, 'sampling/importance_sampling_ratio/min': 0.3075996935367584, 'sampling/importance_sampling_ratio/mean': 1.0056933164596558, 'sampling/importance_sampling_ratio/max': 1.8513555526733398, 'entropy': 0.24777152389287949, 'clip_ratio/low_mean': 0.0016891892300918698, 'clip_ratio/low_min': 0.0016891892300918698, 'clip_ratio/high_mean': 0.040843427646905184, 'clip_ratio/high_max': 0.040843427646905184, 'clip_ratio/region_mean': 0.042532616876997054, 'reward_total_mean': 0.9842313528060913, 'reward_meter_mean': 0.9842313528060913, 'reward_meter_std': 0.035786159336566925, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9842313528060913, 'reward_total_composite_std': 0.035786159336566925, 'epoch': 0.05}

 41%|████      | 1358/3300 [1:26:53<3:16:00,  6.06s/it]INFO 04-12 00:03:45 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1359/3300 [1:26:58<3:02:01,  5.63s/it]2026-04-12 00:03:49,376 | INFO | train_grpo_train | metrics_logged mode=train step=1359

                                                       
{'loss': 0.0055, 'grad_norm': 6.726247787475586, 'learning_rate': 5.884848484848486e-06, 'num_tokens': 3062858.0, 'completions/mean_length': 72.125, 'completions/min_length': 71.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.938990592956543, 'rewards/meter/std': 0.15821680426597595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.938990592956543, 'rewards/total_composite/std': 0.15821680426597595, 'reward': 0.938990592956543, 'reward_std': 0.15821681916713715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04102391377091408, 'sampling/sampling_logp_difference/max': 1.6344432830810547, 'sampling/importance_sampling_ratio/min': 0.19506093859672546, 'sampling/importance_sampling_ratio/mean': 1.00753915309906, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31099611707031727, 'clip_ratio/low_mean': 0.006849315017461777, 'clip_ratio/low_min': 0.006849315017461777, 'clip_ratio/high_mean': 0.026044346392154694, 'clip_ratio/high_max': 0.026044346392154694, 'clip_ratio/region_mean': 0.03289366140961647, 'reward_total_mean': 0.938990592956543, 'reward_meter_mean': 0.938990592956543, 'reward_meter_std': 0.15821680426597595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.938990592956543, 'reward_total_composite_std': 0.15821680426597595, 'epoch': 0.05}

 41%|████      | 1359/3300 [1:26:58<3:02:01,  5.63s/it]INFO 04-12 00:03:49 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1360/3300 [1:27:03<2:59:05,  5.54s/it]2026-04-12 00:03:54,712 | INFO | train_grpo_train | metrics_logged mode=train step=1360

                                                       
{'loss': 0.0188, 'grad_norm': 6.566722869873047, 'learning_rate': 5.881818181818182e-06, 'num_tokens': 3064544.0, 'completions/mean_length': 69.75, 'completions/min_length': 65.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.75, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9943356513977051, 'rewards/meter/std': 0.011215281672775745, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943356513977051, 'rewards/total_composite/std': 0.011215281672775745, 'reward': 0.9943356513977051, 'reward_std': 0.011215298436582088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07216734439134598, 'sampling/sampling_logp_difference/max': 1.3258538246154785, 'sampling/importance_sampling_ratio/min': 0.2655761241912842, 'sampling/importance_sampling_ratio/mean': 0.9980753064155579, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3794647231698036, 'clip_ratio/low_mean': 0.0069444444961845875, 'clip_ratio/low_min': 0.0069444444961845875, 'clip_ratio/high_mean': 0.06222078762948513, 'clip_ratio/high_max': 0.06222078762948513, 'clip_ratio/region_mean': 0.06916523212566972, 'reward_total_mean': 0.9943356513977051, 'reward_meter_mean': 0.9943356513977051, 'reward_meter_std': 0.011215281672775745, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943356513977051, 'reward_total_composite_std': 0.011215281672775745, 'epoch': 0.05}

 41%|████      | 1360/3300 [1:27:03<2:59:05,  5.54s/it]INFO 04-12 00:03:54 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1361/3300 [1:27:07<2:46:20,  5.15s/it]2026-04-12 00:03:58,946 | INFO | train_grpo_train | metrics_logged mode=train step=1361

                                                       
{'loss': 0.0262, 'grad_norm': 10.907339096069336, 'learning_rate': 5.878787878787879e-06, 'num_tokens': 3066089.0, 'completions/mean_length': 39.125, 'completions/min_length': 37.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8477911949157715, 'rewards/meter/std': 0.32613706588745117, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8477911949157715, 'rewards/total_composite/std': 0.32613706588745117, 'reward': 0.8477911949157715, 'reward_std': 0.3261370360851288, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07346105575561523, 'sampling/sampling_logp_difference/max': 1.893021583557129, 'sampling/importance_sampling_ratio/min': 0.15061601996421814, 'sampling/importance_sampling_ratio/mean': 1.0094305276870728, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40571761690080166, 'clip_ratio/low_mean': 0.0062500000931322575, 'clip_ratio/low_min': 0.0062500000931322575, 'clip_ratio/high_mean': 0.051507277181372046, 'clip_ratio/high_max': 0.051507277181372046, 'clip_ratio/region_mean': 0.057757277274504304, 'reward_total_mean': 0.8477911949157715, 'reward_meter_mean': 0.8477911949157715, 'reward_meter_std': 0.32613706588745117, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8477911949157715, 'reward_total_composite_std': 0.32613706588745117, 'epoch': 0.05}

 41%|████      | 1361/3300 [1:27:07<2:46:20,  5.15s/it]INFO 04-12 00:03:59 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1362/3300 [1:27:12<2:41:12,  4.99s/it]2026-04-12 00:04:03,573 | INFO | train_grpo_train | metrics_logged mode=train step=1362

                                                       
{'loss': 0.0211, 'grad_norm': 5.575023651123047, 'learning_rate': 5.875757575757576e-06, 'num_tokens': 3068064.0, 'completions/mean_length': 72.875, 'completions/min_length': 67.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8556216955184937, 'rewards/meter/std': 0.27104073762893677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.8141105771064758, 'rewards/total_composite/std': 0.27185216546058655, 'reward': 0.8141105771064758, 'reward_std': 0.27185213565826416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07884806394577026, 'sampling/sampling_logp_difference/max': 1.8866146802902222, 'sampling/importance_sampling_ratio/min': 0.15158410370349884, 'sampling/importance_sampling_ratio/mean': 0.9950591921806335, 'sampling/importance_sampling_ratio/max': 1.9909788370132446, 'entropy': 0.3851870410144329, 'clip_ratio/low_mean': 0.015579710714519024, 'clip_ratio/low_min': 0.015579710714519024, 'clip_ratio/high_mean': 0.06273440551012754, 'clip_ratio/high_max': 0.06273440551012754, 'clip_ratio/region_mean': 0.07831411622464657, 'reward_total_mean': 0.8141105771064758, 'reward_meter_mean': 0.8556216955184937, 'reward_meter_std': 0.27104073762893677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.8141105771064758, 'reward_total_composite_std': 0.27185216546058655, 'epoch': 0.05}

 41%|████▏     | 1362/3300 [1:27:12<2:41:12,  4.99s/it]INFO 04-12 00:04:03 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1363/3300 [1:27:16<2:33:53,  4.77s/it]2026-04-12 00:04:07,816 | INFO | train_grpo_train | metrics_logged mode=train step=1363

                                                       
{'loss': 0.0432, 'grad_norm': 14.677414894104004, 'learning_rate': 5.872727272727273e-06, 'num_tokens': 3069688.0, 'completions/mean_length': 32.0, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.601068377494812, 'rewards/meter/std': 0.3853451907634735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.601068377494812, 'rewards/total_composite/std': 0.3853451907634735, 'reward': 0.601068377494812, 'reward_std': 0.3853451907634735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05407094210386276, 'sampling/sampling_logp_difference/max': 1.9056625366210938, 'sampling/importance_sampling_ratio/min': 0.14872407913208008, 'sampling/importance_sampling_ratio/mean': 1.0056836605072021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2922775577753782, 'clip_ratio/low_mean': 0.010135134682059288, 'clip_ratio/low_min': 0.010135134682059288, 'clip_ratio/high_mean': 0.03211413393728435, 'clip_ratio/high_max': 0.03211413393728435, 'clip_ratio/region_mean': 0.04224926861934364, 'reward_total_mean': 0.601068377494812, 'reward_meter_mean': 0.601068377494812, 'reward_meter_std': 0.3853451907634735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.601068377494812, 'reward_total_composite_std': 0.3853451907634735, 'epoch': 0.05}

 41%|████▏     | 1363/3300 [1:27:16<2:33:53,  4.77s/it]INFO 04-12 00:04:08 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1364/3300 [1:27:21<2:32:49,  4.74s/it]2026-04-12 00:04:12,485 | INFO | train_grpo_train | metrics_logged mode=train step=1364

                                                       
{'loss': -0.0013, 'grad_norm': 3.343250274658203, 'learning_rate': 5.8696969696969694e-06, 'num_tokens': 3071527.0, 'completions/mean_length': 60.875, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.987983763217926, 'rewards/meter/std': 0.010048545897006989, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.987983763217926, 'rewards/total_composite/std': 0.010048545897006989, 'reward': 0.987983763217926, 'reward_std': 0.010048558004200459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020703302696347237, 'sampling/sampling_logp_difference/max': 0.8887186050415039, 'sampling/importance_sampling_ratio/min': 0.41118231415748596, 'sampling/importance_sampling_ratio/mean': 0.9991443157196045, 'sampling/importance_sampling_ratio/max': 1.870787501335144, 'entropy': 0.10629169084131718, 'clip_ratio/low_mean': 0.004132513655349612, 'clip_ratio/low_min': 0.004132513655349612, 'clip_ratio/high_mean': 0.010213951813057065, 'clip_ratio/high_max': 0.010213951813057065, 'clip_ratio/region_mean': 0.014346465468406677, 'reward_total_mean': 0.987983763217926, 'reward_meter_mean': 0.987983763217926, 'reward_meter_std': 0.010048545897006989, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.987983763217926, 'reward_total_composite_std': 0.010048545897006989, 'epoch': 0.05}

 41%|████▏     | 1364/3300 [1:27:21<2:32:49,  4.74s/it]INFO 04-12 00:04:12 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1365/3300 [1:27:26<2:34:52,  4.80s/it]2026-04-12 00:04:17,442 | INFO | train_grpo_train | metrics_logged mode=train step=1365

                                                       
{'loss': 0.0227, 'grad_norm': 5.388495445251465, 'learning_rate': 5.8666666666666675e-06, 'num_tokens': 3073505.0, 'completions/mean_length': 91.25, 'completions/min_length': 88.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.25, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.09923887252807617, 'rewards/meter/std': 0.17276984453201294, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.09825273603200912, 'rewards/total_composite/std': 0.17330721020698547, 'reward': 0.09825273603200912, 'reward_std': 0.17330722510814667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05340426787734032, 'sampling/sampling_logp_difference/max': 1.4949496984481812, 'sampling/importance_sampling_ratio/min': 0.2242598831653595, 'sampling/importance_sampling_ratio/mean': 1.003082036972046, 'sampling/importance_sampling_ratio/max': 1.888500452041626, 'entropy': 0.3218696843832731, 'clip_ratio/low_mean': 0.032539316453039646, 'clip_ratio/low_min': 0.032539316453039646, 'clip_ratio/high_mean': 0.01694969367235899, 'clip_ratio/high_max': 0.01694969367235899, 'clip_ratio/region_mean': 0.049489010125398636, 'reward_total_mean': 0.09825273603200912, 'reward_meter_mean': 0.09923887252807617, 'reward_meter_std': 0.17276984453201294, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.09825273603200912, 'reward_total_composite_std': 0.17330721020698547, 'epoch': 0.05}

 41%|████▏     | 1365/3300 [1:27:26<2:34:52,  4.80s/it]INFO 04-12 00:04:17 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1366/3300 [1:27:31<2:35:18,  4.82s/it]2026-04-12 00:04:22,295 | INFO | train_grpo_train | metrics_logged mode=train step=1366

                                                       
{'loss': 0.0137, 'grad_norm': 8.658242225646973, 'learning_rate': 5.863636363636364e-06, 'num_tokens': 3075341.0, 'completions/mean_length': 74.5, 'completions/min_length': 69.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.9018104076385498, 'rewards/meter/std': 0.15291984379291534, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9018104076385498, 'rewards/total_composite/std': 0.15291984379291534, 'reward': 0.9018104076385498, 'reward_std': 0.15291984379291534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08389697223901749, 'sampling/sampling_logp_difference/max': 1.793445110321045, 'sampling/importance_sampling_ratio/min': 0.1663859635591507, 'sampling/importance_sampling_ratio/mean': 1.0000320672988892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37352898344397545, 'clip_ratio/low_mean': 0.019174665678292513, 'clip_ratio/low_min': 0.019174665678292513, 'clip_ratio/high_mean': 0.055123385740444064, 'clip_ratio/high_max': 0.055123385740444064, 'clip_ratio/region_mean': 0.07429805141873658, 'reward_total_mean': 0.9018104076385498, 'reward_meter_mean': 0.9018104076385498, 'reward_meter_std': 0.15291984379291534, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9018104076385498, 'reward_total_composite_std': 0.15291984379291534, 'epoch': 0.05}

 41%|████▏     | 1366/3300 [1:27:31<2:35:18,  4.82s/it]INFO 04-12 00:04:22 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1367/3300 [1:27:35<2:34:07,  4.78s/it]2026-04-12 00:04:26,998 | INFO | train_grpo_train | metrics_logged mode=train step=1367

                                                       
{'loss': 0.0233, 'grad_norm': 7.315425395965576, 'learning_rate': 5.860606060606061e-06, 'num_tokens': 3076962.0, 'completions/mean_length': 62.625, 'completions/min_length': 61.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9573122262954712, 'rewards/meter/std': 0.04094774276018143, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9573122262954712, 'rewards/total_composite/std': 0.04094774276018143, 'reward': 0.9573122262954712, 'reward_std': 0.04094773158431053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03150878846645355, 'sampling/sampling_logp_difference/max': 1.7324086427688599, 'sampling/importance_sampling_ratio/min': 0.17685790359973907, 'sampling/importance_sampling_ratio/mean': 0.994452714920044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09856067411601543, 'clip_ratio/low_mean': 0.013644688995555043, 'clip_ratio/low_min': 0.013644688995555043, 'clip_ratio/high_mean': 0.014278159011155367, 'clip_ratio/high_max': 0.014278159011155367, 'clip_ratio/region_mean': 0.02792284800671041, 'reward_total_mean': 0.9573122262954712, 'reward_meter_mean': 0.9573122262954712, 'reward_meter_std': 0.04094774276018143, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9573122262954712, 'reward_total_composite_std': 0.04094774276018143, 'epoch': 0.05}

 41%|████▏     | 1367/3300 [1:27:35<2:34:07,  4.78s/it]INFO 04-12 00:04:27 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1368/3300 [1:27:42<2:50:44,  5.30s/it]2026-04-12 00:04:33,512 | INFO | train_grpo_train | metrics_logged mode=train step=1368

                                                       
{'loss': 0.0273, 'grad_norm': 3.8533852100372314, 'learning_rate': 5.8575757575757584e-06, 'num_tokens': 3079903.0, 'completions/mean_length': 186.625, 'completions/min_length': 173.0, 'completions/max_length': 218.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.625, 'completions/min_terminated_length': 173.0, 'completions/max_terminated_length': 218.0, 'rewards/meter/mean': 0.7870528697967529, 'rewards/meter/std': 0.30166095495224, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9494949579238892, 'rewards/repeat_penalty/std': 0.07303925603628159, 'rewards/total_composite/mean': 0.644354522228241, 'rewards/total_composite/std': 0.23647885024547577, 'reward': 0.644354522228241, 'reward_std': 0.23647886514663696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07801329344511032, 'sampling/sampling_logp_difference/max': 3.686089038848877, 'sampling/importance_sampling_ratio/min': 0.025069857016205788, 'sampling/importance_sampling_ratio/mean': 0.9927666783332825, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35142070427536964, 'clip_ratio/low_mean': 0.028853498864918947, 'clip_ratio/low_min': 0.028853498864918947, 'clip_ratio/high_mean': 0.03145654499530792, 'clip_ratio/high_max': 0.03145654499530792, 'clip_ratio/region_mean': 0.06031004386022687, 'reward_total_mean': 0.644354522228241, 'reward_meter_mean': 0.7870528697967529, 'reward_meter_std': 0.30166095495224, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9494949579238892, 'reward_repeat_penalty_std': 0.07303925603628159, 'reward_total_composite_mean': 0.644354522228241, 'reward_total_composite_std': 0.23647885024547577, 'epoch': 0.05}

 41%|████▏     | 1368/3300 [1:27:42<2:50:44,  5.30s/it]INFO 04-12 00:04:33 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1369/3300 [1:27:47<2:50:39,  5.30s/it]2026-04-12 00:04:38,814 | INFO | train_grpo_train | metrics_logged mode=train step=1369

                                                       
{'loss': -0.0105, 'grad_norm': 6.314200401306152, 'learning_rate': 5.854545454545455e-06, 'num_tokens': 3082028.0, 'completions/mean_length': 109.625, 'completions/min_length': 104.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.625, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.09159435331821442, 'rewards/meter/std': 0.1886763721704483, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.09159435331821442, 'rewards/total_composite/std': 0.1886763721704483, 'reward': 0.09159435331821442, 'reward_std': 0.1886763721704483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.056608088314533234, 'sampling/sampling_logp_difference/max': 1.6070945262908936, 'sampling/importance_sampling_ratio/min': 0.2004692256450653, 'sampling/importance_sampling_ratio/mean': 1.008489966392517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2641350254416466, 'clip_ratio/low_mean': 0.04172989120706916, 'clip_ratio/low_min': 0.04172989120706916, 'clip_ratio/high_mean': 0.00881484942510724, 'clip_ratio/high_max': 0.00881484942510724, 'clip_ratio/region_mean': 0.0505447406321764, 'reward_total_mean': 0.09159435331821442, 'reward_meter_mean': 0.09159435331821442, 'reward_meter_std': 0.1886763721704483, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.09159435331821442, 'reward_total_composite_std': 0.1886763721704483, 'epoch': 0.05}

 41%|████▏     | 1369/3300 [1:27:47<2:50:39,  5.30s/it]INFO 04-12 00:04:39 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1370/3300 [1:27:53<2:54:57,  5.44s/it]2026-04-12 00:04:44,571 | INFO | train_grpo_train | metrics_logged mode=train step=1370

                                                       
{'loss': 0.0008, 'grad_norm': 2.687344551086426, 'learning_rate': 5.851515151515152e-06, 'num_tokens': 3084733.0, 'completions/mean_length': 147.125, 'completions/min_length': 129.0, 'completions/max_length': 162.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.125, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.639522135257721, 'rewards/meter/std': 0.19242770969867706, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7559523582458496, 'rewards/repeat_penalty/std': 0.17674170434474945, 'rewards/total_composite/mean': 0.4428256154060364, 'rewards/total_composite/std': 0.15113258361816406, 'reward': 0.4428256154060364, 'reward_std': 0.15113258361816406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030883800238370895, 'sampling/sampling_logp_difference/max': 3.218033790588379, 'sampling/importance_sampling_ratio/min': 0.040033698081970215, 'sampling/importance_sampling_ratio/mean': 0.9980568289756775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1116212410852313, 'clip_ratio/low_mean': 0.013002454303205013, 'clip_ratio/low_min': 0.013002454303205013, 'clip_ratio/high_mean': 0.017049296759068966, 'clip_ratio/high_max': 0.017049296759068966, 'clip_ratio/region_mean': 0.03005175106227398, 'reward_total_mean': 0.4428256154060364, 'reward_meter_mean': 0.639522135257721, 'reward_meter_std': 0.19242770969867706, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7559523582458496, 'reward_repeat_penalty_std': 0.17674170434474945, 'reward_total_composite_mean': 0.4428256154060364, 'reward_total_composite_std': 0.15113258361816406, 'epoch': 0.06}

 42%|████▏     | 1370/3300 [1:27:53<2:54:57,  5.44s/it]INFO 04-12 00:04:44 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1371/3300 [1:27:58<2:54:08,  5.42s/it]2026-04-12 00:04:49,934 | INFO | train_grpo_train | metrics_logged mode=train step=1371

                                                       
{'loss': 0.021, 'grad_norm': 3.544004440307617, 'learning_rate': 5.8484848484848485e-06, 'num_tokens': 3087080.0, 'completions/mean_length': 125.375, 'completions/min_length': 119.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.375, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.7613728642463684, 'rewards/meter/std': 0.35931819677352905, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.727212131023407, 'rewards/total_composite/std': 0.34967657923698425, 'reward': 0.727212131023407, 'reward_std': 0.34967654943466187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034298431128263474, 'sampling/sampling_logp_difference/max': 0.8778483867645264, 'sampling/importance_sampling_ratio/min': 0.4156763255596161, 'sampling/importance_sampling_ratio/mean': 1.009077787399292, 'sampling/importance_sampling_ratio/max': 1.806413173675537, 'entropy': 0.18958070501685143, 'clip_ratio/low_mean': 0.010699023492634296, 'clip_ratio/low_min': 0.010699023492634296, 'clip_ratio/high_mean': 0.022983014350757003, 'clip_ratio/high_max': 0.022983014350757003, 'clip_ratio/region_mean': 0.0336820378433913, 'reward_total_mean': 0.727212131023407, 'reward_meter_mean': 0.7613728642463684, 'reward_meter_std': 0.35931819677352905, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.727212131023407, 'reward_total_composite_std': 0.34967657923698425, 'epoch': 0.06}

 42%|████▏     | 1371/3300 [1:27:58<2:54:08,  5.42s/it]INFO 04-12 00:04:50 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1372/3300 [1:28:04<2:56:45,  5.50s/it]2026-04-12 00:04:55,631 | INFO | train_grpo_train | metrics_logged mode=train step=1372

                                                       
{'loss': 0.0079, 'grad_norm': 5.1783366203308105, 'learning_rate': 5.845454545454547e-06, 'num_tokens': 3089717.0, 'completions/mean_length': 153.625, 'completions/min_length': 146.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 153.625, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.05596386641263962, 'rewards/meter/std': 0.034602247178554535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.05596386641263962, 'rewards/total_composite/std': 0.034602247178554535, 'reward': 0.05596386641263962, 'reward_std': 0.03460225090384483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07258274406194687, 'sampling/sampling_logp_difference/max': 5.408177852630615, 'sampling/importance_sampling_ratio/min': 0.004479795694351196, 'sampling/importance_sampling_ratio/mean': 0.9990571141242981, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26759735494852066, 'clip_ratio/low_mean': 0.009019391611218452, 'clip_ratio/low_min': 0.009019391611218452, 'clip_ratio/high_mean': 0.03647972596809268, 'clip_ratio/high_max': 0.03647972596809268, 'clip_ratio/region_mean': 0.04549911757931113, 'reward_total_mean': 0.05596386641263962, 'reward_meter_mean': 0.05596386641263962, 'reward_meter_std': 0.034602247178554535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.05596386641263962, 'reward_total_composite_std': 0.034602247178554535, 'epoch': 0.06}

 42%|████▏     | 1372/3300 [1:28:04<2:56:45,  5.50s/it]INFO 04-12 00:04:55 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1373/3300 [1:28:11<3:12:38,  6.00s/it]2026-04-12 00:05:02,791 | INFO | train_grpo_train | metrics_logged mode=train step=1373

                                                       
{'loss': 0.0156, 'grad_norm': 2.5243654251098633, 'learning_rate': 5.842424242424243e-06, 'num_tokens': 3093370.0, 'completions/mean_length': 265.625, 'completions/min_length': 259.0, 'completions/max_length': 272.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 265.625, 'completions/min_terminated_length': 259.0, 'completions/max_terminated_length': 272.0, 'rewards/meter/mean': 0.9450531005859375, 'rewards/meter/std': 0.09262127429246902, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7980769872665405, 'rewards/repeat_penalty/std': 0.10019000619649887, 'rewards/total_composite/mean': 0.6546376943588257, 'rewards/total_composite/std': 0.05606427788734436, 'reward': 0.6546376943588257, 'reward_std': 0.05606428161263466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036455702036619186, 'sampling/sampling_logp_difference/max': 3.1315479278564453, 'sampling/importance_sampling_ratio/min': 0.043650176376104355, 'sampling/importance_sampling_ratio/mean': 1.0016841888427734, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22261576727032661, 'clip_ratio/low_mean': 0.0139405254740268, 'clip_ratio/low_min': 0.0139405254740268, 'clip_ratio/high_mean': 0.01714563579298556, 'clip_ratio/high_max': 0.01714563579298556, 'clip_ratio/region_mean': 0.031086161267012358, 'reward_total_mean': 0.6546376943588257, 'reward_meter_mean': 0.9450531005859375, 'reward_meter_std': 0.09262127429246902, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7980769872665405, 'reward_repeat_penalty_std': 0.10019000619649887, 'reward_total_composite_mean': 0.6546376943588257, 'reward_total_composite_std': 0.05606427788734436, 'epoch': 0.06}

 42%|████▏     | 1373/3300 [1:28:11<3:12:38,  6.00s/it]INFO 04-12 00:05:03 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1374/3300 [1:28:16<2:58:34,  5.56s/it]2026-04-12 00:05:07,339 | INFO | train_grpo_train | metrics_logged mode=train step=1374

                                                       
{'loss': 0.0443, 'grad_norm': 5.411123752593994, 'learning_rate': 5.83939393939394e-06, 'num_tokens': 3095138.0, 'completions/mean_length': 63.0, 'completions/min_length': 61.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9784355759620667, 'rewards/meter/std': 0.027008963748812675, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9784355759620667, 'rewards/total_composite/std': 0.027008963748812675, 'reward': 0.9784355759620667, 'reward_std': 0.027008960023522377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03236331045627594, 'sampling/sampling_logp_difference/max': 1.732818365097046, 'sampling/importance_sampling_ratio/min': 0.17678546905517578, 'sampling/importance_sampling_ratio/mean': 1.0000766515731812, 'sampling/importance_sampling_ratio/max': 1.9240130186080933, 'entropy': 0.1000348855741322, 'clip_ratio/low_mean': 0.011315496172755957, 'clip_ratio/low_min': 0.011315496172755957, 'clip_ratio/high_mean': 0.020197488833218813, 'clip_ratio/high_max': 0.020197488833218813, 'clip_ratio/region_mean': 0.03151298500597477, 'reward_total_mean': 0.9784355759620667, 'reward_meter_mean': 0.9784355759620667, 'reward_meter_std': 0.027008963748812675, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9784355759620667, 'reward_total_composite_std': 0.027008963748812675, 'epoch': 0.06}

 42%|████▏     | 1374/3300 [1:28:16<2:58:34,  5.56s/it]INFO 04-12 00:05:07 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1375/3300 [1:28:20<2:45:39,  5.16s/it]2026-04-12 00:05:11,568 | INFO | train_grpo_train | metrics_logged mode=train step=1375

                                                       
{'loss': 0.0116, 'grad_norm': 10.806543350219727, 'learning_rate': 5.836363636363637e-06, 'num_tokens': 3096588.0, 'completions/mean_length': 40.25, 'completions/min_length': 36.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9826058149337769, 'rewards/meter/std': 0.01861639879643917, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9826058149337769, 'rewards/total_composite/std': 0.01861639879643917, 'reward': 0.9826058149337769, 'reward_std': 0.018616393208503723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06725956499576569, 'sampling/sampling_logp_difference/max': 1.3660998344421387, 'sampling/importance_sampling_ratio/min': 0.25509995222091675, 'sampling/importance_sampling_ratio/mean': 1.0129060745239258, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47422025352716446, 'clip_ratio/low_mean': 0.00937500037252903, 'clip_ratio/low_min': 0.00937500037252903, 'clip_ratio/high_mean': 0.04311362677253783, 'clip_ratio/high_max': 0.04311362677253783, 'clip_ratio/region_mean': 0.05248862714506686, 'reward_total_mean': 0.9826058149337769, 'reward_meter_mean': 0.9826058149337769, 'reward_meter_std': 0.01861639879643917, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9826058149337769, 'reward_total_composite_std': 0.01861639879643917, 'epoch': 0.06}

 42%|████▏     | 1375/3300 [1:28:20<2:45:39,  5.16s/it]INFO 04-12 00:05:11 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1376/3300 [1:28:27<3:01:42,  5.67s/it]2026-04-12 00:05:18,411 | INFO | train_grpo_train | metrics_logged mode=train step=1376

                                                       
{'loss': 0.026, 'grad_norm': 4.0455732345581055, 'learning_rate': 5.833333333333334e-06, 'num_tokens': 3100217.0, 'completions/mean_length': 245.625, 'completions/min_length': 241.0, 'completions/max_length': 256.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 245.625, 'completions/min_terminated_length': 241.0, 'completions/max_terminated_length': 256.0, 'rewards/meter/mean': 0.735993504524231, 'rewards/meter/std': 0.3604574203491211, 'rewards/count_adherence/mean': 0.7272727489471436, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.10079052299261093, 'rewards/total_composite/mean': 0.4196595549583435, 'rewards/total_composite/std': 0.19926168024539948, 'reward': 0.4196595549583435, 'reward_std': 0.19926169514656067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06053018942475319, 'sampling/sampling_logp_difference/max': 8.299843788146973, 'sampling/importance_sampling_ratio/min': 0.00024855564697645605, 'sampling/importance_sampling_ratio/mean': 0.9991115927696228, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2056045476347208, 'clip_ratio/low_mean': 0.021116425283253193, 'clip_ratio/low_min': 0.021116425283253193, 'clip_ratio/high_mean': 0.019607282243669033, 'clip_ratio/high_max': 0.019607282243669033, 'clip_ratio/region_mean': 0.040723707526922226, 'reward_total_mean': 0.4196595549583435, 'reward_meter_mean': 0.735993504524231, 'reward_meter_std': 0.3604574203491211, 'reward_count_adherence_mean': 0.7272727489471436, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.10079052299261093, 'reward_total_composite_mean': 0.4196595549583435, 'reward_total_composite_std': 0.19926168024539948, 'epoch': 0.06}

 42%|████▏     | 1376/3300 [1:28:27<3:01:42,  5.67s/it]INFO 04-12 00:05:18 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1377/3300 [1:28:32<2:52:49,  5.39s/it]2026-04-12 00:05:23,164 | INFO | train_grpo_train | metrics_logged mode=train step=1377

                                                       
{'loss': 0.0091, 'grad_norm': 5.782622814178467, 'learning_rate': 5.83030303030303e-06, 'num_tokens': 3102173.0, 'completions/mean_length': 79.5, 'completions/min_length': 76.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9965003728866577, 'rewards/meter/std': 0.0018758656224235892, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9965003728866577, 'rewards/total_composite/std': 0.0018758656224235892, 'reward': 0.9965003728866577, 'reward_std': 0.0018758544465526938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.057900648564100266, 'sampling/sampling_logp_difference/max': 2.173585891723633, 'sampling/importance_sampling_ratio/min': 0.11376892775297165, 'sampling/importance_sampling_ratio/mean': 1.003879189491272, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3501878045499325, 'clip_ratio/low_mean': 0.02467968501150608, 'clip_ratio/low_min': 0.02467968501150608, 'clip_ratio/high_mean': 0.028373429435305297, 'clip_ratio/high_max': 0.028373429435305297, 'clip_ratio/region_mean': 0.05305311444681138, 'reward_total_mean': 0.9965003728866577, 'reward_meter_mean': 0.9965003728866577, 'reward_meter_std': 0.0018758656224235892, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9965003728866577, 'reward_total_composite_std': 0.0018758656224235892, 'epoch': 0.06}

 42%|████▏     | 1377/3300 [1:28:32<2:52:49,  5.39s/it]INFO 04-12 00:05:23 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1378/3300 [1:28:38<2:59:13,  5.59s/it]2026-04-12 00:05:29,234 | INFO | train_grpo_train | metrics_logged mode=train step=1378

                                                       
{'loss': -0.0018, 'grad_norm': 6.239497184753418, 'learning_rate': 5.8272727272727285e-06, 'num_tokens': 3104993.0, 'completions/mean_length': 172.5, 'completions/min_length': 142.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 172.5, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9290133714675903, 'rewards/meter/std': 0.17080947756767273, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.8381317853927612, 'rewards/total_composite/std': 0.152348130941391, 'reward': 0.8381317853927612, 'reward_std': 0.1523481160402298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04662080481648445, 'sampling/sampling_logp_difference/max': 2.5756921768188477, 'sampling/importance_sampling_ratio/min': 0.07610113173723221, 'sampling/importance_sampling_ratio/mean': 1.0066179037094116, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2811085060238838, 'clip_ratio/low_mean': 0.010481631616130471, 'clip_ratio/low_min': 0.010481631616130471, 'clip_ratio/high_mean': 0.022769354982301593, 'clip_ratio/high_max': 0.022769354982301593, 'clip_ratio/region_mean': 0.033250986598432064, 'reward_total_mean': 0.8381317853927612, 'reward_meter_mean': 0.9290133714675903, 'reward_meter_std': 0.17080947756767273, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.8381317853927612, 'reward_total_composite_std': 0.152348130941391, 'epoch': 0.06}

 42%|████▏     | 1378/3300 [1:28:38<2:59:13,  5.59s/it]INFO 04-12 00:05:29 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1379/3300 [1:28:44<3:10:07,  5.94s/it]2026-04-12 00:05:35,971 | INFO | train_grpo_train | metrics_logged mode=train step=1379

                                                       
{'loss': -0.0035, 'grad_norm': 2.3529670238494873, 'learning_rate': 5.824242424242425e-06, 'num_tokens': 3108711.0, 'completions/mean_length': 241.75, 'completions/min_length': 226.0, 'completions/max_length': 248.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 241.75, 'completions/min_terminated_length': 226.0, 'completions/max_terminated_length': 248.0, 'rewards/meter/mean': 0.9438363313674927, 'rewards/meter/std': 0.13904310762882233, 'rewards/count_adherence/mean': 0.7272727489471436, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7250000238418579, 'rewards/repeat_penalty/std': 0.115125872194767, 'rewards/total_composite/mean': 0.4973941147327423, 'rewards/total_composite/std': 0.1098327562212944, 'reward': 0.4973941147327423, 'reward_std': 0.1098327562212944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019148869439959526, 'sampling/sampling_logp_difference/max': 1.7155005931854248, 'sampling/importance_sampling_ratio/min': 0.17987366020679474, 'sampling/importance_sampling_ratio/mean': 1.0025800466537476, 'sampling/importance_sampling_ratio/max': 1.9625972509384155, 'entropy': 0.09616232290863991, 'clip_ratio/low_mean': 0.007781740743666887, 'clip_ratio/low_min': 0.007781740743666887, 'clip_ratio/high_mean': 0.006693349685519934, 'clip_ratio/high_max': 0.006693349685519934, 'clip_ratio/region_mean': 0.014475090429186821, 'reward_total_mean': 0.4973941147327423, 'reward_meter_mean': 0.9438363313674927, 'reward_meter_std': 0.13904310762882233, 'reward_count_adherence_mean': 0.7272727489471436, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7250000238418579, 'reward_repeat_penalty_std': 0.115125872194767, 'reward_total_composite_mean': 0.4973941147327423, 'reward_total_composite_std': 0.1098327562212944, 'epoch': 0.06}

 42%|████▏     | 1379/3300 [1:28:44<3:10:07,  5.94s/it]INFO 04-12 00:05:36 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1380/3300 [1:28:49<2:55:51,  5.50s/it]2026-04-12 00:05:40,431 | INFO | train_grpo_train | metrics_logged mode=train step=1380

                                                       
{'loss': 0.0058, 'grad_norm': 3.860546112060547, 'learning_rate': 5.821212121212122e-06, 'num_tokens': 3110495.0, 'completions/mean_length': 69.0, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.014178031124174595, 'rewards/meter/std': 0.004543017130345106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.014178031124174595, 'rewards/total_composite/std': 0.004543017130345106, 'reward': 0.014178031124174595, 'reward_std': 0.004543016664683819, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02862457185983658, 'sampling/sampling_logp_difference/max': 2.9717161655426025, 'sampling/importance_sampling_ratio/min': 0.051215339452028275, 'sampling/importance_sampling_ratio/mean': 0.9993897676467896, 'sampling/importance_sampling_ratio/max': 1.7626913785934448, 'entropy': 0.1250211326405406, 'clip_ratio/low_mean': 0.01270932296756655, 'clip_ratio/low_min': 0.01270932296756655, 'clip_ratio/high_mean': 0.008928571827709675, 'clip_ratio/high_max': 0.008928571827709675, 'clip_ratio/region_mean': 0.021637894795276225, 'reward_total_mean': 0.014178031124174595, 'reward_meter_mean': 0.014178031124174595, 'reward_meter_std': 0.004543017130345106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.014178031124174595, 'reward_total_composite_std': 0.004543017130345106, 'epoch': 0.06}

 42%|████▏     | 1380/3300 [1:28:49<2:55:51,  5.50s/it]INFO 04-12 00:05:40 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1381/3300 [1:28:54<2:51:47,  5.37s/it]2026-04-12 00:05:45,515 | INFO | train_grpo_train | metrics_logged mode=train step=1381

                                                       
{'loss': 0.0024, 'grad_norm': 1.1249967813491821, 'learning_rate': 5.8181818181818185e-06, 'num_tokens': 3112713.0, 'completions/mean_length': 115.25, 'completions/min_length': 115.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.25, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9941756725311279, 'rewards/meter/std': 0.00028240663232281804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.3214285969734192, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.319552481174469, 'rewards/total_composite/std': 0.0657178983092308, 'reward': 0.319552481174469, 'reward_std': 0.06571788340806961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009904337115585804, 'sampling/sampling_logp_difference/max': 1.0573194026947021, 'sampling/importance_sampling_ratio/min': 0.3473857343196869, 'sampling/importance_sampling_ratio/mean': 0.9989582896232605, 'sampling/importance_sampling_ratio/max': 1.2996612787246704, 'entropy': 0.039332801941782236, 'clip_ratio/low_mean': 0.003251499147154391, 'clip_ratio/low_min': 0.003251499147154391, 'clip_ratio/high_mean': 0.003260869416408241, 'clip_ratio/high_max': 0.003260869416408241, 'clip_ratio/region_mean': 0.006512368563562632, 'reward_total_mean': 0.319552481174469, 'reward_meter_mean': 0.9941756725311279, 'reward_meter_std': 0.00028240663232281804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.3214285969734192, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.319552481174469, 'reward_total_composite_std': 0.0657178983092308, 'epoch': 0.06}

 42%|████▏     | 1381/3300 [1:28:54<2:51:47,  5.37s/it]INFO 04-12 00:05:45 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1382/3300 [1:28:59<2:52:52,  5.41s/it]2026-04-12 00:05:51,006 | INFO | train_grpo_train | metrics_logged mode=train step=1382

                                                       
{'loss': 0.0069, 'grad_norm': 4.528812408447266, 'learning_rate': 5.815151515151516e-06, 'num_tokens': 3115264.0, 'completions/mean_length': 151.875, 'completions/min_length': 149.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.875, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.9926921725273132, 'rewards/meter/std': 0.008148393593728542, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7857142686843872, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.7798627614974976, 'rewards/total_composite/std': 0.07480987906455994, 'reward': 0.7798627614974976, 'reward_std': 0.07480985671281815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030164044350385666, 'sampling/sampling_logp_difference/max': 1.2373886108398438, 'sampling/importance_sampling_ratio/min': 0.29014089703559875, 'sampling/importance_sampling_ratio/mean': 1.0049320459365845, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1818629987537861, 'clip_ratio/low_mean': 0.0058121298789046705, 'clip_ratio/low_min': 0.0058121298789046705, 'clip_ratio/high_mean': 0.017227754462510347, 'clip_ratio/high_max': 0.017227754462510347, 'clip_ratio/region_mean': 0.023039884341415018, 'reward_total_mean': 0.7798627614974976, 'reward_meter_mean': 0.9926921725273132, 'reward_meter_std': 0.008148393593728542, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7857142686843872, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.7798627614974976, 'reward_total_composite_std': 0.07480987906455994, 'epoch': 0.06}

 42%|████▏     | 1382/3300 [1:29:00<2:52:52,  5.41s/it]INFO 04-12 00:05:51 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1383/3300 [1:29:04<2:48:33,  5.28s/it]2026-04-12 00:05:55,976 | INFO | train_grpo_train | metrics_logged mode=train step=1383

                                                       
{'loss': 0.0148, 'grad_norm': 3.192974805831909, 'learning_rate': 5.812121212121212e-06, 'num_tokens': 3117462.0, 'completions/mean_length': 103.75, 'completions/min_length': 102.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.75, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9773154258728027, 'rewards/meter/std': 0.0542588047683239, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9773154258728027, 'rewards/total_composite/std': 0.0542588047683239, 'reward': 0.9773154258728027, 'reward_std': 0.0542587973177433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03143962472677231, 'sampling/sampling_logp_difference/max': 1.3268545866012573, 'sampling/importance_sampling_ratio/min': 0.331994891166687, 'sampling/importance_sampling_ratio/mean': 1.0025712251663208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18817367777228355, 'clip_ratio/low_mean': 0.001168224262073636, 'clip_ratio/low_min': 0.001168224262073636, 'clip_ratio/high_mean': 0.02543990476988256, 'clip_ratio/high_max': 0.02543990476988256, 'clip_ratio/region_mean': 0.026608129031956196, 'reward_total_mean': 0.9773154258728027, 'reward_meter_mean': 0.9773154258728027, 'reward_meter_std': 0.0542588047683239, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9773154258728027, 'reward_total_composite_std': 0.0542588047683239, 'epoch': 0.06}

 42%|████▏     | 1383/3300 [1:29:04<2:48:33,  5.28s/it]INFO 04-12 00:05:56 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1384/3300 [1:29:09<2:42:38,  5.09s/it]2026-04-12 00:06:00,642 | INFO | train_grpo_train | metrics_logged mode=train step=1384

                                                       
{'loss': 0.0122, 'grad_norm': 7.283441543579102, 'learning_rate': 5.8090909090909095e-06, 'num_tokens': 3119369.0, 'completions/mean_length': 69.375, 'completions/min_length': 66.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.99678635597229, 'rewards/meter/std': 0.001158631988801062, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99678635597229, 'rewards/total_composite/std': 0.001158631988801062, 'reward': 0.99678635597229, 'reward_std': 0.001158619881607592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042962342500686646, 'sampling/sampling_logp_difference/max': 4.707744598388672, 'sampling/importance_sampling_ratio/min': 0.009025109931826591, 'sampling/importance_sampling_ratio/mean': 1.0015077590942383, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13980651926249266, 'clip_ratio/low_mean': 0.014534936985000968, 'clip_ratio/low_min': 0.014534936985000968, 'clip_ratio/high_mean': 0.01638579391874373, 'clip_ratio/high_max': 0.01638579391874373, 'clip_ratio/region_mean': 0.030920730903744698, 'reward_total_mean': 0.99678635597229, 'reward_meter_mean': 0.99678635597229, 'reward_meter_std': 0.001158631988801062, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99678635597229, 'reward_total_composite_std': 0.001158631988801062, 'epoch': 0.06}

 42%|████▏     | 1384/3300 [1:29:09<2:42:38,  5.09s/it]INFO 04-12 00:06:00 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1385/3300 [1:29:13<2:35:13,  4.86s/it]2026-04-12 00:06:04,971 | INFO | train_grpo_train | metrics_logged mode=train step=1385

                                                       
{'loss': 0.0196, 'grad_norm': 8.730010032653809, 'learning_rate': 5.806060606060606e-06, 'num_tokens': 3120803.0, 'completions/mean_length': 32.25, 'completions/min_length': 31.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9718137979507446, 'rewards/meter/std': 0.05166900157928467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9718137979507446, 'rewards/total_composite/std': 0.05166900157928467, 'reward': 0.9718137979507446, 'reward_std': 0.051669009029865265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01950211450457573, 'sampling/sampling_logp_difference/max': 0.6880896091461182, 'sampling/importance_sampling_ratio/min': 0.5025351643562317, 'sampling/importance_sampling_ratio/mean': 1.0018343925476074, 'sampling/importance_sampling_ratio/max': 1.604980230331421, 'entropy': 0.1377080613747239, 'clip_ratio/low_mean': 0.007464349502697587, 'clip_ratio/low_min': 0.007464349502697587, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.011370599502697587, 'reward_total_mean': 0.9718137979507446, 'reward_meter_mean': 0.9718137979507446, 'reward_meter_std': 0.05166900157928467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9718137979507446, 'reward_total_composite_std': 0.05166900157928467, 'epoch': 0.06}

 42%|████▏     | 1385/3300 [1:29:13<2:35:13,  4.86s/it]INFO 04-12 00:06:05 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1386/3300 [1:29:18<2:32:47,  4.79s/it]2026-04-12 00:06:09,589 | INFO | train_grpo_train | metrics_logged mode=train step=1386

                                                       
{'loss': -0.0023, 'grad_norm': 5.677158832550049, 'learning_rate': 5.803030303030304e-06, 'num_tokens': 3122583.0, 'completions/mean_length': 58.5, 'completions/min_length': 55.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.307847261428833, 'rewards/meter/std': 0.2652658224105835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.307847261428833, 'rewards/total_composite/std': 0.2652658224105835, 'reward': 0.307847261428833, 'reward_std': 0.2652658224105835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04078377038240433, 'sampling/sampling_logp_difference/max': 1.0301332473754883, 'sampling/importance_sampling_ratio/min': 0.35695940256118774, 'sampling/importance_sampling_ratio/mean': 1.0097620487213135, 'sampling/importance_sampling_ratio/max': 1.7868449687957764, 'entropy': 0.29168289341032505, 'clip_ratio/low_mean': 0.017045454820618033, 'clip_ratio/low_min': 0.017045454820618033, 'clip_ratio/high_mean': 0.010490705259144306, 'clip_ratio/high_max': 0.010490705259144306, 'clip_ratio/region_mean': 0.02753616007976234, 'reward_total_mean': 0.307847261428833, 'reward_meter_mean': 0.307847261428833, 'reward_meter_std': 0.2652658224105835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.307847261428833, 'reward_total_composite_std': 0.2652658224105835, 'epoch': 0.06}

 42%|████▏     | 1386/3300 [1:29:18<2:32:47,  4.79s/it]INFO 04-12 00:06:09 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1387/3300 [1:29:25<2:50:59,  5.36s/it]2026-04-12 00:06:16,288 | INFO | train_grpo_train | metrics_logged mode=train step=1387

                                                       
{'loss': 0.0117, 'grad_norm': 3.5950727462768555, 'learning_rate': 5.8e-06, 'num_tokens': 3125864.0, 'completions/mean_length': 223.125, 'completions/min_length': 206.0, 'completions/max_length': 235.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 223.125, 'completions/min_terminated_length': 206.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.9973956346511841, 'rewards/meter/std': 0.0017723083728924394, 'rewards/count_adherence/mean': 0.8055555820465088, 'rewards/count_adherence/std': 0.05143444985151291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6742216348648071, 'rewards/repeat_penalty/std': 0.1199769675731659, 'rewards/total_composite/mean': 0.5415959358215332, 'rewards/total_composite/std': 0.10306404531002045, 'reward': 0.5415959358215332, 'reward_std': 0.10306404531002045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058670658618211746, 'sampling/sampling_logp_difference/max': 17.84847640991211, 'sampling/importance_sampling_ratio/min': 1.7721692557870483e-08, 'sampling/importance_sampling_ratio/mean': 1.0015687942504883, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16380778048187494, 'clip_ratio/low_mean': 0.011592080816626549, 'clip_ratio/low_min': 0.011592080816626549, 'clip_ratio/high_mean': 0.01465759938582778, 'clip_ratio/high_max': 0.01465759938582778, 'clip_ratio/region_mean': 0.02624968020245433, 'reward_total_mean': 0.5415959358215332, 'reward_meter_mean': 0.9973956346511841, 'reward_meter_std': 0.0017723083728924394, 'reward_count_adherence_mean': 0.8055555820465088, 'reward_count_adherence_std': 0.05143444985151291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6742216348648071, 'reward_repeat_penalty_std': 0.1199769675731659, 'reward_total_composite_mean': 0.5415959358215332, 'reward_total_composite_std': 0.10306404531002045, 'epoch': 0.06}

 42%|████▏     | 1387/3300 [1:29:25<2:50:59,  5.36s/it]INFO 04-12 00:06:16 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1388/3300 [1:29:29<2:44:47,  5.17s/it]2026-04-12 00:06:21,012 | INFO | train_grpo_train | metrics_logged mode=train step=1388

                                                       
{'loss': 0.0122, 'grad_norm': 3.6467227935791016, 'learning_rate': 5.796969696969698e-06, 'num_tokens': 3127767.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9714630842208862, 'rewards/meter/std': 0.009149931371212006, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9714630842208862, 'rewards/total_composite/std': 0.009149931371212006, 'reward': 0.9714630842208862, 'reward_std': 0.009149930439889431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01836930215358734, 'sampling/sampling_logp_difference/max': 0.6224894523620605, 'sampling/importance_sampling_ratio/min': 0.5366069078445435, 'sampling/importance_sampling_ratio/mean': 1.0066375732421875, 'sampling/importance_sampling_ratio/max': 1.4535624980926514, 'entropy': 0.12053523678332567, 'clip_ratio/low_mean': 0.007462686393409967, 'clip_ratio/low_min': 0.007462686393409967, 'clip_ratio/high_mean': 0.009384893695823848, 'clip_ratio/high_max': 0.009384893695823848, 'clip_ratio/region_mean': 0.016847580089233816, 'reward_total_mean': 0.9714630842208862, 'reward_meter_mean': 0.9714630842208862, 'reward_meter_std': 0.009149931371212006, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9714630842208862, 'reward_total_composite_std': 0.009149931371212006, 'epoch': 0.06}

 42%|████▏     | 1388/3300 [1:29:30<2:44:47,  5.17s/it]INFO 04-12 00:06:21 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1389/3300 [1:29:37<3:05:48,  5.83s/it]2026-04-12 00:06:28,393 | INFO | train_grpo_train | metrics_logged mode=train step=1389

                                                       
{'loss': -0.0159, 'grad_norm': 2.405141592025757, 'learning_rate': 5.793939393939394e-06, 'num_tokens': 3131978.0, 'completions/mean_length': 276.375, 'completions/min_length': 259.0, 'completions/max_length': 295.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 276.375, 'completions/min_terminated_length': 259.0, 'completions/max_terminated_length': 295.0, 'rewards/meter/mean': 0.9949830770492554, 'rewards/meter/std': 0.0029864103998988867, 'rewards/count_adherence/mean': 0.7272727489471436, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7510416507720947, 'rewards/repeat_penalty/std': 0.10675939172506332, 'rewards/total_composite/mean': 0.5435611009597778, 'rewards/total_composite/std': 0.07788535207509995, 'reward': 0.5435611009597778, 'reward_std': 0.07788533717393875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028575100004673004, 'sampling/sampling_logp_difference/max': 1.6693830490112305, 'sampling/importance_sampling_ratio/min': 0.1883632391691208, 'sampling/importance_sampling_ratio/mean': 1.002191185951233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1869909018278122, 'clip_ratio/low_mean': 0.010067827010061592, 'clip_ratio/low_min': 0.010067827010061592, 'clip_ratio/high_mean': 0.009672214509919286, 'clip_ratio/high_max': 0.009672214509919286, 'clip_ratio/region_mean': 0.019740041519980878, 'reward_total_mean': 0.5435611009597778, 'reward_meter_mean': 0.9949830770492554, 'reward_meter_std': 0.0029864103998988867, 'reward_count_adherence_mean': 0.7272727489471436, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7510416507720947, 'reward_repeat_penalty_std': 0.10675939172506332, 'reward_total_composite_mean': 0.5435611009597778, 'reward_total_composite_std': 0.07788535207509995, 'epoch': 0.06}

 42%|████▏     | 1389/3300 [1:29:37<3:05:48,  5.83s/it]INFO 04-12 00:06:28 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1390/3300 [1:29:42<2:54:40,  5.49s/it]2026-04-12 00:06:33,070 | INFO | train_grpo_train | metrics_logged mode=train step=1390

                                                       
{'loss': 0.0238, 'grad_norm': 6.391792297363281, 'learning_rate': 5.790909090909091e-06, 'num_tokens': 3133718.0, 'completions/mean_length': 67.5, 'completions/min_length': 64.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.11840285360813141, 'rewards/meter/std': 0.15882401168346405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.11840285360813141, 'rewards/total_composite/std': 0.15882401168346405, 'reward': 0.11840285360813141, 'reward_std': 0.15882402658462524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04550213739275932, 'sampling/sampling_logp_difference/max': 1.3136234283447266, 'sampling/importance_sampling_ratio/min': 0.2688441574573517, 'sampling/importance_sampling_ratio/mean': 1.00211763381958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28015392273664474, 'clip_ratio/low_mean': 0.01820334349758923, 'clip_ratio/low_min': 0.01820334349758923, 'clip_ratio/high_mean': 0.011374081019312143, 'clip_ratio/high_max': 0.011374081019312143, 'clip_ratio/region_mean': 0.029577424516901374, 'reward_total_mean': 0.11840285360813141, 'reward_meter_mean': 0.11840285360813141, 'reward_meter_std': 0.15882401168346405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.11840285360813141, 'reward_total_composite_std': 0.15882401168346405, 'epoch': 0.06}

 42%|████▏     | 1390/3300 [1:29:42<2:54:40,  5.49s/it]INFO 04-12 00:06:33 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1391/3300 [1:29:46<2:47:41,  5.27s/it]2026-04-12 00:06:37,835 | INFO | train_grpo_train | metrics_logged mode=train step=1391

                                                       
{'loss': -0.0043, 'grad_norm': 5.131795406341553, 'learning_rate': 5.787878787878788e-06, 'num_tokens': 3135632.0, 'completions/mean_length': 76.25, 'completions/min_length': 71.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9707291722297668, 'rewards/meter/std': 0.025657914578914642, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9707291722297668, 'rewards/total_composite/std': 0.025657914578914642, 'reward': 0.9707291722297668, 'reward_std': 0.02565791644155979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0554119348526001, 'sampling/sampling_logp_difference/max': 1.9454677104949951, 'sampling/importance_sampling_ratio/min': 0.1429203599691391, 'sampling/importance_sampling_ratio/mean': 0.9987998008728027, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3864719457924366, 'clip_ratio/low_mean': 0.005090707214549184, 'clip_ratio/low_min': 0.005090707214549184, 'clip_ratio/high_mean': 0.04821625351905823, 'clip_ratio/high_max': 0.04821625351905823, 'clip_ratio/region_mean': 0.05330696073360741, 'reward_total_mean': 0.9707291722297668, 'reward_meter_mean': 0.9707291722297668, 'reward_meter_std': 0.025657914578914642, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9707291722297668, 'reward_total_composite_std': 0.025657914578914642, 'epoch': 0.06}

 42%|████▏     | 1391/3300 [1:29:46<2:47:41,  5.27s/it]INFO 04-12 00:06:38 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1392/3300 [1:29:51<2:40:55,  5.06s/it]2026-04-12 00:06:42,406 | INFO | train_grpo_train | metrics_logged mode=train step=1392

                                                       
{'loss': 0.0218, 'grad_norm': 11.628338813781738, 'learning_rate': 5.784848484848486e-06, 'num_tokens': 3137481.0, 'completions/mean_length': 65.125, 'completions/min_length': 63.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9977819919586182, 'rewards/meter/std': 0.0022414319682866335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977819919586182, 'rewards/total_composite/std': 0.0022414319682866335, 'reward': 0.9977819919586182, 'reward_std': 0.0022414233535528183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035027600824832916, 'sampling/sampling_logp_difference/max': 0.8262596130371094, 'sampling/importance_sampling_ratio/min': 0.4376833438873291, 'sampling/importance_sampling_ratio/mean': 1.0126001834869385, 'sampling/importance_sampling_ratio/max': 1.7962960004806519, 'entropy': 0.2185972724109888, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.041855036513879895, 'clip_ratio/high_max': 0.041855036513879895, 'clip_ratio/region_mean': 0.04553150711581111, 'reward_total_mean': 0.9977819919586182, 'reward_meter_mean': 0.9977819919586182, 'reward_meter_std': 0.0022414319682866335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977819919586182, 'reward_total_composite_std': 0.0022414319682866335, 'epoch': 0.06}

 42%|████▏     | 1392/3300 [1:29:51<2:40:55,  5.06s/it]INFO 04-12 00:06:42 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1393/3300 [1:29:55<2:35:30,  4.89s/it]2026-04-12 00:06:46,912 | INFO | train_grpo_train | metrics_logged mode=train step=1393

                                                       
{'loss': -0.0412, 'grad_norm': 8.172205924987793, 'learning_rate': 5.781818181818181e-06, 'num_tokens': 3139228.0, 'completions/mean_length': 58.375, 'completions/min_length': 46.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.48125261068344116, 'rewards/meter/std': 0.2949608266353607, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.45941346883773804, 'rewards/total_composite/std': 0.29778623580932617, 'reward': 0.45941346883773804, 'reward_std': 0.29778623580932617, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07994799315929413, 'sampling/sampling_logp_difference/max': 1.2517375946044922, 'sampling/importance_sampling_ratio/min': 0.2860074043273926, 'sampling/importance_sampling_ratio/mean': 1.0176113843917847, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6722332537174225, 'clip_ratio/low_mean': 0.03256637742742896, 'clip_ratio/low_min': 0.03256637742742896, 'clip_ratio/high_mean': 0.032872630283236504, 'clip_ratio/high_max': 0.032872630283236504, 'clip_ratio/region_mean': 0.06543900771066546, 'reward_total_mean': 0.45941346883773804, 'reward_meter_mean': 0.48125261068344116, 'reward_meter_std': 0.2949608266353607, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.45941346883773804, 'reward_total_composite_std': 0.29778623580932617, 'epoch': 0.06}

 42%|████▏     | 1393/3300 [1:29:55<2:35:30,  4.89s/it]INFO 04-12 00:06:47 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1394/3300 [1:30:00<2:34:29,  4.86s/it]2026-04-12 00:06:51,704 | INFO | train_grpo_train | metrics_logged mode=train step=1394

                                                       
{'loss': 0.1048, 'grad_norm': 5.719612121582031, 'learning_rate': 5.7787878787878795e-06, 'num_tokens': 3141241.0, 'completions/mean_length': 75.625, 'completions/min_length': 66.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.625, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5849511027336121, 'rewards/meter/std': 0.4843460023403168, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.824999988079071, 'rewards/repeat_penalty/std': 0.19820624589920044, 'rewards/total_composite/mean': 0.4432606101036072, 'rewards/total_composite/std': 0.39577049016952515, 'reward': 0.4432606101036072, 'reward_std': 0.39577043056488037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023130640387535095, 'sampling/sampling_logp_difference/max': 0.8899397850036621, 'sampling/importance_sampling_ratio/min': 0.41068050265312195, 'sampling/importance_sampling_ratio/mean': 1.0027085542678833, 'sampling/importance_sampling_ratio/max': 1.5665262937545776, 'entropy': 0.136459331959486, 'clip_ratio/low_mean': 0.011919185984879732, 'clip_ratio/low_min': 0.011919185984879732, 'clip_ratio/high_mean': 0.01639902195893228, 'clip_ratio/high_max': 0.01639902195893228, 'clip_ratio/region_mean': 0.028318207943812013, 'reward_total_mean': 0.4432606101036072, 'reward_meter_mean': 0.5849511027336121, 'reward_meter_std': 0.4843460023403168, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.824999988079071, 'reward_repeat_penalty_std': 0.19820624589920044, 'reward_total_composite_mean': 0.4432606101036072, 'reward_total_composite_std': 0.39577049016952515, 'epoch': 0.06}

 42%|████▏     | 1394/3300 [1:30:00<2:34:29,  4.86s/it]INFO 04-12 00:06:51 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1395/3300 [1:30:05<2:32:16,  4.80s/it]2026-04-12 00:06:56,346 | INFO | train_grpo_train | metrics_logged mode=train step=1395

                                                       
{'loss': 0.0056, 'grad_norm': 4.375842094421387, 'learning_rate': 5.775757575757577e-06, 'num_tokens': 3143152.0, 'completions/mean_length': 69.875, 'completions/min_length': 64.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9944968223571777, 'rewards/meter/std': 0.0024136791471391916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9529762864112854, 'rewards/total_composite/std': 0.11665359884500504, 'reward': 0.9529762864112854, 'reward_std': 0.11665360629558563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04802675172686577, 'sampling/sampling_logp_difference/max': 1.7336769104003906, 'sampling/importance_sampling_ratio/min': 0.17663374543190002, 'sampling/importance_sampling_ratio/mean': 1.009644627571106, 'sampling/importance_sampling_ratio/max': 1.78848135471344, 'entropy': 0.37176400050520897, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.03542683261912316, 'clip_ratio/high_max': 0.03542683261912316, 'clip_ratio/region_mean': 0.0372125469148159, 'reward_total_mean': 0.9529762864112854, 'reward_meter_mean': 0.9944968223571777, 'reward_meter_std': 0.0024136791471391916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9529762864112854, 'reward_total_composite_std': 0.11665359884500504, 'epoch': 0.06}

 42%|████▏     | 1395/3300 [1:30:05<2:32:16,  4.80s/it]INFO 04-12 00:06:56 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1396/3300 [1:30:13<3:02:09,  5.74s/it]2026-04-12 00:07:04,286 | INFO | train_grpo_train | metrics_logged mode=train step=1396

                                                       
{'loss': -0.0037, 'grad_norm': 2.749262809753418, 'learning_rate': 5.772727272727273e-06, 'num_tokens': 3147511.0, 'completions/mean_length': 316.875, 'completions/min_length': 296.0, 'completions/max_length': 346.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 316.875, 'completions/min_terminated_length': 296.0, 'completions/max_terminated_length': 346.0, 'rewards/meter/mean': 0.323222815990448, 'rewards/meter/std': 0.21111111342906952, 'rewards/count_adherence/mean': 0.6111111044883728, 'rewards/count_adherence/std': 0.029695691540837288, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7082297801971436, 'rewards/repeat_penalty/std': 0.0814073234796524, 'rewards/total_composite/mean': 0.13397403061389923, 'rewards/total_composite/std': 0.07767514884471893, 'reward': 0.13397403061389923, 'reward_std': 0.07767514139413834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027990715578198433, 'sampling/sampling_logp_difference/max': 4.105655193328857, 'sampling/importance_sampling_ratio/min': 0.016479218378663063, 'sampling/importance_sampling_ratio/mean': 0.9997315406799316, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0825919434428215, 'clip_ratio/low_mean': 0.010312073398381472, 'clip_ratio/low_min': 0.010312073398381472, 'clip_ratio/high_mean': 0.012935725739225745, 'clip_ratio/high_max': 0.012935725739225745, 'clip_ratio/region_mean': 0.023247799137607217, 'reward_total_mean': 0.13397403061389923, 'reward_meter_mean': 0.323222815990448, 'reward_meter_std': 0.21111111342906952, 'reward_count_adherence_mean': 0.6111111044883728, 'reward_count_adherence_std': 0.029695691540837288, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7082297801971436, 'reward_repeat_penalty_std': 0.0814073234796524, 'reward_total_composite_mean': 0.13397403061389923, 'reward_total_composite_std': 0.07767514884471893, 'epoch': 0.06}

 42%|████▏     | 1396/3300 [1:30:13<3:02:09,  5.74s/it]INFO 04-12 00:07:04 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1397/3300 [1:30:17<2:51:46,  5.42s/it]2026-04-12 00:07:08,945 | INFO | train_grpo_train | metrics_logged mode=train step=1397

                                                       
{'loss': 0.0166, 'grad_norm': 4.050926208496094, 'learning_rate': 5.76969696969697e-06, 'num_tokens': 3149369.0, 'completions/mean_length': 70.25, 'completions/min_length': 66.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9413079023361206, 'rewards/meter/std': 0.009388444945216179, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.753129243850708, 'rewards/total_composite/std': 0.10083267837762833, 'reward': 0.753129243850708, 'reward_std': 0.10083267837762833, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018803365528583527, 'sampling/sampling_logp_difference/max': 0.8203344345092773, 'sampling/importance_sampling_ratio/min': 0.44028440117836, 'sampling/importance_sampling_ratio/mean': 0.9976539611816406, 'sampling/importance_sampling_ratio/max': 1.494532585144043, 'entropy': 0.08694514073431492, 'clip_ratio/low_mean': 0.0034246575087308884, 'clip_ratio/low_min': 0.0034246575087308884, 'clip_ratio/high_mean': 0.00919352809432894, 'clip_ratio/high_max': 0.00919352809432894, 'clip_ratio/region_mean': 0.012618185603059828, 'reward_total_mean': 0.753129243850708, 'reward_meter_mean': 0.9413079023361206, 'reward_meter_std': 0.009388444945216179, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.753129243850708, 'reward_total_composite_std': 0.10083267837762833, 'epoch': 0.06}

 42%|████▏     | 1397/3300 [1:30:17<2:51:46,  5.42s/it]INFO 04-12 00:07:09 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1398/3300 [1:30:22<2:43:23,  5.15s/it]2026-04-12 00:07:13,486 | INFO | train_grpo_train | metrics_logged mode=train step=1398

                                                       
{'loss': -0.0167, 'grad_norm': 4.091942310333252, 'learning_rate': 5.766666666666667e-06, 'num_tokens': 3151224.0, 'completions/mean_length': 58.875, 'completions/min_length': 54.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6508887410163879, 'rewards/meter/std': 0.1606731116771698, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6508887410163879, 'rewards/total_composite/std': 0.1606731116771698, 'reward': 0.6508887410163879, 'reward_std': 0.1606731414794922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02822190895676613, 'sampling/sampling_logp_difference/max': 0.8628159761428833, 'sampling/importance_sampling_ratio/min': 0.4910596013069153, 'sampling/importance_sampling_ratio/mean': 1.0124748945236206, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17077440861612558, 'clip_ratio/low_mean': 0.012992122676223516, 'clip_ratio/low_min': 0.012992122676223516, 'clip_ratio/high_mean': 0.0061509020160883665, 'clip_ratio/high_max': 0.0061509020160883665, 'clip_ratio/region_mean': 0.019143024692311883, 'reward_total_mean': 0.6508887410163879, 'reward_meter_mean': 0.6508887410163879, 'reward_meter_std': 0.1606731116771698, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6508887410163879, 'reward_total_composite_std': 0.1606731116771698, 'epoch': 0.06}

 42%|████▏     | 1398/3300 [1:30:22<2:43:23,  5.15s/it]INFO 04-12 00:07:13 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1399/3300 [1:30:26<2:37:24,  4.97s/it]2026-04-12 00:07:18,023 | INFO | train_grpo_train | metrics_logged mode=train step=1399

                                                       
{'loss': 0.0332, 'grad_norm': 7.159199237823486, 'learning_rate': 5.763636363636365e-06, 'num_tokens': 3152918.0, 'completions/mean_length': 63.75, 'completions/min_length': 61.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9983058571815491, 'rewards/meter/std': 0.00029361830092966557, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9983058571815491, 'rewards/total_composite/std': 0.00029361830092966557, 'reward': 0.9983058571815491, 'reward_std': 0.000293627759674564, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03937718644738197, 'sampling/sampling_logp_difference/max': 2.0187900066375732, 'sampling/importance_sampling_ratio/min': 0.13281607627868652, 'sampling/importance_sampling_ratio/mean': 1.0061930418014526, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1746031977236271, 'clip_ratio/low_mean': 0.00562528264708817, 'clip_ratio/low_min': 0.00562528264708817, 'clip_ratio/high_mean': 0.011937764240428805, 'clip_ratio/high_max': 0.011937764240428805, 'clip_ratio/region_mean': 0.017563046887516975, 'reward_total_mean': 0.9983058571815491, 'reward_meter_mean': 0.9983058571815491, 'reward_meter_std': 0.00029361830092966557, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9983058571815491, 'reward_total_composite_std': 0.00029361830092966557, 'epoch': 0.06}

 42%|████▏     | 1399/3300 [1:30:27<2:37:24,  4.97s/it]INFO 04-12 00:07:18 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1400/3300 [1:30:32<2:44:12,  5.19s/it]2026-04-12 00:07:23,715 | INFO | train_grpo_train | metrics_logged mode=train step=1400

                                                       
{'loss': 0.0063, 'grad_norm': 3.5979604721069336, 'learning_rate': 5.760606060606061e-06, 'num_tokens': 3155594.0, 'completions/mean_length': 165.5, 'completions/min_length': 159.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.5, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.8759403228759766, 'rewards/meter/std': 0.19313931465148926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9027777910232544, 'rewards/repeat_penalty/std': 0.0927247703075409, 'rewards/total_composite/mean': 0.7801923751831055, 'rewards/total_composite/std': 0.1476370245218277, 'reward': 0.7801923751831055, 'reward_std': 0.1476370245218277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05515680089592934, 'sampling/sampling_logp_difference/max': 1.6768007278442383, 'sampling/importance_sampling_ratio/min': 0.18697118759155273, 'sampling/importance_sampling_ratio/mean': 1.0012664794921875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4123094528913498, 'clip_ratio/low_mean': 0.022338663460686803, 'clip_ratio/low_min': 0.022338663460686803, 'clip_ratio/high_mean': 0.027914301492273808, 'clip_ratio/high_max': 0.027914301492273808, 'clip_ratio/region_mean': 0.05025296495296061, 'reward_total_mean': 0.7801923751831055, 'reward_meter_mean': 0.8759403228759766, 'reward_meter_std': 0.19313931465148926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9027777910232544, 'reward_repeat_penalty_std': 0.0927247703075409, 'reward_total_composite_mean': 0.7801923751831055, 'reward_total_composite_std': 0.1476370245218277, 'epoch': 0.06}

 42%|████▏     | 1400/3300 [1:30:32<2:44:12,  5.19s/it]INFO 04-12 00:07:23 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:03<00:21,  2.00s/it]

 23%|██▎       | 3/13 [00:06<00:22,  2.24s/it]

 31%|███       | 4/13 [00:11<00:27,  3.08s/it]

 38%|███▊      | 5/13 [00:14<00:26,  3.37s/it]

 46%|████▌     | 6/13 [00:19<00:25,  3.68s/it]

 54%|█████▍    | 7/13 [00:23<00:22,  3.81s/it]

 62%|██████▏   | 8/13 [00:28<00:21,  4.25s/it]

 69%|██████▉   | 9/13 [00:34<00:18,  4.71s/it]

 77%|███████▋  | 10/13 [00:37<00:13,  4.36s/it]

 85%|████████▍ | 11/13 [00:41<00:08,  4.12s/it]

 92%|█████████▏| 12/13 [00:46<00:04,  4.26s/it]

100%|██████████| 13/13 [00:50<00:00,  4.20s/it]2026-04-12 00:08:18,842 | INFO | train_grpo_train | metrics_logged mode=eval step=1400

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 55.0853, 'eval_samples_per_second': 1.888, 'eval_steps_per_second': 0.236, 'eval_num_tokens': 3155594.0, 'eval_completions/mean_length': 158.23076923076923, 'eval_completions/min_length': 57.92307692307692, 'eval_completions/max_length': 281.6923076923077, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 158.23076923076923, 'eval_completions/min_terminated_length': 57.92307692307692, 'eval_completions/max_terminated_length': 281.6923076923077, 'eval_rewards/meter/mean': 0.6238016211069547, 'eval_rewards/meter/std': 0.41127324333557713, 'eval_rewards/count_adherence/mean': 0.8451297649970422, 'eval_rewards/count_adherence/std': 0.1567458017514302, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.7824494563616239, 'eval_rewards/repeat_penalty/std': 0.1815235666357554, 'eval_rewards/total_composite/mean': 0.42582815427046555, 'eval_rewards/total_composite/std': 0.3518183070879716, 'eval_reward': 0.42582815427046555, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.02079938738965071, 'eval_sampling/sampling_logp_difference/max': 0.9362153823559101, 'eval_sampling/importance_sampling_ratio/min': 0.40005409259062547, 'eval_sampling/importance_sampling_ratio/mean': 1.0046258247815645, 'eval_sampling/importance_sampling_ratio/max': 1.4843606765453632, 'eval_entropy': 0.18456935252134615, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.42582815427046555, 'eval_reward_meter_mean': 0.6238016211069547, 'eval_reward_meter_std': 0.41127324333557713, 'eval_reward_count_adherence_mean': 0.8451297649970422, 'eval_reward_count_adherence_std': 0.1567458017514302, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.7824494563616239, 'eval_reward_repeat_penalty_std': 0.1815235666357554, 'eval_reward_total_composite_mean': 0.42582815427046555, 'eval_reward_total_composite_std': 0.3518183070879716, 'epoch': 0.06}

 42%|████▏     | 1400/3300 [1:31:27<2:44:12,  5.19s/it]

100%|██████████| 13/13 [00:50<00:00,  4.20s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 42%|████▏     | 1401/3300 [1:31:34<11:44:38, 22.26s/it]2026-04-12 00:08:25,828 | INFO | train_grpo_train | metrics_logged mode=train step=1401

                                                        
{'loss': 0.0053, 'grad_norm': 2.925625801086426, 'learning_rate': 5.7575757575757586e-06, 'num_tokens': 3157247.0, 'completions/mean_length': 61.625, 'completions/min_length': 61.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9903584122657776, 'rewards/meter/std': 0.011271141469478607, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9903584122657776, 'rewards/total_composite/std': 0.011271141469478607, 'reward': 0.9903584122657776, 'reward_std': 0.011271136812865734, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015541310422122478, 'sampling/sampling_logp_difference/max': 1.4758706092834473, 'sampling/importance_sampling_ratio/min': 0.22857964038848877, 'sampling/importance_sampling_ratio/mean': 1.0016825199127197, 'sampling/importance_sampling_ratio/max': 1.5843480825424194, 'entropy': 0.07412398885935545, 'clip_ratio/low_mean': 0.004032257944345474, 'clip_ratio/low_min': 0.004032257944345474, 'clip_ratio/high_mean': 0.014245107769966125, 'clip_ratio/high_max': 0.014245107769966125, 'clip_ratio/region_mean': 0.0182773657143116, 'reward_total_mean': 0.9903584122657776, 'reward_meter_mean': 0.9903584122657776, 'reward_meter_std': 0.011271141469478607, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9903584122657776, 'reward_total_composite_std': 0.011271141469478607, 'epoch': 0.06}

 42%|████▏     | 1401/3300 [1:31:34<11:44:38, 22.26s/it]INFO 04-12 00:08:26 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1402/3300 [1:31:40<9:08:29, 17.34s/it] 2026-04-12 00:08:31,677 | INFO | train_grpo_train | metrics_logged mode=train step=1402

                                                       
{'loss': -0.0084, 'grad_norm': 2.7992911338806152, 'learning_rate': 5.754545454545455e-06, 'num_tokens': 3160271.0, 'completions/mean_length': 139.0, 'completions/min_length': 133.0, 'completions/max_length': 147.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 139.0, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 147.0, 'rewards/meter/mean': 0.4930988550186157, 'rewards/meter/std': 0.10791701823472977, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.4305555522441864, 'rewards/repeat_penalty/std': 0.15068919956684113, 'rewards/total_composite/mean': 0.1747434139251709, 'rewards/total_composite/std': 0.07833553850650787, 'reward': 0.1747434139251709, 'reward_std': 0.07833553850650787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02746950089931488, 'sampling/sampling_logp_difference/max': 1.4570236206054688, 'sampling/importance_sampling_ratio/min': 0.23292852938175201, 'sampling/importance_sampling_ratio/mean': 1.0046955347061157, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15692214109003544, 'clip_ratio/low_mean': 0.013523990812245756, 'clip_ratio/low_min': 0.013523990812245756, 'clip_ratio/high_mean': 0.0025862068869173527, 'clip_ratio/high_max': 0.0025862068869173527, 'clip_ratio/region_mean': 0.01611019769916311, 'reward_total_mean': 0.1747434139251709, 'reward_meter_mean': 0.4930988550186157, 'reward_meter_std': 0.10791701823472977, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.4305555522441864, 'reward_repeat_penalty_std': 0.15068919956684113, 'reward_total_composite_mean': 0.1747434139251709, 'reward_total_composite_std': 0.07833553850650787, 'epoch': 0.06}

 42%|████▏     | 1402/3300 [1:31:40<9:08:29, 17.34s/it]INFO 04-12 00:08:31 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1403/3300 [1:31:46<7:23:00, 14.01s/it]2026-04-12 00:08:37,927 | INFO | train_grpo_train | metrics_logged mode=train step=1403

                                                       
{'loss': -0.0093, 'grad_norm': 3.119965076446533, 'learning_rate': 5.751515151515152e-06, 'num_tokens': 3163193.0, 'completions/mean_length': 187.25, 'completions/min_length': 157.0, 'completions/max_length': 202.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 187.25, 'completions/min_terminated_length': 157.0, 'completions/max_terminated_length': 202.0, 'rewards/meter/mean': 0.958865761756897, 'rewards/meter/std': 0.024461213499307632, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.829365074634552, 'rewards/repeat_penalty/std': 0.10101525485515594, 'rewards/total_composite/mean': 0.7755063772201538, 'rewards/total_composite/std': 0.11600048094987869, 'reward': 0.7755063772201538, 'reward_std': 0.1160004660487175, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030590685084462166, 'sampling/sampling_logp_difference/max': 1.2629318237304688, 'sampling/importance_sampling_ratio/min': 0.2828236222267151, 'sampling/importance_sampling_ratio/mean': 1.0046159029006958, 'sampling/importance_sampling_ratio/max': 1.7141207456588745, 'entropy': 0.19072475843131542, 'clip_ratio/low_mean': 0.0072619569837115705, 'clip_ratio/low_min': 0.0072619569837115705, 'clip_ratio/high_mean': 0.007620060350745916, 'clip_ratio/high_max': 0.007620060350745916, 'clip_ratio/region_mean': 0.014882017334457487, 'reward_total_mean': 0.7755063772201538, 'reward_meter_mean': 0.958865761756897, 'reward_meter_std': 0.024461213499307632, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.829365074634552, 'reward_repeat_penalty_std': 0.10101525485515594, 'reward_total_composite_mean': 0.7755063772201538, 'reward_total_composite_std': 0.11600048094987869, 'epoch': 0.06}

 43%|████▎     | 1403/3300 [1:31:46<7:23:00, 14.01s/it]INFO 04-12 00:08:38 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1404/3300 [1:31:51<5:56:30, 11.28s/it]2026-04-12 00:08:42,836 | INFO | train_grpo_train | metrics_logged mode=train step=1404

                                                       
{'loss': 0.0225, 'grad_norm': 4.241501808166504, 'learning_rate': 5.748484848484849e-06, 'num_tokens': 3165135.0, 'completions/mean_length': 77.75, 'completions/min_length': 74.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.75, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9753485321998596, 'rewards/meter/std': 0.02492102049291134, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9753485321998596, 'rewards/total_composite/std': 0.02492102049291134, 'reward': 0.9753485321998596, 'reward_std': 0.024921026080846786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04453163966536522, 'sampling/sampling_logp_difference/max': 1.5494651794433594, 'sampling/importance_sampling_ratio/min': 0.21236151456832886, 'sampling/importance_sampling_ratio/mean': 1.0083844661712646, 'sampling/importance_sampling_ratio/max': 1.8341128826141357, 'entropy': 0.3366150539368391, 'clip_ratio/low_mean': 0.0077174786711111665, 'clip_ratio/low_min': 0.0077174786711111665, 'clip_ratio/high_mean': 0.019914651405997574, 'clip_ratio/high_max': 0.019914651405997574, 'clip_ratio/region_mean': 0.02763213007710874, 'reward_total_mean': 0.9753485321998596, 'reward_meter_mean': 0.9753485321998596, 'reward_meter_std': 0.02492102049291134, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9753485321998596, 'reward_total_composite_std': 0.02492102049291134, 'epoch': 0.06}

 43%|████▎     | 1404/3300 [1:31:51<5:56:30, 11.28s/it]INFO 04-12 00:08:43 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1405/3300 [1:31:56<4:53:35,  9.30s/it]2026-04-12 00:08:47,499 | INFO | train_grpo_train | metrics_logged mode=train step=1405

                                                       
{'loss': 0.0037, 'grad_norm': 3.760753631591797, 'learning_rate': 5.745454545454546e-06, 'num_tokens': 3166898.0, 'completions/mean_length': 67.375, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9491584897041321, 'rewards/meter/std': 0.09555412083864212, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9491584897041321, 'rewards/total_composite/std': 0.09555412083864212, 'reward': 0.9491584897041321, 'reward_std': 0.09555413573980331, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026108134537935257, 'sampling/sampling_logp_difference/max': 1.3400421142578125, 'sampling/importance_sampling_ratio/min': 0.26183465123176575, 'sampling/importance_sampling_ratio/mean': 1.005703091621399, 'sampling/importance_sampling_ratio/max': 1.8443472385406494, 'entropy': 0.20304011879488826, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.0202326342696324, 'clip_ratio/high_max': 0.0202326342696324, 'clip_ratio/region_mean': 0.022070869570598006, 'reward_total_mean': 0.9491584897041321, 'reward_meter_mean': 0.9491584897041321, 'reward_meter_std': 0.09555412083864212, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9491584897041321, 'reward_total_composite_std': 0.09555412083864212, 'epoch': 0.06}

 43%|████▎     | 1405/3300 [1:31:56<4:53:35,  9.30s/it]INFO 04-12 00:08:47 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1406/3300 [1:32:01<4:08:34,  7.87s/it]2026-04-12 00:08:52,060 | INFO | train_grpo_train | metrics_logged mode=train step=1406

                                                       
{'loss': 0.0102, 'grad_norm': 5.033778190612793, 'learning_rate': 5.742424242424242e-06, 'num_tokens': 3168628.0, 'completions/mean_length': 59.25, 'completions/min_length': 55.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.37988948822021484, 'rewards/meter/std': 0.33830785751342773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.37988948822021484, 'rewards/total_composite/std': 0.33830785751342773, 'reward': 0.37988948822021484, 'reward_std': 0.33830785751342773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05973369628190994, 'sampling/sampling_logp_difference/max': 1.5036615133285522, 'sampling/importance_sampling_ratio/min': 0.22231467068195343, 'sampling/importance_sampling_ratio/mean': 1.01566481590271, 'sampling/importance_sampling_ratio/max': 1.7001794576644897, 'entropy': 0.4192809797823429, 'clip_ratio/low_mean': 0.025304300244897604, 'clip_ratio/low_min': 0.025304300244897604, 'clip_ratio/high_mean': 0.03157911077141762, 'clip_ratio/high_max': 0.03157911077141762, 'clip_ratio/region_mean': 0.05688341101631522, 'reward_total_mean': 0.37988948822021484, 'reward_meter_mean': 0.37988948822021484, 'reward_meter_std': 0.33830785751342773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.37988948822021484, 'reward_total_composite_std': 0.33830785751342773, 'epoch': 0.06}

 43%|████▎     | 1406/3300 [1:32:01<4:08:34,  7.87s/it]INFO 04-12 00:08:52 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1407/3300 [1:32:06<3:42:35,  7.06s/it]2026-04-12 00:08:57,203 | INFO | train_grpo_train | metrics_logged mode=train step=1407

                                                       
{'loss': 0.0163, 'grad_norm': 6.2097625732421875, 'learning_rate': 5.73939393939394e-06, 'num_tokens': 3170743.0, 'completions/mean_length': 99.375, 'completions/min_length': 94.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.375, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.8928050398826599, 'rewards/meter/std': 0.15822991728782654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.829431414604187, 'rewards/total_composite/std': 0.1831192672252655, 'reward': 0.829431414604187, 'reward_std': 0.1831192672252655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04797036573290825, 'sampling/sampling_logp_difference/max': 1.2440190315246582, 'sampling/importance_sampling_ratio/min': 0.2882235050201416, 'sampling/importance_sampling_ratio/mean': 1.0000447034835815, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32133867032825947, 'clip_ratio/low_mean': 0.01607399620115757, 'clip_ratio/low_min': 0.01607399620115757, 'clip_ratio/high_mean': 0.0139403420034796, 'clip_ratio/high_max': 0.0139403420034796, 'clip_ratio/region_mean': 0.03001433820463717, 'reward_total_mean': 0.829431414604187, 'reward_meter_mean': 0.8928050398826599, 'reward_meter_std': 0.15822991728782654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.829431414604187, 'reward_total_composite_std': 0.1831192672252655, 'epoch': 0.06}

 43%|████▎     | 1407/3300 [1:32:06<3:42:35,  7.06s/it]INFO 04-12 00:08:57 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1408/3300 [1:32:10<3:16:15,  6.22s/it]2026-04-12 00:09:01,485 | INFO | train_grpo_train | metrics_logged mode=train step=1408

                                                       
{'loss': -0.0044, 'grad_norm': 7.772467136383057, 'learning_rate': 5.736363636363637e-06, 'num_tokens': 3172326.0, 'completions/mean_length': 30.875, 'completions/min_length': 30.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9951827526092529, 'rewards/meter/std': 0.007008199580013752, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9951827526092529, 'rewards/total_composite/std': 0.007008199580013752, 'reward': 0.9951827526092529, 'reward_std': 0.007008187007158995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033370744436979294, 'sampling/sampling_logp_difference/max': 0.8426570892333984, 'sampling/importance_sampling_ratio/min': 0.4305649697780609, 'sampling/importance_sampling_ratio/mean': 0.9993618726730347, 'sampling/importance_sampling_ratio/max': 1.774673581123352, 'entropy': 0.1420626100152731, 'clip_ratio/low_mean': 0.008198924828320742, 'clip_ratio/low_min': 0.008198924828320742, 'clip_ratio/high_mean': 0.01601142482832074, 'clip_ratio/high_max': 0.01601142482832074, 'clip_ratio/region_mean': 0.024210349656641483, 'reward_total_mean': 0.9951827526092529, 'reward_meter_mean': 0.9951827526092529, 'reward_meter_std': 0.007008199580013752, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9951827526092529, 'reward_total_composite_std': 0.007008199580013752, 'epoch': 0.06}

 43%|████▎     | 1408/3300 [1:32:10<3:16:15,  6.22s/it]INFO 04-12 00:09:01 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1409/3300 [1:32:14<3:00:04,  5.71s/it]2026-04-12 00:09:06,010 | INFO | train_grpo_train | metrics_logged mode=train step=1409

                                                       
{'loss': 0.2629, 'grad_norm': 11.023374557495117, 'learning_rate': 5.733333333333334e-06, 'num_tokens': 3173907.0, 'completions/mean_length': 51.625, 'completions/min_length': 36.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.992419958114624, 'rewards/meter/std': 0.00853477232158184, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4936572313308716, 'rewards/total_composite/std': 0.5277820229530334, 'reward': 0.4936572313308716, 'reward_std': 0.5277819633483887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06906002014875412, 'sampling/sampling_logp_difference/max': 3.973472833633423, 'sampling/importance_sampling_ratio/min': 0.018808001652359962, 'sampling/importance_sampling_ratio/mean': 0.9999293088912964, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28332205303013325, 'clip_ratio/low_mean': 0.02108370151836425, 'clip_ratio/low_min': 0.02108370151836425, 'clip_ratio/high_mean': 0.016995614394545555, 'clip_ratio/high_max': 0.016995614394545555, 'clip_ratio/region_mean': 0.038079315912909806, 'reward_total_mean': 0.4936572313308716, 'reward_meter_mean': 0.992419958114624, 'reward_meter_std': 0.00853477232158184, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4936572313308716, 'reward_total_composite_std': 0.5277820229530334, 'epoch': 0.06}

 43%|████▎     | 1409/3300 [1:32:15<3:00:04,  5.71s/it]INFO 04-12 00:09:06 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1410/3300 [1:32:19<2:50:49,  5.42s/it]2026-04-12 00:09:10,756 | INFO | train_grpo_train | metrics_logged mode=train step=1410

                                                       
{'loss': 0.0359, 'grad_norm': 5.875049591064453, 'learning_rate': 5.7303030303030305e-06, 'num_tokens': 3175555.0, 'completions/mean_length': 58.0, 'completions/min_length': 54.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7323773503303528, 'rewards/meter/std': 0.34669190645217896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7323773503303528, 'rewards/total_composite/std': 0.34669190645217896, 'reward': 0.7323773503303528, 'reward_std': 0.34669187664985657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04662606492638588, 'sampling/sampling_logp_difference/max': 2.092195987701416, 'sampling/importance_sampling_ratio/min': 0.12341582030057907, 'sampling/importance_sampling_ratio/mean': 1.0006393194198608, 'sampling/importance_sampling_ratio/max': 1.7754662036895752, 'entropy': 0.20567422546446323, 'clip_ratio/low_mean': 0.01024590153247118, 'clip_ratio/low_min': 0.01024590153247118, 'clip_ratio/high_mean': 0.04356799623928964, 'clip_ratio/high_max': 0.04356799623928964, 'clip_ratio/region_mean': 0.05381389777176082, 'reward_total_mean': 0.7323773503303528, 'reward_meter_mean': 0.7323773503303528, 'reward_meter_std': 0.34669190645217896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7323773503303528, 'reward_total_composite_std': 0.34669190645217896, 'epoch': 0.06}

 43%|████▎     | 1410/3300 [1:32:19<2:50:49,  5.42s/it]INFO 04-12 00:09:10 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1411/3300 [1:32:24<2:43:33,  5.20s/it]2026-04-12 00:09:15,420 | INFO | train_grpo_train | metrics_logged mode=train step=1411

                                                       
{'loss': 0.0404, 'grad_norm': 7.019533157348633, 'learning_rate': 5.727272727272728e-06, 'num_tokens': 3177472.0, 'completions/mean_length': 58.625, 'completions/min_length': 54.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.930023729801178, 'rewards/meter/std': 0.11754162609577179, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.930023729801178, 'rewards/total_composite/std': 0.11754162609577179, 'reward': 0.930023729801178, 'reward_std': 0.1175416111946106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03719957172870636, 'sampling/sampling_logp_difference/max': 0.9408185482025146, 'sampling/importance_sampling_ratio/min': 0.3903082311153412, 'sampling/importance_sampling_ratio/mean': 1.0033373832702637, 'sampling/importance_sampling_ratio/max': 1.9308364391326904, 'entropy': 0.22591773979365826, 'clip_ratio/low_mean': 0.007942708441987634, 'clip_ratio/low_min': 0.007942708441987634, 'clip_ratio/high_mean': 0.02417309512384236, 'clip_ratio/high_max': 0.02417309512384236, 'clip_ratio/region_mean': 0.03211580356582999, 'reward_total_mean': 0.930023729801178, 'reward_meter_mean': 0.930023729801178, 'reward_meter_std': 0.11754162609577179, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.930023729801178, 'reward_total_composite_std': 0.11754162609577179, 'epoch': 0.06}

 43%|████▎     | 1411/3300 [1:32:24<2:43:33,  5.20s/it]INFO 04-12 00:09:15 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1412/3300 [1:32:29<2:44:42,  5.23s/it]2026-04-12 00:09:20,753 | INFO | train_grpo_train | metrics_logged mode=train step=1412

                                                       
{'loss': 0.0095, 'grad_norm': 3.1584293842315674, 'learning_rate': 5.724242424242424e-06, 'num_tokens': 3179678.0, 'completions/mean_length': 121.75, 'completions/min_length': 119.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.75, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.980501115322113, 'rewards/meter/std': 0.04536837711930275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.8404378890991211, 'rewards/total_composite/std': 0.0856878012418747, 'reward': 0.8404378890991211, 'reward_std': 0.0856877937912941, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04456208273768425, 'sampling/sampling_logp_difference/max': 1.8470945358276367, 'sampling/importance_sampling_ratio/min': 0.19929902255535126, 'sampling/importance_sampling_ratio/mean': 1.0000073909759521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.25815106742084026, 'clip_ratio/low_mean': 0.007936508394777775, 'clip_ratio/low_min': 0.007936508394777775, 'clip_ratio/high_mean': 0.03802949539385736, 'clip_ratio/high_max': 0.03802949539385736, 'clip_ratio/region_mean': 0.045966003788635135, 'reward_total_mean': 0.8404378890991211, 'reward_meter_mean': 0.980501115322113, 'reward_meter_std': 0.04536837711930275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.8404378890991211, 'reward_total_composite_std': 0.0856878012418747, 'epoch': 0.06}

 43%|████▎     | 1412/3300 [1:32:29<2:44:42,  5.23s/it]INFO 04-12 00:09:20 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1413/3300 [1:32:35<2:48:08,  5.35s/it]2026-04-12 00:09:26,350 | INFO | train_grpo_train | metrics_logged mode=train step=1413

                                                       
{'loss': -0.0061, 'grad_norm': 2.6527252197265625, 'learning_rate': 5.721212121212122e-06, 'num_tokens': 3182168.0, 'completions/mean_length': 148.25, 'completions/min_length': 143.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.25, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.9954190850257874, 'rewards/meter/std': 0.002940161619335413, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.8531917333602905, 'rewards/total_composite/std': 0.07582316547632217, 'reward': 0.8531917333602905, 'reward_std': 0.07582316547632217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04073692858219147, 'sampling/sampling_logp_difference/max': 1.4822022914886475, 'sampling/importance_sampling_ratio/min': 0.22713692486286163, 'sampling/importance_sampling_ratio/mean': 1.0049986839294434, 'sampling/importance_sampling_ratio/max': 1.6937066316604614, 'entropy': 0.24800713174045086, 'clip_ratio/low_mean': 0.02131642634049058, 'clip_ratio/low_min': 0.02131642634049058, 'clip_ratio/high_mean': 0.019891314674168825, 'clip_ratio/high_max': 0.019891314674168825, 'clip_ratio/region_mean': 0.041207741014659405, 'reward_total_mean': 0.8531917333602905, 'reward_meter_mean': 0.9954190850257874, 'reward_meter_std': 0.002940161619335413, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.8531917333602905, 'reward_total_composite_std': 0.07582316547632217, 'epoch': 0.06}

 43%|████▎     | 1413/3300 [1:32:35<2:48:08,  5.35s/it]INFO 04-12 00:09:26 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1414/3300 [1:32:39<2:40:35,  5.11s/it]2026-04-12 00:09:30,905 | INFO | train_grpo_train | metrics_logged mode=train step=1414

                                                       
{'loss': -0.0031, 'grad_norm': 8.496710777282715, 'learning_rate': 5.718181818181819e-06, 'num_tokens': 3183867.0, 'completions/mean_length': 66.375, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7868467569351196, 'rewards/meter/std': 0.3379902243614197, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7868467569351196, 'rewards/total_composite/std': 0.3379902243614197, 'reward': 0.7868467569351196, 'reward_std': 0.3379901945590973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03707931190729141, 'sampling/sampling_logp_difference/max': 1.9779717922210693, 'sampling/importance_sampling_ratio/min': 0.13834954798221588, 'sampling/importance_sampling_ratio/mean': 0.9952957034111023, 'sampling/importance_sampling_ratio/max': 1.4725927114486694, 'entropy': 0.16014890000224113, 'clip_ratio/low_mean': 0.007663170341402292, 'clip_ratio/low_min': 0.007663170341402292, 'clip_ratio/high_mean': 0.0170001951046288, 'clip_ratio/high_max': 0.0170001951046288, 'clip_ratio/region_mean': 0.024663365446031094, 'reward_total_mean': 0.7868467569351196, 'reward_meter_mean': 0.7868467569351196, 'reward_meter_std': 0.3379902243614197, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7868467569351196, 'reward_total_composite_std': 0.3379902243614197, 'epoch': 0.06}

 43%|████▎     | 1414/3300 [1:32:39<2:40:35,  5.11s/it]INFO 04-12 00:09:31 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1415/3300 [1:32:44<2:32:58,  4.87s/it]2026-04-12 00:09:35,217 | INFO | train_grpo_train | metrics_logged mode=train step=1415

                                                       
{'loss': 0.0077, 'grad_norm': 12.028426170349121, 'learning_rate': 5.715151515151516e-06, 'num_tokens': 3185507.0, 'completions/mean_length': 41.0, 'completions/min_length': 40.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9893122911453247, 'rewards/meter/std': 0.020975705236196518, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9893122911453247, 'rewards/total_composite/std': 0.020975705236196518, 'reward': 0.9893122911453247, 'reward_std': 0.020975695922970772, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07620865851640701, 'sampling/sampling_logp_difference/max': 2.0691659450531006, 'sampling/importance_sampling_ratio/min': 0.12629108130931854, 'sampling/importance_sampling_ratio/mean': 1.0033856630325317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4230155944824219, 'clip_ratio/low_mean': 0.00937500037252903, 'clip_ratio/low_min': 0.00937500037252903, 'clip_ratio/high_mean': 0.04510642075911164, 'clip_ratio/high_max': 0.04510642075911164, 'clip_ratio/region_mean': 0.05448142113164067, 'reward_total_mean': 0.9893122911453247, 'reward_meter_mean': 0.9893122911453247, 'reward_meter_std': 0.020975705236196518, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9893122911453247, 'reward_total_composite_std': 0.020975705236196518, 'epoch': 0.06}

 43%|████▎     | 1415/3300 [1:32:44<2:32:58,  4.87s/it]INFO 04-12 00:09:35 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1416/3300 [1:32:50<2:43:26,  5.20s/it]2026-04-12 00:09:41,207 | INFO | train_grpo_train | metrics_logged mode=train step=1416

                                                       
{'loss': 0.0021, 'grad_norm': 1.9633679389953613, 'learning_rate': 5.712121212121212e-06, 'num_tokens': 3188624.0, 'completions/mean_length': 179.625, 'completions/min_length': 177.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 179.625, 'completions/min_terminated_length': 177.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.9946449995040894, 'rewards/meter/std': 0.006311771925538778, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8055555820465088, 'rewards/repeat_penalty/std': 0.07856741547584534, 'rewards/total_composite/mean': 0.8011291027069092, 'rewards/total_composite/std': 0.07706641405820847, 'reward': 0.8011291027069092, 'reward_std': 0.07706642150878906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025626802816987038, 'sampling/sampling_logp_difference/max': 1.5970072746276855, 'sampling/importance_sampling_ratio/min': 0.20250163972377777, 'sampling/importance_sampling_ratio/mean': 1.0057470798492432, 'sampling/importance_sampling_ratio/max': 1.6300792694091797, 'entropy': 0.15431117359548807, 'clip_ratio/low_mean': 0.009044285747222602, 'clip_ratio/low_min': 0.009044285747222602, 'clip_ratio/high_mean': 0.006235292763449252, 'clip_ratio/high_max': 0.006235292763449252, 'clip_ratio/region_mean': 0.015279578510671854, 'reward_total_mean': 0.8011291027069092, 'reward_meter_mean': 0.9946449995040894, 'reward_meter_std': 0.006311771925538778, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8055555820465088, 'reward_repeat_penalty_std': 0.07856741547584534, 'reward_total_composite_mean': 0.8011291027069092, 'reward_total_composite_std': 0.07706641405820847, 'epoch': 0.06}

 43%|████▎     | 1416/3300 [1:32:50<2:43:26,  5.20s/it]INFO 04-12 00:09:41 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1417/3300 [1:32:54<2:38:09,  5.04s/it]2026-04-12 00:09:45,857 | INFO | train_grpo_train | metrics_logged mode=train step=1417

                                                       
{'loss': -0.0002, 'grad_norm': 3.094541549682617, 'learning_rate': 5.7090909090909096e-06, 'num_tokens': 3190361.0, 'completions/mean_length': 61.125, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.998288631439209, 'rewards/meter/std': 0.00019117545161861926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998288631439209, 'rewards/total_composite/std': 0.00019117545161861926, 'reward': 0.998288631439209, 'reward_std': 0.00019117545161861926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01802898570895195, 'sampling/sampling_logp_difference/max': 1.3229293823242188, 'sampling/importance_sampling_ratio/min': 0.26635390520095825, 'sampling/importance_sampling_ratio/mean': 1.0027981996536255, 'sampling/importance_sampling_ratio/max': 1.6339690685272217, 'entropy': 0.06280876183882356, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.004098360426723957, 'clip_ratio/high_max': 0.004098360426723957, 'clip_ratio/region_mean': 0.006181693868711591, 'reward_total_mean': 0.998288631439209, 'reward_meter_mean': 0.998288631439209, 'reward_meter_std': 0.00019117545161861926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998288631439209, 'reward_total_composite_std': 0.00019117545161861926, 'epoch': 0.06}

 43%|████▎     | 1417/3300 [1:32:54<2:38:09,  5.04s/it]INFO 04-12 00:09:46 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1418/3300 [1:32:59<2:37:03,  5.01s/it]2026-04-12 00:09:50,789 | INFO | train_grpo_train | metrics_logged mode=train step=1418

                                                       
{'loss': 0.0502, 'grad_norm': 10.003430366516113, 'learning_rate': 5.706060606060606e-06, 'num_tokens': 3191990.0, 'completions/mean_length': 45.625, 'completions/min_length': 42.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8391948938369751, 'rewards/meter/std': 0.1604534387588501, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8391948938369751, 'rewards/total_composite/std': 0.1604534387588501, 'reward': 0.8391948938369751, 'reward_std': 0.1604534238576889, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05166694149374962, 'sampling/sampling_logp_difference/max': 1.3240852355957031, 'sampling/importance_sampling_ratio/min': 0.2660462260246277, 'sampling/importance_sampling_ratio/mean': 1.0050475597381592, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32396047934889793, 'clip_ratio/low_mean': 0.013166894670575857, 'clip_ratio/low_min': 0.013166894670575857, 'clip_ratio/high_mean': 0.013707729522138834, 'clip_ratio/high_max': 0.013707729522138834, 'clip_ratio/region_mean': 0.02687462419271469, 'reward_total_mean': 0.8391948938369751, 'reward_meter_mean': 0.8391948938369751, 'reward_meter_std': 0.1604534387588501, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8391948938369751, 'reward_total_composite_std': 0.1604534387588501, 'epoch': 0.06}

 43%|████▎     | 1418/3300 [1:32:59<2:37:03,  5.01s/it]INFO 04-12 00:09:51 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1419/3300 [1:33:03<2:28:57,  4.75s/it]2026-04-12 00:09:54,957 | INFO | train_grpo_train | metrics_logged mode=train step=1419

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.703030303030303e-06, 'num_tokens': 3193518.0, 'completions/mean_length': 31.0, 'completions/min_length': 31.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.9990130662918091, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990130662918091, 'rewards/total_composite/std': 0.0, 'reward': 0.9990130662918091, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.006462951190769672, 'sampling/sampling_logp_difference/max': 0.3100270926952362, 'sampling/importance_sampling_ratio/min': 0.733427107334137, 'sampling/importance_sampling_ratio/mean': 1.0024038553237915, 'sampling/importance_sampling_ratio/max': 1.2404334545135498, 'entropy': 0.038966658525168896, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990130662918091, 'reward_meter_mean': 0.9990130662918091, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990130662918091, 'reward_total_composite_std': 0.0, 'epoch': 0.06}

 43%|████▎     | 1419/3300 [1:33:03<2:28:57,  4.75s/it]INFO 04-12 00:09:55 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1420/3300 [1:33:09<2:33:51,  4.91s/it]2026-04-12 00:10:00,230 | INFO | train_grpo_train | metrics_logged mode=train step=1420

                                                       
{'loss': 0.0034, 'grad_norm': 3.3092191219329834, 'learning_rate': 5.7e-06, 'num_tokens': 3195977.0, 'completions/mean_length': 118.375, 'completions/min_length': 115.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.375, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9984818696975708, 'rewards/meter/std': 8.360291394637898e-05, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8035714030265808, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.6418830156326294, 'rewards/total_composite/std': 0.05907980352640152, 'reward': 0.6418830156326294, 'reward_std': 0.05907979980111122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015416157431900501, 'sampling/sampling_logp_difference/max': 1.7627546787261963, 'sampling/importance_sampling_ratio/min': 0.17157158255577087, 'sampling/importance_sampling_ratio/mean': 1.001312494277954, 'sampling/importance_sampling_ratio/max': 1.6658356189727783, 'entropy': 0.07044978253543377, 'clip_ratio/low_mean': 0.0031779661076143384, 'clip_ratio/low_min': 0.0031779661076143384, 'clip_ratio/high_mean': 0.01176445058081299, 'clip_ratio/high_max': 0.01176445058081299, 'clip_ratio/region_mean': 0.014942416688427329, 'reward_total_mean': 0.6418830156326294, 'reward_meter_mean': 0.9984818696975708, 'reward_meter_std': 8.360291394637898e-05, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8035714030265808, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.6418830156326294, 'reward_total_composite_std': 0.05907980352640152, 'epoch': 0.06}

 43%|████▎     | 1420/3300 [1:33:09<2:33:51,  4.91s/it]INFO 04-12 00:10:00 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1421/3300 [1:33:13<2:30:10,  4.80s/it]2026-04-12 00:10:04,762 | INFO | train_grpo_train | metrics_logged mode=train step=1421

                                                       
{'loss': -0.006, 'grad_norm': 11.100446701049805, 'learning_rate': 5.696969696969698e-06, 'num_tokens': 3197831.0, 'completions/mean_length': 42.75, 'completions/min_length': 39.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7552404403686523, 'rewards/meter/std': 0.31532391905784607, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7552404403686523, 'rewards/total_composite/std': 0.31532391905784607, 'reward': 0.7552404403686523, 'reward_std': 0.31532391905784607, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06012430414557457, 'sampling/sampling_logp_difference/max': 1.009714961051941, 'sampling/importance_sampling_ratio/min': 0.36432281136512756, 'sampling/importance_sampling_ratio/mean': 1.007340669631958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3276716824620962, 'clip_ratio/low_mean': 0.018345543881878257, 'clip_ratio/low_min': 0.018345543881878257, 'clip_ratio/high_mean': 0.049062950536608696, 'clip_ratio/high_max': 0.049062950536608696, 'clip_ratio/region_mean': 0.06740849441848695, 'reward_total_mean': 0.7552404403686523, 'reward_meter_mean': 0.7552404403686523, 'reward_meter_std': 0.31532391905784607, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7552404403686523, 'reward_total_composite_std': 0.31532391905784607, 'epoch': 0.06}

 43%|████▎     | 1421/3300 [1:33:13<2:30:10,  4.80s/it]INFO 04-12 00:10:05 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1422/3300 [1:33:18<2:26:57,  4.70s/it]2026-04-12 00:10:09,216 | INFO | train_grpo_train | metrics_logged mode=train step=1422

                                                       
{'loss': 0.0018, 'grad_norm': 3.4191842079162598, 'learning_rate': 5.693939393939394e-06, 'num_tokens': 3199274.0, 'completions/mean_length': 29.375, 'completions/min_length': 29.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9963138103485107, 'rewards/meter/std': 0.00021444838785100728, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9963138103485107, 'rewards/total_composite/std': 0.00021444838785100728, 'reward': 0.9963138103485107, 'reward_std': 0.00021445513993967324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013728239573538303, 'sampling/sampling_logp_difference/max': 0.6362069845199585, 'sampling/importance_sampling_ratio/min': 0.5292962193489075, 'sampling/importance_sampling_ratio/mean': 1.001716136932373, 'sampling/importance_sampling_ratio/max': 1.1206741333007812, 'entropy': 0.09071109816431999, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004310344811528921, 'reward_total_mean': 0.9963138103485107, 'reward_meter_mean': 0.9963138103485107, 'reward_meter_std': 0.00021444838785100728, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9963138103485107, 'reward_total_composite_std': 0.00021444838785100728, 'epoch': 0.06}

 43%|████▎     | 1422/3300 [1:33:18<2:26:57,  4.70s/it]INFO 04-12 00:10:09 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1423/3300 [1:33:22<2:23:53,  4.60s/it]2026-04-12 00:10:13,603 | INFO | train_grpo_train | metrics_logged mode=train step=1423

                                                       
{'loss': 0.0189, 'grad_norm': 12.835563659667969, 'learning_rate': 5.690909090909091e-06, 'num_tokens': 3200795.0, 'completions/mean_length': 34.125, 'completions/min_length': 33.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9920762777328491, 'rewards/meter/std': 0.0020317891612648964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9920762777328491, 'rewards/total_composite/std': 0.0020317891612648964, 'reward': 0.9920762777328491, 'reward_std': 0.0020317891612648964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02493356168270111, 'sampling/sampling_logp_difference/max': 0.7522921562194824, 'sampling/importance_sampling_ratio/min': 0.47128504514694214, 'sampling/importance_sampling_ratio/mean': 0.9959531426429749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10061774123460054, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.014817290706560016, 'clip_ratio/high_max': 0.014817290706560016, 'clip_ratio/region_mean': 0.014817290706560016, 'reward_total_mean': 0.9920762777328491, 'reward_meter_mean': 0.9920762777328491, 'reward_meter_std': 0.0020317891612648964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9920762777328491, 'reward_total_composite_std': 0.0020317891612648964, 'epoch': 0.06}

 43%|████▎     | 1423/3300 [1:33:22<2:23:53,  4.60s/it]INFO 04-12 00:10:13 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1424/3300 [1:33:27<2:25:41,  4.66s/it]2026-04-12 00:10:18,400 | INFO | train_grpo_train | metrics_logged mode=train step=1424

                                                       
{'loss': -0.0099, 'grad_norm': 5.431046485900879, 'learning_rate': 5.687878787878789e-06, 'num_tokens': 3202519.0, 'completions/mean_length': 55.5, 'completions/min_length': 53.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9895903468132019, 'rewards/meter/std': 0.012036402709782124, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9895903468132019, 'rewards/total_composite/std': 0.012036402709782124, 'reward': 0.9895903468132019, 'reward_std': 0.012036396190524101, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044165343046188354, 'sampling/sampling_logp_difference/max': 1.3421528339385986, 'sampling/importance_sampling_ratio/min': 0.2612825632095337, 'sampling/importance_sampling_ratio/mean': 1.0014959573745728, 'sampling/importance_sampling_ratio/max': 1.5429896116256714, 'entropy': 0.2528437674045563, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0493996306322515, 'clip_ratio/high_max': 0.0493996306322515, 'clip_ratio/region_mean': 0.0493996306322515, 'reward_total_mean': 0.9895903468132019, 'reward_meter_mean': 0.9895903468132019, 'reward_meter_std': 0.012036402709782124, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9895903468132019, 'reward_total_composite_std': 0.012036402709782124, 'epoch': 0.06}

 43%|████▎     | 1424/3300 [1:33:27<2:25:41,  4.66s/it]INFO 04-12 00:10:18 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1425/3300 [1:33:35<2:58:14,  5.70s/it]2026-04-12 00:10:26,540 | INFO | train_grpo_train | metrics_logged mode=train step=1425

                                                       
{'loss': 0.0056, 'grad_norm': 2.6007766723632812, 'learning_rate': 5.684848484848485e-06, 'num_tokens': 3206882.0, 'completions/mean_length': 320.375, 'completions/min_length': 305.0, 'completions/max_length': 331.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 320.375, 'completions/min_terminated_length': 305.0, 'completions/max_terminated_length': 331.0, 'rewards/meter/mean': 0.942986011505127, 'rewards/meter/std': 0.145384281873703, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.033064987510442734, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8428308963775635, 'rewards/repeat_penalty/std': 0.10538350045681, 'rewards/total_composite/mean': 0.4912066161632538, 'rewards/total_composite/std': 0.07578601688146591, 'reward': 0.4912066161632538, 'reward_std': 0.07578601688146591, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03792227804660797, 'sampling/sampling_logp_difference/max': 2.1388535499572754, 'sampling/importance_sampling_ratio/min': 0.11778980493545532, 'sampling/importance_sampling_ratio/mean': 1.0055261850357056, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2617227640002966, 'clip_ratio/low_mean': 0.006616426864638925, 'clip_ratio/low_min': 0.006616426864638925, 'clip_ratio/high_mean': 0.018805116647854447, 'clip_ratio/high_max': 0.018805116647854447, 'clip_ratio/region_mean': 0.025421543512493372, 'reward_total_mean': 0.4912066161632538, 'reward_meter_mean': 0.942986011505127, 'reward_meter_std': 0.145384281873703, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.033064987510442734, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8428308963775635, 'reward_repeat_penalty_std': 0.10538350045681, 'reward_total_composite_mean': 0.4912066161632538, 'reward_total_composite_std': 0.07578601688146591, 'epoch': 0.06}

 43%|████▎     | 1425/3300 [1:33:35<2:58:14,  5.70s/it]INFO 04-12 00:10:26 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1426/3300 [1:33:40<2:48:41,  5.40s/it]2026-04-12 00:10:31,233 | INFO | train_grpo_train | metrics_logged mode=train step=1426

                                                       
{'loss': -0.01, 'grad_norm': 12.311224937438965, 'learning_rate': 5.681818181818183e-06, 'num_tokens': 3208730.0, 'completions/mean_length': 63.0, 'completions/min_length': 60.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.996849000453949, 'rewards/meter/std': 0.0023078599479049444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.996849000453949, 'rewards/total_composite/std': 0.0023078599479049444, 'reward': 0.996849000453949, 'reward_std': 0.002307864371687174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03911422938108444, 'sampling/sampling_logp_difference/max': 1.8360328674316406, 'sampling/importance_sampling_ratio/min': 0.15944872796535492, 'sampling/importance_sampling_ratio/mean': 1.0055710077285767, 'sampling/importance_sampling_ratio/max': 1.7190132141113281, 'entropy': 0.2404075190424919, 'clip_ratio/low_mean': 0.006051587639376521, 'clip_ratio/low_min': 0.006051587639376521, 'clip_ratio/high_mean': 0.043533024145290256, 'clip_ratio/high_max': 0.043533024145290256, 'clip_ratio/region_mean': 0.04958461178466678, 'reward_total_mean': 0.996849000453949, 'reward_meter_mean': 0.996849000453949, 'reward_meter_std': 0.0023078599479049444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.996849000453949, 'reward_total_composite_std': 0.0023078599479049444, 'epoch': 0.06}

 43%|████▎     | 1426/3300 [1:33:40<2:48:41,  5.40s/it]INFO 04-12 00:10:31 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1427/3300 [1:33:45<2:44:51,  5.28s/it]2026-04-12 00:10:36,230 | INFO | train_grpo_train | metrics_logged mode=train step=1427

                                                       
{'loss': -0.0018, 'grad_norm': 2.7146718502044678, 'learning_rate': 5.67878787878788e-06, 'num_tokens': 3210997.0, 'completions/mean_length': 90.375, 'completions/min_length': 89.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.375, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9984976053237915, 'rewards/meter/std': 9.82403289526701e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9485726356506348, 'rewards/total_composite/std': 0.09244248270988464, 'reward': 0.9485726356506348, 'reward_std': 0.09244248270988464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014895378611981869, 'sampling/sampling_logp_difference/max': 1.0964069366455078, 'sampling/importance_sampling_ratio/min': 0.33406925201416016, 'sampling/importance_sampling_ratio/mean': 0.9975816011428833, 'sampling/importance_sampling_ratio/max': 1.3982148170471191, 'entropy': 0.06489505246281624, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.012471178779378533, 'clip_ratio/high_max': 0.012471178779378533, 'clip_ratio/region_mean': 0.012471178779378533, 'reward_total_mean': 0.9485726356506348, 'reward_meter_mean': 0.9984976053237915, 'reward_meter_std': 9.82403289526701e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9485726356506348, 'reward_total_composite_std': 0.09244248270988464, 'epoch': 0.06}

 43%|████▎     | 1427/3300 [1:33:45<2:44:51,  5.28s/it]INFO 04-12 00:10:36 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1428/3300 [1:33:52<3:01:42,  5.82s/it]2026-04-12 00:10:43,329 | INFO | train_grpo_train | metrics_logged mode=train step=1428

                                                       
{'loss': 0.0093, 'grad_norm': 2.3229403495788574, 'learning_rate': 5.675757575757577e-06, 'num_tokens': 3214669.0, 'completions/mean_length': 256.0, 'completions/min_length': 245.0, 'completions/max_length': 272.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 256.0, 'completions/min_terminated_length': 245.0, 'completions/max_terminated_length': 272.0, 'rewards/meter/mean': 0.9840718507766724, 'rewards/meter/std': 0.03672640025615692, 'rewards/count_adherence/mean': 0.699999988079071, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8125, 'rewards/repeat_penalty/std': 0.10492872446775436, 'rewards/total_composite/mean': 0.5593699216842651, 'rewards/total_composite/std': 0.07350600510835648, 'reward': 0.5593699216842651, 'reward_std': 0.07350601255893707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0390043631196022, 'sampling/sampling_logp_difference/max': 3.3374457359313965, 'sampling/importance_sampling_ratio/min': 0.035527586936950684, 'sampling/importance_sampling_ratio/mean': 1.0045963525772095, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24354491382837296, 'clip_ratio/low_mean': 0.01068126189056784, 'clip_ratio/low_min': 0.01068126189056784, 'clip_ratio/high_mean': 0.012853078544139862, 'clip_ratio/high_max': 0.012853078544139862, 'clip_ratio/region_mean': 0.0235343404347077, 'reward_total_mean': 0.5593699216842651, 'reward_meter_mean': 0.9840718507766724, 'reward_meter_std': 0.03672640025615692, 'reward_count_adherence_mean': 0.699999988079071, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8125, 'reward_repeat_penalty_std': 0.10492872446775436, 'reward_total_composite_mean': 0.5593699216842651, 'reward_total_composite_std': 0.07350600510835648, 'epoch': 0.06}

 43%|████▎     | 1428/3300 [1:33:52<3:01:42,  5.82s/it]INFO 04-12 00:10:43 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1429/3300 [1:33:57<2:57:02,  5.68s/it]2026-04-12 00:10:48,654 | INFO | train_grpo_train | metrics_logged mode=train step=1429

                                                       
{'loss': 0.0002, 'grad_norm': 2.9396092891693115, 'learning_rate': 5.672727272727273e-06, 'num_tokens': 3217095.0, 'completions/mean_length': 123.25, 'completions/min_length': 122.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.25, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9978071451187134, 'rewards/meter/std': 0.0011365336831659079, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8035714030265808, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.8018503189086914, 'rewards/total_composite/std': 0.07426349073648453, 'reward': 0.8018503189086914, 'reward_std': 0.07426347583532333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015226030722260475, 'sampling/sampling_logp_difference/max': 1.5297276973724365, 'sampling/importance_sampling_ratio/min': 0.2165946513414383, 'sampling/importance_sampling_ratio/mean': 0.999223530292511, 'sampling/importance_sampling_ratio/max': 1.302078366279602, 'entropy': 0.07903782464563847, 'clip_ratio/low_mean': 0.006106290849857032, 'clip_ratio/low_min': 0.006106290849857032, 'clip_ratio/high_mean': 0.013211918994784355, 'clip_ratio/high_max': 0.013211918994784355, 'clip_ratio/region_mean': 0.019318209844641387, 'reward_total_mean': 0.8018503189086914, 'reward_meter_mean': 0.9978071451187134, 'reward_meter_std': 0.0011365336831659079, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8035714030265808, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.8018503189086914, 'reward_total_composite_std': 0.07426349073648453, 'epoch': 0.06}

 43%|████▎     | 1429/3300 [1:33:57<2:57:02,  5.68s/it]INFO 04-12 00:10:48 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1430/3300 [1:34:02<2:46:52,  5.35s/it]2026-04-12 00:10:53,253 | INFO | train_grpo_train | metrics_logged mode=train step=1430

                                                       
{'loss': 0.0069, 'grad_norm': 5.2827301025390625, 'learning_rate': 5.6696969696969705e-06, 'num_tokens': 3218659.0, 'completions/mean_length': 56.5, 'completions/min_length': 53.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.5, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.43106603622436523, 'rewards/meter/std': 0.3671012222766876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.43106603622436523, 'rewards/total_composite/std': 0.3671012222766876, 'reward': 0.43106603622436523, 'reward_std': 0.36710119247436523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059371624141931534, 'sampling/sampling_logp_difference/max': 1.259765625, 'sampling/importance_sampling_ratio/min': 0.2837205231189728, 'sampling/importance_sampling_ratio/mean': 1.0158692598342896, 'sampling/importance_sampling_ratio/max': 1.4617520570755005, 'entropy': 0.5998508296906948, 'clip_ratio/low_mean': 0.030917756259441376, 'clip_ratio/low_min': 0.030917756259441376, 'clip_ratio/high_mean': 0.008850250858813524, 'clip_ratio/high_max': 0.008850250858813524, 'clip_ratio/region_mean': 0.0397680071182549, 'reward_total_mean': 0.43106603622436523, 'reward_meter_mean': 0.43106603622436523, 'reward_meter_std': 0.3671012222766876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.43106603622436523, 'reward_total_composite_std': 0.3671012222766876, 'epoch': 0.06}

 43%|████▎     | 1430/3300 [1:34:02<2:46:52,  5.35s/it]INFO 04-12 00:10:53 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1431/3300 [1:34:07<2:43:52,  5.26s/it]2026-04-12 00:10:58,295 | INFO | train_grpo_train | metrics_logged mode=train step=1431

                                                       
{'loss': -0.0112, 'grad_norm': 3.8664066791534424, 'learning_rate': 5.666666666666667e-06, 'num_tokens': 3220585.0, 'completions/mean_length': 90.75, 'completions/min_length': 87.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.75, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9018567800521851, 'rewards/meter/std': 0.26490822434425354, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9018567800521851, 'rewards/total_composite/std': 0.26490822434425354, 'reward': 0.9018567800521851, 'reward_std': 0.2649082541465759, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042072996497154236, 'sampling/sampling_logp_difference/max': 1.666844367980957, 'sampling/importance_sampling_ratio/min': 0.1888420283794403, 'sampling/importance_sampling_ratio/mean': 1.0014922618865967, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12930846586823463, 'clip_ratio/low_mean': 0.008522727526724339, 'clip_ratio/low_min': 0.008522727526724339, 'clip_ratio/high_mean': 0.027468874352052808, 'clip_ratio/high_max': 0.027468874352052808, 'clip_ratio/region_mean': 0.035991601878777146, 'reward_total_mean': 0.9018567800521851, 'reward_meter_mean': 0.9018567800521851, 'reward_meter_std': 0.26490822434425354, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9018567800521851, 'reward_total_composite_std': 0.26490822434425354, 'epoch': 0.06}

 43%|████▎     | 1431/3300 [1:34:07<2:43:52,  5.26s/it]INFO 04-12 00:10:58 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1432/3300 [1:34:12<2:43:43,  5.26s/it]2026-04-12 00:11:03,550 | INFO | train_grpo_train | metrics_logged mode=train step=1432

                                                       
{'loss': 0.002, 'grad_norm': 3.0692851543426514, 'learning_rate': 5.663636363636364e-06, 'num_tokens': 3222928.0, 'completions/mean_length': 118.875, 'completions/min_length': 118.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.875, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.9983639717102051, 'rewards/meter/std': 0.00039870149339549243, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.11921755969524384, 'rewards/total_composite/mean': 0.7309507131576538, 'rewards/total_composite/std': 0.11907364428043365, 'reward': 0.7309507131576538, 'reward_std': 0.11907363682985306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022948887199163437, 'sampling/sampling_logp_difference/max': 1.848806619644165, 'sampling/importance_sampling_ratio/min': 0.1574249118566513, 'sampling/importance_sampling_ratio/mean': 1.0022363662719727, 'sampling/importance_sampling_ratio/max': 1.6000510454177856, 'entropy': 0.1475573442876339, 'clip_ratio/low_mean': 0.0031601624796167016, 'clip_ratio/low_min': 0.0031601624796167016, 'clip_ratio/high_mean': 0.007326977560296655, 'clip_ratio/high_max': 0.007326977560296655, 'clip_ratio/region_mean': 0.010487140039913356, 'reward_total_mean': 0.7309507131576538, 'reward_meter_mean': 0.9983639717102051, 'reward_meter_std': 0.00039870149339549243, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.11921755969524384, 'reward_total_composite_mean': 0.7309507131576538, 'reward_total_composite_std': 0.11907364428043365, 'epoch': 0.06}

 43%|████▎     | 1432/3300 [1:34:12<2:43:43,  5.26s/it]INFO 04-12 00:11:03 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1433/3300 [1:34:17<2:38:58,  5.11s/it]2026-04-12 00:11:08,309 | INFO | train_grpo_train | metrics_logged mode=train step=1433

                                                       
{'loss': -0.0086, 'grad_norm': 4.21342658996582, 'learning_rate': 5.6606060606060606e-06, 'num_tokens': 3224674.0, 'completions/mean_length': 64.25, 'completions/min_length': 60.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7867549657821655, 'rewards/meter/std': 0.2720617949962616, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7867549657821655, 'rewards/total_composite/std': 0.2720617949962616, 'reward': 0.7867549657821655, 'reward_std': 0.2720617651939392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.045874159783124924, 'sampling/sampling_logp_difference/max': 1.6135480403900146, 'sampling/importance_sampling_ratio/min': 0.19917964935302734, 'sampling/importance_sampling_ratio/mean': 1.0058132410049438, 'sampling/importance_sampling_ratio/max': 1.7156156301498413, 'entropy': 0.3227595426142216, 'clip_ratio/low_mean': 0.01948191737756133, 'clip_ratio/low_min': 0.01948191737756133, 'clip_ratio/high_mean': 0.017400611890479922, 'clip_ratio/high_max': 0.017400611890479922, 'clip_ratio/region_mean': 0.03688252926804125, 'reward_total_mean': 0.7867549657821655, 'reward_meter_mean': 0.7867549657821655, 'reward_meter_std': 0.2720617949962616, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7867549657821655, 'reward_total_composite_std': 0.2720617949962616, 'epoch': 0.06}

 43%|████▎     | 1433/3300 [1:34:17<2:38:58,  5.11s/it]INFO 04-12 00:11:08 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1434/3300 [1:34:21<2:35:02,  4.99s/it]2026-04-12 00:11:13,006 | INFO | train_grpo_train | metrics_logged mode=train step=1434

                                                       
{'loss': -0.0555, 'grad_norm': 7.206746578216553, 'learning_rate': 5.657575757575759e-06, 'num_tokens': 3226467.0, 'completions/mean_length': 63.125, 'completions/min_length': 54.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9712523221969604, 'rewards/meter/std': 0.06995871663093567, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9712523221969604, 'rewards/total_composite/std': 0.06995871663093567, 'reward': 0.9712523221969604, 'reward_std': 0.06995872408151627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024187438189983368, 'sampling/sampling_logp_difference/max': 1.3712058067321777, 'sampling/importance_sampling_ratio/min': 0.25380071997642517, 'sampling/importance_sampling_ratio/mean': 1.0004560947418213, 'sampling/importance_sampling_ratio/max': 1.8253268003463745, 'entropy': 0.12332060746848583, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0194749265210703, 'clip_ratio/high_max': 0.0194749265210703, 'clip_ratio/region_mean': 0.02178974135313183, 'reward_total_mean': 0.9712523221969604, 'reward_meter_mean': 0.9712523221969604, 'reward_meter_std': 0.06995871663093567, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9712523221969604, 'reward_total_composite_std': 0.06995871663093567, 'epoch': 0.06}

 43%|████▎     | 1434/3300 [1:34:22<2:35:02,  4.99s/it]INFO 04-12 00:11:13 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1435/3300 [1:34:26<2:30:24,  4.84s/it]2026-04-12 00:11:17,504 | INFO | train_grpo_train | metrics_logged mode=train step=1435

                                                       
{'loss': -0.0, 'grad_norm': 6.715077877044678, 'learning_rate': 5.654545454545455e-06, 'num_tokens': 3228023.0, 'completions/mean_length': 45.5, 'completions/min_length': 43.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8577103614807129, 'rewards/meter/std': 0.20525795221328735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8577103614807129, 'rewards/total_composite/std': 0.20525795221328735, 'reward': 0.8577103614807129, 'reward_std': 0.20525793731212616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021754316985607147, 'sampling/sampling_logp_difference/max': 0.6161007881164551, 'sampling/importance_sampling_ratio/min': 0.5400460958480835, 'sampling/importance_sampling_ratio/mean': 0.9992294311523438, 'sampling/importance_sampling_ratio/max': 1.5408525466918945, 'entropy': 0.1447599595412612, 'clip_ratio/low_mean': 0.00872093066573143, 'clip_ratio/low_min': 0.00872093066573143, 'clip_ratio/high_mean': 0.019032032461836934, 'clip_ratio/high_max': 0.019032032461836934, 'clip_ratio/region_mean': 0.027752963127568364, 'reward_total_mean': 0.8577103614807129, 'reward_meter_mean': 0.8577103614807129, 'reward_meter_std': 0.20525795221328735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8577103614807129, 'reward_total_composite_std': 0.20525795221328735, 'epoch': 0.06}

 43%|████▎     | 1435/3300 [1:34:26<2:30:24,  4.84s/it]INFO 04-12 00:11:17 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1436/3300 [1:34:31<2:28:49,  4.79s/it]2026-04-12 00:11:22,183 | INFO | train_grpo_train | metrics_logged mode=train step=1436

                                                       
{'loss': -0.0268, 'grad_norm': 5.0643205642700195, 'learning_rate': 5.651515151515152e-06, 'num_tokens': 3229879.0, 'completions/mean_length': 64.0, 'completions/min_length': 60.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9666499495506287, 'rewards/meter/std': 0.03790951892733574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9666499495506287, 'rewards/total_composite/std': 0.03790951892733574, 'reward': 0.9666499495506287, 'reward_std': 0.03790954127907753, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030510857701301575, 'sampling/sampling_logp_difference/max': 1.486882209777832, 'sampling/importance_sampling_ratio/min': 0.2260764092206955, 'sampling/importance_sampling_ratio/mean': 1.0077619552612305, 'sampling/importance_sampling_ratio/max': 1.844899296760559, 'entropy': 0.13815412484109402, 'clip_ratio/low_mean': 0.008230874547734857, 'clip_ratio/low_min': 0.008230874547734857, 'clip_ratio/high_mean': 0.024871644098311663, 'clip_ratio/high_max': 0.024871644098311663, 'clip_ratio/region_mean': 0.03310251864604652, 'reward_total_mean': 0.9666499495506287, 'reward_meter_mean': 0.9666499495506287, 'reward_meter_std': 0.03790951892733574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9666499495506287, 'reward_total_composite_std': 0.03790951892733574, 'epoch': 0.06}

 44%|████▎     | 1436/3300 [1:34:31<2:28:49,  4.79s/it]INFO 04-12 00:11:22 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1437/3300 [1:34:35<2:27:54,  4.76s/it]2026-04-12 00:11:26,881 | INFO | train_grpo_train | metrics_logged mode=train step=1437

                                                       
{'loss': -0.0032, 'grad_norm': 7.082005500793457, 'learning_rate': 5.648484848484849e-06, 'num_tokens': 3231833.0, 'completions/mean_length': 64.25, 'completions/min_length': 61.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.990328311920166, 'rewards/meter/std': 0.005682698916643858, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.990328311920166, 'rewards/total_composite/std': 0.005682698916643858, 'reward': 0.990328311920166, 'reward_std': 0.005682693794369698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0303493719547987, 'sampling/sampling_logp_difference/max': 2.149282455444336, 'sampling/importance_sampling_ratio/min': 0.11656776815652847, 'sampling/importance_sampling_ratio/mean': 0.9941160082817078, 'sampling/importance_sampling_ratio/max': 1.7731237411499023, 'entropy': 0.1368792960420251, 'clip_ratio/low_mean': 0.005988386110402644, 'clip_ratio/low_min': 0.005988386110402644, 'clip_ratio/high_mean': 0.005740093300119042, 'clip_ratio/high_max': 0.005740093300119042, 'clip_ratio/region_mean': 0.011728479410521686, 'reward_total_mean': 0.990328311920166, 'reward_meter_mean': 0.990328311920166, 'reward_meter_std': 0.005682698916643858, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.990328311920166, 'reward_total_composite_std': 0.005682698916643858, 'epoch': 0.06}

 44%|████▎     | 1437/3300 [1:34:35<2:27:54,  4.76s/it]INFO 04-12 00:11:27 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1438/3300 [1:34:40<2:26:12,  4.71s/it]2026-04-12 00:11:31,472 | INFO | train_grpo_train | metrics_logged mode=train step=1438

                                                       
{'loss': -0.009, 'grad_norm': 4.488318920135498, 'learning_rate': 5.645454545454546e-06, 'num_tokens': 3233673.0, 'completions/mean_length': 54.0, 'completions/min_length': 47.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9756955504417419, 'rewards/meter/std': 0.017869651317596436, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9756955504417419, 'rewards/total_composite/std': 0.017869651317596436, 'reward': 0.9756955504417419, 'reward_std': 0.01786964386701584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03772404417395592, 'sampling/sampling_logp_difference/max': 1.3473641872406006, 'sampling/importance_sampling_ratio/min': 0.2599244713783264, 'sampling/importance_sampling_ratio/mean': 1.008415699005127, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21246233023703098, 'clip_ratio/low_mean': 0.007591875968500972, 'clip_ratio/low_min': 0.007591875968500972, 'clip_ratio/high_mean': 0.018068846315145493, 'clip_ratio/high_max': 0.018068846315145493, 'clip_ratio/region_mean': 0.025660722283646464, 'reward_total_mean': 0.9756955504417419, 'reward_meter_mean': 0.9756955504417419, 'reward_meter_std': 0.017869651317596436, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9756955504417419, 'reward_total_composite_std': 0.017869651317596436, 'epoch': 0.06}

 44%|████▎     | 1438/3300 [1:34:40<2:26:12,  4.71s/it]INFO 04-12 00:11:31 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1439/3300 [1:34:45<2:25:51,  4.70s/it]2026-04-12 00:11:36,153 | INFO | train_grpo_train | metrics_logged mode=train step=1439

                                                       
{'loss': -0.0046, 'grad_norm': 3.8302407264709473, 'learning_rate': 5.642424242424242e-06, 'num_tokens': 3235514.0, 'completions/mean_length': 66.125, 'completions/min_length': 59.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9976571798324585, 'rewards/meter/std': 0.001123868627473712, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976571798324585, 'rewards/total_composite/std': 0.001123868627473712, 'reward': 0.9976571798324585, 'reward_std': 0.001123874681070447, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04735542833805084, 'sampling/sampling_logp_difference/max': 1.1300544738769531, 'sampling/importance_sampling_ratio/min': 0.3230156898498535, 'sampling/importance_sampling_ratio/mean': 1.0110960006713867, 'sampling/importance_sampling_ratio/max': 1.605924129486084, 'entropy': 0.2875422090291977, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.03688650333788246, 'clip_ratio/high_max': 0.03688650333788246, 'clip_ratio/region_mean': 0.044462261139415205, 'reward_total_mean': 0.9976571798324585, 'reward_meter_mean': 0.9976571798324585, 'reward_meter_std': 0.001123868627473712, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976571798324585, 'reward_total_composite_std': 0.001123868627473712, 'epoch': 0.06}

 44%|████▎     | 1439/3300 [1:34:45<2:25:51,  4.70s/it]INFO 04-12 00:11:36 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1440/3300 [1:34:51<2:40:54,  5.19s/it]2026-04-12 00:11:42,483 | INFO | train_grpo_train | metrics_logged mode=train step=1440

                                                       
{'loss': -0.0197, 'grad_norm': 2.74971342086792, 'learning_rate': 5.6393939393939405e-06, 'num_tokens': 3238860.0, 'completions/mean_length': 197.25, 'completions/min_length': 188.0, 'completions/max_length': 211.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.25, 'completions/min_terminated_length': 188.0, 'completions/max_terminated_length': 211.0, 'rewards/meter/mean': 0.9103585481643677, 'rewards/meter/std': 0.11414875090122223, 'rewards/count_adherence/mean': 0.8392857313156128, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8397727012634277, 'rewards/repeat_penalty/std': 0.1051156297326088, 'rewards/total_composite/mean': 0.641140341758728, 'rewards/total_composite/std': 0.11881579458713531, 'reward': 0.641140341758728, 'reward_std': 0.11881577223539352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.052030667662620544, 'sampling/sampling_logp_difference/max': 1.4830026626586914, 'sampling/importance_sampling_ratio/min': 0.22695519030094147, 'sampling/importance_sampling_ratio/mean': 1.0104938745498657, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3795258644968271, 'clip_ratio/low_mean': 0.01373849913943559, 'clip_ratio/low_min': 0.01373849913943559, 'clip_ratio/high_mean': 0.025198507588356733, 'clip_ratio/high_max': 0.025198507588356733, 'clip_ratio/region_mean': 0.03893700672779232, 'reward_total_mean': 0.641140341758728, 'reward_meter_mean': 0.9103585481643677, 'reward_meter_std': 0.11414875090122223, 'reward_count_adherence_mean': 0.8392857313156128, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8397727012634277, 'reward_repeat_penalty_std': 0.1051156297326088, 'reward_total_composite_mean': 0.641140341758728, 'reward_total_composite_std': 0.11881579458713531, 'epoch': 0.06}

 44%|████▎     | 1440/3300 [1:34:51<2:40:54,  5.19s/it]INFO 04-12 00:11:42 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1441/3300 [1:34:55<2:34:29,  4.99s/it]2026-04-12 00:11:46,993 | INFO | train_grpo_train | metrics_logged mode=train step=1441

                                                       
{'loss': -0.0023, 'grad_norm': 5.571752548217773, 'learning_rate': 5.636363636363636e-06, 'num_tokens': 3240498.0, 'completions/mean_length': 54.75, 'completions/min_length': 54.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.75, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9891074895858765, 'rewards/meter/std': 0.0014680020976811647, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9891074895858765, 'rewards/total_composite/std': 0.0014680020976811647, 'reward': 0.9891074895858765, 'reward_std': 0.0014680068707093596, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022303204983472824, 'sampling/sampling_logp_difference/max': 0.8718528747558594, 'sampling/importance_sampling_ratio/min': 0.41817599534988403, 'sampling/importance_sampling_ratio/mean': 1.0070688724517822, 'sampling/importance_sampling_ratio/max': 1.8307318687438965, 'entropy': 0.15620220359414816, 'clip_ratio/low_mean': 0.009132996667176485, 'clip_ratio/low_min': 0.009132996667176485, 'clip_ratio/high_mean': 0.004545454401522875, 'clip_ratio/high_max': 0.004545454401522875, 'clip_ratio/region_mean': 0.01367845106869936, 'reward_total_mean': 0.9891074895858765, 'reward_meter_mean': 0.9891074895858765, 'reward_meter_std': 0.0014680020976811647, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9891074895858765, 'reward_total_composite_std': 0.0014680020976811647, 'epoch': 0.06}

 44%|████▎     | 1441/3300 [1:34:55<2:34:29,  4.99s/it]INFO 04-12 00:11:47 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1442/3300 [1:35:01<2:43:58,  5.30s/it]2026-04-12 00:11:53,009 | INFO | train_grpo_train | metrics_logged mode=train step=1442

                                                       
{'loss': 0.0076, 'grad_norm': 4.191330909729004, 'learning_rate': 5.633333333333334e-06, 'num_tokens': 3243758.0, 'completions/mean_length': 168.5, 'completions/min_length': 145.0, 'completions/max_length': 182.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.5, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.12805184721946716, 'rewards/meter/std': 0.1377500742673874, 'rewards/count_adherence/mean': 0.6250000596046448, 'rewards/count_adherence/std': 0.05750546231865883, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9217171669006348, 'rewards/repeat_penalty/std': 0.1535457819700241, 'rewards/total_composite/mean': 0.07739339768886566, 'rewards/total_composite/std': 0.08249951899051666, 'reward': 0.07739339768886566, 'reward_std': 0.08249951899051666, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.089544378221035, 'sampling/sampling_logp_difference/max': 3.6343159675598145, 'sampling/importance_sampling_ratio/min': 0.02640198916196823, 'sampling/importance_sampling_ratio/mean': 1.0114877223968506, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7288497872650623, 'clip_ratio/low_mean': 0.0417820424772799, 'clip_ratio/low_min': 0.0417820424772799, 'clip_ratio/high_mean': 0.030609328765422106, 'clip_ratio/high_max': 0.030609328765422106, 'clip_ratio/region_mean': 0.07239137124270201, 'reward_total_mean': 0.07739339768886566, 'reward_meter_mean': 0.12805184721946716, 'reward_meter_std': 0.1377500742673874, 'reward_count_adherence_mean': 0.6250000596046448, 'reward_count_adherence_std': 0.05750546231865883, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9217171669006348, 'reward_repeat_penalty_std': 0.1535457819700241, 'reward_total_composite_mean': 0.07739339768886566, 'reward_total_composite_std': 0.08249951899051666, 'epoch': 0.06}

 44%|████▎     | 1442/3300 [1:35:02<2:43:58,  5.30s/it]INFO 04-12 00:11:53 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1443/3300 [1:35:09<3:02:27,  5.90s/it]2026-04-12 00:12:00,303 | INFO | train_grpo_train | metrics_logged mode=train step=1443

                                                       
{'loss': -0.0031, 'grad_norm': 2.3659539222717285, 'learning_rate': 5.630303030303031e-06, 'num_tokens': 3247759.0, 'completions/mean_length': 274.125, 'completions/min_length': 263.0, 'completions/max_length': 293.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 274.125, 'completions/min_terminated_length': 263.0, 'completions/max_terminated_length': 293.0, 'rewards/meter/mean': 0.993506908416748, 'rewards/meter/std': 0.008024541661143303, 'rewards/count_adherence/mean': 0.7250000238418579, 'rewards/count_adherence/std': 0.04629101604223251, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8730769157409668, 'rewards/repeat_penalty/std': 0.05622680485248566, 'rewards/total_composite/mean': 0.6270208954811096, 'rewards/total_composite/std': 0.023548215627670288, 'reward': 0.6270208954811096, 'reward_std': 0.023548215627670288, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04566960409283638, 'sampling/sampling_logp_difference/max': 2.6650161743164062, 'sampling/importance_sampling_ratio/min': 0.0695982277393341, 'sampling/importance_sampling_ratio/mean': 1.0080655813217163, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34477731212973595, 'clip_ratio/low_mean': 0.008868969744071364, 'clip_ratio/low_min': 0.008868969744071364, 'clip_ratio/high_mean': 0.026962524512782693, 'clip_ratio/high_max': 0.026962524512782693, 'clip_ratio/region_mean': 0.03583149425685406, 'reward_total_mean': 0.6270208954811096, 'reward_meter_mean': 0.993506908416748, 'reward_meter_std': 0.008024541661143303, 'reward_count_adherence_mean': 0.7250000238418579, 'reward_count_adherence_std': 0.04629101604223251, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8730769157409668, 'reward_repeat_penalty_std': 0.05622680485248566, 'reward_total_composite_mean': 0.6270208954811096, 'reward_total_composite_std': 0.023548215627670288, 'epoch': 0.06}

 44%|████▎     | 1443/3300 [1:35:09<3:02:27,  5.90s/it]INFO 04-12 00:12:00 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1444/3300 [1:35:14<2:59:40,  5.81s/it]2026-04-12 00:12:05,910 | INFO | train_grpo_train | metrics_logged mode=train step=1444

                                                       
{'loss': -0.0009, 'grad_norm': 3.132058620452881, 'learning_rate': 5.627272727272728e-06, 'num_tokens': 3250457.0, 'completions/mean_length': 150.25, 'completions/min_length': 143.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.25, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.9959889054298401, 'rewards/meter/std': 0.00317049166187644, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9605348110198975, 'rewards/total_composite/std': 0.06770440191030502, 'reward': 0.9605348110198975, 'reward_std': 0.06770440936088562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05538027361035347, 'sampling/sampling_logp_difference/max': 1.4519987106323242, 'sampling/importance_sampling_ratio/min': 0.23410192131996155, 'sampling/importance_sampling_ratio/mean': 1.0069429874420166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3815644532442093, 'clip_ratio/low_mean': 0.006868131808005273, 'clip_ratio/low_min': 0.006868131808005273, 'clip_ratio/high_mean': 0.035727029433473945, 'clip_ratio/high_max': 0.035727029433473945, 'clip_ratio/region_mean': 0.04259516124147922, 'reward_total_mean': 0.9605348110198975, 'reward_meter_mean': 0.9959889054298401, 'reward_meter_std': 0.00317049166187644, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9605348110198975, 'reward_total_composite_std': 0.06770440191030502, 'epoch': 0.06}

 44%|████▍     | 1444/3300 [1:35:14<2:59:40,  5.81s/it]INFO 04-12 00:12:06 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1445/3300 [1:35:21<3:02:55,  5.92s/it]2026-04-12 00:12:12,079 | INFO | train_grpo_train | metrics_logged mode=train step=1445

                                                       
{'loss': 0.0052, 'grad_norm': 3.3944966793060303, 'learning_rate': 5.624242424242424e-06, 'num_tokens': 3253588.0, 'completions/mean_length': 188.375, 'completions/min_length': 182.0, 'completions/max_length': 196.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 188.375, 'completions/min_terminated_length': 182.0, 'completions/max_terminated_length': 196.0, 'rewards/meter/mean': 0.9968078136444092, 'rewards/meter/std': 0.0010806269710883498, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.08399210125207901, 'rewards/total_composite/mean': 0.9414394497871399, 'rewards/total_composite/std': 0.08388549834489822, 'reward': 0.9414394497871399, 'reward_std': 0.08388549834489822, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.050331130623817444, 'sampling/sampling_logp_difference/max': 1.6728553771972656, 'sampling/importance_sampling_ratio/min': 0.18771031498908997, 'sampling/importance_sampling_ratio/mean': 1.013535737991333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.38519781827926636, 'clip_ratio/low_mean': 0.012751690810546279, 'clip_ratio/low_min': 0.012751690810546279, 'clip_ratio/high_mean': 0.02173018571920693, 'clip_ratio/high_max': 0.02173018571920693, 'clip_ratio/region_mean': 0.03448187652975321, 'reward_total_mean': 0.9414394497871399, 'reward_meter_mean': 0.9968078136444092, 'reward_meter_std': 0.0010806269710883498, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.08399210125207901, 'reward_total_composite_mean': 0.9414394497871399, 'reward_total_composite_std': 0.08388549834489822, 'epoch': 0.06}

 44%|████▍     | 1445/3300 [1:35:21<3:02:55,  5.92s/it]INFO 04-12 00:12:12 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1446/3300 [1:35:25<2:51:36,  5.55s/it]2026-04-12 00:12:16,785 | INFO | train_grpo_train | metrics_logged mode=train step=1446

                                                       
{'loss': 0.0122, 'grad_norm': 5.573071479797363, 'learning_rate': 5.6212121212121215e-06, 'num_tokens': 3255370.0, 'completions/mean_length': 70.75, 'completions/min_length': 69.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.994280993938446, 'rewards/meter/std': 0.0039334069006145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.994280993938446, 'rewards/total_composite/std': 0.0039334069006145, 'reward': 0.994280993938446, 'reward_std': 0.003933423198759556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07322768121957779, 'sampling/sampling_logp_difference/max': 1.5738625526428223, 'sampling/importance_sampling_ratio/min': 0.2072431445121765, 'sampling/importance_sampling_ratio/mean': 1.0078636407852173, 'sampling/importance_sampling_ratio/max': 1.8624515533447266, 'entropy': 0.48965371400117874, 'clip_ratio/low_mean': 0.017605633474886417, 'clip_ratio/low_min': 0.017605633474886417, 'clip_ratio/high_mean': 0.05117733031511307, 'clip_ratio/high_max': 0.05117733031511307, 'clip_ratio/region_mean': 0.06878296378999949, 'reward_total_mean': 0.994280993938446, 'reward_meter_mean': 0.994280993938446, 'reward_meter_std': 0.0039334069006145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.994280993938446, 'reward_total_composite_std': 0.0039334069006145, 'epoch': 0.06}

 44%|████▍     | 1446/3300 [1:35:25<2:51:36,  5.55s/it]INFO 04-12 00:12:17 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1447/3300 [1:35:30<2:39:40,  5.17s/it]2026-04-12 00:12:21,062 | INFO | train_grpo_train | metrics_logged mode=train step=1447

                                                       
{'loss': 0.0006, 'grad_norm': 7.160624980926514, 'learning_rate': 5.618181818181818e-06, 'num_tokens': 3256876.0, 'completions/mean_length': 35.25, 'completions/min_length': 34.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9780558943748474, 'rewards/meter/std': 0.022425083443522453, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9780558943748474, 'rewards/total_composite/std': 0.022425083443522453, 'reward': 0.9780558943748474, 'reward_std': 0.022425075992941856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04081781581044197, 'sampling/sampling_logp_difference/max': 0.8824992179870605, 'sampling/importance_sampling_ratio/min': 0.4137475788593292, 'sampling/importance_sampling_ratio/mean': 1.0069624185562134, 'sampling/importance_sampling_ratio/max': 1.6601144075393677, 'entropy': 0.27801981568336487, 'clip_ratio/low_mean': 0.0173611119389534, 'clip_ratio/low_min': 0.0173611119389534, 'clip_ratio/high_mean': 0.03094316739588976, 'clip_ratio/high_max': 0.03094316739588976, 'clip_ratio/region_mean': 0.04830427933484316, 'reward_total_mean': 0.9780558943748474, 'reward_meter_mean': 0.9780558943748474, 'reward_meter_std': 0.022425083443522453, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9780558943748474, 'reward_total_composite_std': 0.022425083443522453, 'epoch': 0.06}

 44%|████▍     | 1447/3300 [1:35:30<2:39:40,  5.17s/it]INFO 04-12 00:12:21 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1448/3300 [1:35:35<2:40:12,  5.19s/it]2026-04-12 00:12:26,298 | INFO | train_grpo_train | metrics_logged mode=train step=1448

                                                       
{'loss': 0.0162, 'grad_norm': 2.8199713230133057, 'learning_rate': 5.615151515151516e-06, 'num_tokens': 3259212.0, 'completions/mean_length': 114.0, 'completions/min_length': 111.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.0, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.9958944320678711, 'rewards/meter/std': 0.0037708207964897156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9709444046020508, 'rewards/total_composite/std': 0.06981504708528519, 'reward': 0.9709444046020508, 'reward_std': 0.06981503963470459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05924813821911812, 'sampling/sampling_logp_difference/max': 1.7622275352478027, 'sampling/importance_sampling_ratio/min': 0.17166204750537872, 'sampling/importance_sampling_ratio/mean': 1.0023696422576904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.348561218008399, 'clip_ratio/low_mean': 0.0032051282469183207, 'clip_ratio/low_min': 0.0032051282469183207, 'clip_ratio/high_mean': 0.04604411777108908, 'clip_ratio/high_max': 0.04604411777108908, 'clip_ratio/region_mean': 0.0492492460180074, 'reward_total_mean': 0.9709444046020508, 'reward_meter_mean': 0.9958944320678711, 'reward_meter_std': 0.0037708207964897156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9709444046020508, 'reward_total_composite_std': 0.06981504708528519, 'epoch': 0.06}

 44%|████▍     | 1448/3300 [1:35:35<2:40:12,  5.19s/it]INFO 04-12 00:12:26 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1449/3300 [1:35:41<2:46:50,  5.41s/it]2026-04-12 00:12:32,214 | INFO | train_grpo_train | metrics_logged mode=train step=1449

                                                       
{'loss': 0.0176, 'grad_norm': 2.3722407817840576, 'learning_rate': 5.612121212121212e-06, 'num_tokens': 3262076.0, 'completions/mean_length': 171.0, 'completions/min_length': 165.0, 'completions/max_length': 178.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.0, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 178.0, 'rewards/meter/mean': 0.4469248056411743, 'rewards/meter/std': 0.35773399472236633, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.07120776921510696, 'rewards/total_composite/mean': 0.29604610800743103, 'rewards/total_composite/std': 0.23469020426273346, 'reward': 0.29604610800743103, 'reward_std': 0.23469018936157227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025468070060014725, 'sampling/sampling_logp_difference/max': 2.328211784362793, 'sampling/importance_sampling_ratio/min': 0.09746988862752914, 'sampling/importance_sampling_ratio/mean': 1.0035395622253418, 'sampling/importance_sampling_ratio/max': 1.7698057889938354, 'entropy': 0.16733690444380045, 'clip_ratio/low_mean': 0.007986687181983143, 'clip_ratio/low_min': 0.007986687181983143, 'clip_ratio/high_mean': 0.011789240641519427, 'clip_ratio/high_max': 0.011789240641519427, 'clip_ratio/region_mean': 0.01977592782350257, 'reward_total_mean': 0.29604610800743103, 'reward_meter_mean': 0.4469248056411743, 'reward_meter_std': 0.35773399472236633, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.07120776921510696, 'reward_total_composite_mean': 0.29604610800743103, 'reward_total_composite_std': 0.23469020426273346, 'epoch': 0.06}

 44%|████▍     | 1449/3300 [1:35:41<2:46:50,  5.41s/it]INFO 04-12 00:12:32 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1450/3300 [1:35:47<2:53:45,  5.64s/it]2026-04-12 00:12:38,380 | INFO | train_grpo_train | metrics_logged mode=train step=1450

                                                       
{'loss': 0.0088, 'grad_norm': 2.1575472354888916, 'learning_rate': 5.60909090909091e-06, 'num_tokens': 3265160.0, 'completions/mean_length': 192.5, 'completions/min_length': 184.0, 'completions/max_length': 197.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 192.5, 'completions/min_terminated_length': 184.0, 'completions/max_terminated_length': 197.0, 'rewards/meter/mean': 0.9949524998664856, 'rewards/meter/std': 0.0050539844669401646, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666269302368, 'rewards/repeat_penalty/std': 0.07856741547584534, 'rewards/total_composite/mean': 0.911919116973877, 'rewards/total_composite/std': 0.0767272487282753, 'reward': 0.911919116973877, 'reward_std': 0.07672726362943649, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04098372161388397, 'sampling/sampling_logp_difference/max': 2.1571097373962402, 'sampling/importance_sampling_ratio/min': 0.11565891653299332, 'sampling/importance_sampling_ratio/mean': 1.0060464143753052, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28528641909360886, 'clip_ratio/low_mean': 0.017584156128577888, 'clip_ratio/low_min': 0.017584156128577888, 'clip_ratio/high_mean': 0.015601763967424631, 'clip_ratio/high_max': 0.015601763967424631, 'clip_ratio/region_mean': 0.03318592009600252, 'reward_total_mean': 0.911919116973877, 'reward_meter_mean': 0.9949524998664856, 'reward_meter_std': 0.0050539844669401646, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666269302368, 'reward_repeat_penalty_std': 0.07856741547584534, 'reward_total_composite_mean': 0.911919116973877, 'reward_total_composite_std': 0.0767272487282753, 'epoch': 0.06}

 44%|████▍     | 1450/3300 [1:35:47<2:53:45,  5.64s/it]INFO 04-12 00:12:38 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:23,  2.14s/it]

 23%|██▎       | 3/13 [00:06<00:23,  2.32s/it]

 31%|███       | 4/13 [00:11<00:27,  3.08s/it]

 38%|███▊      | 5/13 [00:15<00:27,  3.46s/it]

 46%|████▌     | 6/13 [00:20<00:27,  3.86s/it]

 54%|█████▍    | 7/13 [00:23<00:23,  3.86s/it]

 62%|██████▏   | 8/13 [00:28<00:21,  4.24s/it]

 69%|██████▉   | 9/13 [00:34<00:18,  4.65s/it]

 77%|███████▋  | 10/13 [00:38<00:13,  4.43s/it]

 85%|████████▍ | 11/13 [00:42<00:08,  4.17s/it]

 92%|█████████▏| 12/13 [00:46<00:04,  4.21s/it]

100%|██████████| 13/13 [00:50<00:00,  4.28s/it]2026-04-12 00:13:33,718 | INFO | train_grpo_train | metrics_logged mode=eval step=1450

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 55.2927, 'eval_samples_per_second': 1.881, 'eval_steps_per_second': 0.235, 'eval_num_tokens': 3265160.0, 'eval_completions/mean_length': 160.5, 'eval_completions/min_length': 60.46153846153846, 'eval_completions/max_length': 283.0769230769231, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 160.5, 'eval_completions/min_terminated_length': 60.46153846153846, 'eval_completions/max_terminated_length': 283.0769230769231, 'eval_rewards/meter/mean': 0.5874437128122036, 'eval_rewards/meter/std': 0.4289410481086144, 'eval_rewards/count_adherence/mean': 0.8191676002282363, 'eval_rewards/count_adherence/std': 0.17260103730055001, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8410639992127051, 'eval_rewards/repeat_penalty/std': 0.16562061241039863, 'eval_rewards/total_composite/mean': 0.4260200307919429, 'eval_rewards/total_composite/std': 0.3662373045316109, 'eval_reward': 0.4260200307919429, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.02707064416832649, 'eval_sampling/sampling_logp_difference/max': 0.9593030489408053, 'eval_sampling/importance_sampling_ratio/min': 0.3921516973238725, 'eval_sampling/importance_sampling_ratio/mean': 1.0070600234545195, 'eval_sampling/importance_sampling_ratio/max': 1.4355276914743276, 'eval_entropy': 0.26288481056690216, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.4260200307919429, 'eval_reward_meter_mean': 0.5874437128122036, 'eval_reward_meter_std': 0.4289410481086144, 'eval_reward_count_adherence_mean': 0.8191676002282363, 'eval_reward_count_adherence_std': 0.17260103730055001, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8410639992127051, 'eval_reward_repeat_penalty_std': 0.16562061241039863, 'eval_reward_total_composite_mean': 0.4260200307919429, 'eval_reward_total_composite_std': 0.3662373045316109, 'epoch': 0.06}

 44%|████▍     | 1450/3300 [1:36:42<2:53:45,  5.64s/it]

100%|██████████| 13/13 [00:50<00:00,  4.28s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 44%|████▍     | 1451/3300 [1:36:49<11:40:36, 22.73s/it]2026-04-12 00:13:41,035 | INFO | train_grpo_train | metrics_logged mode=train step=1451

                                                        
{'loss': -0.0183, 'grad_norm': 9.900911331176758, 'learning_rate': 5.606060606060606e-06, 'num_tokens': 3266919.0, 'completions/mean_length': 61.875, 'completions/min_length': 55.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.5548748970031738, 'rewards/meter/std': 0.3548371493816376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5548748970031738, 'rewards/total_composite/std': 0.3548371493816376, 'reward': 0.5548748970031738, 'reward_std': 0.3548371493816376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12563803791999817, 'sampling/sampling_logp_difference/max': 1.3014025688171387, 'sampling/importance_sampling_ratio/min': 0.27214980125427246, 'sampling/importance_sampling_ratio/mean': 1.0315823554992676, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.44280257076025, 'clip_ratio/low_mean': 0.04481779085472226, 'clip_ratio/low_min': 0.04481779085472226, 'clip_ratio/high_mean': 0.04324009292759001, 'clip_ratio/high_max': 0.04324009292759001, 'clip_ratio/region_mean': 0.08805788378231227, 'reward_total_mean': 0.5548748970031738, 'reward_meter_mean': 0.5548748970031738, 'reward_meter_std': 0.3548371493816376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5548748970031738, 'reward_total_composite_std': 0.3548371493816376, 'epoch': 0.06}

 44%|████▍     | 1451/3300 [1:36:50<11:40:36, 22.73s/it]INFO 04-12 00:13:41 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1452/3300 [1:36:55<8:59:28, 17.52s/it] 2026-04-12 00:13:46,350 | INFO | train_grpo_train | metrics_logged mode=train step=1452

                                                       
{'loss': 0.0264, 'grad_norm': 3.927211046218872, 'learning_rate': 5.603030303030303e-06, 'num_tokens': 3269067.0, 'completions/mean_length': 102.5, 'completions/min_length': 95.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.5, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9900820255279541, 'rewards/meter/std': 0.011413360014557838, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9900820255279541, 'rewards/total_composite/std': 0.011413360014557838, 'reward': 0.9900820255279541, 'reward_std': 0.011413362808525562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06290001422166824, 'sampling/sampling_logp_difference/max': 1.65745210647583, 'sampling/importance_sampling_ratio/min': 0.19062404334545135, 'sampling/importance_sampling_ratio/mean': 1.0114924907684326, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5281493403017521, 'clip_ratio/low_mean': 0.013940956210717559, 'clip_ratio/low_min': 0.013940956210717559, 'clip_ratio/high_mean': 0.041309412801638246, 'clip_ratio/high_max': 0.041309412801638246, 'clip_ratio/region_mean': 0.055250369012355804, 'reward_total_mean': 0.9900820255279541, 'reward_meter_mean': 0.9900820255279541, 'reward_meter_std': 0.011413360014557838, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9900820255279541, 'reward_total_composite_std': 0.011413360014557838, 'epoch': 0.06}

 44%|████▍     | 1452/3300 [1:36:55<8:59:28, 17.52s/it]INFO 04-12 00:13:46 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1453/3300 [1:36:59<7:00:38, 13.66s/it]2026-04-12 00:13:51,031 | INFO | train_grpo_train | metrics_logged mode=train step=1453

                                                       
{'loss': 0.029, 'grad_norm': 4.87343168258667, 'learning_rate': 5.600000000000001e-06, 'num_tokens': 3271008.0, 'completions/mean_length': 61.625, 'completions/min_length': 58.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.18945728242397308, 'rewards/meter/std': 0.2888922691345215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.18945728242397308, 'rewards/total_composite/std': 0.2888922691345215, 'reward': 0.18945728242397308, 'reward_std': 0.28889229893684387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06080813705921173, 'sampling/sampling_logp_difference/max': 2.205195426940918, 'sampling/importance_sampling_ratio/min': 0.11022898554801941, 'sampling/importance_sampling_ratio/mean': 1.011013150215149, 'sampling/importance_sampling_ratio/max': 1.7248963117599487, 'entropy': 0.5127889458090067, 'clip_ratio/low_mean': 0.024039480136707425, 'clip_ratio/low_min': 0.024039480136707425, 'clip_ratio/high_mean': 0.01592310261912644, 'clip_ratio/high_max': 0.01592310261912644, 'clip_ratio/region_mean': 0.039962582755833864, 'reward_total_mean': 0.18945728242397308, 'reward_meter_mean': 0.18945728242397308, 'reward_meter_std': 0.2888922691345215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.18945728242397308, 'reward_total_composite_std': 0.2888922691345215, 'epoch': 0.06}

 44%|████▍     | 1453/3300 [1:37:00<7:00:38, 13.66s/it]INFO 04-12 00:13:51 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1454/3300 [1:37:04<5:37:08, 10.96s/it]2026-04-12 00:13:55,673 | INFO | train_grpo_train | metrics_logged mode=train step=1454

                                                       
{'loss': 0.0102, 'grad_norm': 3.9825429916381836, 'learning_rate': 5.596969696969697e-06, 'num_tokens': 3272821.0, 'completions/mean_length': 58.625, 'completions/min_length': 57.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9886520504951477, 'rewards/meter/std': 0.006697438657283783, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9886520504951477, 'rewards/total_composite/std': 0.006697438657283783, 'reward': 0.9886520504951477, 'reward_std': 0.00669743912294507, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03104584477841854, 'sampling/sampling_logp_difference/max': 1.1052379608154297, 'sampling/importance_sampling_ratio/min': 0.33113205432891846, 'sampling/importance_sampling_ratio/mean': 0.9944877028465271, 'sampling/importance_sampling_ratio/max': 1.4476277828216553, 'entropy': 0.15899417083710432, 'clip_ratio/low_mean': 0.004204352619126439, 'clip_ratio/low_min': 0.004204352619126439, 'clip_ratio/high_mean': 0.017170886043459177, 'clip_ratio/high_max': 0.017170886043459177, 'clip_ratio/region_mean': 0.021375238662585616, 'reward_total_mean': 0.9886520504951477, 'reward_meter_mean': 0.9886520504951477, 'reward_meter_std': 0.006697438657283783, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9886520504951477, 'reward_total_composite_std': 0.006697438657283783, 'epoch': 0.06}

 44%|████▍     | 1454/3300 [1:37:04<5:37:08, 10.96s/it]INFO 04-12 00:13:55 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1455/3300 [1:37:08<4:36:03,  8.98s/it]2026-04-12 00:14:00,030 | INFO | train_grpo_train | metrics_logged mode=train step=1455

                                                       
{'loss': 0.0227, 'grad_norm': 7.097043514251709, 'learning_rate': 5.593939393939395e-06, 'num_tokens': 3274349.0, 'completions/mean_length': 36.0, 'completions/min_length': 35.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9973074793815613, 'rewards/meter/std': 0.0026173454243689775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973074793815613, 'rewards/total_composite/std': 0.0026173454243689775, 'reward': 0.9973074793815613, 'reward_std': 0.0026173454243689775, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04888688772916794, 'sampling/sampling_logp_difference/max': 0.6760835647583008, 'sampling/importance_sampling_ratio/min': 0.5086050033569336, 'sampling/importance_sampling_ratio/mean': 0.9993876218795776, 'sampling/importance_sampling_ratio/max': 1.5138667821884155, 'entropy': 0.32455965876579285, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.0423718374222517, 'clip_ratio/high_max': 0.0423718374222517, 'clip_ratio/region_mean': 0.045844059670343995, 'reward_total_mean': 0.9973074793815613, 'reward_meter_mean': 0.9973074793815613, 'reward_meter_std': 0.0026173454243689775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973074793815613, 'reward_total_composite_std': 0.0026173454243689775, 'epoch': 0.06}

 44%|████▍     | 1455/3300 [1:37:09<4:36:03,  8.98s/it]INFO 04-12 00:14:00 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1456/3300 [1:37:16<4:19:20,  8.44s/it]2026-04-12 00:14:07,209 | INFO | train_grpo_train | metrics_logged mode=train step=1456

                                                       
{'loss': -0.0485, 'grad_norm': 2.318981170654297, 'learning_rate': 5.5909090909090915e-06, 'num_tokens': 3278247.0, 'completions/mean_length': 263.25, 'completions/min_length': 234.0, 'completions/max_length': 277.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 263.25, 'completions/min_terminated_length': 234.0, 'completions/max_terminated_length': 277.0, 'rewards/meter/mean': 0.9962884187698364, 'rewards/meter/std': 0.0015494668623432517, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.911057710647583, 'rewards/repeat_penalty/std': 0.051787521690130234, 'rewards/total_composite/mean': 0.7669380903244019, 'rewards/total_composite/std': 0.07985159009695053, 'reward': 0.7669380903244019, 'reward_std': 0.07985157519578934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05095827206969261, 'sampling/sampling_logp_difference/max': 1.4401264190673828, 'sampling/importance_sampling_ratio/min': 0.2368977963924408, 'sampling/importance_sampling_ratio/mean': 1.0123025178909302, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41418151557445526, 'clip_ratio/low_mean': 0.012334733735769987, 'clip_ratio/low_min': 0.012334733735769987, 'clip_ratio/high_mean': 0.03280176408588886, 'clip_ratio/high_max': 0.03280176408588886, 'clip_ratio/region_mean': 0.04513649782165885, 'reward_total_mean': 0.7669380903244019, 'reward_meter_mean': 0.9962884187698364, 'reward_meter_std': 0.0015494668623432517, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.911057710647583, 'reward_repeat_penalty_std': 0.051787521690130234, 'reward_total_composite_mean': 0.7669380903244019, 'reward_total_composite_std': 0.07985159009695053, 'epoch': 0.06}

 44%|████▍     | 1456/3300 [1:37:16<4:19:20,  8.44s/it]INFO 04-12 00:14:07 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1457/3300 [1:37:21<3:49:07,  7.46s/it]2026-04-12 00:14:12,385 | INFO | train_grpo_train | metrics_logged mode=train step=1457

                                                       
{'loss': -0.0297, 'grad_norm': 2.4325199127197266, 'learning_rate': 5.587878787878789e-06, 'num_tokens': 3280349.0, 'completions/mean_length': 96.75, 'completions/min_length': 89.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.75, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9892526268959045, 'rewards/meter/std': 0.015538093633949757, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.2070196568965912, 'rewards/total_composite/mean': 0.8396018743515015, 'rewards/total_composite/std': 0.20004859566688538, 'reward': 0.8396018743515015, 'reward_std': 0.20004859566688538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05370476841926575, 'sampling/sampling_logp_difference/max': 2.0252418518066406, 'sampling/importance_sampling_ratio/min': 0.13196192681789398, 'sampling/importance_sampling_ratio/mean': 1.0023468732833862, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3380627855658531, 'clip_ratio/low_mean': 0.00958982715383172, 'clip_ratio/low_min': 0.00958982715383172, 'clip_ratio/high_mean': 0.04439025931060314, 'clip_ratio/high_max': 0.04439025931060314, 'clip_ratio/region_mean': 0.05398008646443486, 'reward_total_mean': 0.8396018743515015, 'reward_meter_mean': 0.9892526268959045, 'reward_meter_std': 0.015538093633949757, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.2070196568965912, 'reward_total_composite_mean': 0.8396018743515015, 'reward_total_composite_std': 0.20004859566688538, 'epoch': 0.06}

 44%|████▍     | 1457/3300 [1:37:21<3:49:07,  7.46s/it]INFO 04-12 00:14:12 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1458/3300 [1:37:26<3:26:42,  6.73s/it]2026-04-12 00:14:17,423 | INFO | train_grpo_train | metrics_logged mode=train step=1458

                                                       
{'loss': -0.0117, 'grad_norm': 4.565279006958008, 'learning_rate': 5.584848484848485e-06, 'num_tokens': 3282489.0, 'completions/mean_length': 93.5, 'completions/min_length': 89.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.5, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9970370531082153, 'rewards/meter/std': 0.0014831003500148654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7750000357627869, 'rewards/repeat_penalty/std': 0.16690459847450256, 'rewards/total_composite/mean': 0.772584080696106, 'rewards/total_composite/std': 0.1656908094882965, 'reward': 0.772584080696106, 'reward_std': 0.1656908094882965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03978914022445679, 'sampling/sampling_logp_difference/max': 1.165329098701477, 'sampling/importance_sampling_ratio/min': 0.31182003021240234, 'sampling/importance_sampling_ratio/mean': 1.0047050714492798, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24885553307831287, 'clip_ratio/low_mean': 0.009437821572646499, 'clip_ratio/low_min': 0.009437821572646499, 'clip_ratio/high_mean': 0.022278774995356798, 'clip_ratio/high_max': 0.022278774995356798, 'clip_ratio/region_mean': 0.0317165965680033, 'reward_total_mean': 0.772584080696106, 'reward_meter_mean': 0.9970370531082153, 'reward_meter_std': 0.0014831003500148654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7750000357627869, 'reward_repeat_penalty_std': 0.16690459847450256, 'reward_total_composite_mean': 0.772584080696106, 'reward_total_composite_std': 0.1656908094882965, 'epoch': 0.06}

 44%|████▍     | 1458/3300 [1:37:26<3:26:42,  6.73s/it]INFO 04-12 00:14:17 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1459/3300 [1:37:34<3:34:56,  7.01s/it]2026-04-12 00:14:25,062 | INFO | train_grpo_train | metrics_logged mode=train step=1459

                                                       
{'loss': -0.0396, 'grad_norm': 3.691830635070801, 'learning_rate': 5.5818181818181824e-06, 'num_tokens': 3285975.0, 'completions/mean_length': 223.75, 'completions/min_length': 200.0, 'completions/max_length': 258.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 223.75, 'completions/min_terminated_length': 200.0, 'completions/max_terminated_length': 258.0, 'rewards/meter/mean': 0.2490365207195282, 'rewards/meter/std': 0.21091364324092865, 'rewards/count_adherence/mean': 0.5576923489570618, 'rewards/count_adherence/std': 0.05439283326268196, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8426282405853271, 'rewards/repeat_penalty/std': 0.1708158552646637, 'rewards/total_composite/mean': 0.12107732146978378, 'rewards/total_composite/std': 0.09244445711374283, 'reward': 0.12107732146978378, 'reward_std': 0.09244445711374283, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0653611496090889, 'sampling/sampling_logp_difference/max': 1.6574034690856934, 'sampling/importance_sampling_ratio/min': 0.19063332676887512, 'sampling/importance_sampling_ratio/mean': 1.0115958452224731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5131779182702303, 'clip_ratio/low_mean': 0.014779910678043962, 'clip_ratio/low_min': 0.014779910678043962, 'clip_ratio/high_mean': 0.03423361713066697, 'clip_ratio/high_max': 0.03423361713066697, 'clip_ratio/region_mean': 0.04901352780871093, 'reward_total_mean': 0.12107732146978378, 'reward_meter_mean': 0.2490365207195282, 'reward_meter_std': 0.21091364324092865, 'reward_count_adherence_mean': 0.5576923489570618, 'reward_count_adherence_std': 0.05439283326268196, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8426282405853271, 'reward_repeat_penalty_std': 0.1708158552646637, 'reward_total_composite_mean': 0.12107732146978378, 'reward_total_composite_std': 0.09244445711374283, 'epoch': 0.06}

 44%|████▍     | 1459/3300 [1:37:34<3:34:56,  7.01s/it]INFO 04-12 00:14:25 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1460/3300 [1:37:38<3:14:01,  6.33s/it]2026-04-12 00:14:29,807 | INFO | train_grpo_train | metrics_logged mode=train step=1460

                                                       
{'loss': 0.0048, 'grad_norm': 4.334770679473877, 'learning_rate': 5.578787878787879e-06, 'num_tokens': 3287646.0, 'completions/mean_length': 65.875, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9952852129936218, 'rewards/meter/std': 0.0014361762441694736, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952852129936218, 'rewards/total_composite/std': 0.0014361762441694736, 'reward': 0.9952852129936218, 'reward_std': 0.0014361703069880605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03258511424064636, 'sampling/sampling_logp_difference/max': 1.3955812454223633, 'sampling/importance_sampling_ratio/min': 0.24768902361392975, 'sampling/importance_sampling_ratio/mean': 1.0027287006378174, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1356327636167407, 'clip_ratio/low_mean': 0.003759611048735678, 'clip_ratio/low_min': 0.003759611048735678, 'clip_ratio/high_mean': 0.026631702319718897, 'clip_ratio/high_max': 0.026631702319718897, 'clip_ratio/region_mean': 0.030391313368454576, 'reward_total_mean': 0.9952852129936218, 'reward_meter_mean': 0.9952852129936218, 'reward_meter_std': 0.0014361762441694736, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952852129936218, 'reward_total_composite_std': 0.0014361762441694736, 'epoch': 0.06}

 44%|████▍     | 1460/3300 [1:37:38<3:14:01,  6.33s/it]INFO 04-12 00:14:30 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1461/3300 [1:37:43<2:59:39,  5.86s/it]2026-04-12 00:14:34,584 | INFO | train_grpo_train | metrics_logged mode=train step=1461

                                                       
{'loss': 0.0375, 'grad_norm': 7.510387420654297, 'learning_rate': 5.575757575757577e-06, 'num_tokens': 3289631.0, 'completions/mean_length': 68.125, 'completions/min_length': 66.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8637334108352661, 'rewards/meter/std': 0.34701576828956604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8637334108352661, 'rewards/total_composite/std': 0.34701576828956604, 'reward': 0.8637334108352661, 'reward_std': 0.34701573848724365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03330973535776138, 'sampling/sampling_logp_difference/max': 1.0436010360717773, 'sampling/importance_sampling_ratio/min': 0.3521841764450073, 'sampling/importance_sampling_ratio/mean': 1.0041791200637817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24057043716311455, 'clip_ratio/low_mean': 0.008333333767950535, 'clip_ratio/low_min': 0.008333333767950535, 'clip_ratio/high_mean': 0.016740256920456886, 'clip_ratio/high_max': 0.016740256920456886, 'clip_ratio/region_mean': 0.02507359068840742, 'reward_total_mean': 0.8637334108352661, 'reward_meter_mean': 0.8637334108352661, 'reward_meter_std': 0.34701576828956604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8637334108352661, 'reward_total_composite_std': 0.34701576828956604, 'epoch': 0.06}

 44%|████▍     | 1461/3300 [1:37:43<2:59:39,  5.86s/it]INFO 04-12 00:14:34 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1462/3300 [1:37:48<2:48:56,  5.52s/it]2026-04-12 00:14:39,292 | INFO | train_grpo_train | metrics_logged mode=train step=1462

                                                       
{'loss': 0.0109, 'grad_norm': 5.110687255859375, 'learning_rate': 5.572727272727273e-06, 'num_tokens': 3291531.0, 'completions/mean_length': 69.5, 'completions/min_length': 65.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9955259561538696, 'rewards/meter/std': 0.0017520119436085224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955259561538696, 'rewards/total_composite/std': 0.0017520119436085224, 'reward': 0.9955259561538696, 'reward_std': 0.001752000767737627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06455224007368088, 'sampling/sampling_logp_difference/max': 1.3436126708984375, 'sampling/importance_sampling_ratio/min': 0.26090142130851746, 'sampling/importance_sampling_ratio/mean': 1.018951654434204, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5490291193127632, 'clip_ratio/low_mean': 0.029018934816122055, 'clip_ratio/low_min': 0.029018934816122055, 'clip_ratio/high_mean': 0.02201478136703372, 'clip_ratio/high_max': 0.02201478136703372, 'clip_ratio/region_mean': 0.051033716183155775, 'reward_total_mean': 0.9955259561538696, 'reward_meter_mean': 0.9955259561538696, 'reward_meter_std': 0.0017520119436085224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955259561538696, 'reward_total_composite_std': 0.0017520119436085224, 'epoch': 0.06}

 44%|████▍     | 1462/3300 [1:37:48<2:48:56,  5.52s/it]INFO 04-12 00:14:39 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1463/3300 [1:37:52<2:40:40,  5.25s/it]2026-04-12 00:14:43,916 | INFO | train_grpo_train | metrics_logged mode=train step=1463

                                                       
{'loss': -0.0018, 'grad_norm': 5.067295551300049, 'learning_rate': 5.569696969696971e-06, 'num_tokens': 3292987.0, 'completions/mean_length': 63.0, 'completions/min_length': 59.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7616782188415527, 'rewards/meter/std': 0.24172598123550415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7616782188415527, 'rewards/total_composite/std': 0.24172598123550415, 'reward': 0.7616782188415527, 'reward_std': 0.24172596633434296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07128675282001495, 'sampling/sampling_logp_difference/max': 1.4407329559326172, 'sampling/importance_sampling_ratio/min': 0.2367541640996933, 'sampling/importance_sampling_ratio/mean': 1.0115406513214111, 'sampling/importance_sampling_ratio/max': 1.8223750591278076, 'entropy': 0.5481097176671028, 'clip_ratio/low_mean': 0.027423894964158535, 'clip_ratio/low_min': 0.027423894964158535, 'clip_ratio/high_mean': 0.02786910650320351, 'clip_ratio/high_max': 0.02786910650320351, 'clip_ratio/region_mean': 0.055293001467362046, 'reward_total_mean': 0.7616782188415527, 'reward_meter_mean': 0.7616782188415527, 'reward_meter_std': 0.24172598123550415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7616782188415527, 'reward_total_composite_std': 0.24172598123550415, 'epoch': 0.06}

 44%|████▍     | 1463/3300 [1:37:52<2:40:40,  5.25s/it]INFO 04-12 00:14:44 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1464/3300 [1:37:57<2:34:57,  5.06s/it]2026-04-12 00:14:48,549 | INFO | train_grpo_train | metrics_logged mode=train step=1464

                                                       
{'loss': -0.0081, 'grad_norm': 5.747054100036621, 'learning_rate': 5.566666666666667e-06, 'num_tokens': 3294759.0, 'completions/mean_length': 59.5, 'completions/min_length': 57.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.33728209137916565, 'rewards/meter/std': 0.2937431037425995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.33728209137916565, 'rewards/total_composite/std': 0.2937431037425995, 'reward': 0.33728209137916565, 'reward_std': 0.2937430739402771, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06646484136581421, 'sampling/sampling_logp_difference/max': 0.8989553451538086, 'sampling/importance_sampling_ratio/min': 0.40699464082717896, 'sampling/importance_sampling_ratio/mean': 1.0084482431411743, 'sampling/importance_sampling_ratio/max': 1.9918129444122314, 'entropy': 0.5435944311320782, 'clip_ratio/low_mean': 0.01695890142582357, 'clip_ratio/low_min': 0.01695890142582357, 'clip_ratio/high_mean': 0.02262367820367217, 'clip_ratio/high_max': 0.02262367820367217, 'clip_ratio/region_mean': 0.03958257962949574, 'reward_total_mean': 0.33728209137916565, 'reward_meter_mean': 0.33728209137916565, 'reward_meter_std': 0.2937431037425995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.33728209137916565, 'reward_total_composite_std': 0.2937431037425995, 'epoch': 0.06}

 44%|████▍     | 1464/3300 [1:37:57<2:34:57,  5.06s/it]INFO 04-12 00:14:48 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1465/3300 [1:38:02<2:31:53,  4.97s/it]2026-04-12 00:14:53,290 | INFO | train_grpo_train | metrics_logged mode=train step=1465

                                                       
{'loss': 0.0259, 'grad_norm': 5.774188041687012, 'learning_rate': 5.563636363636364e-06, 'num_tokens': 3296618.0, 'completions/mean_length': 67.375, 'completions/min_length': 65.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9969590306282043, 'rewards/meter/std': 0.0012544745113700628, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969590306282043, 'rewards/total_composite/std': 0.0012544745113700628, 'reward': 0.9969590306282043, 'reward_std': 0.0012544887140393257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06404515355825424, 'sampling/sampling_logp_difference/max': 2.009974479675293, 'sampling/importance_sampling_ratio/min': 0.13399209082126617, 'sampling/importance_sampling_ratio/mean': 1.0145848989486694, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4691624492406845, 'clip_ratio/low_mean': 0.010546227567829192, 'clip_ratio/low_min': 0.010546227567829192, 'clip_ratio/high_mean': 0.02795016940217465, 'clip_ratio/high_max': 0.02795016940217465, 'clip_ratio/region_mean': 0.03849639697000384, 'reward_total_mean': 0.9969590306282043, 'reward_meter_mean': 0.9969590306282043, 'reward_meter_std': 0.0012544745113700628, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969590306282043, 'reward_total_composite_std': 0.0012544745113700628, 'epoch': 0.06}

 44%|████▍     | 1465/3300 [1:38:02<2:31:53,  4.97s/it]INFO 04-12 00:14:53 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1466/3300 [1:38:06<2:28:46,  4.87s/it]2026-04-12 00:14:57,924 | INFO | train_grpo_train | metrics_logged mode=train step=1466

                                                       
{'loss': 0.0091, 'grad_norm': 6.224666118621826, 'learning_rate': 5.560606060606061e-06, 'num_tokens': 3298395.0, 'completions/mean_length': 59.125, 'completions/min_length': 56.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9083940982818604, 'rewards/meter/std': 0.1951943337917328, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9083940982818604, 'rewards/total_composite/std': 0.1951943337917328, 'reward': 0.9083940982818604, 'reward_std': 0.1951943188905716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030300971120595932, 'sampling/sampling_logp_difference/max': 0.9119729995727539, 'sampling/importance_sampling_ratio/min': 0.40173083543777466, 'sampling/importance_sampling_ratio/mean': 1.0091822147369385, 'sampling/importance_sampling_ratio/max': 1.7400184869766235, 'entropy': 0.2567995712161064, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.025272560073062778, 'clip_ratio/high_max': 0.025272560073062778, 'clip_ratio/region_mean': 0.02735589351505041, 'reward_total_mean': 0.9083940982818604, 'reward_meter_mean': 0.9083940982818604, 'reward_meter_std': 0.1951943337917328, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9083940982818604, 'reward_total_composite_std': 0.1951943337917328, 'epoch': 0.06}

 44%|████▍     | 1466/3300 [1:38:06<2:28:46,  4.87s/it]INFO 04-12 00:14:58 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1467/3300 [1:38:13<2:45:19,  5.41s/it]2026-04-12 00:15:04,606 | INFO | train_grpo_train | metrics_logged mode=train step=1467

                                                       
{'loss': -0.0184, 'grad_norm': 3.2179627418518066, 'learning_rate': 5.557575757575758e-06, 'num_tokens': 3301728.0, 'completions/mean_length': 220.625, 'completions/min_length': 201.0, 'completions/max_length': 236.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 220.625, 'completions/min_terminated_length': 201.0, 'completions/max_terminated_length': 236.0, 'rewards/meter/mean': 0.3030664324760437, 'rewards/meter/std': 0.29540300369262695, 'rewards/count_adherence/mean': 0.5288461446762085, 'rewards/count_adherence/std': 0.027196424081921577, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8133013248443604, 'rewards/repeat_penalty/std': 0.10961552709341049, 'rewards/total_composite/mean': 0.1352500319480896, 'rewards/total_composite/std': 0.13207849860191345, 'reward': 0.1352500319480896, 'reward_std': 0.13207849860191345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05384691804647446, 'sampling/sampling_logp_difference/max': 1.9810237884521484, 'sampling/importance_sampling_ratio/min': 0.13792794942855835, 'sampling/importance_sampling_ratio/mean': 1.006094217300415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3614166211336851, 'clip_ratio/low_mean': 0.024785209679976106, 'clip_ratio/low_min': 0.024785209679976106, 'clip_ratio/high_mean': 0.014436864759773016, 'clip_ratio/high_max': 0.014436864759773016, 'clip_ratio/region_mean': 0.03922207443974912, 'reward_total_mean': 0.1352500319480896, 'reward_meter_mean': 0.3030664324760437, 'reward_meter_std': 0.29540300369262695, 'reward_count_adherence_mean': 0.5288461446762085, 'reward_count_adherence_std': 0.027196424081921577, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8133013248443604, 'reward_repeat_penalty_std': 0.10961552709341049, 'reward_total_composite_mean': 0.1352500319480896, 'reward_total_composite_std': 0.13207849860191345, 'epoch': 0.06}

 44%|████▍     | 1467/3300 [1:38:13<2:45:19,  5.41s/it]INFO 04-12 00:15:04 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1468/3300 [1:38:18<2:38:50,  5.20s/it]2026-04-12 00:15:09,319 | INFO | train_grpo_train | metrics_logged mode=train step=1468

                                                       
{'loss': 0.0394, 'grad_norm': 3.850555419921875, 'learning_rate': 5.554545454545454e-06, 'num_tokens': 3303612.0, 'completions/mean_length': 70.5, 'completions/min_length': 62.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.993515133857727, 'rewards/meter/std': 0.002871233271434903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.993515133857727, 'rewards/total_composite/std': 0.002871233271434903, 'reward': 0.993515133857727, 'reward_std': 0.0028712216299027205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06810689717531204, 'sampling/sampling_logp_difference/max': 1.0864152908325195, 'sampling/importance_sampling_ratio/min': 0.3374238908290863, 'sampling/importance_sampling_ratio/mean': 1.0134084224700928, 'sampling/importance_sampling_ratio/max': 1.9790068864822388, 'entropy': 0.5614041350781918, 'clip_ratio/low_mean': 0.0190093262353912, 'clip_ratio/low_min': 0.0190093262353912, 'clip_ratio/high_mean': 0.023276393418200314, 'clip_ratio/high_max': 0.023276393418200314, 'clip_ratio/region_mean': 0.042285719653591514, 'reward_total_mean': 0.993515133857727, 'reward_meter_mean': 0.993515133857727, 'reward_meter_std': 0.002871233271434903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.993515133857727, 'reward_total_composite_std': 0.002871233271434903, 'epoch': 0.06}

 44%|████▍     | 1468/3300 [1:38:18<2:38:50,  5.20s/it]INFO 04-12 00:15:09 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1469/3300 [1:38:23<2:37:01,  5.15s/it]2026-04-12 00:15:14,333 | INFO | train_grpo_train | metrics_logged mode=train step=1469

                                                       
{'loss': -0.0341, 'grad_norm': 3.2438156604766846, 'learning_rate': 5.5515151515151524e-06, 'num_tokens': 3305697.0, 'completions/mean_length': 92.625, 'completions/min_length': 87.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.625, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.7984445691108704, 'rewards/meter/std': 0.2091599404811859, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.17728103697299957, 'rewards/total_composite/mean': 0.7041343450546265, 'rewards/total_composite/std': 0.271832674741745, 'reward': 0.7041343450546265, 'reward_std': 0.2718327045440674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041127223521471024, 'sampling/sampling_logp_difference/max': 1.2640190124511719, 'sampling/importance_sampling_ratio/min': 0.2825163006782532, 'sampling/importance_sampling_ratio/mean': 1.0063135623931885, 'sampling/importance_sampling_ratio/max': 1.4045096635818481, 'entropy': 0.2813885472714901, 'clip_ratio/low_mean': 0.0071839080192148685, 'clip_ratio/low_min': 0.0071839080192148685, 'clip_ratio/high_mean': 0.025160838733427227, 'clip_ratio/high_max': 0.025160838733427227, 'clip_ratio/region_mean': 0.032344746752642095, 'reward_total_mean': 0.7041343450546265, 'reward_meter_mean': 0.7984445691108704, 'reward_meter_std': 0.2091599404811859, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.17728103697299957, 'reward_total_composite_mean': 0.7041343450546265, 'reward_total_composite_std': 0.271832674741745, 'epoch': 0.06}

 45%|████▍     | 1469/3300 [1:38:23<2:37:01,  5.15s/it]INFO 04-12 00:15:14 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1470/3300 [1:38:28<2:33:27,  5.03s/it]2026-04-12 00:15:19,098 | INFO | train_grpo_train | metrics_logged mode=train step=1470

                                                       
{'loss': -0.0266, 'grad_norm': 6.198017120361328, 'learning_rate': 5.548484848484849e-06, 'num_tokens': 3307476.0, 'completions/mean_length': 70.375, 'completions/min_length': 64.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8553205132484436, 'rewards/meter/std': 0.22217348217964172, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8553205132484436, 'rewards/total_composite/std': 0.22217348217964172, 'reward': 0.8553205132484436, 'reward_std': 0.22217348217964172, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07493096590042114, 'sampling/sampling_logp_difference/max': 1.1873397827148438, 'sampling/importance_sampling_ratio/min': 0.3050316274166107, 'sampling/importance_sampling_ratio/mean': 1.0185297727584839, 'sampling/importance_sampling_ratio/max': 1.9982537031173706, 'entropy': 0.7106457650661469, 'clip_ratio/low_mean': 0.013322061393409967, 'clip_ratio/low_min': 0.013322061393409967, 'clip_ratio/high_mean': 0.04982595471665263, 'clip_ratio/high_max': 0.04982595471665263, 'clip_ratio/region_mean': 0.0631480161100626, 'reward_total_mean': 0.8553205132484436, 'reward_meter_mean': 0.8553205132484436, 'reward_meter_std': 0.22217348217964172, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8553205132484436, 'reward_total_composite_std': 0.22217348217964172, 'epoch': 0.06}

 45%|████▍     | 1470/3300 [1:38:28<2:33:27,  5.03s/it]INFO 04-12 00:15:19 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1471/3300 [1:38:32<2:29:00,  4.89s/it]2026-04-12 00:15:23,652 | INFO | train_grpo_train | metrics_logged mode=train step=1471

                                                       
{'loss': -0.0016, 'grad_norm': 4.865560054779053, 'learning_rate': 5.545454545454546e-06, 'num_tokens': 3309226.0, 'completions/mean_length': 56.75, 'completions/min_length': 53.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9437645673751831, 'rewards/meter/std': 0.12552663683891296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9023252725601196, 'rewards/total_composite/std': 0.15711426734924316, 'reward': 0.9023252725601196, 'reward_std': 0.15711423754692078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03652970865368843, 'sampling/sampling_logp_difference/max': 1.6262216567993164, 'sampling/importance_sampling_ratio/min': 0.196671262383461, 'sampling/importance_sampling_ratio/mean': 1.0094138383865356, 'sampling/importance_sampling_ratio/max': 1.7097090482711792, 'entropy': 0.237571744248271, 'clip_ratio/low_mean': 0.011254789307713509, 'clip_ratio/low_min': 0.011254789307713509, 'clip_ratio/high_mean': 0.021573547972366214, 'clip_ratio/high_max': 0.021573547972366214, 'clip_ratio/region_mean': 0.03282833728007972, 'reward_total_mean': 0.9023252725601196, 'reward_meter_mean': 0.9437645673751831, 'reward_meter_std': 0.12552663683891296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9023252725601196, 'reward_total_composite_std': 0.15711426734924316, 'epoch': 0.06}

 45%|████▍     | 1471/3300 [1:38:32<2:29:00,  4.89s/it]INFO 04-12 00:15:23 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1472/3300 [1:38:38<2:37:17,  5.16s/it]2026-04-12 00:15:29,458 | INFO | train_grpo_train | metrics_logged mode=train step=1472

                                                       
{'loss': -0.0156, 'grad_norm': 2.8264143466949463, 'learning_rate': 5.5424242424242425e-06, 'num_tokens': 3311622.0, 'completions/mean_length': 123.5, 'completions/min_length': 114.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.5, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.7211212515830994, 'rewards/meter/std': 0.2165415734052658, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428060531616, 'rewards/repeat_penalty/std': 0.1937432438135147, 'rewards/total_composite/mean': 0.5241959095001221, 'rewards/total_composite/std': 0.20148435235023499, 'reward': 0.5241959095001221, 'reward_std': 0.20148435235023499, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03288319706916809, 'sampling/sampling_logp_difference/max': 1.2717273235321045, 'sampling/importance_sampling_ratio/min': 0.28034695982933044, 'sampling/importance_sampling_ratio/mean': 1.0071460008621216, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21886500529944897, 'clip_ratio/low_mean': 0.005906250094994903, 'clip_ratio/low_min': 0.005906250094994903, 'clip_ratio/high_mean': 0.009076492046006024, 'clip_ratio/high_max': 0.009076492046006024, 'clip_ratio/region_mean': 0.014982742141000926, 'reward_total_mean': 0.5241959095001221, 'reward_meter_mean': 0.7211212515830994, 'reward_meter_std': 0.2165415734052658, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428060531616, 'reward_repeat_penalty_std': 0.1937432438135147, 'reward_total_composite_mean': 0.5241959095001221, 'reward_total_composite_std': 0.20148435235023499, 'epoch': 0.06}

 45%|████▍     | 1472/3300 [1:38:38<2:37:17,  5.16s/it]INFO 04-12 00:15:29 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1473/3300 [1:38:44<2:46:19,  5.46s/it]2026-04-12 00:15:35,618 | INFO | train_grpo_train | metrics_logged mode=train step=1473

                                                       
{'loss': 0.0149, 'grad_norm': 2.3183512687683105, 'learning_rate': 5.53939393939394e-06, 'num_tokens': 3314535.0, 'completions/mean_length': 184.125, 'completions/min_length': 164.0, 'completions/max_length': 202.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 184.125, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 202.0, 'rewards/meter/mean': 0.8684302568435669, 'rewards/meter/std': 0.16792094707489014, 'rewards/count_adherence/mean': 0.7857142686843872, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7905303239822388, 'rewards/repeat_penalty/std': 0.09503685683012009, 'rewards/total_composite/mean': 0.5370750427246094, 'rewards/total_composite/std': 0.12571245431900024, 'reward': 0.5370750427246094, 'reward_std': 0.12571245431900024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022778350859880447, 'sampling/sampling_logp_difference/max': 1.592844009399414, 'sampling/importance_sampling_ratio/min': 0.20334647595882416, 'sampling/importance_sampling_ratio/mean': 1.0009946823120117, 'sampling/importance_sampling_ratio/max': 1.7813695669174194, 'entropy': 0.12613181956112385, 'clip_ratio/low_mean': 0.005232378258369863, 'clip_ratio/low_min': 0.005232378258369863, 'clip_ratio/high_mean': 0.006139094999525696, 'clip_ratio/high_max': 0.006139094999525696, 'clip_ratio/region_mean': 0.011371473257895559, 'reward_total_mean': 0.5370750427246094, 'reward_meter_mean': 0.8684302568435669, 'reward_meter_std': 0.16792094707489014, 'reward_count_adherence_mean': 0.7857142686843872, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7905303239822388, 'reward_repeat_penalty_std': 0.09503685683012009, 'reward_total_composite_mean': 0.5370750427246094, 'reward_total_composite_std': 0.12571245431900024, 'epoch': 0.06}

 45%|████▍     | 1473/3300 [1:38:44<2:46:19,  5.46s/it]INFO 04-12 00:15:35 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1474/3300 [1:38:50<2:49:17,  5.56s/it]2026-04-12 00:15:41,415 | INFO | train_grpo_train | metrics_logged mode=train step=1474

                                                       
{'loss': 0.0044, 'grad_norm': 2.7291924953460693, 'learning_rate': 5.536363636363636e-06, 'num_tokens': 3317453.0, 'completions/mean_length': 166.75, 'completions/min_length': 162.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.75, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.959197998046875, 'rewards/meter/std': 0.10039316117763519, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8055555820465088, 'rewards/repeat_penalty/std': 0.11501092463731766, 'rewards/total_composite/mean': 0.6380866169929504, 'rewards/total_composite/std': 0.07123729586601257, 'reward': 0.6380866169929504, 'reward_std': 0.07123728096485138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027119241654872894, 'sampling/sampling_logp_difference/max': 1.337611198425293, 'sampling/importance_sampling_ratio/min': 0.26247191429138184, 'sampling/importance_sampling_ratio/mean': 1.0012325048446655, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15955681633204222, 'clip_ratio/low_mean': 0.008916479535400867, 'clip_ratio/low_min': 0.008916479535400867, 'clip_ratio/high_mean': 0.018076194741297513, 'clip_ratio/high_max': 0.018076194741297513, 'clip_ratio/region_mean': 0.02699267427669838, 'reward_total_mean': 0.6380866169929504, 'reward_meter_mean': 0.959197998046875, 'reward_meter_std': 0.10039316117763519, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8055555820465088, 'reward_repeat_penalty_std': 0.11501092463731766, 'reward_total_composite_mean': 0.6380866169929504, 'reward_total_composite_std': 0.07123729586601257, 'epoch': 0.06}

 45%|████▍     | 1474/3300 [1:38:50<2:49:17,  5.56s/it]INFO 04-12 00:15:41 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1475/3300 [1:38:55<2:44:52,  5.42s/it]2026-04-12 00:15:46,502 | INFO | train_grpo_train | metrics_logged mode=train step=1475

                                                       
{'loss': 0.0032, 'grad_norm': 7.203929424285889, 'learning_rate': 5.533333333333334e-06, 'num_tokens': 3319564.0, 'completions/mean_length': 102.875, 'completions/min_length': 96.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.875, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9892958402633667, 'rewards/meter/std': 0.02243867516517639, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9892958402633667, 'rewards/total_composite/std': 0.02243867516517639, 'reward': 0.9892958402633667, 'reward_std': 0.022438665851950645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06217264011502266, 'sampling/sampling_logp_difference/max': 2.3590099811553955, 'sampling/importance_sampling_ratio/min': 0.09451375156641006, 'sampling/importance_sampling_ratio/mean': 1.0040569305419922, 'sampling/importance_sampling_ratio/max': 1.710269570350647, 'entropy': 0.4008924439549446, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.044174039736390114, 'clip_ratio/high_max': 0.044174039736390114, 'clip_ratio/region_mean': 0.0477454683277756, 'reward_total_mean': 0.9892958402633667, 'reward_meter_mean': 0.9892958402633667, 'reward_meter_std': 0.02243867516517639, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9892958402633667, 'reward_total_composite_std': 0.02243867516517639, 'epoch': 0.06}

 45%|████▍     | 1475/3300 [1:38:55<2:44:52,  5.42s/it]INFO 04-12 00:15:46 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1476/3300 [1:39:00<2:43:36,  5.38s/it]2026-04-12 00:15:51,796 | INFO | train_grpo_train | metrics_logged mode=train step=1476

                                                       
{'loss': -0.0149, 'grad_norm': 4.188331604003906, 'learning_rate': 5.530303030303031e-06, 'num_tokens': 3321999.0, 'completions/mean_length': 120.375, 'completions/min_length': 113.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.375, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.7211596369743347, 'rewards/meter/std': 0.29947173595428467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8214285373687744, 'rewards/repeat_penalty/std': 0.1266293078660965, 'rewards/total_composite/mean': 0.5741269588470459, 'rewards/total_composite/std': 0.22667214274406433, 'reward': 0.5741269588470459, 'reward_std': 0.22667212784290314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05203632265329361, 'sampling/sampling_logp_difference/max': 1.7595195770263672, 'sampling/importance_sampling_ratio/min': 0.2374485731124878, 'sampling/importance_sampling_ratio/mean': 1.0193830728530884, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45302187465131283, 'clip_ratio/low_mean': 0.013878762954846025, 'clip_ratio/low_min': 0.013878762954846025, 'clip_ratio/high_mean': 0.015247002593241632, 'clip_ratio/high_max': 0.015247002593241632, 'clip_ratio/region_mean': 0.029125765548087656, 'reward_total_mean': 0.5741269588470459, 'reward_meter_mean': 0.7211596369743347, 'reward_meter_std': 0.29947173595428467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8214285373687744, 'reward_repeat_penalty_std': 0.1266293078660965, 'reward_total_composite_mean': 0.5741269588470459, 'reward_total_composite_std': 0.22667214274406433, 'epoch': 0.06}

 45%|████▍     | 1476/3300 [1:39:00<2:43:36,  5.38s/it]INFO 04-12 00:15:52 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1477/3300 [1:39:05<2:37:16,  5.18s/it]2026-04-12 00:15:56,493 | INFO | train_grpo_train | metrics_logged mode=train step=1477

                                                       
{'loss': 0.0153, 'grad_norm': 6.804332733154297, 'learning_rate': 5.527272727272728e-06, 'num_tokens': 3323777.0, 'completions/mean_length': 68.25, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9923118352890015, 'rewards/meter/std': 0.008320425637066364, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9923118352890015, 'rewards/total_composite/std': 0.008320425637066364, 'reward': 0.9923118352890015, 'reward_std': 0.008320420049130917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0801224485039711, 'sampling/sampling_logp_difference/max': 1.6460561752319336, 'sampling/importance_sampling_ratio/min': 0.19280880689620972, 'sampling/importance_sampling_ratio/mean': 1.0044664144515991, 'sampling/importance_sampling_ratio/max': 1.8288415670394897, 'entropy': 0.6124710254371166, 'clip_ratio/low_mean': 0.01433747448027134, 'clip_ratio/low_min': 0.01433747448027134, 'clip_ratio/high_mean': 0.055169664323329926, 'clip_ratio/high_max': 0.055169664323329926, 'clip_ratio/region_mean': 0.06950713880360126, 'reward_total_mean': 0.9923118352890015, 'reward_meter_mean': 0.9923118352890015, 'reward_meter_std': 0.008320425637066364, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9923118352890015, 'reward_total_composite_std': 0.008320425637066364, 'epoch': 0.06}

 45%|████▍     | 1477/3300 [1:39:05<2:37:16,  5.18s/it]INFO 04-12 00:15:56 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1478/3300 [1:39:09<2:28:12,  4.88s/it]2026-04-12 00:16:00,707 | INFO | train_grpo_train | metrics_logged mode=train step=1478

                                                       
{'loss': 0.0228, 'grad_norm': 10.884235382080078, 'learning_rate': 5.524242424242424e-06, 'num_tokens': 3325231.0, 'completions/mean_length': 29.75, 'completions/min_length': 29.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9929543733596802, 'rewards/meter/std': 0.0024007554166018963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9929543733596802, 'rewards/total_composite/std': 0.0024007554166018963, 'reward': 0.9929543733596802, 'reward_std': 0.002400748198851943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04087400063872337, 'sampling/sampling_logp_difference/max': 1.8403053283691406, 'sampling/importance_sampling_ratio/min': 0.1587689369916916, 'sampling/importance_sampling_ratio/mean': 1.0106008052825928, 'sampling/importance_sampling_ratio/max': 1.9975814819335938, 'entropy': 0.25689505971968174, 'clip_ratio/low_mean': 0.012122844811528921, 'clip_ratio/low_min': 0.012122844811528921, 'clip_ratio/high_mean': 0.012526939623057842, 'clip_ratio/high_max': 0.012526939623057842, 'clip_ratio/region_mean': 0.024649784434586763, 'reward_total_mean': 0.9929543733596802, 'reward_meter_mean': 0.9929543733596802, 'reward_meter_std': 0.0024007554166018963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9929543733596802, 'reward_total_composite_std': 0.0024007554166018963, 'epoch': 0.06}

 45%|████▍     | 1478/3300 [1:39:09<2:28:12,  4.88s/it]INFO 04-12 00:16:00 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1479/3300 [1:39:14<2:26:20,  4.82s/it]2026-04-12 00:16:05,368 | INFO | train_grpo_train | metrics_logged mode=train step=1479

                                                       
{'loss': 0.021, 'grad_norm': 6.581116676330566, 'learning_rate': 5.521212121212122e-06, 'num_tokens': 3327045.0, 'completions/mean_length': 67.75, 'completions/min_length': 63.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8861526846885681, 'rewards/meter/std': 0.3106532394886017, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8861526846885681, 'rewards/total_composite/std': 0.3106532394886017, 'reward': 0.8861526846885681, 'reward_std': 0.3106532096862793, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06694039702415466, 'sampling/sampling_logp_difference/max': 0.8918218612670898, 'sampling/importance_sampling_ratio/min': 0.4099082946777344, 'sampling/importance_sampling_ratio/mean': 1.0218294858932495, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6052872687578201, 'clip_ratio/low_mean': 0.0036231884732842445, 'clip_ratio/low_min': 0.0036231884732842445, 'clip_ratio/high_mean': 0.052147963899187744, 'clip_ratio/high_max': 0.052147963899187744, 'clip_ratio/region_mean': 0.05577115237247199, 'reward_total_mean': 0.8861526846885681, 'reward_meter_mean': 0.8861526846885681, 'reward_meter_std': 0.3106532394886017, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8861526846885681, 'reward_total_composite_std': 0.3106532394886017, 'epoch': 0.06}

 45%|████▍     | 1479/3300 [1:39:14<2:26:20,  4.82s/it]INFO 04-12 00:16:05 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1480/3300 [1:39:18<2:21:37,  4.67s/it]2026-04-12 00:16:09,681 | INFO | train_grpo_train | metrics_logged mode=train step=1480

                                                       
{'loss': 0.0211, 'grad_norm': 12.08879280090332, 'learning_rate': 5.518181818181818e-06, 'num_tokens': 3328501.0, 'completions/mean_length': 37.0, 'completions/min_length': 35.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.997024655342102, 'rewards/meter/std': 0.001989408629015088, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997024655342102, 'rewards/total_composite/std': 0.001989408629015088, 'reward': 0.997024655342102, 'reward_std': 0.001989412121474743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06934916973114014, 'sampling/sampling_logp_difference/max': 1.2974138259887695, 'sampling/importance_sampling_ratio/min': 0.27323752641677856, 'sampling/importance_sampling_ratio/mean': 1.000410795211792, 'sampling/importance_sampling_ratio/max': 1.6434557437896729, 'entropy': 0.5856506302952766, 'clip_ratio/low_mean': 0.027099420549347997, 'clip_ratio/low_min': 0.027099420549347997, 'clip_ratio/high_mean': 0.023252843879163265, 'clip_ratio/high_max': 0.023252843879163265, 'clip_ratio/region_mean': 0.05035226442851126, 'reward_total_mean': 0.997024655342102, 'reward_meter_mean': 0.997024655342102, 'reward_meter_std': 0.001989408629015088, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997024655342102, 'reward_total_composite_std': 0.001989408629015088, 'epoch': 0.06}

 45%|████▍     | 1480/3300 [1:39:18<2:21:37,  4.67s/it]INFO 04-12 00:16:09 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1481/3300 [1:39:26<2:49:20,  5.59s/it]2026-04-12 00:16:17,404 | INFO | train_grpo_train | metrics_logged mode=train step=1481

                                                       
{'loss': -0.0133, 'grad_norm': 3.2796733379364014, 'learning_rate': 5.515151515151515e-06, 'num_tokens': 3332149.0, 'completions/mean_length': 243.0, 'completions/min_length': 225.0, 'completions/max_length': 267.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 243.0, 'completions/min_terminated_length': 225.0, 'completions/max_terminated_length': 267.0, 'rewards/meter/mean': 0.7210352420806885, 'rewards/meter/std': 0.4204363524913788, 'rewards/count_adherence/mean': 0.5089285373687744, 'rewards/count_adherence/std': 0.025253823027014732, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.848809540271759, 'rewards/repeat_penalty/std': 0.11776864528656006, 'rewards/total_composite/mean': 0.2932712435722351, 'rewards/total_composite/std': 0.1669246405363083, 'reward': 0.2932712435722351, 'reward_std': 0.1669246405363083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06317514181137085, 'sampling/sampling_logp_difference/max': 1.6376686096191406, 'sampling/importance_sampling_ratio/min': 0.1944328099489212, 'sampling/importance_sampling_ratio/mean': 1.0123441219329834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5313003696501255, 'clip_ratio/low_mean': 0.015805913135409355, 'clip_ratio/low_min': 0.015805913135409355, 'clip_ratio/high_mean': 0.023832474602386355, 'clip_ratio/high_max': 0.023832474602386355, 'clip_ratio/region_mean': 0.03963838773779571, 'reward_total_mean': 0.2932712435722351, 'reward_meter_mean': 0.7210352420806885, 'reward_meter_std': 0.4204363524913788, 'reward_count_adherence_mean': 0.5089285373687744, 'reward_count_adherence_std': 0.025253823027014732, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.848809540271759, 'reward_repeat_penalty_std': 0.11776864528656006, 'reward_total_composite_mean': 0.2932712435722351, 'reward_total_composite_std': 0.1669246405363083, 'epoch': 0.06}

 45%|████▍     | 1481/3300 [1:39:26<2:49:20,  5.59s/it]INFO 04-12 00:16:17 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1482/3300 [1:39:33<3:02:44,  6.03s/it]2026-04-12 00:16:24,506 | INFO | train_grpo_train | metrics_logged mode=train step=1482

                                                       
{'loss': -0.0162, 'grad_norm': 4.245050430297852, 'learning_rate': 5.512121212121213e-06, 'num_tokens': 3334931.0, 'completions/mean_length': 157.75, 'completions/min_length': 135.0, 'completions/max_length': 172.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.75, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 172.0, 'rewards/meter/mean': 0.9921485781669617, 'rewards/meter/std': 0.002238625893369317, 'rewards/count_adherence/mean': 0.8035714030265808, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8219696879386902, 'rewards/repeat_penalty/std': 0.08952570706605911, 'rewards/total_composite/mean': 0.6518079042434692, 'rewards/total_composite/std': 0.05643462389707565, 'reward': 0.6518079042434692, 'reward_std': 0.05643462389707565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04733636975288391, 'sampling/sampling_logp_difference/max': 1.2221221923828125, 'sampling/importance_sampling_ratio/min': 0.29460427165031433, 'sampling/importance_sampling_ratio/mean': 1.0066455602645874, 'sampling/importance_sampling_ratio/max': 1.8663012981414795, 'entropy': 0.415604081004858, 'clip_ratio/low_mean': 0.016570850741118193, 'clip_ratio/low_min': 0.016570850741118193, 'clip_ratio/high_mean': 0.023007763549685478, 'clip_ratio/high_max': 0.023007763549685478, 'clip_ratio/region_mean': 0.03957861429080367, 'reward_total_mean': 0.6518079042434692, 'reward_meter_mean': 0.9921485781669617, 'reward_meter_std': 0.002238625893369317, 'reward_count_adherence_mean': 0.8035714030265808, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8219696879386902, 'reward_repeat_penalty_std': 0.08952570706605911, 'reward_total_composite_mean': 0.6518079042434692, 'reward_total_composite_std': 0.05643462389707565, 'epoch': 0.06}

 45%|████▍     | 1482/3300 [1:39:33<3:02:44,  6.03s/it]INFO 04-12 00:16:24 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1483/3300 [1:39:39<3:01:52,  6.01s/it]2026-04-12 00:16:30,425 | INFO | train_grpo_train | metrics_logged mode=train step=1483

                                                       
{'loss': 0.0024, 'grad_norm': 4.041107177734375, 'learning_rate': 5.50909090909091e-06, 'num_tokens': 3337913.0, 'completions/mean_length': 159.75, 'completions/min_length': 145.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.75, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.8998898267745972, 'rewards/meter/std': 0.24152173101902008, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8854166865348816, 'rewards/repeat_penalty/std': 0.10751881450414658, 'rewards/total_composite/mean': 0.6401326656341553, 'rewards/total_composite/std': 0.18946325778961182, 'reward': 0.6401326656341553, 'reward_std': 0.18946325778961182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06710459291934967, 'sampling/sampling_logp_difference/max': 2.618948221206665, 'sampling/importance_sampling_ratio/min': 0.07287947088479996, 'sampling/importance_sampling_ratio/mean': 1.0176057815551758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6857388708740473, 'clip_ratio/low_mean': 0.014218881260603666, 'clip_ratio/low_min': 0.014218881260603666, 'clip_ratio/high_mean': 0.03336897538974881, 'clip_ratio/high_max': 0.03336897538974881, 'clip_ratio/region_mean': 0.04758785665035248, 'reward_total_mean': 0.6401326656341553, 'reward_meter_mean': 0.8998898267745972, 'reward_meter_std': 0.24152173101902008, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8854166865348816, 'reward_repeat_penalty_std': 0.10751881450414658, 'reward_total_composite_mean': 0.6401326656341553, 'reward_total_composite_std': 0.18946325778961182, 'epoch': 0.06}

 45%|████▍     | 1483/3300 [1:39:39<3:01:52,  6.01s/it]INFO 04-12 00:16:30 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1484/3300 [1:39:43<2:48:39,  5.57s/it]2026-04-12 00:16:34,990 | INFO | train_grpo_train | metrics_logged mode=train step=1484

                                                       
{'loss': 0.0206, 'grad_norm': 6.229678153991699, 'learning_rate': 5.506060606060607e-06, 'num_tokens': 3339388.0, 'completions/mean_length': 33.375, 'completions/min_length': 32.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9635177850723267, 'rewards/meter/std': 0.020120054483413696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9635177850723267, 'rewards/total_composite/std': 0.020120054483413696, 'reward': 0.9635177850723267, 'reward_std': 0.02012004144489765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058272585272789, 'sampling/sampling_logp_difference/max': 1.0730462074279785, 'sampling/importance_sampling_ratio/min': 0.5093486905097961, 'sampling/importance_sampling_ratio/mean': 1.0330264568328857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5462967716157436, 'clip_ratio/low_mean': 0.007464349502697587, 'clip_ratio/low_min': 0.007464349502697587, 'clip_ratio/high_mean': 0.018196672899648547, 'clip_ratio/high_max': 0.018196672899648547, 'clip_ratio/region_mean': 0.025661022402346134, 'reward_total_mean': 0.9635177850723267, 'reward_meter_mean': 0.9635177850723267, 'reward_meter_std': 0.020120054483413696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9635177850723267, 'reward_total_composite_std': 0.020120054483413696, 'epoch': 0.06}

 45%|████▍     | 1484/3300 [1:39:43<2:48:39,  5.57s/it]INFO 04-12 00:16:35 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1485/3300 [1:39:49<2:46:36,  5.51s/it]2026-04-12 00:16:40,342 | INFO | train_grpo_train | metrics_logged mode=train step=1485

                                                       
{'loss': 0.0024, 'grad_norm': 4.295280456542969, 'learning_rate': 5.5030303030303034e-06, 'num_tokens': 3341484.0, 'completions/mean_length': 90.0, 'completions/min_length': 87.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.0, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.6956532597541809, 'rewards/meter/std': 0.21607494354248047, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.6436759233474731, 'rewards/total_composite/std': 0.27830082178115845, 'reward': 0.6436759233474731, 'reward_std': 0.27830079197883606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07055231183767319, 'sampling/sampling_logp_difference/max': 1.5324945449829102, 'sampling/importance_sampling_ratio/min': 0.21599619090557098, 'sampling/importance_sampling_ratio/mean': 1.0068004131317139, 'sampling/importance_sampling_ratio/max': 1.6940116882324219, 'entropy': 0.5103895664215088, 'clip_ratio/low_mean': 0.006699939724057913, 'clip_ratio/low_min': 0.006699939724057913, 'clip_ratio/high_mean': 0.025170997832901776, 'clip_ratio/high_max': 0.025170997832901776, 'clip_ratio/region_mean': 0.03187093755695969, 'reward_total_mean': 0.6436759233474731, 'reward_meter_mean': 0.6956532597541809, 'reward_meter_std': 0.21607494354248047, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.6436759233474731, 'reward_total_composite_std': 0.27830082178115845, 'epoch': 0.06}

 45%|████▌     | 1485/3300 [1:39:49<2:46:36,  5.51s/it]INFO 04-12 00:16:40 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1486/3300 [1:39:54<2:43:45,  5.42s/it]2026-04-12 00:16:45,547 | INFO | train_grpo_train | metrics_logged mode=train step=1486

                                                       
{'loss': 0.0174, 'grad_norm': 5.7835164070129395, 'learning_rate': 5.500000000000001e-06, 'num_tokens': 3343427.0, 'completions/mean_length': 85.875, 'completions/min_length': 81.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.875, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.5018972158432007, 'rewards/meter/std': 0.3999512195587158, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.41748714447021484, 'rewards/total_composite/std': 0.3111114799976349, 'reward': 0.41748714447021484, 'reward_std': 0.3111114799976349, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0595453642308712, 'sampling/sampling_logp_difference/max': 1.609797477722168, 'sampling/importance_sampling_ratio/min': 0.19992808997631073, 'sampling/importance_sampling_ratio/mean': 1.0213876962661743, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6251698546111584, 'clip_ratio/low_mean': 0.021903468994423747, 'clip_ratio/low_min': 0.021903468994423747, 'clip_ratio/high_mean': 0.010355054982937872, 'clip_ratio/high_max': 0.010355054982937872, 'clip_ratio/region_mean': 0.03225852397736162, 'reward_total_mean': 0.41748714447021484, 'reward_meter_mean': 0.5018972158432007, 'reward_meter_std': 0.3999512195587158, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.41748714447021484, 'reward_total_composite_std': 0.3111114799976349, 'epoch': 0.06}

 45%|████▌     | 1486/3300 [1:39:54<2:43:45,  5.42s/it]INFO 04-12 00:16:45 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1487/3300 [1:39:59<2:40:56,  5.33s/it]2026-04-12 00:16:50,666 | INFO | train_grpo_train | metrics_logged mode=train step=1487

                                                       
{'loss': 0.0425, 'grad_norm': 18.545215606689453, 'learning_rate': 5.496969696969697e-06, 'num_tokens': 3345165.0, 'completions/mean_length': 55.25, 'completions/min_length': 53.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.25, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9714435338973999, 'rewards/meter/std': 0.04950811713933945, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9714435338973999, 'rewards/total_composite/std': 0.04950811713933945, 'reward': 0.9714435338973999, 'reward_std': 0.04950810596346855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06842397898435593, 'sampling/sampling_logp_difference/max': 1.205331802368164, 'sampling/importance_sampling_ratio/min': 0.2995925843715668, 'sampling/importance_sampling_ratio/mean': 1.014945149421692, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33666370064020157, 'clip_ratio/low_mean': 0.0132503192871809, 'clip_ratio/low_min': 0.0132503192871809, 'clip_ratio/high_mean': 0.02728484314866364, 'clip_ratio/high_max': 0.02728484314866364, 'clip_ratio/region_mean': 0.04053516243584454, 'reward_total_mean': 0.9714435338973999, 'reward_meter_mean': 0.9714435338973999, 'reward_meter_std': 0.04950811713933945, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9714435338973999, 'reward_total_composite_std': 0.04950811713933945, 'epoch': 0.06}

 45%|████▌     | 1487/3300 [1:39:59<2:40:56,  5.33s/it]INFO 04-12 00:16:50 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1488/3300 [1:40:04<2:39:49,  5.29s/it]2026-04-12 00:16:55,876 | INFO | train_grpo_train | metrics_logged mode=train step=1488

                                                       
{'loss': -0.0096, 'grad_norm': 4.115540504455566, 'learning_rate': 5.493939393939395e-06, 'num_tokens': 3346850.0, 'completions/mean_length': 55.625, 'completions/min_length': 54.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.992435097694397, 'rewards/meter/std': 0.006868826691061258, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.992435097694397, 'rewards/total_composite/std': 0.006868826691061258, 'reward': 0.992435097694397, 'reward_std': 0.006868841592222452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0397869236767292, 'sampling/sampling_logp_difference/max': 1.0334584712982178, 'sampling/importance_sampling_ratio/min': 0.3714584410190582, 'sampling/importance_sampling_ratio/mean': 1.0044257640838623, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2505888659507036, 'clip_ratio/low_mean': 0.0069444444961845875, 'clip_ratio/low_min': 0.0069444444961845875, 'clip_ratio/high_mean': 0.020023792050778866, 'clip_ratio/high_max': 0.020023792050778866, 'clip_ratio/region_mean': 0.026968236546963453, 'reward_total_mean': 0.992435097694397, 'reward_meter_mean': 0.992435097694397, 'reward_meter_std': 0.006868826691061258, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.992435097694397, 'reward_total_composite_std': 0.006868826691061258, 'epoch': 0.06}

 45%|████▌     | 1488/3300 [1:40:04<2:39:49,  5.29s/it]INFO 04-12 00:16:56 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1489/3300 [1:40:10<2:40:23,  5.31s/it]2026-04-12 00:17:01,241 | INFO | train_grpo_train | metrics_logged mode=train step=1489

                                                       
{'loss': 0.0041, 'grad_norm': 3.699056386947632, 'learning_rate': 5.490909090909091e-06, 'num_tokens': 3348897.0, 'completions/mean_length': 90.875, 'completions/min_length': 87.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.875, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7553209066390991, 'rewards/meter/std': 0.27557846903800964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.7304862141609192, 'rewards/total_composite/std': 0.2595451772212982, 'reward': 0.7304862141609192, 'reward_std': 0.2595451772212982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07237864285707474, 'sampling/sampling_logp_difference/max': 1.377182960510254, 'sampling/importance_sampling_ratio/min': 0.252288281917572, 'sampling/importance_sampling_ratio/mean': 1.0141518115997314, 'sampling/importance_sampling_ratio/max': 1.844269871711731, 'entropy': 0.6783212311565876, 'clip_ratio/low_mean': 0.013681166106835008, 'clip_ratio/low_min': 0.013681166106835008, 'clip_ratio/high_mean': 0.04566098749637604, 'clip_ratio/high_max': 0.04566098749637604, 'clip_ratio/region_mean': 0.059342153603211045, 'reward_total_mean': 0.7304862141609192, 'reward_meter_mean': 0.7553209066390991, 'reward_meter_std': 0.27557846903800964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.7304862141609192, 'reward_total_composite_std': 0.2595451772212982, 'epoch': 0.06}

 45%|████▌     | 1489/3300 [1:40:10<2:40:23,  5.31s/it]INFO 04-12 00:17:01 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1490/3300 [1:40:16<2:47:36,  5.56s/it]2026-04-12 00:17:07,364 | INFO | train_grpo_train | metrics_logged mode=train step=1490

                                                       
{'loss': 0.019, 'grad_norm': 3.8729865550994873, 'learning_rate': 5.487878787878789e-06, 'num_tokens': 3351734.0, 'completions/mean_length': 155.625, 'completions/min_length': 151.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 155.625, 'completions/min_terminated_length': 151.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.8863186240196228, 'rewards/meter/std': 0.23323532938957214, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8888888955116272, 'rewards/repeat_penalty/std': 0.10286889225244522, 'rewards/total_composite/mean': 0.6489295363426208, 'rewards/total_composite/std': 0.17390431463718414, 'reward': 0.6489295363426208, 'reward_std': 0.17390431463718414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08673720061779022, 'sampling/sampling_logp_difference/max': 1.3975944519042969, 'sampling/importance_sampling_ratio/min': 0.24719087779521942, 'sampling/importance_sampling_ratio/mean': 1.026315450668335, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8372621536254883, 'clip_ratio/low_mean': 0.03902717446908355, 'clip_ratio/low_min': 0.03902717446908355, 'clip_ratio/high_mean': 0.03319170791655779, 'clip_ratio/high_max': 0.03319170791655779, 'clip_ratio/region_mean': 0.07221888238564134, 'reward_total_mean': 0.6489295363426208, 'reward_meter_mean': 0.8863186240196228, 'reward_meter_std': 0.23323532938957214, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8888888955116272, 'reward_repeat_penalty_std': 0.10286889225244522, 'reward_total_composite_mean': 0.6489295363426208, 'reward_total_composite_std': 0.17390431463718414, 'epoch': 0.06}

 45%|████▌     | 1490/3300 [1:40:16<2:47:36,  5.56s/it]INFO 04-12 00:17:07 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1491/3300 [1:40:21<2:41:48,  5.37s/it]2026-04-12 00:17:12,282 | INFO | train_grpo_train | metrics_logged mode=train step=1491

                                                       
{'loss': -0.0074, 'grad_norm': 5.608306884765625, 'learning_rate': 5.484848484848485e-06, 'num_tokens': 3353757.0, 'completions/mean_length': 65.875, 'completions/min_length': 64.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9181324243545532, 'rewards/meter/std': 0.15554846823215485, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9181324243545532, 'rewards/total_composite/std': 0.15554846823215485, 'reward': 0.9181324243545532, 'reward_std': 0.15554848313331604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04674442857503891, 'sampling/sampling_logp_difference/max': 1.3651247024536133, 'sampling/importance_sampling_ratio/min': 0.2553488314151764, 'sampling/importance_sampling_ratio/mean': 1.010719895362854, 'sampling/importance_sampling_ratio/max': 1.6148149967193604, 'entropy': 0.31265123561024666, 'clip_ratio/low_mean': 0.013612689450383186, 'clip_ratio/low_min': 0.013612689450383186, 'clip_ratio/high_mean': 0.02461273316293955, 'clip_ratio/high_max': 0.02461273316293955, 'clip_ratio/region_mean': 0.038225422613322735, 'reward_total_mean': 0.9181324243545532, 'reward_meter_mean': 0.9181324243545532, 'reward_meter_std': 0.15554846823215485, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9181324243545532, 'reward_total_composite_std': 0.15554846823215485, 'epoch': 0.06}

 45%|████▌     | 1491/3300 [1:40:21<2:41:48,  5.37s/it]INFO 04-12 00:17:12 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1492/3300 [1:40:25<2:35:02,  5.15s/it]2026-04-12 00:17:16,912 | INFO | train_grpo_train | metrics_logged mode=train step=1492

                                                       
{'loss': -0.0022, 'grad_norm': 7.214973449707031, 'learning_rate': 5.4818181818181825e-06, 'num_tokens': 3355426.0, 'completions/mean_length': 56.625, 'completions/min_length': 53.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.625, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7181947231292725, 'rewards/meter/std': 0.3745432198047638, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7181947231292725, 'rewards/total_composite/std': 0.3745432198047638, 'reward': 0.7181947231292725, 'reward_std': 0.3745432198047638, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07947485893964767, 'sampling/sampling_logp_difference/max': 1.292196273803711, 'sampling/importance_sampling_ratio/min': 0.2746668756008148, 'sampling/importance_sampling_ratio/mean': 1.0328161716461182, 'sampling/importance_sampling_ratio/max': 1.654712200164795, 'entropy': 0.9116378463804722, 'clip_ratio/low_mean': 0.025138093391433358, 'clip_ratio/low_min': 0.025138093391433358, 'clip_ratio/high_mean': 0.019893484190106392, 'clip_ratio/high_max': 0.019893484190106392, 'clip_ratio/region_mean': 0.04503157758153975, 'reward_total_mean': 0.7181947231292725, 'reward_meter_mean': 0.7181947231292725, 'reward_meter_std': 0.3745432198047638, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7181947231292725, 'reward_total_composite_std': 0.3745432198047638, 'epoch': 0.06}

 45%|████▌     | 1492/3300 [1:40:25<2:35:02,  5.15s/it]INFO 04-12 00:17:17 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1493/3300 [1:40:32<2:50:38,  5.67s/it]2026-04-12 00:17:23,793 | INFO | train_grpo_train | metrics_logged mode=train step=1493

                                                       
{'loss': 0.0529, 'grad_norm': 4.562529563903809, 'learning_rate': 5.478787878787879e-06, 'num_tokens': 3358873.0, 'completions/mean_length': 235.875, 'completions/min_length': 211.0, 'completions/max_length': 259.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 235.875, 'completions/min_terminated_length': 211.0, 'completions/max_terminated_length': 259.0, 'rewards/meter/mean': 0.9776490926742554, 'rewards/meter/std': 0.024429455399513245, 'rewards/count_adherence/mean': 0.7638888955116272, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.8934294581413269, 'rewards/repeat_penalty/std': 0.1227281242609024, 'rewards/total_composite/mean': 0.5754737257957458, 'rewards/total_composite/std': 0.2489534169435501, 'reward': 0.5754737257957458, 'reward_std': 0.2489534169435501, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07718866318464279, 'sampling/sampling_logp_difference/max': 1.9858770370483398, 'sampling/importance_sampling_ratio/min': 0.13726016879081726, 'sampling/importance_sampling_ratio/mean': 1.0195549726486206, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8409713134169579, 'clip_ratio/low_mean': 0.011852287454530597, 'clip_ratio/low_min': 0.011852287454530597, 'clip_ratio/high_mean': 0.03274800395593047, 'clip_ratio/high_max': 0.03274800395593047, 'clip_ratio/region_mean': 0.04460029141046107, 'reward_total_mean': 0.5754737257957458, 'reward_meter_mean': 0.9776490926742554, 'reward_meter_std': 0.024429455399513245, 'reward_count_adherence_mean': 0.7638888955116272, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.8934294581413269, 'reward_repeat_penalty_std': 0.1227281242609024, 'reward_total_composite_mean': 0.5754737257957458, 'reward_total_composite_std': 0.2489534169435501, 'epoch': 0.06}

 45%|████▌     | 1493/3300 [1:40:32<2:50:38,  5.67s/it]INFO 04-12 00:17:24 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1494/3300 [1:40:37<2:44:47,  5.47s/it]2026-04-12 00:17:28,824 | INFO | train_grpo_train | metrics_logged mode=train step=1494

                                                       
{'loss': 0.0076, 'grad_norm': 7.282406806945801, 'learning_rate': 5.475757575757576e-06, 'num_tokens': 3360884.0, 'completions/mean_length': 85.375, 'completions/min_length': 78.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.375, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.665711522102356, 'rewards/meter/std': 0.4568917751312256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.6164090037345886, 'rewards/total_composite/std': 0.42542141675949097, 'reward': 0.6164090037345886, 'reward_std': 0.4254213869571686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09246491640806198, 'sampling/sampling_logp_difference/max': 1.6489763259887695, 'sampling/importance_sampling_ratio/min': 0.19224660098552704, 'sampling/importance_sampling_ratio/mean': 1.025902509689331, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0446771308779716, 'clip_ratio/low_mean': 0.028380894218571484, 'clip_ratio/low_min': 0.028380894218571484, 'clip_ratio/high_mean': 0.014402368804439902, 'clip_ratio/high_max': 0.014402368804439902, 'clip_ratio/region_mean': 0.042783263023011386, 'reward_total_mean': 0.6164090037345886, 'reward_meter_mean': 0.665711522102356, 'reward_meter_std': 0.4568917751312256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.6164090037345886, 'reward_total_composite_std': 0.42542141675949097, 'epoch': 0.06}

 45%|████▌     | 1494/3300 [1:40:37<2:44:47,  5.47s/it]INFO 04-12 00:17:29 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1495/3300 [1:40:42<2:38:11,  5.26s/it]2026-04-12 00:17:33,580 | INFO | train_grpo_train | metrics_logged mode=train step=1495

                                                       
{'loss': 0.0083, 'grad_norm': 3.7021563053131104, 'learning_rate': 5.472727272727273e-06, 'num_tokens': 3362761.0, 'completions/mean_length': 65.625, 'completions/min_length': 64.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9978386163711548, 'rewards/meter/std': 0.0011219180887565017, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978386163711548, 'rewards/total_composite/std': 0.0011219180887565017, 'reward': 0.9978386163711548, 'reward_std': 0.0011219230946153402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019154906272888184, 'sampling/sampling_logp_difference/max': 0.6693814992904663, 'sampling/importance_sampling_ratio/min': 0.5120251774787903, 'sampling/importance_sampling_ratio/mean': 1.000672698020935, 'sampling/importance_sampling_ratio/max': 1.4048590660095215, 'entropy': 0.10151981003582478, 'clip_ratio/low_mean': 0.007462686393409967, 'clip_ratio/low_min': 0.007462686393409967, 'clip_ratio/high_mean': 0.011480186600238085, 'clip_ratio/high_max': 0.011480186600238085, 'clip_ratio/region_mean': 0.018942872993648052, 'reward_total_mean': 0.9978386163711548, 'reward_meter_mean': 0.9978386163711548, 'reward_meter_std': 0.0011219180887565017, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978386163711548, 'reward_total_composite_std': 0.0011219180887565017, 'epoch': 0.06}

 45%|████▌     | 1495/3300 [1:40:42<2:38:11,  5.26s/it]INFO 04-12 00:17:33 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1496/3300 [1:40:47<2:34:25,  5.14s/it]2026-04-12 00:17:38,432 | INFO | train_grpo_train | metrics_logged mode=train step=1496

                                                       
{'loss': 0.0026, 'grad_norm': 4.584413528442383, 'learning_rate': 5.469696969696971e-06, 'num_tokens': 3364575.0, 'completions/mean_length': 65.75, 'completions/min_length': 63.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9560348987579346, 'rewards/meter/std': 0.09728128463029861, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9560348987579346, 'rewards/total_composite/std': 0.09728128463029861, 'reward': 0.9560348987579346, 'reward_std': 0.09728129208087921, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04986903816461563, 'sampling/sampling_logp_difference/max': 1.1417694091796875, 'sampling/importance_sampling_ratio/min': 0.3192536234855652, 'sampling/importance_sampling_ratio/mean': 1.0161854028701782, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3818973954766989, 'clip_ratio/low_mean': 0.001923076924867928, 'clip_ratio/low_min': 0.001923076924867928, 'clip_ratio/high_mean': 0.02847922092769295, 'clip_ratio/high_max': 0.02847922092769295, 'clip_ratio/region_mean': 0.030402297852560878, 'reward_total_mean': 0.9560348987579346, 'reward_meter_mean': 0.9560348987579346, 'reward_meter_std': 0.09728128463029861, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9560348987579346, 'reward_total_composite_std': 0.09728128463029861, 'epoch': 0.06}

 45%|████▌     | 1496/3300 [1:40:47<2:34:25,  5.14s/it]INFO 04-12 00:17:38 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1497/3300 [1:40:52<2:32:49,  5.09s/it]2026-04-12 00:17:43,395 | INFO | train_grpo_train | metrics_logged mode=train step=1497

                                                       
{'loss': 0.0391, 'grad_norm': 6.656605243682861, 'learning_rate': 5.466666666666667e-06, 'num_tokens': 3366568.0, 'completions/mean_length': 89.125, 'completions/min_length': 84.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.125, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.6094886064529419, 'rewards/meter/std': 0.4335397183895111, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.5929021239280701, 'rewards/total_composite/std': 0.4219600558280945, 'reward': 0.5929021239280701, 'reward_std': 0.4219600260257721, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08898783475160599, 'sampling/sampling_logp_difference/max': 1.6831512451171875, 'sampling/importance_sampling_ratio/min': 0.18578758835792542, 'sampling/importance_sampling_ratio/mean': 1.0115280151367188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7317456901073456, 'clip_ratio/low_mean': 0.01920653972774744, 'clip_ratio/low_min': 0.01920653972774744, 'clip_ratio/high_mean': 0.05977132357656956, 'clip_ratio/high_max': 0.05977132357656956, 'clip_ratio/region_mean': 0.078977863304317, 'reward_total_mean': 0.5929021239280701, 'reward_meter_mean': 0.6094886064529419, 'reward_meter_std': 0.4335397183895111, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.5929021239280701, 'reward_total_composite_std': 0.4219600558280945, 'epoch': 0.06}

 45%|████▌     | 1497/3300 [1:40:52<2:32:49,  5.09s/it]INFO 04-12 00:17:43 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1498/3300 [1:40:56<2:25:36,  4.85s/it]2026-04-12 00:17:47,695 | INFO | train_grpo_train | metrics_logged mode=train step=1498

                                                       
{'loss': 0.0331, 'grad_norm': 9.448737144470215, 'learning_rate': 5.463636363636364e-06, 'num_tokens': 3368045.0, 'completions/mean_length': 34.625, 'completions/min_length': 33.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9881667494773865, 'rewards/meter/std': 0.003932104911655188, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9881667494773865, 'rewards/total_composite/std': 0.003932104911655188, 'reward': 0.9881667494773865, 'reward_std': 0.003932103049010038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04901129752397537, 'sampling/sampling_logp_difference/max': 1.6449816226959229, 'sampling/importance_sampling_ratio/min': 0.19301611185073853, 'sampling/importance_sampling_ratio/mean': 1.021955132484436, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3656330443918705, 'clip_ratio/low_mean': 0.03597753681242466, 'clip_ratio/low_min': 0.03597753681242466, 'clip_ratio/high_mean': 0.017676767893135548, 'clip_ratio/high_max': 0.017676767893135548, 'clip_ratio/region_mean': 0.05365430470556021, 'reward_total_mean': 0.9881667494773865, 'reward_meter_mean': 0.9881667494773865, 'reward_meter_std': 0.003932104911655188, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9881667494773865, 'reward_total_composite_std': 0.003932104911655188, 'epoch': 0.06}

 45%|████▌     | 1498/3300 [1:40:56<2:25:36,  4.85s/it]INFO 04-12 00:17:47 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1499/3300 [1:41:02<2:30:54,  5.03s/it]2026-04-12 00:17:53,134 | INFO | train_grpo_train | metrics_logged mode=train step=1499

                                                       
{'loss': 0.0209, 'grad_norm': 3.978285551071167, 'learning_rate': 5.460606060606061e-06, 'num_tokens': 3370801.0, 'completions/mean_length': 128.5, 'completions/min_length': 120.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.5, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.9624004364013672, 'rewards/meter/std': 0.06776177138090134, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9302494525909424, 'rewards/total_composite/std': 0.10968229919672012, 'reward': 0.9302494525909424, 'reward_std': 0.10968230664730072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09097252786159515, 'sampling/sampling_logp_difference/max': 1.2857685089111328, 'sampling/importance_sampling_ratio/min': 0.27643805742263794, 'sampling/importance_sampling_ratio/mean': 1.029836893081665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9432215169072151, 'clip_ratio/low_mean': 0.024429237004369497, 'clip_ratio/low_min': 0.024429237004369497, 'clip_ratio/high_mean': 0.05129538010805845, 'clip_ratio/high_max': 0.05129538010805845, 'clip_ratio/region_mean': 0.07572461711242795, 'reward_total_mean': 0.9302494525909424, 'reward_meter_mean': 0.9624004364013672, 'reward_meter_std': 0.06776177138090134, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9302494525909424, 'reward_total_composite_std': 0.10968229919672012, 'epoch': 0.06}

 45%|████▌     | 1499/3300 [1:41:02<2:30:54,  5.03s/it]INFO 04-12 00:17:53 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1500/3300 [1:41:06<2:29:17,  4.98s/it]2026-04-12 00:17:57,999 | INFO | train_grpo_train | metrics_logged mode=train step=1500

                                                       
{'loss': -0.0122, 'grad_norm': 5.225851058959961, 'learning_rate': 5.457575757575758e-06, 'num_tokens': 3372511.0, 'completions/mean_length': 63.75, 'completions/min_length': 60.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9959254264831543, 'rewards/meter/std': 0.0024275805335491896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9959254264831543, 'rewards/total_composite/std': 0.0024275805335491896, 'reward': 0.9959254264831543, 'reward_std': 0.002427596366032958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07604724913835526, 'sampling/sampling_logp_difference/max': 0.9057703018188477, 'sampling/importance_sampling_ratio/min': 0.40423041582107544, 'sampling/importance_sampling_ratio/mean': 1.0106977224349976, 'sampling/importance_sampling_ratio/max': 1.7771986722946167, 'entropy': 0.7506604716181755, 'clip_ratio/low_mean': 0.02421415294520557, 'clip_ratio/low_min': 0.02421415294520557, 'clip_ratio/high_mean': 0.01567198208067566, 'clip_ratio/high_max': 0.01567198208067566, 'clip_ratio/region_mean': 0.03988613502588123, 'reward_total_mean': 0.9959254264831543, 'reward_meter_mean': 0.9959254264831543, 'reward_meter_std': 0.0024275805335491896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9959254264831543, 'reward_total_composite_std': 0.0024275805335491896, 'epoch': 0.06}

 45%|████▌     | 1500/3300 [1:41:06<2:29:17,  4.98s/it]INFO 04-12 00:17:58 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:24,  2.19s/it]

 23%|██▎       | 3/13 [00:07<00:24,  2.45s/it]

 31%|███       | 4/13 [00:11<00:27,  3.01s/it]

 38%|███▊      | 5/13 [00:15<00:27,  3.39s/it]

 46%|████▌     | 6/13 [00:22<00:32,  4.61s/it]

 54%|█████▍    | 7/13 [00:26<00:26,  4.41s/it]

 62%|██████▏   | 8/13 [00:31<00:22,  4.54s/it]

 69%|██████▉   | 9/13 [00:38<00:21,  5.39s/it]

 77%|███████▋  | 10/13 [00:42<00:14,  4.93s/it]

 85%|████████▍ | 11/13 [00:46<00:09,  4.58s/it]

 92%|█████████▏| 12/13 [00:50<00:04,  4.51s/it]

100%|██████████| 13/13 [00:54<00:00,  4.41s/it]2026-04-12 00:18:57,302 | INFO | train_grpo_train | metrics_logged mode=eval step=1500

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 59.2586, 'eval_samples_per_second': 1.755, 'eval_steps_per_second': 0.219, 'eval_num_tokens': 3372511.0, 'eval_completions/mean_length': 163.32692307692307, 'eval_completions/min_length': 57.92307692307692, 'eval_completions/max_length': 304.38461538461536, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 156.44917766864484, 'eval_completions/min_terminated_length': 57.92307692307692, 'eval_completions/max_terminated_length': 271.9230769230769, 'eval_rewards/meter/mean': 0.6285286339429709, 'eval_rewards/meter/std': 0.44250402083763707, 'eval_rewards/count_adherence/mean': 0.8234805831542382, 'eval_rewards/count_adherence/std': 0.17093561990902975, 'eval_rewards/arabic_clean/mean': 0.9711538461538461, 'eval_rewards/arabic_clean/std': 0.08158924258672275, 'eval_rewards/repeat_penalty/mean': 0.8920627649013813, 'eval_rewards/repeat_penalty/std': 0.12081348953338769, 'eval_rewards/total_composite/mean': 0.4787046152811784, 'eval_rewards/total_composite/std': 0.3746733757165762, 'eval_reward': 0.4787046152811784, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.04665847968023557, 'eval_sampling/sampling_logp_difference/max': 1.12705293068519, 'eval_sampling/importance_sampling_ratio/min': 0.32900666961303127, 'eval_sampling/importance_sampling_ratio/mean': 1.0148254266152015, 'eval_sampling/importance_sampling_ratio/max': 1.5625533782518828, 'eval_entropy': 0.5243608149198385, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.4787046152811784, 'eval_reward_meter_mean': 0.6285286339429709, 'eval_reward_meter_std': 0.44250402083763707, 'eval_reward_count_adherence_mean': 0.8234805831542382, 'eval_reward_count_adherence_std': 0.17093561990902975, 'eval_reward_arabic_clean_mean': 0.9711538461538461, 'eval_reward_arabic_clean_std': 0.08158924258672275, 'eval_reward_repeat_penalty_mean': 0.8920627649013813, 'eval_reward_repeat_penalty_std': 0.12081348953338769, 'eval_reward_total_composite_mean': 0.4787046152811784, 'eval_reward_total_composite_std': 0.3746733757165762, 'epoch': 0.06}

 45%|████▌     | 1500/3300 [1:42:06<2:29:17,  4.98s/it]

100%|██████████| 13/13 [00:54<00:00,  4.41s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 45%|████▌     | 1501/3300 [1:42:13<11:39:15, 23.32s/it]2026-04-12 00:19:04,123 | INFO | train_grpo_train | metrics_logged mode=train step=1501

                                                        
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.4545454545454545e-06, 'num_tokens': 3374031.0, 'completions/mean_length': 33.0, 'completions/min_length': 33.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9988001585006714, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988001585006714, 'rewards/total_composite/std': 0.0, 'reward': 0.9988001585006714, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0034284384455531836, 'sampling/sampling_logp_difference/max': 0.05314040184020996, 'sampling/importance_sampling_ratio/min': 0.9900038242340088, 'sampling/importance_sampling_ratio/mean': 1.0031137466430664, 'sampling/importance_sampling_ratio/max': 1.0545777082443237, 'entropy': 0.031635227147489786, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9988001585006714, 'reward_meter_mean': 0.9988001585006714, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988001585006714, 'reward_total_composite_std': 0.0, 'epoch': 0.06}

 45%|████▌     | 1501/3300 [1:42:13<11:39:15, 23.32s/it]INFO 04-12 00:19:04 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1502/3300 [1:42:18<8:56:54, 17.92s/it] 2026-04-12 00:19:09,429 | INFO | train_grpo_train | metrics_logged mode=train step=1502

                                                       
{'loss': 0.0069, 'grad_norm': 5.628778457641602, 'learning_rate': 5.451515151515152e-06, 'num_tokens': 3376224.0, 'completions/mean_length': 98.125, 'completions/min_length': 93.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.125, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.8838797807693481, 'rewards/meter/std': 0.2619064450263977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8838797807693481, 'rewards/total_composite/std': 0.2619064450263977, 'reward': 0.8838797807693481, 'reward_std': 0.2619064748287201, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06008461117744446, 'sampling/sampling_logp_difference/max': 1.558868408203125, 'sampling/importance_sampling_ratio/min': 0.21037398278713226, 'sampling/importance_sampling_ratio/mean': 1.0124897956848145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5762354806065559, 'clip_ratio/low_mean': 0.020408162847161293, 'clip_ratio/low_min': 0.020408162847161293, 'clip_ratio/high_mean': 0.028100816532969475, 'clip_ratio/high_max': 0.028100816532969475, 'clip_ratio/region_mean': 0.04850897938013077, 'reward_total_mean': 0.8838797807693481, 'reward_meter_mean': 0.8838797807693481, 'reward_meter_std': 0.2619064450263977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8838797807693481, 'reward_total_composite_std': 0.2619064450263977, 'epoch': 0.06}

 46%|████▌     | 1502/3300 [1:42:18<8:56:54, 17.92s/it]INFO 04-12 00:19:09 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1503/3300 [1:42:23<6:59:25, 14.00s/it]2026-04-12 00:19:14,304 | INFO | train_grpo_train | metrics_logged mode=train step=1503

                                                       
{'loss': -0.0001, 'grad_norm': 2.2546792030334473, 'learning_rate': 5.448484848484848e-06, 'num_tokens': 3378172.0, 'completions/mean_length': 65.5, 'completions/min_length': 63.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.998626708984375, 'rewards/meter/std': 0.00024546196800656617, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998626708984375, 'rewards/total_composite/std': 0.00024546196800656617, 'reward': 0.998626708984375, 'reward_std': 0.0002454565546941012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021490152925252914, 'sampling/sampling_logp_difference/max': 1.5800909996032715, 'sampling/importance_sampling_ratio/min': 0.2059563547372818, 'sampling/importance_sampling_ratio/mean': 1.001434326171875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10324481781572104, 'clip_ratio/low_mean': 0.009538664249703288, 'clip_ratio/low_min': 0.009538664249703288, 'clip_ratio/high_mean': 0.001923076924867928, 'clip_ratio/high_max': 0.001923076924867928, 'clip_ratio/region_mean': 0.011461741174571216, 'reward_total_mean': 0.998626708984375, 'reward_meter_mean': 0.998626708984375, 'reward_meter_std': 0.00024546196800656617, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998626708984375, 'reward_total_composite_std': 0.00024546196800656617, 'epoch': 0.06}

 46%|████▌     | 1503/3300 [1:42:23<6:59:25, 14.00s/it]INFO 04-12 00:19:14 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1504/3300 [1:42:32<6:14:53, 12.52s/it]2026-04-12 00:19:23,374 | INFO | train_grpo_train | metrics_logged mode=train step=1504

                                                       
{'loss': -0.0139, 'grad_norm': 2.0835001468658447, 'learning_rate': 5.445454545454546e-06, 'num_tokens': 3383481.0, 'completions/mean_length': 384.625, 'completions/min_length': 366.0, 'completions/max_length': 408.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 384.625, 'completions/min_terminated_length': 366.0, 'completions/max_terminated_length': 408.0, 'rewards/meter/mean': 0.9722083806991577, 'rewards/meter/std': 0.05760227516293526, 'rewards/count_adherence/mean': 0.6041666269302368, 'rewards/count_adherence/std': 0.01964184269309044, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8138812780380249, 'rewards/repeat_penalty/std': 0.13613860309123993, 'rewards/total_composite/mean': 0.4811665415763855, 'rewards/total_composite/std': 0.10122059285640717, 'reward': 0.4811665415763855, 'reward_std': 0.10122059285640717, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04671372473239899, 'sampling/sampling_logp_difference/max': 1.6484105587005615, 'sampling/importance_sampling_ratio/min': 0.19235540926456451, 'sampling/importance_sampling_ratio/mean': 1.0164600610733032, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44889700040221214, 'clip_ratio/low_mean': 0.008970056660473347, 'clip_ratio/low_min': 0.008970056660473347, 'clip_ratio/high_mean': 0.019740989664569497, 'clip_ratio/high_max': 0.019740989664569497, 'clip_ratio/region_mean': 0.028711046325042844, 'reward_total_mean': 0.4811665415763855, 'reward_meter_mean': 0.9722083806991577, 'reward_meter_std': 0.05760227516293526, 'reward_count_adherence_mean': 0.6041666269302368, 'reward_count_adherence_std': 0.01964184269309044, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8138812780380249, 'reward_repeat_penalty_std': 0.13613860309123993, 'reward_total_composite_mean': 0.4811665415763855, 'reward_total_composite_std': 0.10122059285640717, 'epoch': 0.06}

 46%|████▌     | 1504/3300 [1:42:32<6:14:53, 12.52s/it]INFO 04-12 00:19:23 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1505/3300 [1:42:39<5:27:06, 10.93s/it]2026-04-12 00:19:30,600 | INFO | train_grpo_train | metrics_logged mode=train step=1505

                                                       
{'loss': -0.0283, 'grad_norm': 2.243826389312744, 'learning_rate': 5.442424242424243e-06, 'num_tokens': 3387080.0, 'completions/mean_length': 249.875, 'completions/min_length': 232.0, 'completions/max_length': 267.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 249.875, 'completions/min_terminated_length': 232.0, 'completions/max_terminated_length': 267.0, 'rewards/meter/mean': 0.9895928502082825, 'rewards/meter/std': 0.008481908589601517, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0534522607922554, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.8847527503967285, 'rewards/repeat_penalty/std': 0.05292898043990135, 'rewards/total_composite/mean': 0.584118127822876, 'rewards/total_composite/std': 0.24487483501434326, 'reward': 0.584118127822876, 'reward_std': 0.24487483501434326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.045951321721076965, 'sampling/sampling_logp_difference/max': 1.492568016052246, 'sampling/importance_sampling_ratio/min': 0.2247946411371231, 'sampling/importance_sampling_ratio/mean': 1.0099492073059082, 'sampling/importance_sampling_ratio/max': 1.8827742338180542, 'entropy': 0.4347769767045975, 'clip_ratio/low_mean': 0.006312201963737607, 'clip_ratio/low_min': 0.006312201963737607, 'clip_ratio/high_mean': 0.035130204167217016, 'clip_ratio/high_max': 0.035130204167217016, 'clip_ratio/region_mean': 0.04144240613095462, 'reward_total_mean': 0.584118127822876, 'reward_meter_mean': 0.9895928502082825, 'reward_meter_std': 0.008481908589601517, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0534522607922554, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.8847527503967285, 'reward_repeat_penalty_std': 0.05292898043990135, 'reward_total_composite_mean': 0.584118127822876, 'reward_total_composite_std': 0.24487483501434326, 'epoch': 0.06}

 46%|████▌     | 1505/3300 [1:42:39<5:27:06, 10.93s/it]INFO 04-12 00:19:30 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1506/3300 [1:42:44<4:29:58,  9.03s/it]2026-04-12 00:19:35,183 | INFO | train_grpo_train | metrics_logged mode=train step=1506

                                                       
{'loss': 0.0015, 'grad_norm': 2.0182154178619385, 'learning_rate': 5.43939393939394e-06, 'num_tokens': 3388903.0, 'completions/mean_length': 65.875, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9988133907318115, 'rewards/meter/std': 0.00020309248066041619, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988133907318115, 'rewards/total_composite/std': 0.00020309248066041619, 'reward': 0.9988133907318115, 'reward_std': 0.00020310519903432578, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009212362580001354, 'sampling/sampling_logp_difference/max': 0.6020421981811523, 'sampling/importance_sampling_ratio/min': 0.5476920008659363, 'sampling/importance_sampling_ratio/mean': 1.0014187097549438, 'sampling/importance_sampling_ratio/max': 1.8056702613830566, 'entropy': 0.06401140103116632, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.005710955825634301, 'clip_ratio/high_max': 0.005710955825634301, 'clip_ratio/region_mean': 0.007604895276017487, 'reward_total_mean': 0.9988133907318115, 'reward_meter_mean': 0.9988133907318115, 'reward_meter_std': 0.00020309248066041619, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988133907318115, 'reward_total_composite_std': 0.00020309248066041619, 'epoch': 0.06}

 46%|████▌     | 1506/3300 [1:42:44<4:29:58,  9.03s/it]INFO 04-12 00:19:35 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1507/3300 [1:42:50<4:07:46,  8.29s/it]2026-04-12 00:19:41,780 | INFO | train_grpo_train | metrics_logged mode=train step=1507

                                                       
{'loss': 0.0083, 'grad_norm': 3.6379973888397217, 'learning_rate': 5.436363636363636e-06, 'num_tokens': 3391499.0, 'completions/mean_length': 125.5, 'completions/min_length': 123.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.5, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9976436495780945, 'rewards/meter/std': 0.0009433169034309685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9441857933998108, 'rewards/total_composite/std': 0.07358824461698532, 'reward': 0.9441857933998108, 'reward_std': 0.07358825951814651, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07620169967412949, 'sampling/sampling_logp_difference/max': 1.429516315460205, 'sampling/importance_sampling_ratio/min': 0.2394247055053711, 'sampling/importance_sampling_ratio/mean': 1.0165196657180786, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6917183268815279, 'clip_ratio/low_mean': 0.006993599818088114, 'clip_ratio/low_min': 0.006993599818088114, 'clip_ratio/high_mean': 0.04194163717329502, 'clip_ratio/high_max': 0.04194163717329502, 'clip_ratio/region_mean': 0.048935236991383135, 'reward_total_mean': 0.9441857933998108, 'reward_meter_mean': 0.9976436495780945, 'reward_meter_std': 0.0009433169034309685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9441857933998108, 'reward_total_composite_std': 0.07358824461698532, 'epoch': 0.06}

 46%|████▌     | 1507/3300 [1:42:50<4:07:46,  8.29s/it]INFO 04-12 00:19:42 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1508/3300 [1:42:57<3:50:10,  7.71s/it]2026-04-12 00:19:48,094 | INFO | train_grpo_train | metrics_logged mode=train step=1508

                                                       
{'loss': 0.0189, 'grad_norm': 5.184098243713379, 'learning_rate': 5.4333333333333335e-06, 'num_tokens': 3393400.0, 'completions/mean_length': 90.625, 'completions/min_length': 86.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.625, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9280390739440918, 'rewards/meter/std': 0.12549975514411926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9280390739440918, 'rewards/total_composite/std': 0.12549975514411926, 'reward': 0.9280390739440918, 'reward_std': 0.12549975514411926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11991255730390549, 'sampling/sampling_logp_difference/max': 1.5914983749389648, 'sampling/importance_sampling_ratio/min': 0.2036202847957611, 'sampling/importance_sampling_ratio/mean': 1.030521035194397, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3208992555737495, 'clip_ratio/low_mean': 0.018996416125446558, 'clip_ratio/low_min': 0.018996416125446558, 'clip_ratio/high_mean': 0.050818526186048985, 'clip_ratio/high_max': 0.050818526186048985, 'clip_ratio/region_mean': 0.06981494231149554, 'reward_total_mean': 0.9280390739440918, 'reward_meter_mean': 0.9280390739440918, 'reward_meter_std': 0.12549975514411926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9280390739440918, 'reward_total_composite_std': 0.12549975514411926, 'epoch': 0.06}

 46%|████▌     | 1508/3300 [1:42:57<3:50:10,  7.71s/it]INFO 04-12 00:19:48 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1509/3300 [1:43:02<3:27:47,  6.96s/it]2026-04-12 00:19:53,318 | INFO | train_grpo_train | metrics_logged mode=train step=1509

                                                       
{'loss': -0.0186, 'grad_norm': 5.793516159057617, 'learning_rate': 5.430303030303032e-06, 'num_tokens': 3395804.0, 'completions/mean_length': 113.5, 'completions/min_length': 104.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.5, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9283153414726257, 'rewards/meter/std': 0.17925596237182617, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.1079898476600647, 'rewards/total_composite/mean': 0.8047507405281067, 'rewards/total_composite/std': 0.20624437928199768, 'reward': 0.8047507405281067, 'reward_std': 0.20624437928199768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03325022757053375, 'sampling/sampling_logp_difference/max': 1.20695161819458, 'sampling/importance_sampling_ratio/min': 0.29910770058631897, 'sampling/importance_sampling_ratio/mean': 1.0089442729949951, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22751092724502087, 'clip_ratio/low_mean': 0.008413461968302727, 'clip_ratio/low_min': 0.008413461968302727, 'clip_ratio/high_mean': 0.015109836007468402, 'clip_ratio/high_max': 0.015109836007468402, 'clip_ratio/region_mean': 0.02352329797577113, 'reward_total_mean': 0.8047507405281067, 'reward_meter_mean': 0.9283153414726257, 'reward_meter_std': 0.17925596237182617, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.1079898476600647, 'reward_total_composite_mean': 0.8047507405281067, 'reward_total_composite_std': 0.20624437928199768, 'epoch': 0.06}

 46%|████▌     | 1509/3300 [1:43:02<3:27:47,  6.96s/it]INFO 04-12 00:19:53 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1510/3300 [1:43:07<3:08:32,  6.32s/it]2026-04-12 00:19:58,139 | INFO | train_grpo_train | metrics_logged mode=train step=1510

                                                       
{'loss': 0.0003, 'grad_norm': 6.951155662536621, 'learning_rate': 5.427272727272728e-06, 'num_tokens': 3397499.0, 'completions/mean_length': 57.875, 'completions/min_length': 55.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7507508397102356, 'rewards/meter/std': 0.3587440252304077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7507508397102356, 'rewards/total_composite/std': 0.3587440252304077, 'reward': 0.7507508397102356, 'reward_std': 0.3587440550327301, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.053551964461803436, 'sampling/sampling_logp_difference/max': 1.5763216018676758, 'sampling/importance_sampling_ratio/min': 0.20673415064811707, 'sampling/importance_sampling_ratio/mean': 1.015701174736023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4132865481078625, 'clip_ratio/low_mean': 0.010469448054209352, 'clip_ratio/low_min': 0.010469448054209352, 'clip_ratio/high_mean': 0.023895947029814124, 'clip_ratio/high_max': 0.023895947029814124, 'clip_ratio/region_mean': 0.034365395084023476, 'reward_total_mean': 0.7507508397102356, 'reward_meter_mean': 0.7507508397102356, 'reward_meter_std': 0.3587440252304077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7507508397102356, 'reward_total_composite_std': 0.3587440252304077, 'epoch': 0.06}

 46%|████▌     | 1510/3300 [1:43:07<3:08:32,  6.32s/it]INFO 04-12 00:19:58 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1511/3300 [1:43:16<3:36:32,  7.26s/it]2026-04-12 00:20:07,597 | INFO | train_grpo_train | metrics_logged mode=train step=1511

                                                       
{'loss': -0.0464, 'grad_norm': 3.092373847961426, 'learning_rate': 5.424242424242425e-06, 'num_tokens': 3399092.0, 'completions/mean_length': 107.125, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.28571701049805, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.42876946926116943, 'rewards/meter/std': 0.4184267520904541, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.42717820405960083, 'rewards/total_composite/std': 0.4202551543712616, 'reward': 0.42717820405960083, 'reward_std': 0.420255184173584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07508043199777603, 'sampling/sampling_logp_difference/max': 1.4143075942993164, 'sampling/importance_sampling_ratio/min': 0.24309387803077698, 'sampling/importance_sampling_ratio/mean': 1.0123780965805054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5180366039276123, 'clip_ratio/low_mean': 0.02230392163619399, 'clip_ratio/low_min': 0.02230392163619399, 'clip_ratio/high_mean': 0.022818502504378557, 'clip_ratio/high_max': 0.022818502504378557, 'clip_ratio/region_mean': 0.04512242414057255, 'reward_total_mean': 0.42717820405960083, 'reward_meter_mean': 0.42876946926116943, 'reward_meter_std': 0.4184267520904541, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.42717820405960083, 'reward_total_composite_std': 0.4202551543712616, 'epoch': 0.06}

 46%|████▌     | 1511/3300 [1:43:16<3:36:32,  7.26s/it]INFO 04-12 00:20:07 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1512/3300 [1:43:26<4:01:07,  8.09s/it]2026-04-12 00:20:17,625 | INFO | train_grpo_train | metrics_logged mode=train step=1512

                                                       
{'loss': -0.3845, 'grad_norm': 1.4069581031799316, 'learning_rate': 5.421212121212122e-06, 'num_tokens': 3402818.0, 'completions/mean_length': 405.75, 'completions/min_length': 309.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 342.0, 'completions/min_terminated_length': 309.0, 'completions/max_terminated_length': 367.0, 'rewards/meter/mean': 0.7066450119018555, 'rewards/meter/std': 0.3806198835372925, 'rewards/count_adherence/mean': 0.6102941036224365, 'rewards/count_adherence/std': 0.08858475089073181, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/repeat_penalty/mean': 0.9273183345794678, 'rewards/repeat_penalty/std': 0.11418548226356506, 'rewards/total_composite/mean': 0.33205440640449524, 'rewards/total_composite/std': 0.27961450815200806, 'reward': 0.33205440640449524, 'reward_std': 0.27961447834968567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08264101296663284, 'sampling/sampling_logp_difference/max': 1.9105606079101562, 'sampling/importance_sampling_ratio/min': 0.1479973942041397, 'sampling/importance_sampling_ratio/mean': 1.020935297012329, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5392166823148727, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.031007058219984174, 'clip_ratio/high_max': 0.031007058219984174, 'clip_ratio/region_mean': 0.031007058219984174, 'reward_total_mean': 0.33205440640449524, 'reward_meter_mean': 0.7066450119018555, 'reward_meter_std': 0.3806198835372925, 'reward_count_adherence_mean': 0.6102941036224365, 'reward_count_adherence_std': 0.08858475089073181, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_repeat_penalty_mean': 0.9273183345794678, 'reward_repeat_penalty_std': 0.11418548226356506, 'reward_total_composite_mean': 0.33205440640449524, 'reward_total_composite_std': 0.27961450815200806, 'epoch': 0.06}

 46%|████▌     | 1512/3300 [1:43:26<4:01:07,  8.09s/it]INFO 04-12 00:20:17 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1513/3300 [1:43:31<3:30:04,  7.05s/it]2026-04-12 00:20:22,254 | INFO | train_grpo_train | metrics_logged mode=train step=1513

                                                       
{'loss': 0.0171, 'grad_norm': 4.46673583984375, 'learning_rate': 5.418181818181819e-06, 'num_tokens': 3404619.0, 'completions/mean_length': 67.125, 'completions/min_length': 66.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9849165081977844, 'rewards/meter/std': 0.019791634753346443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9849165081977844, 'rewards/total_composite/std': 0.019791634753346443, 'reward': 0.9849165081977844, 'reward_std': 0.019791632890701294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025447804480791092, 'sampling/sampling_logp_difference/max': 0.6372613906860352, 'sampling/importance_sampling_ratio/min': 0.5287384390830994, 'sampling/importance_sampling_ratio/mean': 1.0061801671981812, 'sampling/importance_sampling_ratio/max': 1.5401774644851685, 'entropy': 0.2136769648641348, 'clip_ratio/low_mean': 0.005488860071636736, 'clip_ratio/low_min': 0.005488860071636736, 'clip_ratio/high_mean': 0.014931121026165783, 'clip_ratio/high_max': 0.014931121026165783, 'clip_ratio/region_mean': 0.02041998109780252, 'reward_total_mean': 0.9849165081977844, 'reward_meter_mean': 0.9849165081977844, 'reward_meter_std': 0.019791634753346443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9849165081977844, 'reward_total_composite_std': 0.019791634753346443, 'epoch': 0.06}

 46%|████▌     | 1513/3300 [1:43:31<3:30:04,  7.05s/it]INFO 04-12 00:20:22 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1514/3300 [1:43:37<3:26:53,  6.95s/it]2026-04-12 00:20:28,984 | INFO | train_grpo_train | metrics_logged mode=train step=1514

                                                       
{'loss': -0.0047, 'grad_norm': 2.8653032779693604, 'learning_rate': 5.415151515151515e-06, 'num_tokens': 3408083.0, 'completions/mean_length': 228.0, 'completions/min_length': 218.0, 'completions/max_length': 236.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 228.0, 'completions/min_terminated_length': 218.0, 'completions/max_terminated_length': 236.0, 'rewards/meter/mean': 0.9303399324417114, 'rewards/meter/std': 0.1539093255996704, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9519230723381042, 'rewards/repeat_penalty/std': 0.05723259598016739, 'rewards/total_composite/mean': 0.7759657502174377, 'rewards/total_composite/std': 0.14068926870822906, 'reward': 0.7759657502174377, 'reward_std': 0.14068925380706787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08908653259277344, 'sampling/sampling_logp_difference/max': 1.7966899871826172, 'sampling/importance_sampling_ratio/min': 0.16584692895412445, 'sampling/importance_sampling_ratio/mean': 1.0278751850128174, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0589845478534698, 'clip_ratio/low_mean': 0.012164860963821411, 'clip_ratio/low_min': 0.012164860963821411, 'clip_ratio/high_mean': 0.04533929843455553, 'clip_ratio/high_max': 0.04533929843455553, 'clip_ratio/region_mean': 0.05750415939837694, 'reward_total_mean': 0.7759657502174377, 'reward_meter_mean': 0.9303399324417114, 'reward_meter_std': 0.1539093255996704, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9519230723381042, 'reward_repeat_penalty_std': 0.05723259598016739, 'reward_total_composite_mean': 0.7759657502174377, 'reward_total_composite_std': 0.14068926870822906, 'epoch': 0.06}

 46%|████▌     | 1514/3300 [1:43:37<3:26:53,  6.95s/it]INFO 04-12 00:20:29 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1515/3300 [1:43:42<3:06:43,  6.28s/it]2026-04-12 00:20:33,671 | INFO | train_grpo_train | metrics_logged mode=train step=1515

                                                       
{'loss': -0.0027, 'grad_norm': 3.14272403717041, 'learning_rate': 5.412121212121213e-06, 'num_tokens': 3410076.0, 'completions/mean_length': 68.125, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9943439960479736, 'rewards/meter/std': 0.0008108045440167189, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943439960479736, 'rewards/total_composite/std': 0.0008108045440167189, 'reward': 0.9943439960479736, 'reward_std': 0.0008108008187264204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03039509244263172, 'sampling/sampling_logp_difference/max': 0.9262442588806152, 'sampling/importance_sampling_ratio/min': 0.39603835344314575, 'sampling/importance_sampling_ratio/mean': 1.0022821426391602, 'sampling/importance_sampling_ratio/max': 1.5258408784866333, 'entropy': 0.2124041486531496, 'clip_ratio/low_mean': 0.0055147059028968215, 'clip_ratio/low_min': 0.0055147059028968215, 'clip_ratio/high_mean': 0.0221144916722551, 'clip_ratio/high_max': 0.0221144916722551, 'clip_ratio/region_mean': 0.02762919757515192, 'reward_total_mean': 0.9943439960479736, 'reward_meter_mean': 0.9943439960479736, 'reward_meter_std': 0.0008108045440167189, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943439960479736, 'reward_total_composite_std': 0.0008108045440167189, 'epoch': 0.06}

 46%|████▌     | 1515/3300 [1:43:42<3:06:43,  6.28s/it]INFO 04-12 00:20:33 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1516/3300 [1:43:47<2:53:25,  5.83s/it]2026-04-12 00:20:38,466 | INFO | train_grpo_train | metrics_logged mode=train step=1516

                                                       
{'loss': -0.0086, 'grad_norm': 6.165884017944336, 'learning_rate': 5.409090909090909e-06, 'num_tokens': 3412012.0, 'completions/mean_length': 67.0, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9924481511116028, 'rewards/meter/std': 0.005331422667950392, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924481511116028, 'rewards/total_composite/std': 0.005331422667950392, 'reward': 0.9924481511116028, 'reward_std': 0.0053314100950956345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028799764811992645, 'sampling/sampling_logp_difference/max': 1.173213005065918, 'sampling/importance_sampling_ratio/min': 0.30937132239341736, 'sampling/importance_sampling_ratio/mean': 1.0072860717773438, 'sampling/importance_sampling_ratio/max': 1.659652829170227, 'entropy': 0.19451193884015083, 'clip_ratio/low_mean': 0.003818796598352492, 'clip_ratio/low_min': 0.003818796598352492, 'clip_ratio/high_mean': 0.009246049099601805, 'clip_ratio/high_max': 0.009246049099601805, 'clip_ratio/region_mean': 0.013064845697954297, 'reward_total_mean': 0.9924481511116028, 'reward_meter_mean': 0.9924481511116028, 'reward_meter_std': 0.005331422667950392, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924481511116028, 'reward_total_composite_std': 0.005331422667950392, 'epoch': 0.06}

 46%|████▌     | 1516/3300 [1:43:47<2:53:25,  5.83s/it]INFO 04-12 00:20:38 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1517/3300 [1:43:51<2:41:20,  5.43s/it]2026-04-12 00:20:42,956 | INFO | train_grpo_train | metrics_logged mode=train step=1517

                                                       
{'loss': -0.015, 'grad_norm': 2.4998230934143066, 'learning_rate': 5.406060606060607e-06, 'num_tokens': 3413722.0, 'completions/mean_length': 57.75, 'completions/min_length': 55.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9230245351791382, 'rewards/meter/std': 0.1628594845533371, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9230245351791382, 'rewards/total_composite/std': 0.1628594845533371, 'reward': 0.9230245351791382, 'reward_std': 0.1628594696521759, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04387671872973442, 'sampling/sampling_logp_difference/max': 1.2724885940551758, 'sampling/importance_sampling_ratio/min': 0.28013360500335693, 'sampling/importance_sampling_ratio/mean': 1.0011502504348755, 'sampling/importance_sampling_ratio/max': 1.4794626235961914, 'entropy': 0.31292806193232536, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.027737777214497328, 'clip_ratio/high_max': 0.027737777214497328, 'clip_ratio/region_mean': 0.027737777214497328, 'reward_total_mean': 0.9230245351791382, 'reward_meter_mean': 0.9230245351791382, 'reward_meter_std': 0.1628594845533371, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9230245351791382, 'reward_total_composite_std': 0.1628594845533371, 'epoch': 0.06}

 46%|████▌     | 1517/3300 [1:43:51<2:41:20,  5.43s/it]INFO 04-12 00:20:43 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1518/3300 [1:43:56<2:33:52,  5.18s/it]2026-04-12 00:20:47,557 | INFO | train_grpo_train | metrics_logged mode=train step=1518

                                                       
{'loss': 0.0093, 'grad_norm': 6.170195579528809, 'learning_rate': 5.4030303030303036e-06, 'num_tokens': 3415375.0, 'completions/mean_length': 60.625, 'completions/min_length': 57.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5306375622749329, 'rewards/meter/std': 0.289710134267807, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5306375622749329, 'rewards/total_composite/std': 0.289710134267807, 'reward': 0.5306375622749329, 'reward_std': 0.289710134267807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07036992907524109, 'sampling/sampling_logp_difference/max': 1.4188556671142578, 'sampling/importance_sampling_ratio/min': 0.24199077486991882, 'sampling/importance_sampling_ratio/mean': 1.0169787406921387, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6950614899396896, 'clip_ratio/low_mean': 0.030917395371943712, 'clip_ratio/low_min': 0.030917395371943712, 'clip_ratio/high_mean': 0.04732227721251547, 'clip_ratio/high_max': 0.04732227721251547, 'clip_ratio/region_mean': 0.07823967258445919, 'reward_total_mean': 0.5306375622749329, 'reward_meter_mean': 0.5306375622749329, 'reward_meter_std': 0.289710134267807, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5306375622749329, 'reward_total_composite_std': 0.289710134267807, 'epoch': 0.06}

 46%|████▌     | 1518/3300 [1:43:56<2:33:52,  5.18s/it]INFO 04-12 00:20:47 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1519/3300 [1:44:04<2:55:04,  5.90s/it]2026-04-12 00:20:55,127 | INFO | train_grpo_train | metrics_logged mode=train step=1519

                                                       
{'loss': -0.0293, 'grad_norm': 3.1494743824005127, 'learning_rate': 5.400000000000001e-06, 'num_tokens': 3419414.0, 'completions/mean_length': 290.875, 'completions/min_length': 274.0, 'completions/max_length': 312.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 290.875, 'completions/min_terminated_length': 274.0, 'completions/max_terminated_length': 312.0, 'rewards/meter/mean': 0.9781413078308105, 'rewards/meter/std': 0.051752228289842606, 'rewards/count_adherence/mean': 0.7045454978942871, 'rewards/count_adherence/std': 0.04208274558186531, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9208333492279053, 'rewards/repeat_penalty/std': 0.12346728891134262, 'rewards/total_composite/mean': 0.5596551299095154, 'rewards/total_composite/std': 0.2521139085292816, 'reward': 0.5596551299095154, 'reward_std': 0.2521139085292816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07325157523155212, 'sampling/sampling_logp_difference/max': 1.6432204246520996, 'sampling/importance_sampling_ratio/min': 0.19335635006427765, 'sampling/importance_sampling_ratio/mean': 1.0232716798782349, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7886361554265022, 'clip_ratio/low_mean': 0.020800751633942127, 'clip_ratio/low_min': 0.020800751633942127, 'clip_ratio/high_mean': 0.028688129736110568, 'clip_ratio/high_max': 0.028688129736110568, 'clip_ratio/region_mean': 0.049488881370052695, 'reward_total_mean': 0.5596551299095154, 'reward_meter_mean': 0.9781413078308105, 'reward_meter_std': 0.051752228289842606, 'reward_count_adherence_mean': 0.7045454978942871, 'reward_count_adherence_std': 0.04208274558186531, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9208333492279053, 'reward_repeat_penalty_std': 0.12346728891134262, 'reward_total_composite_mean': 0.5596551299095154, 'reward_total_composite_std': 0.2521139085292816, 'epoch': 0.06}

 46%|████▌     | 1519/3300 [1:44:04<2:55:04,  5.90s/it]INFO 04-12 00:20:55 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1520/3300 [1:44:09<2:46:27,  5.61s/it]2026-04-12 00:21:00,070 | INFO | train_grpo_train | metrics_logged mode=train step=1520

                                                       
{'loss': 0.0163, 'grad_norm': 2.657116413116455, 'learning_rate': 5.396969696969697e-06, 'num_tokens': 3421603.0, 'completions/mean_length': 88.625, 'completions/min_length': 86.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.625, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9171510934829712, 'rewards/meter/std': 0.1984940767288208, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8182051777839661, 'rewards/total_composite/std': 0.1850285530090332, 'reward': 0.8182051777839661, 'reward_std': 0.1850285530090332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01846635341644287, 'sampling/sampling_logp_difference/max': 0.759364128112793, 'sampling/importance_sampling_ratio/min': 0.46796390414237976, 'sampling/importance_sampling_ratio/mean': 1.0104413032531738, 'sampling/importance_sampling_ratio/max': 1.6496562957763672, 'entropy': 0.1767467763274908, 'clip_ratio/low_mean': 0.0028572361916303635, 'clip_ratio/low_min': 0.0028572361916303635, 'clip_ratio/high_mean': 0.004295865655876696, 'clip_ratio/high_max': 0.004295865655876696, 'clip_ratio/region_mean': 0.00715310184750706, 'reward_total_mean': 0.8182051777839661, 'reward_meter_mean': 0.9171510934829712, 'reward_meter_std': 0.1984940767288208, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8182051777839661, 'reward_total_composite_std': 0.1850285530090332, 'epoch': 0.06}

 46%|████▌     | 1520/3300 [1:44:09<2:46:27,  5.61s/it]INFO 04-12 00:21:00 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1521/3300 [1:44:13<2:39:15,  5.37s/it]2026-04-12 00:21:04,881 | INFO | train_grpo_train | metrics_logged mode=train step=1521

                                                       
{'loss': 0.0033, 'grad_norm': 1.7797563076019287, 'learning_rate': 5.3939393939393945e-06, 'num_tokens': 3423423.0, 'completions/mean_length': 68.5, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9934801459312439, 'rewards/meter/std': 0.0036428512539714575, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9934801459312439, 'rewards/total_composite/std': 0.0036428512539714575, 'reward': 0.9934801459312439, 'reward_std': 0.00364283611997962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01598731242120266, 'sampling/sampling_logp_difference/max': 1.9260679483413696, 'sampling/importance_sampling_ratio/min': 0.14572004973888397, 'sampling/importance_sampling_ratio/mean': 1.0017439126968384, 'sampling/importance_sampling_ratio/max': 1.2535090446472168, 'entropy': 0.09944538865238428, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010897728730924428, 'clip_ratio/high_max': 0.010897728730924428, 'clip_ratio/region_mean': 0.010897728730924428, 'reward_total_mean': 0.9934801459312439, 'reward_meter_mean': 0.9934801459312439, 'reward_meter_std': 0.0036428512539714575, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9934801459312439, 'reward_total_composite_std': 0.0036428512539714575, 'epoch': 0.06}

 46%|████▌     | 1521/3300 [1:44:13<2:39:15,  5.37s/it]INFO 04-12 00:21:05 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1522/3300 [1:44:18<2:33:38,  5.18s/it]2026-04-12 00:21:09,634 | INFO | train_grpo_train | metrics_logged mode=train step=1522

                                                       
{'loss': -0.0022, 'grad_norm': 7.3915534019470215, 'learning_rate': 5.390909090909091e-06, 'num_tokens': 3425367.0, 'completions/mean_length': 65.0, 'completions/min_length': 63.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9976125955581665, 'rewards/meter/std': 0.0013877107994630933, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976125955581665, 'rewards/total_composite/std': 0.0013877107994630933, 'reward': 0.9976125955581665, 'reward_std': 0.0013877045130357146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040052663534879684, 'sampling/sampling_logp_difference/max': 0.6700773239135742, 'sampling/importance_sampling_ratio/min': 0.5116690397262573, 'sampling/importance_sampling_ratio/mean': 1.0058274269104004, 'sampling/importance_sampling_ratio/max': 1.817367672920227, 'entropy': 0.37458037585020065, 'clip_ratio/low_mean': 0.0058302809484303, 'clip_ratio/low_min': 0.0058302809484303, 'clip_ratio/high_mean': 0.019181564333848655, 'clip_ratio/high_max': 0.019181564333848655, 'clip_ratio/region_mean': 0.025011845282278955, 'reward_total_mean': 0.9976125955581665, 'reward_meter_mean': 0.9976125955581665, 'reward_meter_std': 0.0013877107994630933, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976125955581665, 'reward_total_composite_std': 0.0013877107994630933, 'epoch': 0.06}

 46%|████▌     | 1522/3300 [1:44:18<2:33:38,  5.18s/it]INFO 04-12 00:21:09 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1523/3300 [1:44:22<2:26:37,  4.95s/it]2026-04-12 00:21:14,037 | INFO | train_grpo_train | metrics_logged mode=train step=1523

                                                       
{'loss': 0.0106, 'grad_norm': 9.1239013671875, 'learning_rate': 5.387878787878789e-06, 'num_tokens': 3426883.0, 'completions/mean_length': 34.5, 'completions/min_length': 32.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9722611904144287, 'rewards/meter/std': 0.0555478073656559, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9722611904144287, 'rewards/total_composite/std': 0.0555478073656559, 'reward': 0.9722611904144287, 'reward_std': 0.05554782226681709, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07421485334634781, 'sampling/sampling_logp_difference/max': 1.0678462982177734, 'sampling/importance_sampling_ratio/min': 0.3437480628490448, 'sampling/importance_sampling_ratio/mean': 1.0279422998428345, 'sampling/importance_sampling_ratio/max': 1.5298467874526978, 'entropy': 0.7306949496269226, 'clip_ratio/low_mean': 0.010714286006987095, 'clip_ratio/low_min': 0.010714286006987095, 'clip_ratio/high_mean': 0.04335171659477055, 'clip_ratio/high_max': 0.04335171659477055, 'clip_ratio/region_mean': 0.054066002601757646, 'reward_total_mean': 0.9722611904144287, 'reward_meter_mean': 0.9722611904144287, 'reward_meter_std': 0.0555478073656559, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9722611904144287, 'reward_total_composite_std': 0.0555478073656559, 'epoch': 0.06}

 46%|████▌     | 1523/3300 [1:44:23<2:26:37,  4.95s/it]INFO 04-12 00:21:14 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1524/3300 [1:44:28<2:28:49,  5.03s/it]2026-04-12 00:21:19,242 | INFO | train_grpo_train | metrics_logged mode=train step=1524

                                                       
{'loss': 0.0095, 'grad_norm': 5.130848407745361, 'learning_rate': 5.384848484848485e-06, 'num_tokens': 3429089.0, 'completions/mean_length': 108.75, 'completions/min_length': 99.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.75, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.9915084838867188, 'rewards/meter/std': 0.007808292284607887, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9915084838867188, 'rewards/total_composite/std': 0.007808292284607887, 'reward': 0.9915084838867188, 'reward_std': 0.007808296009898186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08183825016021729, 'sampling/sampling_logp_difference/max': 1.6284008026123047, 'sampling/importance_sampling_ratio/min': 0.19624315202236176, 'sampling/importance_sampling_ratio/mean': 1.0207182168960571, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.894000805914402, 'clip_ratio/low_mean': 0.01814199541695416, 'clip_ratio/low_min': 0.01814199541695416, 'clip_ratio/high_mean': 0.03156230039894581, 'clip_ratio/high_max': 0.03156230039894581, 'clip_ratio/region_mean': 0.04970429581589997, 'reward_total_mean': 0.9915084838867188, 'reward_meter_mean': 0.9915084838867188, 'reward_meter_std': 0.007808292284607887, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9915084838867188, 'reward_total_composite_std': 0.007808292284607887, 'epoch': 0.06}

 46%|████▌     | 1524/3300 [1:44:28<2:28:49,  5.03s/it]INFO 04-12 00:21:19 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1525/3300 [1:44:33<2:27:36,  4.99s/it]2026-04-12 00:21:24,148 | INFO | train_grpo_train | metrics_logged mode=train step=1525

                                                       
{'loss': 0.0026, 'grad_norm': 2.808112382888794, 'learning_rate': 5.381818181818183e-06, 'num_tokens': 3431040.0, 'completions/mean_length': 86.875, 'completions/min_length': 86.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.875, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.99290531873703, 'rewards/meter/std': 0.0034807121846824884, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9431794881820679, 'rewards/total_composite/std': 0.09114136546850204, 'reward': 0.9431794881820679, 'reward_std': 0.09114135801792145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017489619553089142, 'sampling/sampling_logp_difference/max': 0.7286503314971924, 'sampling/importance_sampling_ratio/min': 0.48398464918136597, 'sampling/importance_sampling_ratio/mean': 1.0029078722000122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09439879935234785, 'clip_ratio/low_mean': 0.0028735632076859474, 'clip_ratio/low_min': 0.0028735632076859474, 'clip_ratio/high_mean': 0.012915466912090778, 'clip_ratio/high_max': 0.012915466912090778, 'clip_ratio/region_mean': 0.015789030119776726, 'reward_total_mean': 0.9431794881820679, 'reward_meter_mean': 0.99290531873703, 'reward_meter_std': 0.0034807121846824884, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9431794881820679, 'reward_total_composite_std': 0.09114136546850204, 'epoch': 0.06}

 46%|████▌     | 1525/3300 [1:44:33<2:27:36,  4.99s/it]INFO 04-12 00:21:24 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1526/3300 [1:44:38<2:28:02,  5.01s/it]2026-04-12 00:21:29,196 | INFO | train_grpo_train | metrics_logged mode=train step=1526

                                                       
{'loss': -0.0074, 'grad_norm': 2.34173321723938, 'learning_rate': 5.378787878787879e-06, 'num_tokens': 3433169.0, 'completions/mean_length': 95.125, 'completions/min_length': 92.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.125, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9933592081069946, 'rewards/meter/std': 0.00307702855207026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8195037841796875, 'rewards/total_composite/std': 0.07000674307346344, 'reward': 0.8195037841796875, 'reward_std': 0.07000675052404404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023438630625605583, 'sampling/sampling_logp_difference/max': 1.2709712982177734, 'sampling/importance_sampling_ratio/min': 0.2805590033531189, 'sampling/importance_sampling_ratio/mean': 1.006750464439392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13843786902725697, 'clip_ratio/low_mean': 0.017182219657115638, 'clip_ratio/low_min': 0.017182219657115638, 'clip_ratio/high_mean': 0.0025510203558951616, 'clip_ratio/high_max': 0.0025510203558951616, 'clip_ratio/region_mean': 0.0197332400130108, 'reward_total_mean': 0.8195037841796875, 'reward_meter_mean': 0.9933592081069946, 'reward_meter_std': 0.00307702855207026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8195037841796875, 'reward_total_composite_std': 0.07000674307346344, 'epoch': 0.06}

 46%|████▌     | 1526/3300 [1:44:38<2:28:02,  5.01s/it]INFO 04-12 00:21:29 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1527/3300 [1:44:43<2:34:37,  5.23s/it]2026-04-12 00:21:34,968 | INFO | train_grpo_train | metrics_logged mode=train step=1527

                                                       
{'loss': 0.0164, 'grad_norm': 3.2660653591156006, 'learning_rate': 5.375757575757576e-06, 'num_tokens': 3435703.0, 'completions/mean_length': 142.75, 'completions/min_length': 134.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.75, 'completions/min_terminated_length': 134.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.994535505771637, 'rewards/meter/std': 0.0031923020724207163, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.10628911107778549, 'rewards/total_composite/mean': 0.9411816596984863, 'rewards/total_composite/std': 0.10509292781352997, 'reward': 0.9411816596984863, 'reward_std': 0.10509292781352997, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07457228004932404, 'sampling/sampling_logp_difference/max': 1.6383848190307617, 'sampling/importance_sampling_ratio/min': 0.19429361820220947, 'sampling/importance_sampling_ratio/mean': 1.02262282371521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7129531875252724, 'clip_ratio/low_mean': 0.006133304443210363, 'clip_ratio/low_min': 0.006133304443210363, 'clip_ratio/high_mean': 0.044742210768163204, 'clip_ratio/high_max': 0.044742210768163204, 'clip_ratio/region_mean': 0.05087551521137357, 'reward_total_mean': 0.9411816596984863, 'reward_meter_mean': 0.994535505771637, 'reward_meter_std': 0.0031923020724207163, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.10628911107778549, 'reward_total_composite_mean': 0.9411816596984863, 'reward_total_composite_std': 0.10509292781352997, 'epoch': 0.06}

 46%|████▋     | 1527/3300 [1:44:43<2:34:37,  5.23s/it]INFO 04-12 00:21:35 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1528/3300 [1:44:48<2:32:48,  5.17s/it]2026-04-12 00:21:39,992 | INFO | train_grpo_train | metrics_logged mode=train step=1528

                                                       
{'loss': -0.0161, 'grad_norm': 5.586496353149414, 'learning_rate': 5.372727272727273e-06, 'num_tokens': 3437794.0, 'completions/mean_length': 93.375, 'completions/min_length': 89.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.375, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.8450676202774048, 'rewards/meter/std': 0.220209538936615, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8450676202774048, 'rewards/total_composite/std': 0.220209538936615, 'reward': 0.8450676202774048, 'reward_std': 0.2202095091342926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05679401755332947, 'sampling/sampling_logp_difference/max': 0.9821338653564453, 'sampling/importance_sampling_ratio/min': 0.37451109290122986, 'sampling/importance_sampling_ratio/mean': 1.0161728858947754, 'sampling/importance_sampling_ratio/max': 1.8447113037109375, 'entropy': 0.5462022610008717, 'clip_ratio/low_mean': 0.009738856926560402, 'clip_ratio/low_min': 0.009738856926560402, 'clip_ratio/high_mean': 0.02779570873826742, 'clip_ratio/high_max': 0.02779570873826742, 'clip_ratio/region_mean': 0.037534565664827824, 'reward_total_mean': 0.8450676202774048, 'reward_meter_mean': 0.8450676202774048, 'reward_meter_std': 0.220209538936615, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8450676202774048, 'reward_total_composite_std': 0.220209538936615, 'epoch': 0.06}

 46%|████▋     | 1528/3300 [1:44:48<2:32:48,  5.17s/it]INFO 04-12 00:21:40 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1529/3300 [1:44:53<2:24:24,  4.89s/it]2026-04-12 00:21:44,227 | INFO | train_grpo_train | metrics_logged mode=train step=1529

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.36969696969697e-06, 'num_tokens': 3439074.0, 'completions/mean_length': 33.0, 'completions/min_length': 33.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9988001585006714, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988001585006714, 'rewards/total_composite/std': 0.0, 'reward': 0.9988001585006714, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0021949801594018936, 'sampling/sampling_logp_difference/max': 0.04461796581745148, 'sampling/importance_sampling_ratio/min': 0.9977096915245056, 'sampling/importance_sampling_ratio/mean': 1.0021947622299194, 'sampling/importance_sampling_ratio/max': 1.045628309249878, 'entropy': 0.017181317321956158, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9988001585006714, 'reward_meter_mean': 0.9988001585006714, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988001585006714, 'reward_total_composite_std': 0.0, 'epoch': 0.06}

 46%|████▋     | 1529/3300 [1:44:53<2:24:24,  4.89s/it]INFO 04-12 00:21:44 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1530/3300 [1:44:57<2:22:57,  4.85s/it]2026-04-12 00:21:48,965 | INFO | train_grpo_train | metrics_logged mode=train step=1530

                                                       
{'loss': 0.0132, 'grad_norm': 3.0035605430603027, 'learning_rate': 5.366666666666666e-06, 'num_tokens': 3440826.0, 'completions/mean_length': 66.0, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9936773180961609, 'rewards/meter/std': 0.0019825000781565905, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9936773180961609, 'rewards/total_composite/std': 0.0019825000781565905, 'reward': 0.9936773180961609, 'reward_std': 0.001982480986043811, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02317051589488983, 'sampling/sampling_logp_difference/max': 0.8188035488128662, 'sampling/importance_sampling_ratio/min': 0.4409589171409607, 'sampling/importance_sampling_ratio/mean': 1.0015543699264526, 'sampling/importance_sampling_ratio/max': 1.562909722328186, 'entropy': 0.1264043264091015, 'clip_ratio/low_mean': 0.003791360300965607, 'clip_ratio/low_min': 0.003791360300965607, 'clip_ratio/high_mean': 0.015155904809944332, 'clip_ratio/high_max': 0.015155904809944332, 'clip_ratio/region_mean': 0.01894726511090994, 'reward_total_mean': 0.9936773180961609, 'reward_meter_mean': 0.9936773180961609, 'reward_meter_std': 0.0019825000781565905, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9936773180961609, 'reward_total_composite_std': 0.0019825000781565905, 'epoch': 0.06}

 46%|████▋     | 1530/3300 [1:44:57<2:22:57,  4.85s/it]INFO 04-12 00:21:49 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1531/3300 [1:45:03<2:25:34,  4.94s/it]2026-04-12 00:21:54,118 | INFO | train_grpo_train | metrics_logged mode=train step=1531

                                                       
{'loss': -0.0098, 'grad_norm': 5.306570053100586, 'learning_rate': 5.3636363636363645e-06, 'num_tokens': 3443097.0, 'completions/mean_length': 106.875, 'completions/min_length': 104.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.875, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.980573296546936, 'rewards/meter/std': 0.03288979455828667, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8351579904556274, 'rewards/total_composite/std': 0.3450130224227905, 'reward': 0.8351579904556274, 'reward_std': 0.3450130224227905, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08104944229125977, 'sampling/sampling_logp_difference/max': 2.037365436553955, 'sampling/importance_sampling_ratio/min': 0.13037173449993134, 'sampling/importance_sampling_ratio/mean': 1.011287808418274, 'sampling/importance_sampling_ratio/max': 1.9533883333206177, 'entropy': 0.799922339618206, 'clip_ratio/low_mean': 0.007166183087974787, 'clip_ratio/low_min': 0.007166183087974787, 'clip_ratio/high_mean': 0.045307138469070196, 'clip_ratio/high_max': 0.045307138469070196, 'clip_ratio/region_mean': 0.05247332155704498, 'reward_total_mean': 0.8351579904556274, 'reward_meter_mean': 0.980573296546936, 'reward_meter_std': 0.03288979455828667, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8351579904556274, 'reward_total_composite_std': 0.3450130224227905, 'epoch': 0.06}

 46%|████▋     | 1531/3300 [1:45:03<2:25:34,  4.94s/it]INFO 04-12 00:21:54 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1532/3300 [1:45:08<2:33:08,  5.20s/it]2026-04-12 00:21:59,927 | INFO | train_grpo_train | metrics_logged mode=train step=1532

                                                       
{'loss': -0.0044, 'grad_norm': 3.583285331726074, 'learning_rate': 5.360606060606061e-06, 'num_tokens': 3445821.0, 'completions/mean_length': 160.5, 'completions/min_length': 155.0, 'completions/max_length': 165.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.5, 'completions/min_terminated_length': 155.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.8937799334526062, 'rewards/meter/std': 0.1616331785917282, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333134651184, 'rewards/repeat_penalty/std': 0.11878276616334915, 'rewards/total_composite/mean': 0.7505040168762207, 'rewards/total_composite/std': 0.18041202425956726, 'reward': 0.7505040168762207, 'reward_std': 0.18041202425956726, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06113876402378082, 'sampling/sampling_logp_difference/max': 1.3376035690307617, 'sampling/importance_sampling_ratio/min': 0.2624739110469818, 'sampling/importance_sampling_ratio/mean': 1.012855052947998, 'sampling/importance_sampling_ratio/max': 1.947861909866333, 'entropy': 0.5020967051386833, 'clip_ratio/low_mean': 0.01080156397074461, 'clip_ratio/low_min': 0.01080156397074461, 'clip_ratio/high_mean': 0.035965551156550646, 'clip_ratio/high_max': 0.035965551156550646, 'clip_ratio/region_mean': 0.046767115127295256, 'reward_total_mean': 0.7505040168762207, 'reward_meter_mean': 0.8937799334526062, 'reward_meter_std': 0.1616331785917282, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333134651184, 'reward_repeat_penalty_std': 0.11878276616334915, 'reward_total_composite_mean': 0.7505040168762207, 'reward_total_composite_std': 0.18041202425956726, 'epoch': 0.06}

 46%|████▋     | 1532/3300 [1:45:08<2:33:08,  5.20s/it]INFO 04-12 00:22:00 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1533/3300 [1:45:13<2:26:10,  4.96s/it]2026-04-12 00:22:04,343 | INFO | train_grpo_train | metrics_logged mode=train step=1533

                                                       
{'loss': 0.0012, 'grad_norm': 4.299396514892578, 'learning_rate': 5.357575757575758e-06, 'num_tokens': 3447376.0, 'completions/mean_length': 34.375, 'completions/min_length': 33.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9954771995544434, 'rewards/meter/std': 0.002512240083888173, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9954771995544434, 'rewards/total_composite/std': 0.002512240083888173, 'reward': 0.9954771995544434, 'reward_std': 0.00251222332008183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01558676641434431, 'sampling/sampling_logp_difference/max': 0.46267223358154297, 'sampling/importance_sampling_ratio/min': 0.7345414757728577, 'sampling/importance_sampling_ratio/mean': 1.009590983390808, 'sampling/importance_sampling_ratio/max': 1.5883126258850098, 'entropy': 0.13886073511093855, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.010924369795247912, 'clip_ratio/high_max': 0.010924369795247912, 'clip_ratio/region_mean': 0.014712248696014285, 'reward_total_mean': 0.9954771995544434, 'reward_meter_mean': 0.9954771995544434, 'reward_meter_std': 0.002512240083888173, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9954771995544434, 'reward_total_composite_std': 0.002512240083888173, 'epoch': 0.06}

 46%|████▋     | 1533/3300 [1:45:13<2:26:10,  4.96s/it]INFO 04-12 00:22:04 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1534/3300 [1:45:22<3:06:19,  6.33s/it]2026-04-12 00:22:13,858 | INFO | train_grpo_train | metrics_logged mode=train step=1534

                                                       
{'loss': -0.1403, 'grad_norm': 1.0674489736557007, 'learning_rate': 5.3545454545454546e-06, 'num_tokens': 3449061.0, 'completions/mean_length': 106.625, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.71428680419922, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8021861910820007, 'rewards/meter/std': 0.325016587972641, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.80141282081604, 'rewards/total_composite/std': 0.327181339263916, 'reward': 0.80141282081604, 'reward_std': 0.32718130946159363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038483839482069016, 'sampling/sampling_logp_difference/max': 2.473947525024414, 'sampling/importance_sampling_ratio/min': 0.08425161987543106, 'sampling/importance_sampling_ratio/mean': 1.0069955587387085, 'sampling/importance_sampling_ratio/max': 1.8218151330947876, 'entropy': 0.16696601919829845, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.023747874423861504, 'clip_ratio/high_max': 0.023747874423861504, 'clip_ratio/region_mean': 0.023747874423861504, 'reward_total_mean': 0.80141282081604, 'reward_meter_mean': 0.8021861910820007, 'reward_meter_std': 0.325016587972641, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.80141282081604, 'reward_total_composite_std': 0.327181339263916, 'epoch': 0.06}

 46%|████▋     | 1534/3300 [1:45:22<3:06:19,  6.33s/it]INFO 04-12 00:22:14 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1535/3300 [1:45:28<3:03:48,  6.25s/it]2026-04-12 00:22:19,911 | INFO | train_grpo_train | metrics_logged mode=train step=1535

                                                       
{'loss': -0.0071, 'grad_norm': 4.45563268661499, 'learning_rate': 5.351515151515152e-06, 'num_tokens': 3451653.0, 'completions/mean_length': 153.0, 'completions/min_length': 130.0, 'completions/max_length': 174.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 153.0, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 174.0, 'rewards/meter/mean': 0.7749242782592773, 'rewards/meter/std': 0.3511018455028534, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.07856741547584534, 'rewards/total_composite/mean': 0.6905776858329773, 'rewards/total_composite/std': 0.3377683758735657, 'reward': 0.6905776858329773, 'reward_std': 0.3377683758735657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05948156490921974, 'sampling/sampling_logp_difference/max': 1.3245697021484375, 'sampling/importance_sampling_ratio/min': 0.2659173607826233, 'sampling/importance_sampling_ratio/mean': 1.0109103918075562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5641567148268223, 'clip_ratio/low_mean': 0.018939394503831863, 'clip_ratio/low_min': 0.018939394503831863, 'clip_ratio/high_mean': 0.0292887045070529, 'clip_ratio/high_max': 0.0292887045070529, 'clip_ratio/region_mean': 0.04822809901088476, 'reward_total_mean': 0.6905776858329773, 'reward_meter_mean': 0.7749242782592773, 'reward_meter_std': 0.3511018455028534, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.07856741547584534, 'reward_total_composite_mean': 0.6905776858329773, 'reward_total_composite_std': 0.3377683758735657, 'epoch': 0.06}

 47%|████▋     | 1535/3300 [1:45:28<3:03:48,  6.25s/it]INFO 04-12 00:22:20 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1536/3300 [1:45:33<2:50:15,  5.79s/it]2026-04-12 00:22:24,637 | INFO | train_grpo_train | metrics_logged mode=train step=1536

                                                       
{'loss': -0.0035, 'grad_norm': 4.13173770904541, 'learning_rate': 5.348484848484848e-06, 'num_tokens': 3453309.0, 'completions/mean_length': 68.0, 'completions/min_length': 66.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.8998743295669556, 'rewards/meter/std': 0.1591634452342987, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8998743295669556, 'rewards/total_composite/std': 0.1591634452342987, 'reward': 0.8998743295669556, 'reward_std': 0.1591634303331375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05023188889026642, 'sampling/sampling_logp_difference/max': 1.0234405994415283, 'sampling/importance_sampling_ratio/min': 0.35935643315315247, 'sampling/importance_sampling_ratio/mean': 1.0069003105163574, 'sampling/importance_sampling_ratio/max': 1.8585302829742432, 'entropy': 0.4786563478410244, 'clip_ratio/low_mean': 0.016904115676879883, 'clip_ratio/low_min': 0.016904115676879883, 'clip_ratio/high_mean': 0.03292302007321268, 'clip_ratio/high_max': 0.03292302007321268, 'clip_ratio/region_mean': 0.049827135750092566, 'reward_total_mean': 0.8998743295669556, 'reward_meter_mean': 0.8998743295669556, 'reward_meter_std': 0.1591634452342987, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8998743295669556, 'reward_total_composite_std': 0.1591634452342987, 'epoch': 0.06}

 47%|████▋     | 1536/3300 [1:45:33<2:50:15,  5.79s/it]INFO 04-12 00:22:24 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1537/3300 [1:45:37<2:36:32,  5.33s/it]2026-04-12 00:22:28,885 | INFO | train_grpo_train | metrics_logged mode=train step=1537

                                                       
{'loss': 0.0252, 'grad_norm': 8.413439750671387, 'learning_rate': 5.3454545454545455e-06, 'num_tokens': 3454809.0, 'completions/mean_length': 30.5, 'completions/min_length': 29.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9885121583938599, 'rewards/meter/std': 0.006495221517980099, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9885121583938599, 'rewards/total_composite/std': 0.006495221517980099, 'reward': 0.9885121583938599, 'reward_std': 0.006495220120996237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03234010562300682, 'sampling/sampling_logp_difference/max': 0.9072532653808594, 'sampling/importance_sampling_ratio/min': 0.40363138914108276, 'sampling/importance_sampling_ratio/mean': 1.0021849870681763, 'sampling/importance_sampling_ratio/max': 1.6943132877349854, 'entropy': 0.229267543181777, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.01667593652382493, 'clip_ratio/high_max': 0.01667593652382493, 'clip_ratio/region_mean': 0.02058218652382493, 'reward_total_mean': 0.9885121583938599, 'reward_meter_mean': 0.9885121583938599, 'reward_meter_std': 0.006495221517980099, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9885121583938599, 'reward_total_composite_std': 0.006495221517980099, 'epoch': 0.06}

 47%|████▋     | 1537/3300 [1:45:37<2:36:32,  5.33s/it]INFO 04-12 00:22:29 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1538/3300 [1:45:42<2:30:36,  5.13s/it]2026-04-12 00:22:33,548 | INFO | train_grpo_train | metrics_logged mode=train step=1538

                                                       
{'loss': -0.0059, 'grad_norm': 3.6924545764923096, 'learning_rate': 5.342424242424244e-06, 'num_tokens': 3456577.0, 'completions/mean_length': 67.0, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9983944892883301, 'rewards/meter/std': 0.0004069809801876545, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9983944892883301, 'rewards/total_composite/std': 0.0004069809801876545, 'reward': 0.9983944892883301, 'reward_std': 0.00040697952499613166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0381365530192852, 'sampling/sampling_logp_difference/max': 1.0573501586914062, 'sampling/importance_sampling_ratio/min': 0.3473750650882721, 'sampling/importance_sampling_ratio/mean': 1.0129139423370361, 'sampling/importance_sampling_ratio/max': 1.5474638938903809, 'entropy': 0.3516472205519676, 'clip_ratio/low_mean': 0.007634902372956276, 'clip_ratio/low_min': 0.007634902372956276, 'clip_ratio/high_mean': 0.018546971026808023, 'clip_ratio/high_max': 0.018546971026808023, 'clip_ratio/region_mean': 0.0261818733997643, 'reward_total_mean': 0.9983944892883301, 'reward_meter_mean': 0.9983944892883301, 'reward_meter_std': 0.0004069809801876545, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9983944892883301, 'reward_total_composite_std': 0.0004069809801876545, 'epoch': 0.06}

 47%|████▋     | 1538/3300 [1:45:42<2:30:36,  5.13s/it]INFO 04-12 00:22:33 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1539/3300 [1:45:47<2:28:42,  5.07s/it]2026-04-12 00:22:38,476 | INFO | train_grpo_train | metrics_logged mode=train step=1539

                                                       
{'loss': -0.0048, 'grad_norm': 2.03717303276062, 'learning_rate': 5.33939393939394e-06, 'num_tokens': 3458618.0, 'completions/mean_length': 87.125, 'completions/min_length': 86.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.125, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9949196577072144, 'rewards/meter/std': 0.00045097857946529984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949196577072144, 'rewards/total_composite/std': 0.00045097857946529984, 'reward': 0.9949196577072144, 'reward_std': 0.00045097683323547244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0125720901414752, 'sampling/sampling_logp_difference/max': 1.0620005130767822, 'sampling/importance_sampling_ratio/min': 0.3457634150981903, 'sampling/importance_sampling_ratio/mean': 1.0028852224349976, 'sampling/importance_sampling_ratio/max': 1.4961625337600708, 'entropy': 0.08502828422933817, 'clip_ratio/low_mean': 0.0014367816038429737, 'clip_ratio/low_min': 0.0014367816038429737, 'clip_ratio/high_mean': 0.009995099389925599, 'clip_ratio/high_max': 0.009995099389925599, 'clip_ratio/region_mean': 0.011431880993768573, 'reward_total_mean': 0.9949196577072144, 'reward_meter_mean': 0.9949196577072144, 'reward_meter_std': 0.00045097857946529984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949196577072144, 'reward_total_composite_std': 0.00045097857946529984, 'epoch': 0.06}

 47%|████▋     | 1539/3300 [1:45:47<2:28:42,  5.07s/it]INFO 04-12 00:22:38 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1540/3300 [1:45:51<2:23:02,  4.88s/it]2026-04-12 00:22:42,898 | INFO | train_grpo_train | metrics_logged mode=train step=1540

                                                       
{'loss': -0.0382, 'grad_norm': 12.248023986816406, 'learning_rate': 5.336363636363637e-06, 'num_tokens': 3460405.0, 'completions/mean_length': 48.375, 'completions/min_length': 41.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8395511507987976, 'rewards/meter/std': 0.27874672412872314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8395511507987976, 'rewards/total_composite/std': 0.27874672412872314, 'reward': 0.8395511507987976, 'reward_std': 0.27874669432640076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04395786300301552, 'sampling/sampling_logp_difference/max': 1.4850921630859375, 'sampling/importance_sampling_ratio/min': 0.35930031538009644, 'sampling/importance_sampling_ratio/mean': 1.0040042400360107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2700903480872512, 'clip_ratio/low_mean': 0.009146341122686863, 'clip_ratio/low_min': 0.009146341122686863, 'clip_ratio/high_mean': 0.023029743460938334, 'clip_ratio/high_max': 0.023029743460938334, 'clip_ratio/region_mean': 0.0321760845836252, 'reward_total_mean': 0.8395511507987976, 'reward_meter_mean': 0.8395511507987976, 'reward_meter_std': 0.27874672412872314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8395511507987976, 'reward_total_composite_std': 0.27874672412872314, 'epoch': 0.06}

 47%|████▋     | 1540/3300 [1:45:51<2:23:02,  4.88s/it]INFO 04-12 00:22:43 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1541/3300 [1:45:56<2:21:00,  4.81s/it]2026-04-12 00:22:47,555 | INFO | train_grpo_train | metrics_logged mode=train step=1541

                                                       
{'loss': 0.0077, 'grad_norm': 7.992240905761719, 'learning_rate': 5.333333333333334e-06, 'num_tokens': 3462213.0, 'completions/mean_length': 60.0, 'completions/min_length': 58.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9441971778869629, 'rewards/meter/std': 0.13217569887638092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9441971778869629, 'rewards/total_composite/std': 0.13217569887638092, 'reward': 0.9441971778869629, 'reward_std': 0.13217571377754211, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06318628787994385, 'sampling/sampling_logp_difference/max': 1.5191354751586914, 'sampling/importance_sampling_ratio/min': 0.21890106797218323, 'sampling/importance_sampling_ratio/mean': 1.0145553350448608, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47805059514939785, 'clip_ratio/low_mean': 0.008196720853447914, 'clip_ratio/low_min': 0.008196720853447914, 'clip_ratio/high_mean': 0.0310292961075902, 'clip_ratio/high_max': 0.0310292961075902, 'clip_ratio/region_mean': 0.03922601696103811, 'reward_total_mean': 0.9441971778869629, 'reward_meter_mean': 0.9441971778869629, 'reward_meter_std': 0.13217569887638092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9441971778869629, 'reward_total_composite_std': 0.13217569887638092, 'epoch': 0.06}

 47%|████▋     | 1541/3300 [1:45:56<2:21:00,  4.81s/it]INFO 04-12 00:22:47 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1542/3300 [1:46:01<2:26:48,  5.01s/it]2026-04-12 00:22:53,039 | INFO | train_grpo_train | metrics_logged mode=train step=1542

                                                       
{'loss': 0.0082, 'grad_norm': 3.4401650428771973, 'learning_rate': 5.330303030303031e-06, 'num_tokens': 3464650.0, 'completions/mean_length': 127.625, 'completions/min_length': 124.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.625, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9891485571861267, 'rewards/meter/std': 0.004043546039611101, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285373687744, 'rewards/repeat_penalty/std': 0.10628911107778549, 'rewards/total_composite/mean': 0.9360376596450806, 'rewards/total_composite/std': 0.10430651903152466, 'reward': 0.9360376596450806, 'reward_std': 0.10430650413036346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07292276620864868, 'sampling/sampling_logp_difference/max': 1.6305036544799805, 'sampling/importance_sampling_ratio/min': 0.19583091139793396, 'sampling/importance_sampling_ratio/mean': 1.0177416801452637, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6336969211697578, 'clip_ratio/low_mean': 0.005937984678894281, 'clip_ratio/low_min': 0.005937984678894281, 'clip_ratio/high_mean': 0.04202882433310151, 'clip_ratio/high_max': 0.04202882433310151, 'clip_ratio/region_mean': 0.04796680901199579, 'reward_total_mean': 0.9360376596450806, 'reward_meter_mean': 0.9891485571861267, 'reward_meter_std': 0.004043546039611101, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285373687744, 'reward_repeat_penalty_std': 0.10628911107778549, 'reward_total_composite_mean': 0.9360376596450806, 'reward_total_composite_std': 0.10430651903152466, 'epoch': 0.06}

 47%|████▋     | 1542/3300 [1:46:02<2:26:48,  5.01s/it]INFO 04-12 00:22:53 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1543/3300 [1:46:07<2:27:58,  5.05s/it]2026-04-12 00:22:58,190 | INFO | train_grpo_train | metrics_logged mode=train step=1543

                                                       
{'loss': -0.0036, 'grad_norm': 5.007582664489746, 'learning_rate': 5.327272727272727e-06, 'num_tokens': 3466723.0, 'completions/mean_length': 96.125, 'completions/min_length': 93.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.125, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.7870810627937317, 'rewards/meter/std': 0.3985043168067932, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7870810627937317, 'rewards/total_composite/std': 0.3985043168067932, 'reward': 0.7870810627937317, 'reward_std': 0.3985043168067932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026287445798516273, 'sampling/sampling_logp_difference/max': 1.5934228897094727, 'sampling/importance_sampling_ratio/min': 0.20322878658771515, 'sampling/importance_sampling_ratio/mean': 0.9980475306510925, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08730749879032373, 'clip_ratio/low_mean': 0.00649872503709048, 'clip_ratio/low_min': 0.00649872503709048, 'clip_ratio/high_mean': 0.009101442527025938, 'clip_ratio/high_max': 0.009101442527025938, 'clip_ratio/region_mean': 0.015600167564116418, 'reward_total_mean': 0.7870810627937317, 'reward_meter_mean': 0.7870810627937317, 'reward_meter_std': 0.3985043168067932, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7870810627937317, 'reward_total_composite_std': 0.3985043168067932, 'epoch': 0.06}

 47%|████▋     | 1543/3300 [1:46:07<2:27:58,  5.05s/it]INFO 04-12 00:22:58 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1544/3300 [1:46:11<2:25:39,  4.98s/it]2026-04-12 00:23:02,988 | INFO | train_grpo_train | metrics_logged mode=train step=1544

                                                       
{'loss': -0.0189, 'grad_norm': 4.138071060180664, 'learning_rate': 5.324242424242425e-06, 'num_tokens': 3468547.0, 'completions/mean_length': 65.0, 'completions/min_length': 62.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9653145670890808, 'rewards/meter/std': 0.03330826759338379, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9653145670890808, 'rewards/total_composite/std': 0.03330826759338379, 'reward': 0.9653145670890808, 'reward_std': 0.03330826386809349, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025457728654146194, 'sampling/sampling_logp_difference/max': 0.9563794136047363, 'sampling/importance_sampling_ratio/min': 0.4773949682712555, 'sampling/importance_sampling_ratio/mean': 1.0086570978164673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14403609465807676, 'clip_ratio/low_mean': 0.007875503972172737, 'clip_ratio/low_min': 0.007875503972172737, 'clip_ratio/high_mean': 0.01856476883403957, 'clip_ratio/high_max': 0.01856476883403957, 'clip_ratio/region_mean': 0.026440272806212306, 'reward_total_mean': 0.9653145670890808, 'reward_meter_mean': 0.9653145670890808, 'reward_meter_std': 0.03330826759338379, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9653145670890808, 'reward_total_composite_std': 0.03330826759338379, 'epoch': 0.06}

 47%|████▋     | 1544/3300 [1:46:11<2:25:39,  4.98s/it]INFO 04-12 00:23:03 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1545/3300 [1:46:16<2:23:09,  4.89s/it]2026-04-12 00:23:07,690 | INFO | train_grpo_train | metrics_logged mode=train step=1545

                                                       
{'loss': -0.0078, 'grad_norm': 5.288153648376465, 'learning_rate': 5.321212121212122e-06, 'num_tokens': 3470357.0, 'completions/mean_length': 67.25, 'completions/min_length': 62.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9962025284767151, 'rewards/meter/std': 0.00486297020688653, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9962025284767151, 'rewards/total_composite/std': 0.00486297020688653, 'reward': 0.9962025284767151, 'reward_std': 0.004862963687628508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038305725902318954, 'sampling/sampling_logp_difference/max': 1.3244657516479492, 'sampling/importance_sampling_ratio/min': 0.2659450173377991, 'sampling/importance_sampling_ratio/mean': 1.0034021139144897, 'sampling/importance_sampling_ratio/max': 1.475104570388794, 'entropy': 0.33554743882268667, 'clip_ratio/low_mean': 0.01173020526766777, 'clip_ratio/low_min': 0.01173020526766777, 'clip_ratio/high_mean': 0.02582784171681851, 'clip_ratio/high_max': 0.02582784171681851, 'clip_ratio/region_mean': 0.03755804698448628, 'reward_total_mean': 0.9962025284767151, 'reward_meter_mean': 0.9962025284767151, 'reward_meter_std': 0.00486297020688653, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9962025284767151, 'reward_total_composite_std': 0.00486297020688653, 'epoch': 0.06}

 47%|████▋     | 1545/3300 [1:46:16<2:23:09,  4.89s/it]INFO 04-12 00:23:07 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1546/3300 [1:46:22<2:28:37,  5.08s/it]2026-04-12 00:23:13,218 | INFO | train_grpo_train | metrics_logged mode=train step=1546

                                                       
{'loss': -0.0044, 'grad_norm': 2.919037342071533, 'learning_rate': 5.318181818181819e-06, 'num_tokens': 3472779.0, 'completions/mean_length': 127.75, 'completions/min_length': 122.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.75, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9427036046981812, 'rewards/meter/std': 0.14980448782444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.6911537647247314, 'rewards/total_composite/std': 0.12467769533395767, 'reward': 0.6911537647247314, 'reward_std': 0.12467768788337708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022355277091264725, 'sampling/sampling_logp_difference/max': 1.9683294296264648, 'sampling/importance_sampling_ratio/min': 0.13969002664089203, 'sampling/importance_sampling_ratio/mean': 1.0017569065093994, 'sampling/importance_sampling_ratio/max': 1.6036888360977173, 'entropy': 0.14278420340269804, 'clip_ratio/low_mean': 0.005859375, 'clip_ratio/low_min': 0.005859375, 'clip_ratio/high_mean': 0.013565883273258805, 'clip_ratio/high_max': 0.013565883273258805, 'clip_ratio/region_mean': 0.019425258273258805, 'reward_total_mean': 0.6911537647247314, 'reward_meter_mean': 0.9427036046981812, 'reward_meter_std': 0.14980448782444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.6911537647247314, 'reward_total_composite_std': 0.12467769533395767, 'epoch': 0.06}

 47%|████▋     | 1546/3300 [1:46:22<2:28:37,  5.08s/it]INFO 04-12 00:23:13 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1547/3300 [1:46:26<2:25:17,  4.97s/it]2026-04-12 00:23:17,933 | INFO | train_grpo_train | metrics_logged mode=train step=1547

                                                       
{'loss': -0.0066, 'grad_norm': 6.456394195556641, 'learning_rate': 5.3151515151515155e-06, 'num_tokens': 3474543.0, 'completions/mean_length': 60.5, 'completions/min_length': 57.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.796026885509491, 'rewards/meter/std': 0.2907305955886841, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6942555904388428, 'rewards/total_composite/std': 0.4039342701435089, 'reward': 0.6942555904388428, 'reward_std': 0.4039342701435089, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07712767273187637, 'sampling/sampling_logp_difference/max': 1.3398246765136719, 'sampling/importance_sampling_ratio/min': 0.26189157366752625, 'sampling/importance_sampling_ratio/mean': 1.0131072998046875, 'sampling/importance_sampling_ratio/max': 1.528890609741211, 'entropy': 0.7294313032180071, 'clip_ratio/low_mean': 0.015021929983049631, 'clip_ratio/low_min': 0.015021929983049631, 'clip_ratio/high_mean': 0.04056679271161556, 'clip_ratio/high_max': 0.04056679271161556, 'clip_ratio/region_mean': 0.055588722694665194, 'reward_total_mean': 0.6942555904388428, 'reward_meter_mean': 0.796026885509491, 'reward_meter_std': 0.2907305955886841, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6942555904388428, 'reward_total_composite_std': 0.4039342701435089, 'epoch': 0.06}

 47%|████▋     | 1547/3300 [1:46:26<2:25:17,  4.97s/it]INFO 04-12 00:23:18 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1548/3300 [1:46:31<2:20:10,  4.80s/it]2026-04-12 00:23:22,341 | INFO | train_grpo_train | metrics_logged mode=train step=1548

                                                       
{'loss': 0.0213, 'grad_norm': 11.082610130310059, 'learning_rate': 5.312121212121213e-06, 'num_tokens': 3476009.0, 'completions/mean_length': 34.25, 'completions/min_length': 32.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.8179029822349548, 'rewards/meter/std': 0.24513599276542664, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8179029822349548, 'rewards/total_composite/std': 0.24513599276542664, 'reward': 0.8179029822349548, 'reward_std': 0.24513597786426544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07286211103200912, 'sampling/sampling_logp_difference/max': 1.024322509765625, 'sampling/importance_sampling_ratio/min': 0.35903966426849365, 'sampling/importance_sampling_ratio/mean': 1.0309749841690063, 'sampling/importance_sampling_ratio/max': 1.7788945436477661, 'entropy': 0.7250464260578156, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.03996537090279162, 'clip_ratio/high_max': 0.03996537090279162, 'clip_ratio/region_mean': 0.047541128704324365, 'reward_total_mean': 0.8179029822349548, 'reward_meter_mean': 0.8179029822349548, 'reward_meter_std': 0.24513599276542664, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8179029822349548, 'reward_total_composite_std': 0.24513599276542664, 'epoch': 0.06}

 47%|████▋     | 1548/3300 [1:46:31<2:20:10,  4.80s/it]INFO 04-12 00:23:22 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1549/3300 [1:46:35<2:15:01,  4.63s/it]2026-04-12 00:23:26,561 | INFO | train_grpo_train | metrics_logged mode=train step=1549

                                                       
{'loss': -0.0072, 'grad_norm': 8.7092924118042, 'learning_rate': 5.309090909090909e-06, 'num_tokens': 3477433.0, 'completions/mean_length': 30.0, 'completions/min_length': 30.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9894814491271973, 'rewards/meter/std': 0.00789269246160984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9894814491271973, 'rewards/total_composite/std': 0.00789269246160984, 'reward': 0.9894814491271973, 'reward_std': 0.007892684079706669, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014411688782274723, 'sampling/sampling_logp_difference/max': 1.21435546875, 'sampling/importance_sampling_ratio/min': 0.29690131545066833, 'sampling/importance_sampling_ratio/mean': 0.9964293241500854, 'sampling/importance_sampling_ratio/max': 1.0616073608398438, 'entropy': 0.06372858118265867, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9894814491271973, 'reward_meter_mean': 0.9894814491271973, 'reward_meter_std': 0.00789269246160984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9894814491271973, 'reward_total_composite_std': 0.00789269246160984, 'epoch': 0.06}

 47%|████▋     | 1549/3300 [1:46:35<2:15:01,  4.63s/it]INFO 04-12 00:23:26 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1550/3300 [1:46:40<2:15:13,  4.64s/it]2026-04-12 00:23:31,213 | INFO | train_grpo_train | metrics_logged mode=train step=1550

                                                       
{'loss': 0.0078, 'grad_norm': 4.305907249450684, 'learning_rate': 5.306060606060606e-06, 'num_tokens': 3479115.0, 'completions/mean_length': 58.25, 'completions/min_length': 56.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9904994964599609, 'rewards/meter/std': 0.009607957676053047, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.17251639068126678, 'rewards/total_composite/mean': 0.7832661867141724, 'rewards/total_composite/std': 0.1659567952156067, 'reward': 0.7832661867141724, 'reward_std': 0.1659567803144455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01842333748936653, 'sampling/sampling_logp_difference/max': 0.9088270664215088, 'sampling/importance_sampling_ratio/min': 0.40299662947654724, 'sampling/importance_sampling_ratio/mean': 1.000774621963501, 'sampling/importance_sampling_ratio/max': 1.4585957527160645, 'entropy': 0.09011476766318083, 'clip_ratio/low_mean': 0.004273816477507353, 'clip_ratio/low_min': 0.004273816477507353, 'clip_ratio/high_mean': 0.010820218129083514, 'clip_ratio/high_max': 0.010820218129083514, 'clip_ratio/region_mean': 0.015094034606590867, 'reward_total_mean': 0.7832661867141724, 'reward_meter_mean': 0.9904994964599609, 'reward_meter_std': 0.009607957676053047, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.17251639068126678, 'reward_total_composite_mean': 0.7832661867141724, 'reward_total_composite_std': 0.1659567952156067, 'epoch': 0.06}

 47%|████▋     | 1550/3300 [1:46:40<2:15:13,  4.64s/it]INFO 04-12 00:23:31 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:26,  2.39s/it]

 23%|██▎       | 3/13 [00:08<00:28,  2.82s/it]

 31%|███       | 4/13 [00:12<00:31,  3.51s/it]

 38%|███▊      | 5/13 [00:16<00:29,  3.67s/it]

 46%|████▌     | 6/13 [00:21<00:27,  3.97s/it]

 54%|█████▍    | 7/13 [00:25<00:23,  3.99s/it]

 62%|██████▏   | 8/13 [00:30<00:22,  4.45s/it]

 69%|██████▉   | 9/13 [00:36<00:18,  4.67s/it]

 77%|███████▋  | 10/13 [00:40<00:14,  4.69s/it]

 85%|████████▍ | 11/13 [00:45<00:09,  4.56s/it]

 92%|█████████▏| 12/13 [00:49<00:04,  4.62s/it]

100%|██████████| 13/13 [00:54<00:00,  4.53s/it]2026-04-12 00:24:30,693 | INFO | train_grpo_train | metrics_logged mode=eval step=1550

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 59.4299, 'eval_samples_per_second': 1.75, 'eval_steps_per_second': 0.219, 'eval_num_tokens': 3479115.0, 'eval_completions/mean_length': 174.25, 'eval_completions/min_length': 62.30769230769231, 'eval_completions/max_length': 299.7692307692308, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 174.25, 'eval_completions/min_terminated_length': 62.30769230769231, 'eval_completions/max_terminated_length': 299.7692307692308, 'eval_rewards/meter/mean': 0.7260088416246268, 'eval_rewards/meter/std': 0.3893573456085645, 'eval_rewards/count_adherence/mean': 0.8626743646768423, 'eval_rewards/count_adherence/std': 0.15295347571372986, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.8659125016285822, 'eval_rewards/repeat_penalty/std': 0.14054490625858307, 'eval_rewards/total_composite/mean': 0.5367342164883246, 'eval_rewards/total_composite/std': 0.3409125472490604, 'eval_reward': 0.5367342164883246, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.032458307078251473, 'eval_sampling/sampling_logp_difference/max': 1.1005325317382812, 'eval_sampling/importance_sampling_ratio/min': 0.34075071719976574, 'eval_sampling/importance_sampling_ratio/mean': 1.0097044522945697, 'eval_sampling/importance_sampling_ratio/max': 1.5303159401966975, 'eval_entropy': 0.3527725820357983, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5367342164883246, 'eval_reward_meter_mean': 0.7260088416246268, 'eval_reward_meter_std': 0.3893573456085645, 'eval_reward_count_adherence_mean': 0.8626743646768423, 'eval_reward_count_adherence_std': 0.15295347571372986, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.8659125016285822, 'eval_reward_repeat_penalty_std': 0.14054490625858307, 'eval_reward_total_composite_mean': 0.5367342164883246, 'eval_reward_total_composite_std': 0.3409125472490604, 'epoch': 0.06}

 47%|████▋     | 1550/3300 [1:47:39<2:15:13,  4.64s/it]

100%|██████████| 13/13 [00:54<00:00,  4.53s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 47%|████▋     | 1551/3300 [1:47:48<11:28:23, 23.62s/it]2026-04-12 00:24:39,113 | INFO | train_grpo_train | metrics_logged mode=train step=1551

                                                        
{'loss': 0.0162, 'grad_norm': 1.714144229888916, 'learning_rate': 5.303030303030303e-06, 'num_tokens': 3481876.0, 'completions/mean_length': 168.125, 'completions/min_length': 159.0, 'completions/max_length': 174.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.125, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 174.0, 'rewards/meter/mean': 0.9984342455863953, 'rewards/meter/std': 0.00023047976719681174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.0927247703075409, 'rewards/total_composite/mean': 0.7904214859008789, 'rewards/total_composite/std': 0.09253095835447311, 'reward': 0.7904214859008789, 'reward_std': 0.09253095090389252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02464020438492298, 'sampling/sampling_logp_difference/max': 2.094449281692505, 'sampling/importance_sampling_ratio/min': 0.12313804030418396, 'sampling/importance_sampling_ratio/mean': 1.002803921699524, 'sampling/importance_sampling_ratio/max': 1.999273419380188, 'entropy': 0.14000887237489223, 'clip_ratio/low_mean': 0.006647313362918794, 'clip_ratio/low_min': 0.006647313362918794, 'clip_ratio/high_mean': 0.012758038472384214, 'clip_ratio/high_max': 0.012758038472384214, 'clip_ratio/region_mean': 0.01940535183530301, 'reward_total_mean': 0.7904214859008789, 'reward_meter_mean': 0.9984342455863953, 'reward_meter_std': 0.00023047976719681174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.0927247703075409, 'reward_total_composite_mean': 0.7904214859008789, 'reward_total_composite_std': 0.09253095835447311, 'epoch': 0.06}

 47%|████▋     | 1551/3300 [1:47:48<11:28:23, 23.62s/it]INFO 04-12 00:24:39 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1552/3300 [1:47:54<9:00:41, 18.56s/it] 2026-04-12 00:24:45,876 | INFO | train_grpo_train | metrics_logged mode=train step=1552

                                                       
{'loss': -0.01, 'grad_norm': 2.4818339347839355, 'learning_rate': 5.300000000000001e-06, 'num_tokens': 3485227.0, 'completions/mean_length': 212.875, 'completions/min_length': 197.0, 'completions/max_length': 223.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 212.875, 'completions/min_terminated_length': 197.0, 'completions/max_terminated_length': 223.0, 'rewards/meter/mean': 0.9701088666915894, 'rewards/meter/std': 0.02329552173614502, 'rewards/count_adherence/mean': 0.828125, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8910256624221802, 'rewards/repeat_penalty/std': 0.1113983765244484, 'rewards/total_composite/mean': 0.7158641219139099, 'rewards/total_composite/std': 0.10737311840057373, 'reward': 0.7158641219139099, 'reward_std': 0.10737311094999313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044115372002124786, 'sampling/sampling_logp_difference/max': 1.539175033569336, 'sampling/importance_sampling_ratio/min': 0.21455803513526917, 'sampling/importance_sampling_ratio/mean': 1.0085928440093994, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4306017607450485, 'clip_ratio/low_mean': 0.009651749860495329, 'clip_ratio/low_min': 0.009651749860495329, 'clip_ratio/high_mean': 0.026524552376940846, 'clip_ratio/high_max': 0.026524552376940846, 'clip_ratio/region_mean': 0.036176302237436175, 'reward_total_mean': 0.7158641219139099, 'reward_meter_mean': 0.9701088666915894, 'reward_meter_std': 0.02329552173614502, 'reward_count_adherence_mean': 0.828125, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8910256624221802, 'reward_repeat_penalty_std': 0.1113983765244484, 'reward_total_composite_mean': 0.7158641219139099, 'reward_total_composite_std': 0.10737311840057373, 'epoch': 0.06}

 47%|████▋     | 1552/3300 [1:47:54<9:00:41, 18.56s/it]INFO 04-12 00:24:46 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1553/3300 [1:47:59<7:01:51, 14.49s/it]2026-04-12 00:24:50,864 | INFO | train_grpo_train | metrics_logged mode=train step=1553

                                                       
{'loss': -0.0033, 'grad_norm': 7.115604400634766, 'learning_rate': 5.296969696969697e-06, 'num_tokens': 3487126.0, 'completions/mean_length': 65.375, 'completions/min_length': 63.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.375, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.977942705154419, 'rewards/meter/std': 0.007237757556140423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.977942705154419, 'rewards/total_composite/std': 0.007237757556140423, 'reward': 0.977942705154419, 'reward_std': 0.007237738464027643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03986657038331032, 'sampling/sampling_logp_difference/max': 0.8986415863037109, 'sampling/importance_sampling_ratio/min': 0.40712234377861023, 'sampling/importance_sampling_ratio/mean': 1.0108517408370972, 'sampling/importance_sampling_ratio/max': 1.691331386566162, 'entropy': 0.40439430996775627, 'clip_ratio/low_mean': 0.0174350953893736, 'clip_ratio/low_min': 0.0174350953893736, 'clip_ratio/high_mean': 0.022814685944467783, 'clip_ratio/high_max': 0.022814685944467783, 'clip_ratio/region_mean': 0.04024978133384138, 'reward_total_mean': 0.977942705154419, 'reward_meter_mean': 0.977942705154419, 'reward_meter_std': 0.007237757556140423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.977942705154419, 'reward_total_composite_std': 0.007237757556140423, 'epoch': 0.06}

 47%|████▋     | 1553/3300 [1:47:59<7:01:51, 14.49s/it]INFO 04-12 00:24:51 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1554/3300 [1:48:07<5:58:17, 12.31s/it]2026-04-12 00:24:58,094 | INFO | train_grpo_train | metrics_logged mode=train step=1554

                                                       
{'loss': -0.0326, 'grad_norm': 2.9659626483917236, 'learning_rate': 5.293939393939395e-06, 'num_tokens': 3490583.0, 'completions/mean_length': 229.125, 'completions/min_length': 206.0, 'completions/max_length': 240.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 229.125, 'completions/min_terminated_length': 206.0, 'completions/max_terminated_length': 240.0, 'rewards/meter/mean': 0.9962524771690369, 'rewards/meter/std': 0.002178115537390113, 'rewards/count_adherence/mean': 0.859375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8135302066802979, 'rewards/repeat_penalty/std': 0.05918560549616814, 'rewards/total_composite/mean': 0.6974078416824341, 'rewards/total_composite/std': 0.07142963260412216, 'reward': 0.6974078416824341, 'reward_std': 0.07142962515354156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013559546321630478, 'sampling/sampling_logp_difference/max': 2.815333366394043, 'sampling/importance_sampling_ratio/min': 0.059884749352931976, 'sampling/importance_sampling_ratio/mean': 1.000474452972412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0460505741648376, 'clip_ratio/low_mean': 0.0021819902467541397, 'clip_ratio/low_min': 0.0021819902467541397, 'clip_ratio/high_mean': 0.004815331194549799, 'clip_ratio/high_max': 0.004815331194549799, 'clip_ratio/region_mean': 0.006997321441303939, 'reward_total_mean': 0.6974078416824341, 'reward_meter_mean': 0.9962524771690369, 'reward_meter_std': 0.002178115537390113, 'reward_count_adherence_mean': 0.859375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8135302066802979, 'reward_repeat_penalty_std': 0.05918560549616814, 'reward_total_composite_mean': 0.6974078416824341, 'reward_total_composite_std': 0.07142963260412216, 'epoch': 0.06}

 47%|████▋     | 1554/3300 [1:48:07<5:58:17, 12.31s/it]INFO 04-12 00:24:58 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1555/3300 [1:48:13<5:06:07, 10.53s/it]2026-04-12 00:25:04,452 | INFO | train_grpo_train | metrics_logged mode=train step=1555

                                                       
{'loss': 0.0094, 'grad_norm': 4.632443428039551, 'learning_rate': 5.290909090909091e-06, 'num_tokens': 3493632.0, 'completions/mean_length': 193.125, 'completions/min_length': 188.0, 'completions/max_length': 204.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 193.125, 'completions/min_terminated_length': 188.0, 'completions/max_terminated_length': 204.0, 'rewards/meter/mean': 0.9984503984451294, 'rewards/meter/std': 0.00041957091889344156, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8125, 'rewards/repeat_penalty/std': 0.08256731927394867, 'rewards/total_composite/mean': 0.6953567266464233, 'rewards/total_composite/std': 0.0707344338297844, 'reward': 0.6953567266464233, 'reward_std': 0.0707344338297844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015556903555989265, 'sampling/sampling_logp_difference/max': 1.6991368532180786, 'sampling/importance_sampling_ratio/min': 0.18284127116203308, 'sampling/importance_sampling_ratio/mean': 0.9994859099388123, 'sampling/importance_sampling_ratio/max': 1.899730920791626, 'entropy': 0.05383288115262985, 'clip_ratio/low_mean': 0.0032157512614503503, 'clip_ratio/low_min': 0.0032157512614503503, 'clip_ratio/high_mean': 0.00838999031111598, 'clip_ratio/high_max': 0.00838999031111598, 'clip_ratio/region_mean': 0.01160574157256633, 'reward_total_mean': 0.6953567266464233, 'reward_meter_mean': 0.9984503984451294, 'reward_meter_std': 0.00041957091889344156, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8125, 'reward_repeat_penalty_std': 0.08256731927394867, 'reward_total_composite_mean': 0.6953567266464233, 'reward_total_composite_std': 0.0707344338297844, 'epoch': 0.06}

 47%|████▋     | 1555/3300 [1:48:13<5:06:07, 10.53s/it]INFO 04-12 00:25:04 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1556/3300 [1:48:18<4:18:50,  8.91s/it]2026-04-12 00:25:09,575 | INFO | train_grpo_train | metrics_logged mode=train step=1556

                                                       
{'loss': 0.0201, 'grad_norm': 3.2562649250030518, 'learning_rate': 5.287878787878788e-06, 'num_tokens': 3495501.0, 'completions/mean_length': 71.625, 'completions/min_length': 69.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9960213303565979, 'rewards/meter/std': 0.0024177224840968847, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9960213303565979, 'rewards/total_composite/std': 0.0024177224840968847, 'reward': 0.9960213303565979, 'reward_std': 0.0024177224840968847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02655681222677231, 'sampling/sampling_logp_difference/max': 1.301830768585205, 'sampling/importance_sampling_ratio/min': 0.27203330397605896, 'sampling/importance_sampling_ratio/mean': 1.0080082416534424, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11559192929416895, 'clip_ratio/low_mean': 0.010138889076188207, 'clip_ratio/low_min': 0.010138889076188207, 'clip_ratio/high_mean': 0.014115164172835648, 'clip_ratio/high_max': 0.014115164172835648, 'clip_ratio/region_mean': 0.024254053249023855, 'reward_total_mean': 0.9960213303565979, 'reward_meter_mean': 0.9960213303565979, 'reward_meter_std': 0.0024177224840968847, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9960213303565979, 'reward_total_composite_std': 0.0024177224840968847, 'epoch': 0.06}

 47%|████▋     | 1556/3300 [1:48:18<4:18:50,  8.91s/it]INFO 04-12 00:25:09 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1557/3300 [1:48:23<3:40:54,  7.60s/it]2026-04-12 00:25:14,145 | INFO | train_grpo_train | metrics_logged mode=train step=1557

                                                       
{'loss': -0.0016, 'grad_norm': 1.289771556854248, 'learning_rate': 5.284848484848485e-06, 'num_tokens': 3497301.0, 'completions/mean_length': 68.0, 'completions/min_length': 68.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9974825382232666, 'rewards/meter/std': 0.0009680635412223637, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974825382232666, 'rewards/total_composite/std': 0.0009680635412223637, 'reward': 0.9974825382232666, 'reward_std': 0.000968057313002646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007437488529831171, 'sampling/sampling_logp_difference/max': 0.3382420539855957, 'sampling/importance_sampling_ratio/min': 0.7130227088928223, 'sampling/importance_sampling_ratio/mean': 1.0034716129302979, 'sampling/importance_sampling_ratio/max': 1.1250584125518799, 'entropy': 0.06252996809780598, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.0036764706019312143, 'reward_total_mean': 0.9974825382232666, 'reward_meter_mean': 0.9974825382232666, 'reward_meter_std': 0.0009680635412223637, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974825382232666, 'reward_total_composite_std': 0.0009680635412223637, 'epoch': 0.06}

 47%|████▋     | 1557/3300 [1:48:23<3:40:54,  7.60s/it]INFO 04-12 00:25:14 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1558/3300 [1:48:27<3:11:28,  6.59s/it]2026-04-12 00:25:18,382 | INFO | train_grpo_train | metrics_logged mode=train step=1558

                                                       
{'loss': 0.0063, 'grad_norm': 1.997252345085144, 'learning_rate': 5.281818181818183e-06, 'num_tokens': 3498783.0, 'completions/mean_length': 35.25, 'completions/min_length': 35.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9977306127548218, 'rewards/meter/std': 4.97752262162976e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977306127548218, 'rewards/total_composite/std': 4.97752262162976e-05, 'reward': 0.9977306127548218, 'reward_std': 4.97752262162976e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00860452838242054, 'sampling/sampling_logp_difference/max': 0.33144044876098633, 'sampling/importance_sampling_ratio/min': 0.7421995401382446, 'sampling/importance_sampling_ratio/mean': 1.0048854351043701, 'sampling/importance_sampling_ratio/max': 1.3929730653762817, 'entropy': 0.06664726650342345, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010714285774156451, 'clip_ratio/high_max': 0.010714285774156451, 'clip_ratio/region_mean': 0.010714285774156451, 'reward_total_mean': 0.9977306127548218, 'reward_meter_mean': 0.9977306127548218, 'reward_meter_std': 4.97752262162976e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977306127548218, 'reward_total_composite_std': 4.97752262162976e-05, 'epoch': 0.06}

 47%|████▋     | 1558/3300 [1:48:27<3:11:28,  6.59s/it]INFO 04-12 00:25:18 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1559/3300 [1:48:32<3:00:39,  6.23s/it]2026-04-12 00:25:23,774 | INFO | train_grpo_train | metrics_logged mode=train step=1559

                                                       
{'loss': -0.0004, 'grad_norm': 2.993905782699585, 'learning_rate': 5.278787878787879e-06, 'num_tokens': 3500559.0, 'completions/mean_length': 62.0, 'completions/min_length': 60.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9898895025253296, 'rewards/meter/std': 0.006041497457772493, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9484682083129883, 'rewards/total_composite/std': 0.11559666693210602, 'reward': 0.9484682083129883, 'reward_std': 0.11559665948152542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0489659458398819, 'sampling/sampling_logp_difference/max': 0.9915828704833984, 'sampling/importance_sampling_ratio/min': 0.37098902463912964, 'sampling/importance_sampling_ratio/mean': 1.0192768573760986, 'sampling/importance_sampling_ratio/max': 1.9443280696868896, 'entropy': 0.4067396577447653, 'clip_ratio/low_mean': 0.002016128972172737, 'clip_ratio/low_min': 0.002016128972172737, 'clip_ratio/high_mean': 0.0341619870159775, 'clip_ratio/high_max': 0.0341619870159775, 'clip_ratio/region_mean': 0.03617811598815024, 'reward_total_mean': 0.9484682083129883, 'reward_meter_mean': 0.9898895025253296, 'reward_meter_std': 0.006041497457772493, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9484682083129883, 'reward_total_composite_std': 0.11559666693210602, 'epoch': 0.06}

 47%|████▋     | 1559/3300 [1:48:32<3:00:39,  6.23s/it]INFO 04-12 00:25:24 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1560/3300 [1:48:38<2:58:27,  6.15s/it]2026-04-12 00:25:29,734 | INFO | train_grpo_train | metrics_logged mode=train step=1560

                                                       
{'loss': 0.0004, 'grad_norm': 3.2149598598480225, 'learning_rate': 5.2757575757575764e-06, 'num_tokens': 3503371.0, 'completions/mean_length': 164.5, 'completions/min_length': 155.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 164.5, 'completions/min_terminated_length': 155.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.9614261388778687, 'rewards/meter/std': 0.03620919957756996, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.9209266901016235, 'rewards/total_composite/std': 0.058246713131666183, 'reward': 0.9209266901016235, 'reward_std': 0.05824670195579529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06021958962082863, 'sampling/sampling_logp_difference/max': 1.890352487564087, 'sampling/importance_sampling_ratio/min': 0.15101857483386993, 'sampling/importance_sampling_ratio/mean': 1.010610580444336, 'sampling/importance_sampling_ratio/max': 1.8964616060256958, 'entropy': 0.6054803691804409, 'clip_ratio/low_mean': 0.020772642455995083, 'clip_ratio/low_min': 0.020772642455995083, 'clip_ratio/high_mean': 0.030295601580291986, 'clip_ratio/high_max': 0.030295601580291986, 'clip_ratio/region_mean': 0.05106824403628707, 'reward_total_mean': 0.9209266901016235, 'reward_meter_mean': 0.9614261388778687, 'reward_meter_std': 0.03620919957756996, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.9209266901016235, 'reward_total_composite_std': 0.058246713131666183, 'epoch': 0.06}

 47%|████▋     | 1560/3300 [1:48:38<2:58:27,  6.15s/it]INFO 04-12 00:25:29 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1561/3300 [1:48:44<2:56:02,  6.07s/it]2026-04-12 00:25:35,623 | INFO | train_grpo_train | metrics_logged mode=train step=1561

                                                       
{'loss': -0.0154, 'grad_norm': 3.249744176864624, 'learning_rate': 5.272727272727273e-06, 'num_tokens': 3506277.0, 'completions/mean_length': 172.25, 'completions/min_length': 161.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 172.25, 'completions/min_terminated_length': 161.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.9807374477386475, 'rewards/meter/std': 0.04032743349671364, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9305555820465088, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.9116497039794922, 'rewards/total_composite/std': 0.07851860672235489, 'reward': 0.9116497039794922, 'reward_std': 0.07851860672235489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06819657236337662, 'sampling/sampling_logp_difference/max': 2.131275177001953, 'sampling/importance_sampling_ratio/min': 0.11868584156036377, 'sampling/importance_sampling_ratio/mean': 1.0178204774856567, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5600994266569614, 'clip_ratio/low_mean': 0.026302700862288475, 'clip_ratio/low_min': 0.026302700862288475, 'clip_ratio/high_mean': 0.01949126087129116, 'clip_ratio/high_max': 0.01949126087129116, 'clip_ratio/region_mean': 0.045793961733579636, 'reward_total_mean': 0.9116497039794922, 'reward_meter_mean': 0.9807374477386475, 'reward_meter_std': 0.04032743349671364, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9305555820465088, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.9116497039794922, 'reward_total_composite_std': 0.07851860672235489, 'epoch': 0.06}

 47%|████▋     | 1561/3300 [1:48:44<2:56:02,  6.07s/it]INFO 04-12 00:25:35 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1562/3300 [1:48:49<2:44:42,  5.69s/it]2026-04-12 00:25:40,414 | INFO | train_grpo_train | metrics_logged mode=train step=1562

                                                       
{'loss': 0.0416, 'grad_norm': 7.787070274353027, 'learning_rate': 5.26969696969697e-06, 'num_tokens': 3507984.0, 'completions/mean_length': 66.375, 'completions/min_length': 63.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9753642678260803, 'rewards/meter/std': 0.012682100757956505, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9753642678260803, 'rewards/total_composite/std': 0.012682100757956505, 'reward': 0.9753642678260803, 'reward_std': 0.012682083994150162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05309029668569565, 'sampling/sampling_logp_difference/max': 0.7393133640289307, 'sampling/importance_sampling_ratio/min': 0.4774416387081146, 'sampling/importance_sampling_ratio/mean': 1.020971417427063, 'sampling/importance_sampling_ratio/max': 1.8774487972259521, 'entropy': 0.4796396978199482, 'clip_ratio/low_mean': 0.016098485328257084, 'clip_ratio/low_min': 0.016098485328257084, 'clip_ratio/high_mean': 0.017139121424406767, 'clip_ratio/high_max': 0.017139121424406767, 'clip_ratio/region_mean': 0.03323760675266385, 'reward_total_mean': 0.9753642678260803, 'reward_meter_mean': 0.9753642678260803, 'reward_meter_std': 0.012682100757956505, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9753642678260803, 'reward_total_composite_std': 0.012682100757956505, 'epoch': 0.06}

 47%|████▋     | 1562/3300 [1:48:49<2:44:42,  5.69s/it]INFO 04-12 00:25:40 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1563/3300 [1:48:54<2:43:56,  5.66s/it]2026-04-12 00:25:46,033 | INFO | train_grpo_train | metrics_logged mode=train step=1563

                                                       
{'loss': -0.0149, 'grad_norm': 2.302091121673584, 'learning_rate': 5.2666666666666665e-06, 'num_tokens': 3510443.0, 'completions/mean_length': 137.375, 'completions/min_length': 133.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.375, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9985888004302979, 'rewards/meter/std': 0.0002428983716527, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7321428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.7311152815818787, 'rewards/total_composite/std': 0.05054369568824768, 'reward': 0.7311152815818787, 'reward_std': 0.05054369196295738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01936904713511467, 'sampling/sampling_logp_difference/max': 0.9639506340026855, 'sampling/importance_sampling_ratio/min': 0.3813832104206085, 'sampling/importance_sampling_ratio/mean': 1.0065100193023682, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15207071043550968, 'clip_ratio/low_mean': 0.012661085231229663, 'clip_ratio/low_min': 0.012661085231229663, 'clip_ratio/high_mean': 0.0008680555620230734, 'clip_ratio/high_max': 0.0008680555620230734, 'clip_ratio/region_mean': 0.013529140793252736, 'reward_total_mean': 0.7311152815818787, 'reward_meter_mean': 0.9985888004302979, 'reward_meter_std': 0.0002428983716527, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7321428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.7311152815818787, 'reward_total_composite_std': 0.05054369568824768, 'epoch': 0.06}

 47%|████▋     | 1563/3300 [1:48:55<2:43:56,  5.66s/it]INFO 04-12 00:25:46 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1564/3300 [1:49:00<2:43:10,  5.64s/it]2026-04-12 00:25:51,606 | INFO | train_grpo_train | metrics_logged mode=train step=1564

                                                       
{'loss': -0.0126, 'grad_norm': 5.161999702453613, 'learning_rate': 5.263636363636364e-06, 'num_tokens': 3512785.0, 'completions/mean_length': 131.75, 'completions/min_length': 119.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.75, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.718358039855957, 'rewards/meter/std': 0.32531455159187317, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.692800760269165, 'rewards/total_composite/std': 0.32817986607551575, 'reward': 0.692800760269165, 'reward_std': 0.32817986607551575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07563211023807526, 'sampling/sampling_logp_difference/max': 1.0854854583740234, 'sampling/importance_sampling_ratio/min': 0.3377377986907959, 'sampling/importance_sampling_ratio/mean': 1.01358962059021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7511504292488098, 'clip_ratio/low_mean': 0.02909582480788231, 'clip_ratio/low_min': 0.02909582480788231, 'clip_ratio/high_mean': 0.033232659101486206, 'clip_ratio/high_max': 0.033232659101486206, 'clip_ratio/region_mean': 0.062328483909368515, 'reward_total_mean': 0.692800760269165, 'reward_meter_mean': 0.718358039855957, 'reward_meter_std': 0.32531455159187317, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.692800760269165, 'reward_total_composite_std': 0.32817986607551575, 'epoch': 0.06}

 47%|████▋     | 1564/3300 [1:49:00<2:43:10,  5.64s/it]INFO 04-12 00:25:51 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1565/3300 [1:49:05<2:35:19,  5.37s/it]2026-04-12 00:25:56,364 | INFO | train_grpo_train | metrics_logged mode=train step=1565

                                                       
{'loss': -0.0189, 'grad_norm': 6.407342433929443, 'learning_rate': 5.26060606060606e-06, 'num_tokens': 3514505.0, 'completions/mean_length': 66.0, 'completions/min_length': 58.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6359885931015015, 'rewards/meter/std': 0.39930614829063416, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6359885931015015, 'rewards/total_composite/std': 0.39930614829063416, 'reward': 0.6359885931015015, 'reward_std': 0.39930611848831177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07870090752840042, 'sampling/sampling_logp_difference/max': 1.5872907638549805, 'sampling/importance_sampling_ratio/min': 0.20447883009910583, 'sampling/importance_sampling_ratio/mean': 1.0197231769561768, 'sampling/importance_sampling_ratio/max': 1.9221627712249756, 'entropy': 0.7624497786164284, 'clip_ratio/low_mean': 0.025564054027199745, 'clip_ratio/low_min': 0.025564054027199745, 'clip_ratio/high_mean': 0.041696065571159124, 'clip_ratio/high_max': 0.041696065571159124, 'clip_ratio/region_mean': 0.06726011959835887, 'reward_total_mean': 0.6359885931015015, 'reward_meter_mean': 0.6359885931015015, 'reward_meter_std': 0.39930614829063416, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6359885931015015, 'reward_total_composite_std': 0.39930614829063416, 'epoch': 0.06}

 47%|████▋     | 1565/3300 [1:49:05<2:35:19,  5.37s/it]INFO 04-12 00:25:56 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1566/3300 [1:49:10<2:29:45,  5.18s/it]2026-04-12 00:26:01,091 | INFO | train_grpo_train | metrics_logged mode=train step=1566

                                                       
{'loss': -0.0005, 'grad_norm': 4.360117435455322, 'learning_rate': 5.257575757575758e-06, 'num_tokens': 3516356.0, 'completions/mean_length': 71.375, 'completions/min_length': 65.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9724572896957397, 'rewards/meter/std': 0.05403033271431923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9724572896957397, 'rewards/total_composite/std': 0.05403033271431923, 'reward': 0.9724572896957397, 'reward_std': 0.054030340164899826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06206238269805908, 'sampling/sampling_logp_difference/max': 1.6887876987457275, 'sampling/importance_sampling_ratio/min': 0.18474335968494415, 'sampling/importance_sampling_ratio/mean': 1.0121489763259888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6091113425791264, 'clip_ratio/low_mean': 0.0053571430034935474, 'clip_ratio/low_min': 0.0053571430034935474, 'clip_ratio/high_mean': 0.05206052586436272, 'clip_ratio/high_max': 0.05206052586436272, 'clip_ratio/region_mean': 0.057417668867856264, 'reward_total_mean': 0.9724572896957397, 'reward_meter_mean': 0.9724572896957397, 'reward_meter_std': 0.05403033271431923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9724572896957397, 'reward_total_composite_std': 0.05403033271431923, 'epoch': 0.06}

 47%|████▋     | 1566/3300 [1:49:10<2:29:45,  5.18s/it]INFO 04-12 00:26:01 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1567/3300 [1:49:14<2:25:07,  5.02s/it]2026-04-12 00:26:05,754 | INFO | train_grpo_train | metrics_logged mode=train step=1567

                                                       
{'loss': 0.0001, 'grad_norm': 0.27344831824302673, 'learning_rate': 5.2545454545454555e-06, 'num_tokens': 3518144.0, 'completions/mean_length': 68.5, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9978020191192627, 'rewards/meter/std': 2.9081666070851497e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978020191192627, 'rewards/total_composite/std': 2.9081666070851497e-05, 'reward': 0.9978020191192627, 'reward_std': 2.9075883503537625e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010215016081929207, 'sampling/sampling_logp_difference/max': 0.7412929534912109, 'sampling/importance_sampling_ratio/min': 0.47649744153022766, 'sampling/importance_sampling_ratio/mean': 1.0018134117126465, 'sampling/importance_sampling_ratio/max': 1.357559084892273, 'entropy': 0.0569656016305089, 'clip_ratio/low_mean': 0.007247899193316698, 'clip_ratio/low_min': 0.007247899193316698, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.007247899193316698, 'reward_total_mean': 0.9978020191192627, 'reward_meter_mean': 0.9978020191192627, 'reward_meter_std': 2.9081666070851497e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978020191192627, 'reward_total_composite_std': 2.9081666070851497e-05, 'epoch': 0.06}

 47%|████▋     | 1567/3300 [1:49:14<2:25:07,  5.02s/it]INFO 04-12 00:26:05 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1568/3300 [1:49:19<2:25:32,  5.04s/it]2026-04-12 00:26:10,831 | INFO | train_grpo_train | metrics_logged mode=train step=1568

                                                       
{'loss': -0.0006, 'grad_norm': 0.3887041211128235, 'learning_rate': 5.251515151515152e-06, 'num_tokens': 3520369.0, 'completions/mean_length': 100.125, 'completions/min_length': 100.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9977670311927795, 'rewards/meter/std': 4.242736758897081e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977670311927795, 'rewards/total_composite/std': 4.242736758897081e-05, 'reward': 0.9977670311927795, 'reward_std': 4.243615330778994e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007036585360765457, 'sampling/sampling_logp_difference/max': 0.741631031036377, 'sampling/importance_sampling_ratio/min': 0.47633635997772217, 'sampling/importance_sampling_ratio/mean': 1.00063157081604, 'sampling/importance_sampling_ratio/max': 1.1633974313735962, 'entropy': 0.04269820963963866, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9977670311927795, 'reward_meter_mean': 0.9977670311927795, 'reward_meter_std': 4.242736758897081e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977670311927795, 'reward_total_composite_std': 4.242736758897081e-05, 'epoch': 0.06}

 48%|████▊     | 1568/3300 [1:49:19<2:25:32,  5.04s/it]INFO 04-12 00:26:11 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1569/3300 [1:49:24<2:22:57,  4.96s/it]2026-04-12 00:26:15,581 | INFO | train_grpo_train | metrics_logged mode=train step=1569

                                                       
{'loss': 0.0094, 'grad_norm': 19.87078857421875, 'learning_rate': 5.248484848484849e-06, 'num_tokens': 3522282.0, 'completions/mean_length': 63.125, 'completions/min_length': 62.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.125, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.998197078704834, 'rewards/meter/std': 0.00027716331533156335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998197078704834, 'rewards/total_composite/std': 0.00027716331533156335, 'reward': 0.998197078704834, 'reward_std': 0.00027716669137589633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010409246198832989, 'sampling/sampling_logp_difference/max': 0.8942767381668091, 'sampling/importance_sampling_ratio/min': 0.40890324115753174, 'sampling/importance_sampling_ratio/mean': 0.9992884993553162, 'sampling/importance_sampling_ratio/max': 1.3447296619415283, 'entropy': 0.033813337329775095, 'clip_ratio/low_mean': 0.002016128972172737, 'clip_ratio/low_min': 0.002016128972172737, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.003969253972172737, 'reward_total_mean': 0.998197078704834, 'reward_meter_mean': 0.998197078704834, 'reward_meter_std': 0.00027716331533156335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998197078704834, 'reward_total_composite_std': 0.00027716331533156335, 'epoch': 0.06}

 48%|████▊     | 1569/3300 [1:49:24<2:22:57,  4.96s/it]INFO 04-12 00:26:15 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1570/3300 [1:49:29<2:22:35,  4.95s/it]2026-04-12 00:26:20,503 | INFO | train_grpo_train | metrics_logged mode=train step=1570

                                                       
{'loss': 0.0363, 'grad_norm': 6.926992893218994, 'learning_rate': 5.245454545454546e-06, 'num_tokens': 3524067.0, 'completions/mean_length': 65.125, 'completions/min_length': 60.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7359851598739624, 'rewards/meter/std': 0.4039470851421356, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6135989427566528, 'rewards/total_composite/std': 0.46367478370666504, 'reward': 0.6135989427566528, 'reward_std': 0.46367475390434265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08760727196931839, 'sampling/sampling_logp_difference/max': 1.2285175323486328, 'sampling/importance_sampling_ratio/min': 0.29272621870040894, 'sampling/importance_sampling_ratio/mean': 1.0079141855239868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8175845369696617, 'clip_ratio/low_mean': 0.028282265178859234, 'clip_ratio/low_min': 0.028282265178859234, 'clip_ratio/high_mean': 0.04102045390754938, 'clip_ratio/high_max': 0.04102045390754938, 'clip_ratio/region_mean': 0.06930271908640862, 'reward_total_mean': 0.6135989427566528, 'reward_meter_mean': 0.7359851598739624, 'reward_meter_std': 0.4039470851421356, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6135989427566528, 'reward_total_composite_std': 0.46367478370666504, 'epoch': 0.06}

 48%|████▊     | 1570/3300 [1:49:29<2:22:35,  4.95s/it]INFO 04-12 00:26:20 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1571/3300 [1:49:35<2:33:22,  5.32s/it]2026-04-12 00:26:26,707 | INFO | train_grpo_train | metrics_logged mode=train step=1571

                                                       
{'loss': 0.0237, 'grad_norm': 2.949753999710083, 'learning_rate': 5.242424242424244e-06, 'num_tokens': 3527020.0, 'completions/mean_length': 171.125, 'completions/min_length': 168.0, 'completions/max_length': 185.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.125, 'completions/min_terminated_length': 168.0, 'completions/max_terminated_length': 185.0, 'rewards/meter/mean': 0.9974609613418579, 'rewards/meter/std': 0.0009353406494483352, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7897727489471436, 'rewards/repeat_penalty/std': 0.03682740405201912, 'rewards/total_composite/mean': 0.767343282699585, 'rewards/total_composite/std': 0.05765299126505852, 'reward': 0.767343282699585, 'reward_std': 0.05765299126505852, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014616935513913631, 'sampling/sampling_logp_difference/max': 1.0565416812896729, 'sampling/importance_sampling_ratio/min': 0.46636128425598145, 'sampling/importance_sampling_ratio/mean': 0.9989144206047058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0667440458200872, 'clip_ratio/low_mean': 0.006336958380416036, 'clip_ratio/low_min': 0.006336958380416036, 'clip_ratio/high_mean': 0.011825977824628353, 'clip_ratio/high_max': 0.011825977824628353, 'clip_ratio/region_mean': 0.01816293620504439, 'reward_total_mean': 0.767343282699585, 'reward_meter_mean': 0.9974609613418579, 'reward_meter_std': 0.0009353406494483352, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7897727489471436, 'reward_repeat_penalty_std': 0.03682740405201912, 'reward_total_composite_mean': 0.767343282699585, 'reward_total_composite_std': 0.05765299126505852, 'epoch': 0.06}

 48%|████▊     | 1571/3300 [1:49:35<2:33:22,  5.32s/it]INFO 04-12 00:26:26 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1572/3300 [1:49:45<3:12:28,  6.68s/it]2026-04-12 00:26:36,563 | INFO | train_grpo_train | metrics_logged mode=train step=1572

                                                       
{'loss': -0.218, 'grad_norm': 1.639418125152588, 'learning_rate': 5.23939393939394e-06, 'num_tokens': 3530638.0, 'completions/mean_length': 308.25, 'completions/min_length': 265.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 279.14288330078125, 'completions/min_terminated_length': 265.0, 'completions/max_terminated_length': 297.0, 'rewards/meter/mean': 0.6877518892288208, 'rewards/meter/std': 0.3425425887107849, 'rewards/count_adherence/mean': 0.862500011920929, 'rewards/count_adherence/std': 0.0517548993229866, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9172793626785278, 'rewards/repeat_penalty/std': 0.07193652540445328, 'rewards/total_composite/mean': 0.5252456665039062, 'rewards/total_composite/std': 0.28777360916137695, 'reward': 0.5252456665039062, 'reward_std': 0.28777357935905457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0495113804936409, 'sampling/sampling_logp_difference/max': 2.470219373703003, 'sampling/importance_sampling_ratio/min': 0.08456631004810333, 'sampling/importance_sampling_ratio/mean': 1.012639045715332, 'sampling/importance_sampling_ratio/max': 1.8714102506637573, 'entropy': 0.41537686437368393, 'clip_ratio/low_mean': 0.015551732387393713, 'clip_ratio/low_min': 0.015551732387393713, 'clip_ratio/high_mean': 0.012186205829493701, 'clip_ratio/high_max': 0.012186205829493701, 'clip_ratio/region_mean': 0.027737938216887414, 'reward_total_mean': 0.5252456665039062, 'reward_meter_mean': 0.6877518892288208, 'reward_meter_std': 0.3425425887107849, 'reward_count_adherence_mean': 0.862500011920929, 'reward_count_adherence_std': 0.0517548993229866, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9172793626785278, 'reward_repeat_penalty_std': 0.07193652540445328, 'reward_total_composite_mean': 0.5252456665039062, 'reward_total_composite_std': 0.28777360916137695, 'epoch': 0.06}

 48%|████▊     | 1572/3300 [1:49:45<3:12:28,  6.68s/it]INFO 04-12 00:26:36 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1573/3300 [1:49:50<2:55:56,  6.11s/it]2026-04-12 00:26:41,344 | INFO | train_grpo_train | metrics_logged mode=train step=1573

                                                       
{'loss': 0.0007, 'grad_norm': 5.333371162414551, 'learning_rate': 5.236363636363637e-06, 'num_tokens': 3532465.0, 'completions/mean_length': 64.375, 'completions/min_length': 61.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6559607982635498, 'rewards/meter/std': 0.3895518183708191, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6559607982635498, 'rewards/total_composite/std': 0.3895518183708191, 'reward': 0.6559607982635498, 'reward_std': 0.3895518183708191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07227914035320282, 'sampling/sampling_logp_difference/max': 1.8713045120239258, 'sampling/importance_sampling_ratio/min': 0.15392273664474487, 'sampling/importance_sampling_ratio/mean': 1.0150277614593506, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7125434167683125, 'clip_ratio/low_mean': 0.033473065122962, 'clip_ratio/low_min': 0.033473065122962, 'clip_ratio/high_mean': 0.047869643196463585, 'clip_ratio/high_max': 0.047869643196463585, 'clip_ratio/region_mean': 0.08134270831942558, 'reward_total_mean': 0.6559607982635498, 'reward_meter_mean': 0.6559607982635498, 'reward_meter_std': 0.3895518183708191, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6559607982635498, 'reward_total_composite_std': 0.3895518183708191, 'epoch': 0.06}

 48%|████▊     | 1573/3300 [1:49:50<2:55:56,  6.11s/it]INFO 04-12 00:26:41 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1574/3300 [1:49:55<2:47:59,  5.84s/it]2026-04-12 00:26:46,549 | INFO | train_grpo_train | metrics_logged mode=train step=1574

                                                       
{'loss': -0.0041, 'grad_norm': 6.877980709075928, 'learning_rate': 5.233333333333334e-06, 'num_tokens': 3534610.0, 'completions/mean_length': 94.125, 'completions/min_length': 88.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.125, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.8490567207336426, 'rewards/meter/std': 0.342186838388443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8243577480316162, 'rewards/total_composite/std': 0.3378320336341858, 'reward': 0.8243577480316162, 'reward_std': 0.3378320336341858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0785253494977951, 'sampling/sampling_logp_difference/max': 2.2510461807250977, 'sampling/importance_sampling_ratio/min': 0.10528901219367981, 'sampling/importance_sampling_ratio/mean': 1.00845468044281, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8282850012183189, 'clip_ratio/low_mean': 0.00818392145447433, 'clip_ratio/low_min': 0.00818392145447433, 'clip_ratio/high_mean': 0.04302619933150709, 'clip_ratio/high_max': 0.04302619933150709, 'clip_ratio/region_mean': 0.05121012078598142, 'reward_total_mean': 0.8243577480316162, 'reward_meter_mean': 0.8490567207336426, 'reward_meter_std': 0.342186838388443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8243577480316162, 'reward_total_composite_std': 0.3378320336341858, 'epoch': 0.06}

 48%|████▊     | 1574/3300 [1:49:55<2:47:59,  5.84s/it]INFO 04-12 00:26:46 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1575/3300 [1:50:00<2:42:26,  5.65s/it]2026-04-12 00:26:51,756 | INFO | train_grpo_train | metrics_logged mode=train step=1575

                                                       
{'loss': -0.0016, 'grad_norm': 6.546207427978516, 'learning_rate': 5.230303030303031e-06, 'num_tokens': 3536318.0, 'completions/mean_length': 61.5, 'completions/min_length': 59.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9349583387374878, 'rewards/meter/std': 0.10497313737869263, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9349583387374878, 'rewards/total_composite/std': 0.10497313737869263, 'reward': 0.9349583387374878, 'reward_std': 0.10497313737869263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04767481982707977, 'sampling/sampling_logp_difference/max': 1.3624334335327148, 'sampling/importance_sampling_ratio/min': 0.2560369670391083, 'sampling/importance_sampling_ratio/mean': 1.0037705898284912, 'sampling/importance_sampling_ratio/max': 1.711548924446106, 'entropy': 0.3381665423512459, 'clip_ratio/low_mean': 0.016170635353773832, 'clip_ratio/low_min': 0.016170635353773832, 'clip_ratio/high_mean': 0.02857396099716425, 'clip_ratio/high_max': 0.02857396099716425, 'clip_ratio/region_mean': 0.04474459635093808, 'reward_total_mean': 0.9349583387374878, 'reward_meter_mean': 0.9349583387374878, 'reward_meter_std': 0.10497313737869263, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9349583387374878, 'reward_total_composite_std': 0.10497313737869263, 'epoch': 0.06}

 48%|████▊     | 1575/3300 [1:50:00<2:42:26,  5.65s/it]INFO 04-12 00:26:52 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1576/3300 [1:50:06<2:39:36,  5.55s/it]2026-04-12 00:26:57,087 | INFO | train_grpo_train | metrics_logged mode=train step=1576

                                                       
{'loss': 0.0111, 'grad_norm': 3.9961159229278564, 'learning_rate': 5.2272727272727274e-06, 'num_tokens': 3538180.0, 'completions/mean_length': 61.75, 'completions/min_length': 58.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.75, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9147671461105347, 'rewards/meter/std': 0.21282289922237396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9147671461105347, 'rewards/total_composite/std': 0.21282289922237396, 'reward': 0.9147671461105347, 'reward_std': 0.21282285451889038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05491922050714493, 'sampling/sampling_logp_difference/max': 1.6048030853271484, 'sampling/importance_sampling_ratio/min': 0.20092912018299103, 'sampling/importance_sampling_ratio/mean': 1.0070737600326538, 'sampling/importance_sampling_ratio/max': 1.9876987934112549, 'entropy': 0.42180035449564457, 'clip_ratio/low_mean': 0.001953125, 'clip_ratio/low_min': 0.001953125, 'clip_ratio/high_mean': 0.03661064524203539, 'clip_ratio/high_max': 0.03661064524203539, 'clip_ratio/region_mean': 0.03856377024203539, 'reward_total_mean': 0.9147671461105347, 'reward_meter_mean': 0.9147671461105347, 'reward_meter_std': 0.21282289922237396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9147671461105347, 'reward_total_composite_std': 0.21282289922237396, 'epoch': 0.06}

 48%|████▊     | 1576/3300 [1:50:06<2:39:36,  5.55s/it]INFO 04-12 00:26:57 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1577/3300 [1:50:11<2:38:05,  5.51s/it]2026-04-12 00:27:02,514 | INFO | train_grpo_train | metrics_logged mode=train step=1577

                                                       
{'loss': 0.0008, 'grad_norm': 4.388180732727051, 'learning_rate': 5.224242424242425e-06, 'num_tokens': 3540188.0, 'completions/mean_length': 92.0, 'completions/min_length': 89.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.0, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.9917011260986328, 'rewards/meter/std': 0.006606912240386009, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8181818723678589, 'rewards/total_composite/std': 0.07076939195394516, 'reward': 0.8181818723678589, 'reward_std': 0.07076937705278397, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.037422213703393936, 'sampling/sampling_logp_difference/max': 1.1900215148925781, 'sampling/importance_sampling_ratio/min': 0.3042147159576416, 'sampling/importance_sampling_ratio/mean': 1.0047838687896729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24610379338264465, 'clip_ratio/low_mean': 0.02302604599390179, 'clip_ratio/low_min': 0.02302604599390179, 'clip_ratio/high_mean': 0.005319148767739534, 'clip_ratio/high_max': 0.005319148767739534, 'clip_ratio/region_mean': 0.028345194761641324, 'reward_total_mean': 0.8181818723678589, 'reward_meter_mean': 0.9917011260986328, 'reward_meter_std': 0.006606912240386009, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8181818723678589, 'reward_total_composite_std': 0.07076939195394516, 'epoch': 0.06}

 48%|████▊     | 1577/3300 [1:50:11<2:38:05,  5.51s/it]INFO 04-12 00:27:02 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1578/3300 [1:50:18<2:48:19,  5.86s/it]2026-04-12 00:27:09,188 | INFO | train_grpo_train | metrics_logged mode=train step=1578

                                                       
{'loss': 0.0638, 'grad_norm': 4.709424018859863, 'learning_rate': 5.221212121212121e-06, 'num_tokens': 3543292.0, 'completions/mean_length': 191.0, 'completions/min_length': 175.0, 'completions/max_length': 218.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 191.0, 'completions/min_terminated_length': 175.0, 'completions/max_terminated_length': 218.0, 'rewards/meter/mean': 0.9755741953849792, 'rewards/meter/std': 0.027564238756895065, 'rewards/count_adherence/mean': 0.925000011920929, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9154040217399597, 'rewards/repeat_penalty/std': 0.10342530906200409, 'rewards/total_composite/mean': 0.8337088823318481, 'rewards/total_composite/std': 0.17133085429668427, 'reward': 0.8337088823318481, 'reward_std': 0.17133086919784546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05211658403277397, 'sampling/sampling_logp_difference/max': 2.6808853149414062, 'sampling/importance_sampling_ratio/min': 0.06850247830152512, 'sampling/importance_sampling_ratio/mean': 1.0094478130340576, 'sampling/importance_sampling_ratio/max': 1.927268385887146, 'entropy': 0.4086650311946869, 'clip_ratio/low_mean': 0.012475505471229553, 'clip_ratio/low_min': 0.012475505471229553, 'clip_ratio/high_mean': 0.02987705054692924, 'clip_ratio/high_max': 0.02987705054692924, 'clip_ratio/region_mean': 0.04235255601815879, 'reward_total_mean': 0.8337088823318481, 'reward_meter_mean': 0.9755741953849792, 'reward_meter_std': 0.027564238756895065, 'reward_count_adherence_mean': 0.925000011920929, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9154040217399597, 'reward_repeat_penalty_std': 0.10342530906200409, 'reward_total_composite_mean': 0.8337088823318481, 'reward_total_composite_std': 0.17133085429668427, 'epoch': 0.06}

 48%|████▊     | 1578/3300 [1:50:18<2:48:19,  5.86s/it]INFO 04-12 00:27:09 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1579/3300 [1:50:23<2:44:13,  5.73s/it]2026-04-12 00:27:14,582 | INFO | train_grpo_train | metrics_logged mode=train step=1579

                                                       
{'loss': 0.008, 'grad_norm': 2.8309895992279053, 'learning_rate': 5.218181818181819e-06, 'num_tokens': 3545353.0, 'completions/mean_length': 100.625, 'completions/min_length': 100.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.625, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9977493286132812, 'rewards/meter/std': 0.00013334653340280056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9478673934936523, 'rewards/total_composite/std': 0.09243174642324448, 'reward': 0.9478673934936523, 'reward_std': 0.09243172407150269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007003183010965586, 'sampling/sampling_logp_difference/max': 0.8175020217895508, 'sampling/importance_sampling_ratio/min': 0.5253335237503052, 'sampling/importance_sampling_ratio/mean': 1.0029453039169312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.042127607855945826, 'clip_ratio/low_mean': 0.003676470718346536, 'clip_ratio/low_min': 0.003676470718346536, 'clip_ratio/high_mean': 0.0012376237427815795, 'clip_ratio/high_max': 0.0012376237427815795, 'clip_ratio/region_mean': 0.004914094461128116, 'reward_total_mean': 0.9478673934936523, 'reward_meter_mean': 0.9977493286132812, 'reward_meter_std': 0.00013334653340280056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9478673934936523, 'reward_total_composite_std': 0.09243174642324448, 'epoch': 0.06}

 48%|████▊     | 1579/3300 [1:50:23<2:44:13,  5.73s/it]INFO 04-12 00:27:14 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1580/3300 [1:50:30<2:54:43,  6.10s/it]2026-04-12 00:27:21,540 | INFO | train_grpo_train | metrics_logged mode=train step=1580

                                                       
{'loss': 0.017, 'grad_norm': 3.488912343978882, 'learning_rate': 5.215151515151516e-06, 'num_tokens': 3548859.0, 'completions/mean_length': 233.25, 'completions/min_length': 217.0, 'completions/max_length': 243.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 233.25, 'completions/min_terminated_length': 217.0, 'completions/max_terminated_length': 243.0, 'rewards/meter/mean': 0.6562398672103882, 'rewards/meter/std': 0.2839363217353821, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9615384340286255, 'rewards/repeat_penalty/std': 0.082234226167202, 'rewards/total_composite/mean': 0.6222666501998901, 'rewards/total_composite/std': 0.2602792978286743, 'reward': 0.6222666501998901, 'reward_std': 0.2602792978286743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09925277531147003, 'sampling/sampling_logp_difference/max': 1.4719266891479492, 'sampling/importance_sampling_ratio/min': 0.22948291897773743, 'sampling/importance_sampling_ratio/mean': 1.0219330787658691, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0062834583222866, 'clip_ratio/low_mean': 0.028440887574106455, 'clip_ratio/low_min': 0.028440887574106455, 'clip_ratio/high_mean': 0.03686255170032382, 'clip_ratio/high_max': 0.03686255170032382, 'clip_ratio/region_mean': 0.06530343927443027, 'reward_total_mean': 0.6222666501998901, 'reward_meter_mean': 0.6562398672103882, 'reward_meter_std': 0.2839363217353821, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9615384340286255, 'reward_repeat_penalty_std': 0.082234226167202, 'reward_total_composite_mean': 0.6222666501998901, 'reward_total_composite_std': 0.2602792978286743, 'epoch': 0.06}

 48%|████▊     | 1580/3300 [1:50:30<2:54:43,  6.10s/it]INFO 04-12 00:27:21 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1581/3300 [1:50:37<3:02:02,  6.35s/it]2026-04-12 00:27:28,492 | INFO | train_grpo_train | metrics_logged mode=train step=1581

                                                       
{'loss': -0.0056, 'grad_norm': 1.136604905128479, 'learning_rate': 5.212121212121213e-06, 'num_tokens': 3552445.0, 'completions/mean_length': 250.25, 'completions/min_length': 240.0, 'completions/max_length': 255.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 250.25, 'completions/min_terminated_length': 240.0, 'completions/max_terminated_length': 255.0, 'rewards/meter/mean': 0.9991568922996521, 'rewards/meter/std': 0.00013457036402542144, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7211538553237915, 'rewards/repeat_penalty/std': 0.039811473339796066, 'rewards/total_composite/mean': 0.7205430865287781, 'rewards/total_composite/std': 0.039719030261039734, 'reward': 0.7205430865287781, 'reward_std': 0.03971904516220093, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017896855250000954, 'sampling/sampling_logp_difference/max': 1.916792631149292, 'sampling/importance_sampling_ratio/min': 0.14707794785499573, 'sampling/importance_sampling_ratio/mean': 1.0050798654556274, 'sampling/importance_sampling_ratio/max': 1.794761300086975, 'entropy': 0.12574958708137274, 'clip_ratio/low_mean': 0.00662941113114357, 'clip_ratio/low_min': 0.00662941113114357, 'clip_ratio/high_mean': 0.0039564905455335975, 'clip_ratio/high_max': 0.0039564905455335975, 'clip_ratio/region_mean': 0.010585901676677167, 'reward_total_mean': 0.7205430865287781, 'reward_meter_mean': 0.9991568922996521, 'reward_meter_std': 0.00013457036402542144, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7211538553237915, 'reward_repeat_penalty_std': 0.039811473339796066, 'reward_total_composite_mean': 0.7205430865287781, 'reward_total_composite_std': 0.039719030261039734, 'epoch': 0.06}

 48%|████▊     | 1581/3300 [1:50:37<3:02:02,  6.35s/it]INFO 04-12 00:27:28 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1582/3300 [1:50:42<2:52:46,  6.03s/it]2026-04-12 00:27:33,780 | INFO | train_grpo_train | metrics_logged mode=train step=1582

                                                       
{'loss': 0.0002, 'grad_norm': 4.802741527557373, 'learning_rate': 5.209090909090909e-06, 'num_tokens': 3554625.0, 'completions/mean_length': 107.5, 'completions/min_length': 103.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.5, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.989787757396698, 'rewards/meter/std': 0.00834440253674984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.989787757396698, 'rewards/total_composite/std': 0.00834440253674984, 'reward': 0.989787757396698, 'reward_std': 0.008344405330717564, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05708528682589531, 'sampling/sampling_logp_difference/max': 1.5647554397583008, 'sampling/importance_sampling_ratio/min': 0.24133498966693878, 'sampling/importance_sampling_ratio/mean': 1.015655517578125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4601043425500393, 'clip_ratio/low_mean': 0.010625278111547232, 'clip_ratio/low_min': 0.010625278111547232, 'clip_ratio/high_mean': 0.034577128011733294, 'clip_ratio/high_max': 0.034577128011733294, 'clip_ratio/region_mean': 0.045202406123280525, 'reward_total_mean': 0.989787757396698, 'reward_meter_mean': 0.989787757396698, 'reward_meter_std': 0.00834440253674984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.989787757396698, 'reward_total_composite_std': 0.00834440253674984, 'epoch': 0.06}

 48%|████▊     | 1582/3300 [1:50:42<2:52:46,  6.03s/it]INFO 04-12 00:27:34 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1583/3300 [1:50:47<2:41:47,  5.65s/it]2026-04-12 00:27:38,567 | INFO | train_grpo_train | metrics_logged mode=train step=1583

                                                       
{'loss': 0.0415, 'grad_norm': 13.880090713500977, 'learning_rate': 5.2060606060606065e-06, 'num_tokens': 3556523.0, 'completions/mean_length': 67.25, 'completions/min_length': 64.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7919849157333374, 'rewards/meter/std': 0.3320651650428772, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7919849157333374, 'rewards/total_composite/std': 0.3320651650428772, 'reward': 0.7919849157333374, 'reward_std': 0.3320651352405548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.060586389154195786, 'sampling/sampling_logp_difference/max': 1.4995098114013672, 'sampling/importance_sampling_ratio/min': 0.22323958575725555, 'sampling/importance_sampling_ratio/mean': 1.0017415285110474, 'sampling/importance_sampling_ratio/max': 1.9709203243255615, 'entropy': 0.5009849853813648, 'clip_ratio/low_mean': 0.010714286006987095, 'clip_ratio/low_min': 0.010714286006987095, 'clip_ratio/high_mean': 0.03560801479034126, 'clip_ratio/high_max': 0.03560801479034126, 'clip_ratio/region_mean': 0.04632230079732835, 'reward_total_mean': 0.7919849157333374, 'reward_meter_mean': 0.7919849157333374, 'reward_meter_std': 0.3320651650428772, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7919849157333374, 'reward_total_composite_std': 0.3320651650428772, 'epoch': 0.06}

 48%|████▊     | 1583/3300 [1:50:47<2:41:47,  5.65s/it]INFO 04-12 00:27:38 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1584/3300 [1:50:52<2:33:46,  5.38s/it]2026-04-12 00:27:43,288 | INFO | train_grpo_train | metrics_logged mode=train step=1584

                                                       
{'loss': 0.0011, 'grad_norm': 4.862657070159912, 'learning_rate': 5.203030303030303e-06, 'num_tokens': 3558329.0, 'completions/mean_length': 60.75, 'completions/min_length': 58.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.75, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9940457940101624, 'rewards/meter/std': 0.0014534186339005828, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9940457940101624, 'rewards/total_composite/std': 0.0014534186339005828, 'reward': 0.9940457940101624, 'reward_std': 0.0014534194488078356, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02881665527820587, 'sampling/sampling_logp_difference/max': 1.269845962524414, 'sampling/importance_sampling_ratio/min': 0.28087490797042847, 'sampling/importance_sampling_ratio/mean': 1.0062899589538574, 'sampling/importance_sampling_ratio/max': 1.5038725137710571, 'entropy': 0.22835318371653557, 'clip_ratio/low_mean': 0.006222632946446538, 'clip_ratio/low_min': 0.006222632946446538, 'clip_ratio/high_mean': 0.010113696102052927, 'clip_ratio/high_max': 0.010113696102052927, 'clip_ratio/region_mean': 0.016336329048499465, 'reward_total_mean': 0.9940457940101624, 'reward_meter_mean': 0.9940457940101624, 'reward_meter_std': 0.0014534186339005828, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9940457940101624, 'reward_total_composite_std': 0.0014534186339005828, 'epoch': 0.06}

 48%|████▊     | 1584/3300 [1:50:52<2:33:46,  5.38s/it]INFO 04-12 00:27:43 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1585/3300 [1:50:57<2:35:16,  5.43s/it]2026-04-12 00:27:48,840 | INFO | train_grpo_train | metrics_logged mode=train step=1585

                                                       
{'loss': 0.0019, 'grad_norm': 2.1006689071655273, 'learning_rate': 5.2e-06, 'num_tokens': 3560777.0, 'completions/mean_length': 144.0, 'completions/min_length': 144.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 144.0, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9989944696426392, 'rewards/meter/std': 1.850287117122207e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8392857313156128, 'rewards/repeat_penalty/std': 0.09155284613370895, 'rewards/total_composite/mean': 0.8384406566619873, 'rewards/total_composite/std': 0.09144823998212814, 'reward': 0.8384406566619873, 'reward_std': 0.09144823998212814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010658186860382557, 'sampling/sampling_logp_difference/max': 0.8828556537628174, 'sampling/importance_sampling_ratio/min': 0.4136001169681549, 'sampling/importance_sampling_ratio/mean': 1.0028645992279053, 'sampling/importance_sampling_ratio/max': 1.262144923210144, 'entropy': 0.0726482942700386, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007812500116415322, 'clip_ratio/high_max': 0.007812500116415322, 'clip_ratio/region_mean': 0.007812500116415322, 'reward_total_mean': 0.8384406566619873, 'reward_meter_mean': 0.9989944696426392, 'reward_meter_std': 1.850287117122207e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8392857313156128, 'reward_repeat_penalty_std': 0.09155284613370895, 'reward_total_composite_mean': 0.8384406566619873, 'reward_total_composite_std': 0.09144823998212814, 'epoch': 0.06}

 48%|████▊     | 1585/3300 [1:50:57<2:35:16,  5.43s/it]INFO 04-12 00:27:49 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1586/3300 [1:51:02<2:30:01,  5.25s/it]2026-04-12 00:27:53,671 | INFO | train_grpo_train | metrics_logged mode=train step=1586

                                                       
{'loss': 0.0059, 'grad_norm': 6.0716328620910645, 'learning_rate': 5.196969696969697e-06, 'num_tokens': 3562502.0, 'completions/mean_length': 66.625, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.8443154096603394, 'rewards/meter/std': 0.19778741896152496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8443154096603394, 'rewards/total_composite/std': 0.19778741896152496, 'reward': 0.8443154096603394, 'reward_std': 0.19778741896152496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0697539895772934, 'sampling/sampling_logp_difference/max': 1.1605567932128906, 'sampling/importance_sampling_ratio/min': 0.3133116662502289, 'sampling/importance_sampling_ratio/mean': 1.0118776559829712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6144770756363869, 'clip_ratio/low_mean': 0.017196210101246834, 'clip_ratio/low_min': 0.017196210101246834, 'clip_ratio/high_mean': 0.03553921659477055, 'clip_ratio/high_max': 0.03553921659477055, 'clip_ratio/region_mean': 0.052735426696017385, 'reward_total_mean': 0.8443154096603394, 'reward_meter_mean': 0.8443154096603394, 'reward_meter_std': 0.19778741896152496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8443154096603394, 'reward_total_composite_std': 0.19778741896152496, 'epoch': 0.06}

 48%|████▊     | 1586/3300 [1:51:02<2:30:01,  5.25s/it]INFO 04-12 00:27:53 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1587/3300 [1:51:07<2:25:28,  5.10s/it]2026-04-12 00:27:58,402 | INFO | train_grpo_train | metrics_logged mode=train step=1587

                                                       
{'loss': -0.0017, 'grad_norm': 5.30000114440918, 'learning_rate': 5.193939393939395e-06, 'num_tokens': 3564369.0, 'completions/mean_length': 66.375, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9182997941970825, 'rewards/meter/std': 0.08735460042953491, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9182997941970825, 'rewards/total_composite/std': 0.08735460042953491, 'reward': 0.9182997941970825, 'reward_std': 0.08735460788011551, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06299924105405807, 'sampling/sampling_logp_difference/max': 1.8149499893188477, 'sampling/importance_sampling_ratio/min': 0.16284604370594025, 'sampling/importance_sampling_ratio/mean': 1.0096309185028076, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5464729219675064, 'clip_ratio/low_mean': 0.01317360415123403, 'clip_ratio/low_min': 0.01317360415123403, 'clip_ratio/high_mean': 0.03026041854172945, 'clip_ratio/high_max': 0.03026041854172945, 'clip_ratio/region_mean': 0.04343402269296348, 'reward_total_mean': 0.9182997941970825, 'reward_meter_mean': 0.9182997941970825, 'reward_meter_std': 0.08735460042953491, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9182997941970825, 'reward_total_composite_std': 0.08735460042953491, 'epoch': 0.06}

 48%|████▊     | 1587/3300 [1:51:07<2:25:28,  5.10s/it]INFO 04-12 00:27:58 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1588/3300 [1:51:12<2:25:54,  5.11s/it]2026-04-12 00:28:03,557 | INFO | train_grpo_train | metrics_logged mode=train step=1588

                                                       
{'loss': 0.0024, 'grad_norm': 5.146764755249023, 'learning_rate': 5.190909090909091e-06, 'num_tokens': 3566296.0, 'completions/mean_length': 69.875, 'completions/min_length': 66.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6954999566078186, 'rewards/meter/std': 0.4150845408439636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6954999566078186, 'rewards/total_composite/std': 0.4150845408439636, 'reward': 0.6954999566078186, 'reward_std': 0.41508448123931885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0591333732008934, 'sampling/sampling_logp_difference/max': 1.3553905487060547, 'sampling/importance_sampling_ratio/min': 0.2578465938568115, 'sampling/importance_sampling_ratio/mean': 1.014699101448059, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4779812954366207, 'clip_ratio/low_mean': 0.007130191195756197, 'clip_ratio/low_min': 0.007130191195756197, 'clip_ratio/high_mean': 0.025014446233399212, 'clip_ratio/high_max': 0.025014446233399212, 'clip_ratio/region_mean': 0.03214463742915541, 'reward_total_mean': 0.6954999566078186, 'reward_meter_mean': 0.6954999566078186, 'reward_meter_std': 0.4150845408439636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6954999566078186, 'reward_total_composite_std': 0.4150845408439636, 'epoch': 0.06}

 48%|████▊     | 1588/3300 [1:51:12<2:25:54,  5.11s/it]INFO 04-12 00:28:03 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1589/3300 [1:51:17<2:22:00,  4.98s/it]2026-04-12 00:28:08,224 | INFO | train_grpo_train | metrics_logged mode=train step=1589

                                                       
{'loss': -0.0052, 'grad_norm': 5.699836730957031, 'learning_rate': 5.187878787878788e-06, 'num_tokens': 3568219.0, 'completions/mean_length': 71.375, 'completions/min_length': 69.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.8743665218353271, 'rewards/meter/std': 0.338323175907135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8743665218353271, 'rewards/total_composite/std': 0.338323175907135, 'reward': 0.8743665218353271, 'reward_std': 0.3383232057094574, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.060609057545661926, 'sampling/sampling_logp_difference/max': 1.2876157760620117, 'sampling/importance_sampling_ratio/min': 0.275927871465683, 'sampling/importance_sampling_ratio/mean': 1.0109214782714844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4802176021039486, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.04341738054063171, 'clip_ratio/high_max': 0.04341738054063171, 'clip_ratio/region_mean': 0.04341738054063171, 'reward_total_mean': 0.8743665218353271, 'reward_meter_mean': 0.8743665218353271, 'reward_meter_std': 0.338323175907135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8743665218353271, 'reward_total_composite_std': 0.338323175907135, 'epoch': 0.06}

 48%|████▊     | 1589/3300 [1:51:17<2:22:00,  4.98s/it]INFO 04-12 00:28:08 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1590/3300 [1:51:24<2:40:52,  5.64s/it]2026-04-12 00:28:15,420 | INFO | train_grpo_train | metrics_logged mode=train step=1590

                                                       
{'loss': -0.0196, 'grad_norm': 1.8666749000549316, 'learning_rate': 5.184848484848485e-06, 'num_tokens': 3571993.0, 'completions/mean_length': 271.75, 'completions/min_length': 248.0, 'completions/max_length': 287.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 271.75, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 287.0, 'rewards/meter/mean': 0.9924838542938232, 'rewards/meter/std': 0.004395514260977507, 'rewards/count_adherence/mean': 0.8611111640930176, 'rewards/count_adherence/std': 0.05143444612622261, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6185267567634583, 'rewards/repeat_penalty/std': 0.05406184867024422, 'rewards/total_composite/mean': 0.5278421640396118, 'rewards/total_composite/std': 0.04846978932619095, 'reward': 0.5278421640396118, 'reward_std': 0.048469796776771545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018268149346113205, 'sampling/sampling_logp_difference/max': 5.459444046020508, 'sampling/importance_sampling_ratio/min': 0.004255921114236116, 'sampling/importance_sampling_ratio/mean': 1.0014569759368896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08188144443556666, 'clip_ratio/low_mean': 0.007355856709182262, 'clip_ratio/low_min': 0.007355856709182262, 'clip_ratio/high_mean': 0.004474195709917694, 'clip_ratio/high_max': 0.004474195709917694, 'clip_ratio/region_mean': 0.011830052419099957, 'reward_total_mean': 0.5278421640396118, 'reward_meter_mean': 0.9924838542938232, 'reward_meter_std': 0.004395514260977507, 'reward_count_adherence_mean': 0.8611111640930176, 'reward_count_adherence_std': 0.05143444612622261, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6185267567634583, 'reward_repeat_penalty_std': 0.05406184867024422, 'reward_total_composite_mean': 0.5278421640396118, 'reward_total_composite_std': 0.04846978932619095, 'epoch': 0.06}

 48%|████▊     | 1590/3300 [1:51:24<2:40:52,  5.64s/it]INFO 04-12 00:28:15 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1591/3300 [1:51:28<2:31:39,  5.32s/it]2026-04-12 00:28:20,008 | INFO | train_grpo_train | metrics_logged mode=train step=1591

                                                       
{'loss': 0.0162, 'grad_norm': 6.771106719970703, 'learning_rate': 5.181818181818182e-06, 'num_tokens': 3573831.0, 'completions/mean_length': 67.75, 'completions/min_length': 66.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.89401775598526, 'rewards/meter/std': 0.16673576831817627, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.89401775598526, 'rewards/total_composite/std': 0.16673576831817627, 'reward': 0.89401775598526, 'reward_std': 0.16673575341701508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05256752669811249, 'sampling/sampling_logp_difference/max': 1.1782063245773315, 'sampling/importance_sampling_ratio/min': 0.40265539288520813, 'sampling/importance_sampling_ratio/mean': 1.0108517408370972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48275283351540565, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03321157908067107, 'clip_ratio/high_max': 0.03321157908067107, 'clip_ratio/region_mean': 0.03321157908067107, 'reward_total_mean': 0.89401775598526, 'reward_meter_mean': 0.89401775598526, 'reward_meter_std': 0.16673576831817627, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.89401775598526, 'reward_total_composite_std': 0.16673576831817627, 'epoch': 0.06}

 48%|████▊     | 1591/3300 [1:51:29<2:31:39,  5.32s/it]INFO 04-12 00:28:20 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1592/3300 [1:51:33<2:26:45,  5.16s/it]2026-04-12 00:28:24,762 | INFO | train_grpo_train | metrics_logged mode=train step=1592

                                                       
{'loss': 0.0026, 'grad_norm': 1.81292724609375, 'learning_rate': 5.1787878787878784e-06, 'num_tokens': 3575592.0, 'completions/mean_length': 58.125, 'completions/min_length': 58.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9933634400367737, 'rewards/meter/std': 0.00030929798958823085, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9933634400367737, 'rewards/total_composite/std': 0.00030929798958823085, 'reward': 0.9933634400367737, 'reward_std': 0.00030929522472433746, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004956495948135853, 'sampling/sampling_logp_difference/max': 0.7400164604187012, 'sampling/importance_sampling_ratio/min': 0.4771060645580292, 'sampling/importance_sampling_ratio/mean': 0.9998379349708557, 'sampling/importance_sampling_ratio/max': 1.1543108224868774, 'entropy': 0.020813994109630585, 'clip_ratio/low_mean': 0.0021186440717428923, 'clip_ratio/low_min': 0.0021186440717428923, 'clip_ratio/high_mean': 0.006465517217293382, 'clip_ratio/high_max': 0.006465517217293382, 'clip_ratio/region_mean': 0.008584161289036274, 'reward_total_mean': 0.9933634400367737, 'reward_meter_mean': 0.9933634400367737, 'reward_meter_std': 0.00030929798958823085, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9933634400367737, 'reward_total_composite_std': 0.00030929798958823085, 'epoch': 0.06}

 48%|████▊     | 1592/3300 [1:51:33<2:26:45,  5.16s/it]INFO 04-12 00:28:24 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1593/3300 [1:51:38<2:25:09,  5.10s/it]2026-04-12 00:28:29,739 | INFO | train_grpo_train | metrics_logged mode=train step=1593

                                                       
{'loss': 0.035, 'grad_norm': 10.237826347351074, 'learning_rate': 5.1757575757575765e-06, 'num_tokens': 3577358.0, 'completions/mean_length': 68.75, 'completions/min_length': 67.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7235053181648254, 'rewards/meter/std': 0.3811579644680023, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7235053181648254, 'rewards/total_composite/std': 0.3811579644680023, 'reward': 0.7235053181648254, 'reward_std': 0.3811579942703247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027932489290833473, 'sampling/sampling_logp_difference/max': 1.7486402988433838, 'sampling/importance_sampling_ratio/min': 0.17401038110256195, 'sampling/importance_sampling_ratio/mean': 0.9999585151672363, 'sampling/importance_sampling_ratio/max': 1.7127673625946045, 'entropy': 0.12029994325712323, 'clip_ratio/low_mean': 0.012177230441011488, 'clip_ratio/low_min': 0.012177230441011488, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.01404290203936398, 'reward_total_mean': 0.7235053181648254, 'reward_meter_mean': 0.7235053181648254, 'reward_meter_std': 0.3811579644680023, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7235053181648254, 'reward_total_composite_std': 0.3811579644680023, 'epoch': 0.06}

 48%|████▊     | 1593/3300 [1:51:38<2:25:09,  5.10s/it]INFO 04-12 00:28:29 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1594/3300 [1:51:45<2:39:40,  5.62s/it]2026-04-12 00:28:36,553 | INFO | train_grpo_train | metrics_logged mode=train step=1594

                                                       
{'loss': 0.0095, 'grad_norm': 1.7681658267974854, 'learning_rate': 5.172727272727273e-06, 'num_tokens': 3580425.0, 'completions/mean_length': 195.375, 'completions/min_length': 188.0, 'completions/max_length': 200.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 195.375, 'completions/min_terminated_length': 188.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.9953435659408569, 'rewards/meter/std': 0.0010429132962599397, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7424242496490479, 'rewards/repeat_penalty/std': 0.04169124737381935, 'rewards/total_composite/mean': 0.7389621734619141, 'rewards/total_composite/std': 0.04144872725009918, 'reward': 0.7389621734619141, 'reward_std': 0.041448723524808884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016124166548252106, 'sampling/sampling_logp_difference/max': 2.298177719116211, 'sampling/importance_sampling_ratio/min': 0.10044170916080475, 'sampling/importance_sampling_ratio/mean': 0.9993682503700256, 'sampling/importance_sampling_ratio/max': 1.846528172492981, 'entropy': 0.06844896962866187, 'clip_ratio/low_mean': 0.005208890186622739, 'clip_ratio/low_min': 0.005208890186622739, 'clip_ratio/high_mean': 0.007002223108429462, 'clip_ratio/high_max': 0.007002223108429462, 'clip_ratio/region_mean': 0.0122111132950522, 'reward_total_mean': 0.7389621734619141, 'reward_meter_mean': 0.9953435659408569, 'reward_meter_std': 0.0010429132962599397, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7424242496490479, 'reward_repeat_penalty_std': 0.04169124737381935, 'reward_total_composite_mean': 0.7389621734619141, 'reward_total_composite_std': 0.04144872725009918, 'epoch': 0.06}

 48%|████▊     | 1594/3300 [1:51:45<2:39:40,  5.62s/it]INFO 04-12 00:28:36 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1595/3300 [1:51:51<2:44:53,  5.80s/it]2026-04-12 00:28:42,793 | INFO | train_grpo_train | metrics_logged mode=train step=1595

                                                       
{'loss': 0.0842, 'grad_norm': 4.416578769683838, 'learning_rate': 5.16969696969697e-06, 'num_tokens': 3583190.0, 'completions/mean_length': 161.625, 'completions/min_length': 144.0, 'completions/max_length': 194.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 161.625, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 194.0, 'rewards/meter/mean': 0.9966393113136292, 'rewards/meter/std': 0.0009264187538065016, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9503968358039856, 'rewards/repeat_penalty/std': 0.06915634870529175, 'rewards/total_composite/mean': 0.8883745670318604, 'rewards/total_composite/std': 0.13054178655147552, 'reward': 0.8883745670318604, 'reward_std': 0.13054178655147552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0641026571393013, 'sampling/sampling_logp_difference/max': 2.719486713409424, 'sampling/importance_sampling_ratio/min': 0.06590858101844788, 'sampling/importance_sampling_ratio/mean': 1.0092458724975586, 'sampling/importance_sampling_ratio/max': 1.7064915895462036, 'entropy': 0.49944454059004784, 'clip_ratio/low_mean': 0.018327449448406696, 'clip_ratio/low_min': 0.018327449448406696, 'clip_ratio/high_mean': 0.030346576124429703, 'clip_ratio/high_max': 0.030346576124429703, 'clip_ratio/region_mean': 0.0486740255728364, 'reward_total_mean': 0.8883745670318604, 'reward_meter_mean': 0.9966393113136292, 'reward_meter_std': 0.0009264187538065016, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9503968358039856, 'reward_repeat_penalty_std': 0.06915634870529175, 'reward_total_composite_mean': 0.8883745670318604, 'reward_total_composite_std': 0.13054178655147552, 'epoch': 0.06}

 48%|████▊     | 1595/3300 [1:51:51<2:44:53,  5.80s/it]INFO 04-12 00:28:43 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1596/3300 [1:51:56<2:36:37,  5.52s/it]2026-04-12 00:28:47,642 | INFO | train_grpo_train | metrics_logged mode=train step=1596

                                                       
{'loss': 0.0073, 'grad_norm': 6.151648998260498, 'learning_rate': 5.1666666666666675e-06, 'num_tokens': 3584948.0, 'completions/mean_length': 49.75, 'completions/min_length': 49.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9367859363555908, 'rewards/meter/std': 0.005013443063944578, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9367859363555908, 'rewards/total_composite/std': 0.005013443063944578, 'reward': 0.9367859363555908, 'reward_std': 0.005013437941670418, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024813249707221985, 'sampling/sampling_logp_difference/max': 1.2509276866912842, 'sampling/importance_sampling_ratio/min': 0.2862391471862793, 'sampling/importance_sampling_ratio/mean': 1.0016454458236694, 'sampling/importance_sampling_ratio/max': 1.7191351652145386, 'entropy': 0.11144222784787416, 'clip_ratio/low_mean': 0.010051020188257098, 'clip_ratio/low_min': 0.010051020188257098, 'clip_ratio/high_mean': 0.02265306143090129, 'clip_ratio/high_max': 0.02265306143090129, 'clip_ratio/region_mean': 0.03270408161915839, 'reward_total_mean': 0.9367859363555908, 'reward_meter_mean': 0.9367859363555908, 'reward_meter_std': 0.005013443063944578, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9367859363555908, 'reward_total_composite_std': 0.005013443063944578, 'epoch': 0.06}

 48%|████▊     | 1596/3300 [1:51:56<2:36:37,  5.52s/it]INFO 04-12 00:28:47 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1597/3300 [1:52:01<2:34:01,  5.43s/it]2026-04-12 00:28:52,862 | INFO | train_grpo_train | metrics_logged mode=train step=1597

                                                       
{'loss': 0.0102, 'grad_norm': 4.290965557098389, 'learning_rate': 5.163636363636364e-06, 'num_tokens': 3587243.0, 'completions/mean_length': 111.875, 'completions/min_length': 106.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.875, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.9952096939086914, 'rewards/meter/std': 0.004024881403893232, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.925000011920929, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.9206209182739258, 'rewards/total_composite/std': 0.1035769134759903, 'reward': 0.9206209182739258, 'reward_std': 0.10357693582773209, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.052159715443849564, 'sampling/sampling_logp_difference/max': 1.9255342483520508, 'sampling/importance_sampling_ratio/min': 0.14579784870147705, 'sampling/importance_sampling_ratio/mean': 1.011698842048645, 'sampling/importance_sampling_ratio/max': 1.8452407121658325, 'entropy': 0.40937453880906105, 'clip_ratio/low_mean': 0.011171216145157814, 'clip_ratio/low_min': 0.011171216145157814, 'clip_ratio/high_mean': 0.02771200449205935, 'clip_ratio/high_max': 0.02771200449205935, 'clip_ratio/region_mean': 0.038883220637217164, 'reward_total_mean': 0.9206209182739258, 'reward_meter_mean': 0.9952096939086914, 'reward_meter_std': 0.004024881403893232, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.925000011920929, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.9206209182739258, 'reward_total_composite_std': 0.1035769134759903, 'epoch': 0.06}

 48%|████▊     | 1597/3300 [1:52:01<2:34:01,  5.43s/it]INFO 04-12 00:28:53 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1598/3300 [1:52:06<2:27:46,  5.21s/it]2026-04-12 00:28:57,570 | INFO | train_grpo_train | metrics_logged mode=train step=1598

                                                       
{'loss': 0.003, 'grad_norm': 2.746960401535034, 'learning_rate': 5.160606060606061e-06, 'num_tokens': 3589155.0, 'completions/mean_length': 64.0, 'completions/min_length': 63.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9953999519348145, 'rewards/meter/std': 0.0069043696857988834, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953999519348145, 'rewards/total_composite/std': 0.0069043696857988834, 'reward': 0.9953999519348145, 'reward_std': 0.006904360372573137, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021562010049819946, 'sampling/sampling_logp_difference/max': 0.8483424186706543, 'sampling/importance_sampling_ratio/min': 0.42812401056289673, 'sampling/importance_sampling_ratio/mean': 0.9994773268699646, 'sampling/importance_sampling_ratio/max': 1.8938157558441162, 'entropy': 0.0873101712204516, 'clip_ratio/low_mean': 0.001953125, 'clip_ratio/low_min': 0.001953125, 'clip_ratio/high_mean': 0.01751898298971355, 'clip_ratio/high_max': 0.01751898298971355, 'clip_ratio/region_mean': 0.01947210798971355, 'reward_total_mean': 0.9953999519348145, 'reward_meter_mean': 0.9953999519348145, 'reward_meter_std': 0.0069043696857988834, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953999519348145, 'reward_total_composite_std': 0.0069043696857988834, 'epoch': 0.06}

 48%|████▊     | 1598/3300 [1:52:06<2:27:46,  5.21s/it]INFO 04-12 00:28:57 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1599/3300 [1:52:10<2:20:56,  4.97s/it]2026-04-12 00:29:01,981 | INFO | train_grpo_train | metrics_logged mode=train step=1599

                                                       
{'loss': 0.0135, 'grad_norm': 6.9674506187438965, 'learning_rate': 5.1575757575757575e-06, 'num_tokens': 3590515.0, 'completions/mean_length': 37.0, 'completions/min_length': 36.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9930006265640259, 'rewards/meter/std': 0.00047757019638083875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9930006265640259, 'rewards/total_composite/std': 0.00047757019638083875, 'reward': 0.9930006265640259, 'reward_std': 0.0004775776178576052, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030742382630705833, 'sampling/sampling_logp_difference/max': 0.5062961578369141, 'sampling/importance_sampling_ratio/min': 0.6027238368988037, 'sampling/importance_sampling_ratio/mean': 1.0091636180877686, 'sampling/importance_sampling_ratio/max': 1.5318171977996826, 'entropy': 0.307000532746315, 'clip_ratio/low_mean': 0.006756756920367479, 'clip_ratio/low_min': 0.006756756920367479, 'clip_ratio/high_mean': 0.017178362933918834, 'clip_ratio/high_max': 0.017178362933918834, 'clip_ratio/region_mean': 0.023935119854286313, 'reward_total_mean': 0.9930006265640259, 'reward_meter_mean': 0.9930006265640259, 'reward_meter_std': 0.00047757019638083875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9930006265640259, 'reward_total_composite_std': 0.00047757019638083875, 'epoch': 0.06}

 48%|████▊     | 1599/3300 [1:52:10<2:20:56,  4.97s/it]INFO 04-12 00:29:02 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1600/3300 [1:52:15<2:16:59,  4.83s/it]2026-04-12 00:29:06,496 | INFO | train_grpo_train | metrics_logged mode=train step=1600

                                                       
{'loss': 0.0215, 'grad_norm': 6.287996292114258, 'learning_rate': 5.154545454545456e-06, 'num_tokens': 3592322.0, 'completions/mean_length': 64.875, 'completions/min_length': 64.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9355560541152954, 'rewards/meter/std': 0.06913071125745773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9355560541152954, 'rewards/total_composite/std': 0.06913071125745773, 'reward': 0.9355560541152954, 'reward_std': 0.06913069635629654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03889580816030502, 'sampling/sampling_logp_difference/max': 0.9479434490203857, 'sampling/importance_sampling_ratio/min': 0.3875372111797333, 'sampling/importance_sampling_ratio/mean': 1.003296136856079, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24194820784032345, 'clip_ratio/low_mean': 0.007490954361855984, 'clip_ratio/low_min': 0.007490954361855984, 'clip_ratio/high_mean': 0.01740056835114956, 'clip_ratio/high_max': 0.01740056835114956, 'clip_ratio/region_mean': 0.024891522713005543, 'reward_total_mean': 0.9355560541152954, 'reward_meter_mean': 0.9355560541152954, 'reward_meter_std': 0.06913071125745773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9355560541152954, 'reward_total_composite_std': 0.06913071125745773, 'epoch': 0.06}

 48%|████▊     | 1600/3300 [1:52:15<2:16:59,  4.83s/it]INFO 04-12 00:29:06 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:28,  2.63s/it]

 23%|██▎       | 3/13 [00:08<00:29,  2.92s/it]

 31%|███       | 4/13 [00:13<00:32,  3.63s/it]

 38%|███▊      | 5/13 [00:18<00:32,  4.07s/it]

 46%|████▌     | 6/13 [00:23<00:31,  4.51s/it]

 54%|█████▍    | 7/13 [00:28<00:27,  4.52s/it]

 62%|██████▏   | 8/13 [00:34<00:25,  5.01s/it]

 69%|██████▉   | 9/13 [00:40<00:21,  5.40s/it]

 77%|███████▋  | 10/13 [00:44<00:15,  5.07s/it]

 85%|████████▍ | 11/13 [00:50<00:10,  5.09s/it]

 92%|█████████▏| 12/13 [00:55<00:05,  5.07s/it]

100%|██████████| 13/13 [00:59<00:00,  5.02s/it]2026-04-12 00:30:12,823 | INFO | train_grpo_train | metrics_logged mode=eval step=1600

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 66.2489, 'eval_samples_per_second': 1.57, 'eval_steps_per_second': 0.196, 'eval_num_tokens': 3592322.0, 'eval_completions/mean_length': 192.58653846153845, 'eval_completions/min_length': 61.76923076923077, 'eval_completions/max_length': 347.3076923076923, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 192.58653846153845, 'eval_completions/min_terminated_length': 61.76923076923077, 'eval_completions/max_terminated_length': 347.3076923076923, 'eval_rewards/meter/mean': 0.7114841112723718, 'eval_rewards/meter/std': 0.38725116863273656, 'eval_rewards/count_adherence/mean': 0.9018935148532574, 'eval_rewards/count_adherence/std': 0.12116303610113952, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8302872364337628, 'eval_rewards/repeat_penalty/std': 0.14936749923687714, 'eval_rewards/total_composite/mean': 0.5311165887575883, 'eval_rewards/total_composite/std': 0.34226179122924805, 'eval_reward': 0.5311165887575883, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0201402991436995, 'eval_sampling/sampling_logp_difference/max': 1.060350748208853, 'eval_sampling/importance_sampling_ratio/min': 0.35776989047343916, 'eval_sampling/importance_sampling_ratio/mean': 1.004677598293011, 'eval_sampling/importance_sampling_ratio/max': 1.4436852565178504, 'eval_entropy': 0.19453936872574, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5311165887575883, 'eval_reward_meter_mean': 0.7114841112723718, 'eval_reward_meter_std': 0.38725116863273656, 'eval_reward_count_adherence_mean': 0.9018935148532574, 'eval_reward_count_adherence_std': 0.12116303610113952, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8302872364337628, 'eval_reward_repeat_penalty_std': 0.14936749923687714, 'eval_reward_total_composite_mean': 0.5311165887575883, 'eval_reward_total_composite_std': 0.34226179122924805, 'epoch': 0.06}

 48%|████▊     | 1600/3300 [1:53:21<2:16:59,  4.83s/it]

100%|██████████| 13/13 [01:00<00:00,  5.02s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 49%|████▊     | 1601/3300 [1:53:30<12:11:49, 25.84s/it]2026-04-12 00:30:21,365 | INFO | train_grpo_train | metrics_logged mode=train step=1601

                                                        
{'loss': -0.0132, 'grad_norm': 5.907294273376465, 'learning_rate': 5.151515151515152e-06, 'num_tokens': 3594188.0, 'completions/mean_length': 77.25, 'completions/min_length': 73.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.25, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9925735592842102, 'rewards/meter/std': 0.013519754633307457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9925735592842102, 'rewards/total_composite/std': 0.013519754633307457, 'reward': 0.9925735592842102, 'reward_std': 0.013519756495952606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.061646249145269394, 'sampling/sampling_logp_difference/max': 3.263657808303833, 'sampling/importance_sampling_ratio/min': 0.21933509409427643, 'sampling/importance_sampling_ratio/mean': 1.0113800764083862, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4462718777358532, 'clip_ratio/low_mean': 0.0034246575087308884, 'clip_ratio/low_min': 0.0034246575087308884, 'clip_ratio/high_mean': 0.03826867160387337, 'clip_ratio/high_max': 0.03826867160387337, 'clip_ratio/region_mean': 0.04169332911260426, 'reward_total_mean': 0.9925735592842102, 'reward_meter_mean': 0.9925735592842102, 'reward_meter_std': 0.013519754633307457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9925735592842102, 'reward_total_composite_std': 0.013519754633307457, 'epoch': 0.06}

 49%|████▊     | 1601/3300 [1:53:30<12:11:49, 25.84s/it]INFO 04-12 00:30:21 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1602/3300 [1:53:36<9:22:08, 19.86s/it] 2026-04-12 00:30:27,272 | INFO | train_grpo_train | metrics_logged mode=train step=1602

                                                       
{'loss': -0.0111, 'grad_norm': 3.9016387462615967, 'learning_rate': 5.148484848484849e-06, 'num_tokens': 3596572.0, 'completions/mean_length': 133.0, 'completions/min_length': 125.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.0, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.10229554772377014, 'rewards/meter/std': 0.20073863863945007, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.08772411197423935, 'rewards/total_composite/std': 0.17203770577907562, 'reward': 0.08772411197423935, 'reward_std': 0.17203769087791443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06598487496376038, 'sampling/sampling_logp_difference/max': 1.3837254047393799, 'sampling/importance_sampling_ratio/min': 0.27479445934295654, 'sampling/importance_sampling_ratio/mean': 1.0097557306289673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4912920892238617, 'clip_ratio/low_mean': 0.038138989359140396, 'clip_ratio/low_min': 0.038138989359140396, 'clip_ratio/high_mean': 0.016364703187718987, 'clip_ratio/high_max': 0.016364703187718987, 'clip_ratio/region_mean': 0.054503692546859384, 'reward_total_mean': 0.08772411197423935, 'reward_meter_mean': 0.10229554772377014, 'reward_meter_std': 0.20073863863945007, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.08772411197423935, 'reward_total_composite_std': 0.17203770577907562, 'epoch': 0.06}

 49%|████▊     | 1602/3300 [1:53:36<9:22:08, 19.86s/it]INFO 04-12 00:30:27 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1603/3300 [1:53:42<7:23:32, 15.68s/it]2026-04-12 00:30:33,210 | INFO | train_grpo_train | metrics_logged mode=train step=1603

                                                       
{'loss': -0.0097, 'grad_norm': 2.740659713745117, 'learning_rate': 5.145454545454546e-06, 'num_tokens': 3598835.0, 'completions/mean_length': 104.875, 'completions/min_length': 102.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.875, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9987785816192627, 'rewards/meter/std': 0.00020610357751138508, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8239848613739014, 'rewards/total_composite/std': 0.07050739973783493, 'reward': 0.8239848613739014, 'reward_std': 0.07050740718841553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019803691655397415, 'sampling/sampling_logp_difference/max': 0.6796939373016357, 'sampling/importance_sampling_ratio/min': 0.5668526887893677, 'sampling/importance_sampling_ratio/mean': 1.0108842849731445, 'sampling/importance_sampling_ratio/max': 1.973273754119873, 'entropy': 0.10506064165383577, 'clip_ratio/low_mean': 0.01308055012486875, 'clip_ratio/low_min': 0.01308055012486875, 'clip_ratio/high_mean': 0.002314814832061529, 'clip_ratio/high_max': 0.002314814832061529, 'clip_ratio/region_mean': 0.01539536495693028, 'reward_total_mean': 0.8239848613739014, 'reward_meter_mean': 0.9987785816192627, 'reward_meter_std': 0.00020610357751138508, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8239848613739014, 'reward_total_composite_std': 0.07050739973783493, 'epoch': 0.06}

 49%|████▊     | 1603/3300 [1:53:42<7:23:32, 15.68s/it]INFO 04-12 00:30:33 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1604/3300 [1:53:48<6:00:41, 12.76s/it]2026-04-12 00:30:39,145 | INFO | train_grpo_train | metrics_logged mode=train step=1604

                                                       
{'loss': -0.0024, 'grad_norm': 1.9800543785095215, 'learning_rate': 5.142424242424243e-06, 'num_tokens': 3601733.0, 'completions/mean_length': 178.25, 'completions/min_length': 172.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.25, 'completions/min_terminated_length': 172.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.9987528324127197, 'rewards/meter/std': 0.00028449000092223287, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7361111044883728, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.7351874709129333, 'rewards/total_composite/std': 0.0573553703725338, 'reward': 0.7351874709129333, 'reward_std': 0.0573553666472435, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01709139533340931, 'sampling/sampling_logp_difference/max': 1.673609972000122, 'sampling/importance_sampling_ratio/min': 0.18756872415542603, 'sampling/importance_sampling_ratio/mean': 1.0028082132339478, 'sampling/importance_sampling_ratio/max': 1.682837963104248, 'entropy': 0.08005253970623016, 'clip_ratio/low_mean': 0.00780038780067116, 'clip_ratio/low_min': 0.00780038780067116, 'clip_ratio/high_mean': 0.006316610379144549, 'clip_ratio/high_max': 0.006316610379144549, 'clip_ratio/region_mean': 0.01411699817981571, 'reward_total_mean': 0.7351874709129333, 'reward_meter_mean': 0.9987528324127197, 'reward_meter_std': 0.00028449000092223287, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7361111044883728, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.7351874709129333, 'reward_total_composite_std': 0.0573553703725338, 'epoch': 0.06}

 49%|████▊     | 1604/3300 [1:53:48<6:00:41, 12.76s/it]INFO 04-12 00:30:39 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1605/3300 [1:53:53<4:56:02, 10.48s/it]2026-04-12 00:30:44,311 | INFO | train_grpo_train | metrics_logged mode=train step=1605

                                                       
{'loss': 0.0008, 'grad_norm': 0.7806724905967712, 'learning_rate': 5.139393939393939e-06, 'num_tokens': 3603520.0, 'completions/mean_length': 71.375, 'completions/min_length': 70.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9987327456474304, 'rewards/meter/std': 6.793010834371671e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987327456474304, 'rewards/total_composite/std': 6.793010834371671e-05, 'reward': 0.9987327456474304, 'reward_std': 6.793974171159789e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015499016270041466, 'sampling/sampling_logp_difference/max': 0.9751527309417725, 'sampling/importance_sampling_ratio/min': 0.3771347403526306, 'sampling/importance_sampling_ratio/mean': 1.0033406019210815, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07925032824277878, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.008729460067115724, 'clip_ratio/high_max': 0.008729460067115724, 'clip_ratio/region_mean': 0.012300888658501208, 'reward_total_mean': 0.9987327456474304, 'reward_meter_mean': 0.9987327456474304, 'reward_meter_std': 6.793010834371671e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987327456474304, 'reward_total_composite_std': 6.793010834371671e-05, 'epoch': 0.06}

 49%|████▊     | 1605/3300 [1:53:53<4:56:02, 10.48s/it]INFO 04-12 00:30:44 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1606/3300 [1:53:59<4:16:20,  9.08s/it]2026-04-12 00:30:50,114 | INFO | train_grpo_train | metrics_logged mode=train step=1606

                                                       
{'loss': -0.0107, 'grad_norm': 1.1906522512435913, 'learning_rate': 5.1363636363636375e-06, 'num_tokens': 3605709.0, 'completions/mean_length': 99.625, 'completions/min_length': 99.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.625, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9968845248222351, 'rewards/meter/std': 0.0012903602328151464, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8223729729652405, 'rewards/total_composite/std': 0.06959936022758484, 'reward': 0.8223729729652405, 'reward_std': 0.06959936767816544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006160242948681116, 'sampling/sampling_logp_difference/max': 0.7100238800048828, 'sampling/importance_sampling_ratio/min': 0.49163246154785156, 'sampling/importance_sampling_ratio/mean': 1.002079963684082, 'sampling/importance_sampling_ratio/max': 1.2040398120880127, 'entropy': 0.050112520810216665, 'clip_ratio/low_mean': 0.0012499999720603228, 'clip_ratio/low_min': 0.0012499999720603228, 'clip_ratio/high_mean': 0.0012135922443121672, 'clip_ratio/high_max': 0.0012135922443121672, 'clip_ratio/region_mean': 0.00246359221637249, 'reward_total_mean': 0.8223729729652405, 'reward_meter_mean': 0.9968845248222351, 'reward_meter_std': 0.0012903602328151464, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8223729729652405, 'reward_total_composite_std': 0.06959936022758484, 'epoch': 0.06}

 49%|████▊     | 1606/3300 [1:53:59<4:16:20,  9.08s/it]INFO 04-12 00:30:50 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1607/3300 [1:54:05<3:50:27,  8.17s/it]2026-04-12 00:30:56,158 | INFO | train_grpo_train | metrics_logged mode=train step=1607

                                                       
{'loss': 0.0229, 'grad_norm': 4.926582336425781, 'learning_rate': 5.133333333333334e-06, 'num_tokens': 3608643.0, 'completions/mean_length': 171.75, 'completions/min_length': 163.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.75, 'completions/min_terminated_length': 163.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.937900185585022, 'rewards/meter/std': 0.15882940590381622, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9319444298744202, 'rewards/repeat_penalty/std': 0.08195958286523819, 'rewards/total_composite/mean': 0.8702612519264221, 'rewards/total_composite/std': 0.15232259035110474, 'reward': 0.8702612519264221, 'reward_std': 0.15232259035110474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058821044862270355, 'sampling/sampling_logp_difference/max': 1.4120397567749023, 'sampling/importance_sampling_ratio/min': 0.2436458021402359, 'sampling/importance_sampling_ratio/mean': 1.0150550603866577, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48374373838305473, 'clip_ratio/low_mean': 0.010939412750303745, 'clip_ratio/low_min': 0.010939412750303745, 'clip_ratio/high_mean': 0.035856032744050026, 'clip_ratio/high_max': 0.035856032744050026, 'clip_ratio/region_mean': 0.04679544549435377, 'reward_total_mean': 0.8702612519264221, 'reward_meter_mean': 0.937900185585022, 'reward_meter_std': 0.15882940590381622, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9319444298744202, 'reward_repeat_penalty_std': 0.08195958286523819, 'reward_total_composite_mean': 0.8702612519264221, 'reward_total_composite_std': 0.15232259035110474, 'epoch': 0.06}

 49%|████▊     | 1607/3300 [1:54:05<3:50:27,  8.17s/it]INFO 04-12 00:30:56 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1608/3300 [1:54:11<3:37:27,  7.71s/it]2026-04-12 00:31:02,799 | INFO | train_grpo_train | metrics_logged mode=train step=1608

                                                       
{'loss': -0.02, 'grad_norm': 2.1128990650177, 'learning_rate': 5.130303030303031e-06, 'num_tokens': 3611925.0, 'completions/mean_length': 217.25, 'completions/min_length': 207.0, 'completions/max_length': 239.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 217.25, 'completions/min_terminated_length': 207.0, 'completions/max_terminated_length': 239.0, 'rewards/meter/mean': 0.9952753782272339, 'rewards/meter/std': 0.0022598514333367348, 'rewards/count_adherence/mean': 0.9642857313156128, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250916004180908, 'rewards/repeat_penalty/std': 0.05420750379562378, 'rewards/total_composite/mean': 0.7915375232696533, 'rewards/total_composite/std': 0.07283394038677216, 'reward': 0.7915375232696533, 'reward_std': 0.07283391803503036, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036408815532922745, 'sampling/sampling_logp_difference/max': 8.568058013916016, 'sampling/importance_sampling_ratio/min': 0.00019008143863175064, 'sampling/importance_sampling_ratio/mean': 0.999096691608429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08922967594116926, 'clip_ratio/low_mean': 0.005352868989575654, 'clip_ratio/low_min': 0.005352868989575654, 'clip_ratio/high_mean': 0.014548269449733198, 'clip_ratio/high_max': 0.014548269449733198, 'clip_ratio/region_mean': 0.019901138439308852, 'reward_total_mean': 0.7915375232696533, 'reward_meter_mean': 0.9952753782272339, 'reward_meter_std': 0.0022598514333367348, 'reward_count_adherence_mean': 0.9642857313156128, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250916004180908, 'reward_repeat_penalty_std': 0.05420750379562378, 'reward_total_composite_mean': 0.7915375232696533, 'reward_total_composite_std': 0.07283394038677216, 'epoch': 0.06}

 49%|████▊     | 1608/3300 [1:54:11<3:37:27,  7.71s/it]INFO 04-12 00:31:03 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1609/3300 [1:54:16<3:10:28,  6.76s/it]2026-04-12 00:31:07,334 | INFO | train_grpo_train | metrics_logged mode=train step=1609

                                                       
{'loss': 0.002, 'grad_norm': 3.462772846221924, 'learning_rate': 5.1272727272727275e-06, 'num_tokens': 3613586.0, 'completions/mean_length': 60.625, 'completions/min_length': 59.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9920006990432739, 'rewards/meter/std': 0.00801026076078415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.9506175518035889, 'rewards/total_composite/std': 0.1168404370546341, 'reward': 0.9506175518035889, 'reward_std': 0.11684045940637589, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02973771281540394, 'sampling/sampling_logp_difference/max': 0.9864425659179688, 'sampling/importance_sampling_ratio/min': 0.37290093302726746, 'sampling/importance_sampling_ratio/mean': 1.006797194480896, 'sampling/importance_sampling_ratio/max': 1.5062812566757202, 'entropy': 0.17171061784029007, 'clip_ratio/low_mean': 0.004098360426723957, 'clip_ratio/low_min': 0.004098360426723957, 'clip_ratio/high_mean': 0.016640397254377604, 'clip_ratio/high_max': 0.016640397254377604, 'clip_ratio/region_mean': 0.02073875768110156, 'reward_total_mean': 0.9506175518035889, 'reward_meter_mean': 0.9920006990432739, 'reward_meter_std': 0.00801026076078415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.9506175518035889, 'reward_total_composite_std': 0.1168404370546341, 'epoch': 0.06}

 49%|████▉     | 1609/3300 [1:54:16<3:10:28,  6.76s/it]INFO 04-12 00:31:07 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1610/3300 [1:54:20<2:51:48,  6.10s/it]2026-04-12 00:31:11,893 | INFO | train_grpo_train | metrics_logged mode=train step=1610

                                                       
{'loss': 0.0039, 'grad_norm': 3.0595595836639404, 'learning_rate': 5.124242424242425e-06, 'num_tokens': 3615443.0, 'completions/mean_length': 66.125, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9358171224594116, 'rewards/meter/std': 0.018078165128827095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9358171224594116, 'rewards/total_composite/std': 0.018078165128827095, 'reward': 0.9358171224594116, 'reward_std': 0.01807817816734314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03475549817085266, 'sampling/sampling_logp_difference/max': 1.2117633819580078, 'sampling/importance_sampling_ratio/min': 0.297671914100647, 'sampling/importance_sampling_ratio/mean': 1.0040435791015625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24806034937500954, 'clip_ratio/low_mean': 0.013008192996494472, 'clip_ratio/low_min': 0.013008192996494472, 'clip_ratio/high_mean': 0.013322061393409967, 'clip_ratio/high_max': 0.013322061393409967, 'clip_ratio/region_mean': 0.02633025438990444, 'reward_total_mean': 0.9358171224594116, 'reward_meter_mean': 0.9358171224594116, 'reward_meter_std': 0.018078165128827095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9358171224594116, 'reward_total_composite_std': 0.018078165128827095, 'epoch': 0.06}

 49%|████▉     | 1610/3300 [1:54:20<2:51:48,  6.10s/it]INFO 04-12 00:31:12 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1611/3300 [1:54:26<2:47:46,  5.96s/it]2026-04-12 00:31:17,532 | INFO | train_grpo_train | metrics_logged mode=train step=1611

                                                       
{'loss': 0.0158, 'grad_norm': 4.593719959259033, 'learning_rate': 5.121212121212121e-06, 'num_tokens': 3617986.0, 'completions/mean_length': 141.875, 'completions/min_length': 133.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.875, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.9247380495071411, 'rewards/meter/std': 0.13655264675617218, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.86052405834198, 'rewards/total_composite/std': 0.14909514784812927, 'reward': 0.86052405834198, 'reward_std': 0.14909514784812927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07056324183940887, 'sampling/sampling_logp_difference/max': 7.457266807556152, 'sampling/importance_sampling_ratio/min': 0.0005772316944785416, 'sampling/importance_sampling_ratio/mean': 1.0124247074127197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5591295659542084, 'clip_ratio/low_mean': 0.014760382240638137, 'clip_ratio/low_min': 0.014760382240638137, 'clip_ratio/high_mean': 0.016017435351386666, 'clip_ratio/high_max': 0.016017435351386666, 'clip_ratio/region_mean': 0.030777817592024803, 'reward_total_mean': 0.86052405834198, 'reward_meter_mean': 0.9247380495071411, 'reward_meter_std': 0.13655264675617218, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.86052405834198, 'reward_total_composite_std': 0.14909514784812927, 'epoch': 0.06}

 49%|████▉     | 1611/3300 [1:54:26<2:47:46,  5.96s/it]INFO 04-12 00:31:17 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1612/3300 [1:54:30<2:35:18,  5.52s/it]2026-04-12 00:31:22,031 | INFO | train_grpo_train | metrics_logged mode=train step=1612

                                                       
{'loss': -0.0037, 'grad_norm': 1.593169927597046, 'learning_rate': 5.1181818181818185e-06, 'num_tokens': 3619751.0, 'completions/mean_length': 57.625, 'completions/min_length': 55.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.625, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9934520721435547, 'rewards/meter/std': 3.5445533285383135e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9934520721435547, 'rewards/total_composite/std': 3.5445533285383135e-05, 'reward': 0.9934520721435547, 'reward_std': 3.5451525036478415e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003705526702105999, 'sampling/sampling_logp_difference/max': 0.46508365869522095, 'sampling/importance_sampling_ratio/min': 0.6644558906555176, 'sampling/importance_sampling_ratio/mean': 1.0014033317565918, 'sampling/importance_sampling_ratio/max': 1.5921474695205688, 'entropy': 0.018967063282616436, 'clip_ratio/low_mean': 0.0022727272007614374, 'clip_ratio/low_min': 0.0022727272007614374, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0022727272007614374, 'reward_total_mean': 0.9934520721435547, 'reward_meter_mean': 0.9934520721435547, 'reward_meter_std': 3.5445533285383135e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9934520721435547, 'reward_total_composite_std': 3.5445533285383135e-05, 'epoch': 0.06}

 49%|████▉     | 1612/3300 [1:54:31<2:35:18,  5.52s/it]INFO 04-12 00:31:22 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1613/3300 [1:54:36<2:31:18,  5.38s/it]2026-04-12 00:31:27,082 | INFO | train_grpo_train | metrics_logged mode=train step=1613

                                                       
{'loss': -0.0043, 'grad_norm': 7.702324390411377, 'learning_rate': 5.115151515151515e-06, 'num_tokens': 3621517.0, 'completions/mean_length': 63.75, 'completions/min_length': 63.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9974002838134766, 'rewards/meter/std': 0.0021041962318122387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974002838134766, 'rewards/total_composite/std': 0.0021041962318122387, 'reward': 0.9974002838134766, 'reward_std': 0.002104192739352584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01148869190365076, 'sampling/sampling_logp_difference/max': 0.7638870477676392, 'sampling/importance_sampling_ratio/min': 0.46585214138031006, 'sampling/importance_sampling_ratio/mean': 1.0017887353897095, 'sampling/importance_sampling_ratio/max': 1.317487359046936, 'entropy': 0.056046845857053995, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.001953125, 'reward_total_mean': 0.9974002838134766, 'reward_meter_mean': 0.9974002838134766, 'reward_meter_std': 0.0021041962318122387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974002838134766, 'reward_total_composite_std': 0.0021041962318122387, 'epoch': 0.06}

 49%|████▉     | 1613/3300 [1:54:36<2:31:18,  5.38s/it]INFO 04-12 00:31:27 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1614/3300 [1:54:42<2:37:09,  5.59s/it]2026-04-12 00:31:33,170 | INFO | train_grpo_train | metrics_logged mode=train step=1614

                                                       
{'loss': 0.0583, 'grad_norm': 3.499596118927002, 'learning_rate': 5.112121212121213e-06, 'num_tokens': 3624361.0, 'completions/mean_length': 169.5, 'completions/min_length': 154.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 169.5, 'completions/min_terminated_length': 154.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.9390597343444824, 'rewards/meter/std': 0.12540623545646667, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7876983880996704, 'rewards/repeat_penalty/std': 0.13440276682376862, 'rewards/total_composite/mean': 0.6842913627624512, 'rewards/total_composite/std': 0.13273297250270844, 'reward': 0.6842913627624512, 'reward_std': 0.13273297250270844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05148252472281456, 'sampling/sampling_logp_difference/max': 3.8801069259643555, 'sampling/importance_sampling_ratio/min': 0.020648617297410965, 'sampling/importance_sampling_ratio/mean': 1.012567162513733, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3938692733645439, 'clip_ratio/low_mean': 0.018685899674892426, 'clip_ratio/low_min': 0.018685899674892426, 'clip_ratio/high_mean': 0.02137809677515179, 'clip_ratio/high_max': 0.02137809677515179, 'clip_ratio/region_mean': 0.040063996450044215, 'reward_total_mean': 0.6842913627624512, 'reward_meter_mean': 0.9390597343444824, 'reward_meter_std': 0.12540623545646667, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7876983880996704, 'reward_repeat_penalty_std': 0.13440276682376862, 'reward_total_composite_mean': 0.6842913627624512, 'reward_total_composite_std': 0.13273297250270844, 'epoch': 0.06}

 49%|████▉     | 1614/3300 [1:54:42<2:37:09,  5.59s/it]INFO 04-12 00:31:33 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1615/3300 [1:54:46<2:30:33,  5.36s/it]2026-04-12 00:31:38,016 | INFO | train_grpo_train | metrics_logged mode=train step=1615

                                                       
{'loss': 0.0083, 'grad_norm': 9.031867027282715, 'learning_rate': 5.109090909090909e-06, 'num_tokens': 3626009.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.992936372756958, 'rewards/meter/std': 0.0014942112611606717, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.992936372756958, 'rewards/total_composite/std': 0.0014942112611606717, 'reward': 0.992936372756958, 'reward_std': 0.0014942261623218656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008410189300775528, 'sampling/sampling_logp_difference/max': 0.7060290575027466, 'sampling/importance_sampling_ratio/min': 0.5572132468223572, 'sampling/importance_sampling_ratio/mean': 1.0016261339187622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.02835727483034134, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006465517217293382, 'clip_ratio/high_max': 0.006465517217293382, 'clip_ratio/region_mean': 0.006465517217293382, 'reward_total_mean': 0.992936372756958, 'reward_meter_mean': 0.992936372756958, 'reward_meter_std': 0.0014942112611606717, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.992936372756958, 'reward_total_composite_std': 0.0014942112611606717, 'epoch': 0.06}

 49%|████▉     | 1615/3300 [1:54:47<2:30:33,  5.36s/it]INFO 04-12 00:31:38 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1616/3300 [1:54:51<2:25:43,  5.19s/it]2026-04-12 00:31:42,813 | INFO | train_grpo_train | metrics_logged mode=train step=1616

                                                       
{'loss': 0.02, 'grad_norm': 3.926830530166626, 'learning_rate': 5.106060606060607e-06, 'num_tokens': 3627780.0, 'completions/mean_length': 71.375, 'completions/min_length': 68.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9946064352989197, 'rewards/meter/std': 0.0011085572186857462, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946064352989197, 'rewards/total_composite/std': 0.0011085572186857462, 'reward': 0.9946064352989197, 'reward_std': 0.0011085484875366092, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009593191556632519, 'sampling/sampling_logp_difference/max': 0.8193645477294922, 'sampling/importance_sampling_ratio/min': 0.4407116174697876, 'sampling/importance_sampling_ratio/mean': 1.0016788244247437, 'sampling/importance_sampling_ratio/max': 1.385042428970337, 'entropy': 0.07067765574902296, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0017361111240461469, 'reward_total_mean': 0.9946064352989197, 'reward_meter_mean': 0.9946064352989197, 'reward_meter_std': 0.0011085572186857462, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946064352989197, 'reward_total_composite_std': 0.0011085572186857462, 'epoch': 0.06}

 49%|████▉     | 1616/3300 [1:54:51<2:25:43,  5.19s/it]INFO 04-12 00:31:43 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1617/3300 [1:55:02<3:14:20,  6.93s/it]2026-04-12 00:31:53,769 | INFO | train_grpo_train | metrics_logged mode=train step=1617

                                                       
{'loss': -0.1647, 'grad_norm': 1.3416348695755005, 'learning_rate': 5.103030303030303e-06, 'num_tokens': 3633068.0, 'completions/mean_length': 445.0, 'completions/min_length': 421.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 435.4285888671875, 'completions/min_terminated_length': 421.0, 'completions/max_terminated_length': 459.0, 'rewards/meter/mean': 0.9931340217590332, 'rewards/meter/std': 0.005852298345416784, 'rewards/count_adherence/mean': 0.671875, 'rewards/count_adherence/std': 0.07281029969453812, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.7910888195037842, 'rewards/repeat_penalty/std': 0.1950831413269043, 'rewards/total_composite/mean': 0.4080624282360077, 'rewards/total_composite/std': 0.282764196395874, 'reward': 0.4080624282360077, 'reward_std': 0.282764196395874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04456779360771179, 'sampling/sampling_logp_difference/max': 3.5916011333465576, 'sampling/importance_sampling_ratio/min': 0.02755417674779892, 'sampling/importance_sampling_ratio/mean': 1.0104022026062012, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30827769078314304, 'clip_ratio/low_mean': 0.003343421034514904, 'clip_ratio/low_min': 0.003343421034514904, 'clip_ratio/high_mean': 0.023898042272776365, 'clip_ratio/high_max': 0.023898042272776365, 'clip_ratio/region_mean': 0.02724146330729127, 'reward_total_mean': 0.4080624282360077, 'reward_meter_mean': 0.9931340217590332, 'reward_meter_std': 0.005852298345416784, 'reward_count_adherence_mean': 0.671875, 'reward_count_adherence_std': 0.07281029969453812, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.7910888195037842, 'reward_repeat_penalty_std': 0.1950831413269043, 'reward_total_composite_mean': 0.4080624282360077, 'reward_total_composite_std': 0.282764196395874, 'epoch': 0.06}

 49%|████▉     | 1617/3300 [1:55:02<3:14:20,  6.93s/it]INFO 04-12 00:31:54 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1618/3300 [1:55:11<3:31:18,  7.54s/it]2026-04-12 00:32:02,728 | INFO | train_grpo_train | metrics_logged mode=train step=1618

                                                       
{'loss': -0.0118, 'grad_norm': 1.6642099618911743, 'learning_rate': 5.1e-06, 'num_tokens': 3637584.0, 'completions/mean_length': 389.5, 'completions/min_length': 376.0, 'completions/max_length': 402.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 389.5, 'completions/min_terminated_length': 376.0, 'completions/max_terminated_length': 402.0, 'rewards/meter/mean': 0.9942715167999268, 'rewards/meter/std': 0.005808908957988024, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8421052694320679, 'rewards/repeat_penalty/std': 0.14886459708213806, 'rewards/total_composite/mean': 0.597885012626648, 'rewards/total_composite/std': 0.10485604405403137, 'reward': 0.597885012626648, 'reward_std': 0.10485603660345078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044363927096128464, 'sampling/sampling_logp_difference/max': 2.440901279449463, 'sampling/importance_sampling_ratio/min': 0.08708232641220093, 'sampling/importance_sampling_ratio/mean': 1.0039002895355225, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3395629785954952, 'clip_ratio/low_mean': 0.009485365822911263, 'clip_ratio/low_min': 0.009485365822911263, 'clip_ratio/high_mean': 0.02332985820248723, 'clip_ratio/high_max': 0.02332985820248723, 'clip_ratio/region_mean': 0.03281522402539849, 'reward_total_mean': 0.597885012626648, 'reward_meter_mean': 0.9942715167999268, 'reward_meter_std': 0.005808908957988024, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8421052694320679, 'reward_repeat_penalty_std': 0.14886459708213806, 'reward_total_composite_mean': 0.597885012626648, 'reward_total_composite_std': 0.10485604405403137, 'epoch': 0.06}

 49%|████▉     | 1618/3300 [1:55:11<3:31:18,  7.54s/it]INFO 04-12 00:32:02 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1619/3300 [1:55:16<3:06:00,  6.64s/it]2026-04-12 00:32:07,265 | INFO | train_grpo_train | metrics_logged mode=train step=1619

                                                       
{'loss': 0.0001, 'grad_norm': 10.757038116455078, 'learning_rate': 5.096969696969697e-06, 'num_tokens': 3639232.0, 'completions/mean_length': 59.0, 'completions/min_length': 57.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.0, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9416102766990662, 'rewards/meter/std': 0.15220597386360168, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9416102766990662, 'rewards/total_composite/std': 0.15220597386360168, 'reward': 0.9416102766990662, 'reward_std': 0.1522059589624405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018271297216415405, 'sampling/sampling_logp_difference/max': 0.5301923751831055, 'sampling/importance_sampling_ratio/min': 0.5884917974472046, 'sampling/importance_sampling_ratio/mean': 1.007680892944336, 'sampling/importance_sampling_ratio/max': 1.6727921962738037, 'entropy': 0.12759912386536598, 'clip_ratio/low_mean': 0.0021186440717428923, 'clip_ratio/low_min': 0.0021186440717428923, 'clip_ratio/high_mean': 0.004385964944958687, 'clip_ratio/high_max': 0.004385964944958687, 'clip_ratio/region_mean': 0.006504609016701579, 'reward_total_mean': 0.9416102766990662, 'reward_meter_mean': 0.9416102766990662, 'reward_meter_std': 0.15220597386360168, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9416102766990662, 'reward_total_composite_std': 0.15220597386360168, 'epoch': 0.07}

 49%|████▉     | 1619/3300 [1:55:16<3:06:00,  6.64s/it]INFO 04-12 00:32:07 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1620/3300 [1:55:21<2:51:41,  6.13s/it]2026-04-12 00:32:12,214 | INFO | train_grpo_train | metrics_logged mode=train step=1620

                                                       
{'loss': 0.032, 'grad_norm': 4.442444801330566, 'learning_rate': 5.093939393939395e-06, 'num_tokens': 3641223.0, 'completions/mean_length': 77.875, 'completions/min_length': 75.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9966658353805542, 'rewards/meter/std': 0.0016763400053605437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9966658353805542, 'rewards/total_composite/std': 0.0016763400053605437, 'reward': 0.9966658353805542, 'reward_std': 0.0016763500170782208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06199686974287033, 'sampling/sampling_logp_difference/max': 1.449620246887207, 'sampling/importance_sampling_ratio/min': 0.23465938866138458, 'sampling/importance_sampling_ratio/mean': 1.0116773843765259, 'sampling/importance_sampling_ratio/max': 1.952990174293518, 'entropy': 0.5332462079823017, 'clip_ratio/low_mean': 0.01416793093085289, 'clip_ratio/low_min': 0.01416793093085289, 'clip_ratio/high_mean': 0.02794372313655913, 'clip_ratio/high_max': 0.02794372313655913, 'clip_ratio/region_mean': 0.04211165406741202, 'reward_total_mean': 0.9966658353805542, 'reward_meter_mean': 0.9966658353805542, 'reward_meter_std': 0.0016763400053605437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9966658353805542, 'reward_total_composite_std': 0.0016763400053605437, 'epoch': 0.07}

 49%|████▉     | 1620/3300 [1:55:21<2:51:41,  6.13s/it]INFO 04-12 00:32:12 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1621/3300 [1:55:25<2:40:40,  5.74s/it]2026-04-12 00:32:17,046 | INFO | train_grpo_train | metrics_logged mode=train step=1621

                                                       
{'loss': 0.2585, 'grad_norm': 9.210960388183594, 'learning_rate': 5.090909090909091e-06, 'num_tokens': 3642899.0, 'completions/mean_length': 46.5, 'completions/min_length': 39.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.959256649017334, 'rewards/meter/std': 0.07027505338191986, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.959256649017334, 'rewards/total_composite/std': 0.07027505338191986, 'reward': 0.959256649017334, 'reward_std': 0.07027503848075867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0790342465043068, 'sampling/sampling_logp_difference/max': 1.2574310302734375, 'sampling/importance_sampling_ratio/min': 0.2843836545944214, 'sampling/importance_sampling_ratio/mean': 1.0092580318450928, 'sampling/importance_sampling_ratio/max': 1.8360445499420166, 'entropy': 0.7850339710712433, 'clip_ratio/low_mean': 0.02026098920032382, 'clip_ratio/low_min': 0.02026098920032382, 'clip_ratio/high_mean': 0.0513387790415436, 'clip_ratio/high_max': 0.0513387790415436, 'clip_ratio/region_mean': 0.07159976824186742, 'reward_total_mean': 0.959256649017334, 'reward_meter_mean': 0.959256649017334, 'reward_meter_std': 0.07027505338191986, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.959256649017334, 'reward_total_composite_std': 0.07027505338191986, 'epoch': 0.07}

 49%|████▉     | 1621/3300 [1:55:26<2:40:40,  5.74s/it]INFO 04-12 00:32:17 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1622/3300 [1:55:30<2:33:50,  5.50s/it]2026-04-12 00:32:21,989 | INFO | train_grpo_train | metrics_logged mode=train step=1622

                                                       
{'loss': -0.1515, 'grad_norm': 5.405808925628662, 'learning_rate': 5.0878787878787885e-06, 'num_tokens': 3644592.0, 'completions/mean_length': 55.625, 'completions/min_length': 31.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.993726372718811, 'rewards/meter/std': 0.0034743899013847113, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9314955472946167, 'rewards/total_composite/std': 0.17525316774845123, 'reward': 0.9314955472946167, 'reward_std': 0.17525316774845123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02844316139817238, 'sampling/sampling_logp_difference/max': 1.0102331638336182, 'sampling/importance_sampling_ratio/min': 0.3641340732574463, 'sampling/importance_sampling_ratio/mean': 1.0055111646652222, 'sampling/importance_sampling_ratio/max': 1.6441330909729004, 'entropy': 0.16878793388605118, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.01896783267147839, 'clip_ratio/high_max': 0.01896783267147839, 'clip_ratio/region_mean': 0.01896783267147839, 'reward_total_mean': 0.9314955472946167, 'reward_meter_mean': 0.993726372718811, 'reward_meter_std': 0.0034743899013847113, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9314955472946167, 'reward_total_composite_std': 0.17525316774845123, 'epoch': 0.07}

 49%|████▉     | 1622/3300 [1:55:30<2:33:50,  5.50s/it]INFO 04-12 00:32:22 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1623/3300 [1:55:35<2:27:45,  5.29s/it]2026-04-12 00:32:26,775 | INFO | train_grpo_train | metrics_logged mode=train step=1623

                                                       
{'loss': 0.0053, 'grad_norm': 4.352970123291016, 'learning_rate': 5.084848484848486e-06, 'num_tokens': 3646304.0, 'completions/mean_length': 72.0, 'completions/min_length': 69.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8835433125495911, 'rewards/meter/std': 0.2843259274959564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8835433125495911, 'rewards/total_composite/std': 0.2843259274959564, 'reward': 0.8835433125495911, 'reward_std': 0.2843259274959564, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.061763398349285126, 'sampling/sampling_logp_difference/max': 2.2054567337036133, 'sampling/importance_sampling_ratio/min': 0.1102001741528511, 'sampling/importance_sampling_ratio/mean': 1.0209115743637085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6054626666009426, 'clip_ratio/low_mean': 0.0051369862630963326, 'clip_ratio/low_min': 0.0051369862630963326, 'clip_ratio/high_mean': 0.04867990920320153, 'clip_ratio/high_max': 0.04867990920320153, 'clip_ratio/region_mean': 0.053816895466297865, 'reward_total_mean': 0.8835433125495911, 'reward_meter_mean': 0.8835433125495911, 'reward_meter_std': 0.2843259274959564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8835433125495911, 'reward_total_composite_std': 0.2843259274959564, 'epoch': 0.07}

 49%|████▉     | 1623/3300 [1:55:35<2:27:45,  5.29s/it]INFO 04-12 00:32:27 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1624/3300 [1:55:44<2:57:39,  6.36s/it]2026-04-12 00:32:35,637 | INFO | train_grpo_train | metrics_logged mode=train step=1624

                                                       
{'loss': -0.0003, 'grad_norm': 2.14442777633667, 'learning_rate': 5.081818181818182e-06, 'num_tokens': 3650768.0, 'completions/mean_length': 336.0, 'completions/min_length': 306.0, 'completions/max_length': 347.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 336.0, 'completions/min_terminated_length': 306.0, 'completions/max_terminated_length': 347.0, 'rewards/meter/mean': 0.9964303374290466, 'rewards/meter/std': 0.0028279528487473726, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.046290989965200424, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9401960372924805, 'rewards/repeat_penalty/std': 0.04455278813838959, 'rewards/total_composite/mean': 0.8190490007400513, 'rewards/total_composite/std': 0.04663979262113571, 'reward': 0.8190490007400513, 'reward_std': 0.04663977771997452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058607399463653564, 'sampling/sampling_logp_difference/max': 1.395674467086792, 'sampling/importance_sampling_ratio/min': 0.24766594171524048, 'sampling/importance_sampling_ratio/mean': 1.01429283618927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5225833132863045, 'clip_ratio/low_mean': 0.016402944223955274, 'clip_ratio/low_min': 0.016402944223955274, 'clip_ratio/high_mean': 0.022054610773921013, 'clip_ratio/high_max': 0.022054610773921013, 'clip_ratio/region_mean': 0.038457554997876287, 'reward_total_mean': 0.8190490007400513, 'reward_meter_mean': 0.9964303374290466, 'reward_meter_std': 0.0028279528487473726, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.046290989965200424, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9401960372924805, 'reward_repeat_penalty_std': 0.04455278813838959, 'reward_total_composite_mean': 0.8190490007400513, 'reward_total_composite_std': 0.04663979262113571, 'epoch': 0.07}

 49%|████▉     | 1624/3300 [1:55:44<2:57:39,  6.36s/it]INFO 04-12 00:32:35 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1625/3300 [1:55:50<2:55:50,  6.30s/it]2026-04-12 00:32:41,819 | INFO | train_grpo_train | metrics_logged mode=train step=1625

                                                       
{'loss': -0.0375, 'grad_norm': 4.290684223175049, 'learning_rate': 5.078787878787879e-06, 'num_tokens': 3653643.0, 'completions/mean_length': 171.375, 'completions/min_length': 142.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.375, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9575715065002441, 'rewards/meter/std': 0.10626022517681122, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9682539701461792, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.8372185230255127, 'rewards/total_composite/std': 0.15995363891124725, 'reward': 0.8372185230255127, 'reward_std': 0.15995362401008606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09049484878778458, 'sampling/sampling_logp_difference/max': 3.279475450515747, 'sampling/importance_sampling_ratio/min': 0.037647999823093414, 'sampling/importance_sampling_ratio/mean': 1.009111762046814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7438356392085552, 'clip_ratio/low_mean': 0.02405611891299486, 'clip_ratio/low_min': 0.02405611891299486, 'clip_ratio/high_mean': 0.04015351925045252, 'clip_ratio/high_max': 0.04015351925045252, 'clip_ratio/region_mean': 0.06420963816344738, 'reward_total_mean': 0.8372185230255127, 'reward_meter_mean': 0.9575715065002441, 'reward_meter_std': 0.10626022517681122, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9682539701461792, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.8372185230255127, 'reward_total_composite_std': 0.15995363891124725, 'epoch': 0.07}

 49%|████▉     | 1625/3300 [1:55:50<2:55:50,  6.30s/it]INFO 04-12 00:32:42 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1626/3300 [1:55:55<2:45:33,  5.93s/it]2026-04-12 00:32:46,905 | INFO | train_grpo_train | metrics_logged mode=train step=1626

                                                       
{'loss': 0.0078, 'grad_norm': 3.822707414627075, 'learning_rate': 5.075757575757576e-06, 'num_tokens': 3655935.0, 'completions/mean_length': 96.5, 'completions/min_length': 92.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.5, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9368202686309814, 'rewards/meter/std': 0.026791444048285484, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9368202686309814, 'rewards/total_composite/std': 0.026791444048285484, 'reward': 0.9368202686309814, 'reward_std': 0.02679145336151123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05230702832341194, 'sampling/sampling_logp_difference/max': 0.9677839279174805, 'sampling/importance_sampling_ratio/min': 0.37992405891418457, 'sampling/importance_sampling_ratio/mean': 1.0122920274734497, 'sampling/importance_sampling_ratio/max': 1.796831727027893, 'entropy': 0.45825108513236046, 'clip_ratio/low_mean': 0.017135416390374303, 'clip_ratio/low_min': 0.017135416390374303, 'clip_ratio/high_mean': 0.01032647315878421, 'clip_ratio/high_max': 0.01032647315878421, 'clip_ratio/region_mean': 0.027461889549158514, 'reward_total_mean': 0.9368202686309814, 'reward_meter_mean': 0.9368202686309814, 'reward_meter_std': 0.026791444048285484, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9368202686309814, 'reward_total_composite_std': 0.026791444048285484, 'epoch': 0.07}

 49%|████▉     | 1626/3300 [1:55:55<2:45:33,  5.93s/it]INFO 04-12 00:32:47 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1627/3300 [1:56:00<2:35:00,  5.56s/it]2026-04-12 00:32:51,561 | INFO | train_grpo_train | metrics_logged mode=train step=1627

                                                       
{'loss': 0.0042, 'grad_norm': 6.298683166503906, 'learning_rate': 5.072727272727274e-06, 'num_tokens': 3657666.0, 'completions/mean_length': 59.375, 'completions/min_length': 57.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.375, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8852929472923279, 'rewards/meter/std': 0.30933257937431335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8852929472923279, 'rewards/total_composite/std': 0.30933257937431335, 'reward': 0.8852929472923279, 'reward_std': 0.30933254957199097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025638654828071594, 'sampling/sampling_logp_difference/max': 0.9846286773681641, 'sampling/importance_sampling_ratio/min': 0.37357792258262634, 'sampling/importance_sampling_ratio/mean': 1.0004394054412842, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1392963044345379, 'clip_ratio/low_mean': 0.004098360426723957, 'clip_ratio/low_min': 0.004098360426723957, 'clip_ratio/high_mean': 0.014909721445292234, 'clip_ratio/high_max': 0.014909721445292234, 'clip_ratio/region_mean': 0.01900808187201619, 'reward_total_mean': 0.8852929472923279, 'reward_meter_mean': 0.8852929472923279, 'reward_meter_std': 0.30933257937431335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8852929472923279, 'reward_total_composite_std': 0.30933257937431335, 'epoch': 0.07}

 49%|████▉     | 1627/3300 [1:56:00<2:35:00,  5.56s/it]INFO 04-12 00:32:51 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1628/3300 [1:56:04<2:23:52,  5.16s/it]2026-04-12 00:32:55,797 | INFO | train_grpo_train | metrics_logged mode=train step=1628

                                                       
{'loss': 0.0099, 'grad_norm': 3.436232328414917, 'learning_rate': 5.06969696969697e-06, 'num_tokens': 3659244.0, 'completions/mean_length': 35.25, 'completions/min_length': 35.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9971714019775391, 'rewards/meter/std': 0.0016039953334257007, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971714019775391, 'rewards/total_composite/std': 0.0016039953334257007, 'reward': 0.9971714019775391, 'reward_std': 0.0016039892798289657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0144999660551548, 'sampling/sampling_logp_difference/max': 0.8395180702209473, 'sampling/importance_sampling_ratio/min': 0.4319186508655548, 'sampling/importance_sampling_ratio/mean': 0.9952727556228638, 'sampling/importance_sampling_ratio/max': 1.063112735748291, 'entropy': 0.060785280307754874, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.0035714285913854837, 'clip_ratio/high_max': 0.0035714285913854837, 'clip_ratio/region_mean': 0.006949807051569223, 'reward_total_mean': 0.9971714019775391, 'reward_meter_mean': 0.9971714019775391, 'reward_meter_std': 0.0016039953334257007, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971714019775391, 'reward_total_composite_std': 0.0016039953334257007, 'epoch': 0.07}

 49%|████▉     | 1628/3300 [1:56:04<2:23:52,  5.16s/it]INFO 04-12 00:32:56 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1629/3300 [1:56:12<2:48:21,  6.05s/it]2026-04-12 00:33:03,902 | INFO | train_grpo_train | metrics_logged mode=train step=1629

                                                       
{'loss': -0.0053, 'grad_norm': 3.441396474838257, 'learning_rate': 5.0666666666666676e-06, 'num_tokens': 3663869.0, 'completions/mean_length': 339.125, 'completions/min_length': 321.0, 'completions/max_length': 360.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 339.125, 'completions/min_terminated_length': 321.0, 'completions/max_terminated_length': 360.0, 'rewards/meter/mean': 0.9954209923744202, 'rewards/meter/std': 0.0042012035846710205, 'rewards/count_adherence/mean': 0.8068181276321411, 'rewards/count_adherence/std': 0.03214120864868164, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9562908411026001, 'rewards/repeat_penalty/std': 0.060786258429288864, 'rewards/total_composite/mean': 0.7674725651741028, 'rewards/total_composite/std': 0.04974528029561043, 'reward': 0.7674725651741028, 'reward_std': 0.04974529147148132, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06467485427856445, 'sampling/sampling_logp_difference/max': 1.5753440856933594, 'sampling/importance_sampling_ratio/min': 0.2069363296031952, 'sampling/importance_sampling_ratio/mean': 1.0179475545883179, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.622920136898756, 'clip_ratio/low_mean': 0.004980936297215521, 'clip_ratio/low_min': 0.004980936297215521, 'clip_ratio/high_mean': 0.03499942785128951, 'clip_ratio/high_max': 0.03499942785128951, 'clip_ratio/region_mean': 0.03998036414850503, 'reward_total_mean': 0.7674725651741028, 'reward_meter_mean': 0.9954209923744202, 'reward_meter_std': 0.0042012035846710205, 'reward_count_adherence_mean': 0.8068181276321411, 'reward_count_adherence_std': 0.03214120864868164, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9562908411026001, 'reward_repeat_penalty_std': 0.060786258429288864, 'reward_total_composite_mean': 0.7674725651741028, 'reward_total_composite_std': 0.04974528029561043, 'epoch': 0.07}

 49%|████▉     | 1629/3300 [1:56:12<2:48:21,  6.05s/it]INFO 04-12 00:33:04 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1630/3300 [1:56:21<3:11:56,  6.90s/it]2026-04-12 00:33:12,786 | INFO | train_grpo_train | metrics_logged mode=train step=1630

                                                       
{'loss': 0.01, 'grad_norm': 2.8746509552001953, 'learning_rate': 5.063636363636364e-06, 'num_tokens': 3668262.0, 'completions/mean_length': 343.125, 'completions/min_length': 326.0, 'completions/max_length': 355.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 343.125, 'completions/min_terminated_length': 326.0, 'completions/max_terminated_length': 355.0, 'rewards/meter/mean': 0.9573882222175598, 'rewards/meter/std': 0.10380185395479202, 'rewards/count_adherence/mean': 0.862500011920929, 'rewards/count_adherence/std': 0.0517548993229866, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9774816036224365, 'rewards/repeat_penalty/std': 0.04422420263290405, 'rewards/total_composite/mean': 0.8051279783248901, 'rewards/total_composite/std': 0.09284678101539612, 'reward': 0.8051279783248901, 'reward_std': 0.09284677356481552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07694988697767258, 'sampling/sampling_logp_difference/max': 1.7180042266845703, 'sampling/importance_sampling_ratio/min': 0.17942388355731964, 'sampling/importance_sampling_ratio/mean': 1.016446590423584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7744914256036282, 'clip_ratio/low_mean': 0.031975225545465946, 'clip_ratio/low_min': 0.031975225545465946, 'clip_ratio/high_mean': 0.02247373666614294, 'clip_ratio/high_max': 0.02247373666614294, 'clip_ratio/region_mean': 0.05444896221160889, 'reward_total_mean': 0.8051279783248901, 'reward_meter_mean': 0.9573882222175598, 'reward_meter_std': 0.10380185395479202, 'reward_count_adherence_mean': 0.862500011920929, 'reward_count_adherence_std': 0.0517548993229866, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9774816036224365, 'reward_repeat_penalty_std': 0.04422420263290405, 'reward_total_composite_mean': 0.8051279783248901, 'reward_total_composite_std': 0.09284678101539612, 'epoch': 0.07}

 49%|████▉     | 1630/3300 [1:56:21<3:11:56,  6.90s/it]INFO 04-12 00:33:13 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1631/3300 [1:56:26<2:58:09,  6.41s/it]2026-04-12 00:33:18,042 | INFO | train_grpo_train | metrics_logged mode=train step=1631

                                                       
{'loss': 0.0001, 'grad_norm': 0.9008557200431824, 'learning_rate': 5.060606060606061e-06, 'num_tokens': 3670677.0, 'completions/mean_length': 127.875, 'completions/min_length': 127.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.875, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9939456582069397, 'rewards/meter/std': 0.00018129698582924902, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5963674187660217, 'rewards/total_composite/std': 0.00010878406465053558, 'reward': 0.5963674187660217, 'reward_std': 0.0001087786877178587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007339623291045427, 'sampling/sampling_logp_difference/max': 0.8770111799240112, 'sampling/importance_sampling_ratio/min': 0.41602447628974915, 'sampling/importance_sampling_ratio/mean': 1.0000206232070923, 'sampling/importance_sampling_ratio/max': 1.6699923276901245, 'entropy': 0.02229404100216925, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.006843626964837313, 'clip_ratio/high_max': 0.006843626964837313, 'clip_ratio/region_mean': 0.010749876964837313, 'reward_total_mean': 0.5963674187660217, 'reward_meter_mean': 0.9939456582069397, 'reward_meter_std': 0.00018129698582924902, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5963674187660217, 'reward_total_composite_std': 0.00010878406465053558, 'epoch': 0.07}

 49%|████▉     | 1631/3300 [1:56:27<2:58:09,  6.41s/it]INFO 04-12 00:33:18 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1632/3300 [1:56:31<2:43:49,  5.89s/it]2026-04-12 00:33:22,743 | INFO | train_grpo_train | metrics_logged mode=train step=1632

                                                       
{'loss': -0.0158, 'grad_norm': 4.735329627990723, 'learning_rate': 5.057575757575758e-06, 'num_tokens': 3672574.0, 'completions/mean_length': 73.125, 'completions/min_length': 67.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9905960559844971, 'rewards/meter/std': 0.008815782144665718, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9905960559844971, 'rewards/total_composite/std': 0.008815782144665718, 'reward': 0.9905960559844971, 'reward_std': 0.008815782144665718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07114405184984207, 'sampling/sampling_logp_difference/max': 1.1808457374572754, 'sampling/importance_sampling_ratio/min': 0.3070189952850342, 'sampling/importance_sampling_ratio/mean': 1.0182366371154785, 'sampling/importance_sampling_ratio/max': 1.9160219430923462, 'entropy': 0.7365500181913376, 'clip_ratio/low_mean': 0.027231683605350554, 'clip_ratio/low_min': 0.027231683605350554, 'clip_ratio/high_mean': 0.04253894090652466, 'clip_ratio/high_max': 0.04253894090652466, 'clip_ratio/region_mean': 0.06977062451187521, 'reward_total_mean': 0.9905960559844971, 'reward_meter_mean': 0.9905960559844971, 'reward_meter_std': 0.008815782144665718, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9905960559844971, 'reward_total_composite_std': 0.008815782144665718, 'epoch': 0.07}

 49%|████▉     | 1632/3300 [1:56:31<2:43:49,  5.89s/it]INFO 04-12 00:33:22 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1633/3300 [1:56:35<2:28:05,  5.33s/it]2026-04-12 00:33:26,782 | INFO | train_grpo_train | metrics_logged mode=train step=1633

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.054545454545455e-06, 'num_tokens': 3674102.0, 'completions/mean_length': 30.0, 'completions/min_length': 30.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.992271900177002, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.992271900177002, 'rewards/total_composite/std': 0.0, 'reward': 0.992271900177002, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005566918407566845, 'sampling/sampling_logp_difference/max': 0.005868114531040192, 'sampling/importance_sampling_ratio/min': 0.998640239238739, 'sampling/importance_sampling_ratio/mean': 1.0005319118499756, 'sampling/importance_sampling_ratio/max': 1.0058854818344116, 'entropy': 0.004795511835254729, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.992271900177002, 'reward_meter_mean': 0.992271900177002, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.992271900177002, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 49%|████▉     | 1633/3300 [1:56:35<2:28:05,  5.33s/it]INFO 04-12 00:33:27 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1634/3300 [1:56:40<2:26:26,  5.27s/it]2026-04-12 00:33:31,901 | INFO | train_grpo_train | metrics_logged mode=train step=1634

                                                       
{'loss': 0.0068, 'grad_norm': 5.207538604736328, 'learning_rate': 5.051515151515151e-06, 'num_tokens': 3676259.0, 'completions/mean_length': 108.625, 'completions/min_length': 103.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.625, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.9962839484214783, 'rewards/meter/std': 0.0024302289821207523, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9962839484214783, 'rewards/total_composite/std': 0.0024302289821207523, 'reward': 0.9962839484214783, 'reward_std': 0.0024302229285240173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0648764818906784, 'sampling/sampling_logp_difference/max': 1.4173831939697266, 'sampling/importance_sampling_ratio/min': 0.2423473745584488, 'sampling/importance_sampling_ratio/mean': 1.0127630233764648, 'sampling/importance_sampling_ratio/max': 1.8952765464782715, 'entropy': 0.6268335022032261, 'clip_ratio/low_mean': 0.013648775406181812, 'clip_ratio/low_min': 0.013648775406181812, 'clip_ratio/high_mean': 0.03815131215378642, 'clip_ratio/high_max': 0.03815131215378642, 'clip_ratio/region_mean': 0.05180008755996823, 'reward_total_mean': 0.9962839484214783, 'reward_meter_mean': 0.9962839484214783, 'reward_meter_std': 0.0024302289821207523, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9962839484214783, 'reward_total_composite_std': 0.0024302289821207523, 'epoch': 0.07}

 50%|████▉     | 1634/3300 [1:56:40<2:26:26,  5.27s/it]INFO 04-12 00:33:32 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1635/3300 [1:56:46<2:29:52,  5.40s/it]2026-04-12 00:33:37,598 | INFO | train_grpo_train | metrics_logged mode=train step=1635

                                                       
{'loss': -0.0025, 'grad_norm': 4.150503158569336, 'learning_rate': 5.048484848484849e-06, 'num_tokens': 3679020.0, 'completions/mean_length': 160.125, 'completions/min_length': 153.0, 'completions/max_length': 166.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.125, 'completions/min_terminated_length': 153.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.9920204877853394, 'rewards/meter/std': 0.01226998120546341, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.974323570728302, 'rewards/total_composite/std': 0.05192466080188751, 'reward': 0.974323570728302, 'reward_std': 0.05192466452717781, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08837065100669861, 'sampling/sampling_logp_difference/max': 1.2528629302978516, 'sampling/importance_sampling_ratio/min': 0.2856857180595398, 'sampling/importance_sampling_ratio/mean': 1.0256816148757935, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9452630281448364, 'clip_ratio/low_mean': 0.01331305643543601, 'clip_ratio/low_min': 0.01331305643543601, 'clip_ratio/high_mean': 0.042927493108436465, 'clip_ratio/high_max': 0.042927493108436465, 'clip_ratio/region_mean': 0.056240549543872476, 'reward_total_mean': 0.974323570728302, 'reward_meter_mean': 0.9920204877853394, 'reward_meter_std': 0.01226998120546341, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.974323570728302, 'reward_total_composite_std': 0.05192466080188751, 'epoch': 0.07}

 50%|████▉     | 1635/3300 [1:56:46<2:29:52,  5.40s/it]INFO 04-12 00:33:37 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1636/3300 [1:56:51<2:26:33,  5.28s/it]2026-04-12 00:33:42,635 | INFO | train_grpo_train | metrics_logged mode=train step=1636

                                                       
{'loss': 0.0004, 'grad_norm': 2.887805223464966, 'learning_rate': 5.045454545454546e-06, 'num_tokens': 3680811.0, 'completions/mean_length': 63.875, 'completions/min_length': 63.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9980711936950684, 'rewards/meter/std': 0.0005036595975980163, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980711936950684, 'rewards/total_composite/std': 0.0005036595975980163, 'reward': 0.9980711936950684, 'reward_std': 0.0005036554648540914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019482869654893875, 'sampling/sampling_logp_difference/max': 0.7397842407226562, 'sampling/importance_sampling_ratio/min': 0.477216899394989, 'sampling/importance_sampling_ratio/mean': 0.9980214238166809, 'sampling/importance_sampling_ratio/max': 1.4901469945907593, 'entropy': 0.09926960244774818, 'clip_ratio/low_mean': 0.007905506063252687, 'clip_ratio/low_min': 0.007905506063252687, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.011811756063252687, 'reward_total_mean': 0.9980711936950684, 'reward_meter_mean': 0.9980711936950684, 'reward_meter_std': 0.0005036595975980163, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980711936950684, 'reward_total_composite_std': 0.0005036595975980163, 'epoch': 0.07}

 50%|████▉     | 1636/3300 [1:56:51<2:26:33,  5.28s/it]INFO 04-12 00:33:42 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1637/3300 [1:56:57<2:35:51,  5.62s/it]2026-04-12 00:33:49,029 | INFO | train_grpo_train | metrics_logged mode=train step=1637

                                                       
{'loss': 0.0281, 'grad_norm': 7.507004737854004, 'learning_rate': 5.042424242424243e-06, 'num_tokens': 3682330.0, 'completions/mean_length': 69.875, 'completions/min_length': 63.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6862971782684326, 'rewards/meter/std': 0.2807905673980713, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6862971782684326, 'rewards/total_composite/std': 0.2807905673980713, 'reward': 0.6862971782684326, 'reward_std': 0.2807905673980713, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10850861668586731, 'sampling/sampling_logp_difference/max': 1.5064477920532227, 'sampling/importance_sampling_ratio/min': 0.22169609367847443, 'sampling/importance_sampling_ratio/mean': 1.0039104223251343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0178619101643562, 'clip_ratio/low_mean': 0.026575686410069466, 'clip_ratio/low_min': 0.026575686410069466, 'clip_ratio/high_mean': 0.05837191268801689, 'clip_ratio/high_max': 0.05837191268801689, 'clip_ratio/region_mean': 0.08494759909808636, 'reward_total_mean': 0.6862971782684326, 'reward_meter_mean': 0.6862971782684326, 'reward_meter_std': 0.2807905673980713, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6862971782684326, 'reward_total_composite_std': 0.2807905673980713, 'epoch': 0.07}

 50%|████▉     | 1637/3300 [1:56:58<2:35:51,  5.62s/it]INFO 04-12 00:33:49 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1638/3300 [1:57:02<2:30:12,  5.42s/it]2026-04-12 00:33:53,981 | INFO | train_grpo_train | metrics_logged mode=train step=1638

                                                       
{'loss': 0.0097, 'grad_norm': 3.420053720474243, 'learning_rate': 5.0393939393939395e-06, 'num_tokens': 3684458.0, 'completions/mean_length': 95.0, 'completions/min_length': 94.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.0, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9959671497344971, 'rewards/meter/std': 0.006246180739253759, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8965927362442017, 'rewards/total_composite/std': 0.1087099239230156, 'reward': 0.8965927362442017, 'reward_std': 0.10870993137359619, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01706601306796074, 'sampling/sampling_logp_difference/max': 0.9803454875946045, 'sampling/importance_sampling_ratio/min': 0.6205868124961853, 'sampling/importance_sampling_ratio/mean': 1.0042842626571655, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10523929260671139, 'clip_ratio/low_mean': 0.006431827903725207, 'clip_ratio/low_min': 0.006431827903725207, 'clip_ratio/high_mean': 0.010610593715682626, 'clip_ratio/high_max': 0.010610593715682626, 'clip_ratio/region_mean': 0.017042421619407833, 'reward_total_mean': 0.8965927362442017, 'reward_meter_mean': 0.9959671497344971, 'reward_meter_std': 0.006246180739253759, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8965927362442017, 'reward_total_composite_std': 0.1087099239230156, 'epoch': 0.07}

 50%|████▉     | 1638/3300 [1:57:02<2:30:12,  5.42s/it]INFO 04-12 00:33:54 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1639/3300 [1:57:07<2:22:13,  5.14s/it]2026-04-12 00:33:58,454 | INFO | train_grpo_train | metrics_logged mode=train step=1639

                                                       
{'loss': 0.0003, 'grad_norm': 1.6942023038864136, 'learning_rate': 5.036363636363637e-06, 'num_tokens': 3686354.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9983768463134766, 'rewards/meter/std': 0.00010835672583198175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9983768463134766, 'rewards/total_composite/std': 0.00010835672583198175, 'reward': 0.9983768463134766, 'reward_std': 0.00010834616841748357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01173328422009945, 'sampling/sampling_logp_difference/max': 0.8327422142028809, 'sampling/importance_sampling_ratio/min': 0.434855192899704, 'sampling/importance_sampling_ratio/mean': 1.0026438236236572, 'sampling/importance_sampling_ratio/max': 1.363644003868103, 'entropy': 0.06925144325941801, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.005859375, 'clip_ratio/high_max': 0.005859375, 'clip_ratio/region_mean': 0.009765625, 'reward_total_mean': 0.9983768463134766, 'reward_meter_mean': 0.9983768463134766, 'reward_meter_std': 0.00010835672583198175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9983768463134766, 'reward_total_composite_std': 0.00010835672583198175, 'epoch': 0.07}

 50%|████▉     | 1639/3300 [1:57:07<2:22:13,  5.14s/it]INFO 04-12 00:33:58 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1640/3300 [1:57:12<2:17:46,  4.98s/it]2026-04-12 00:34:03,068 | INFO | train_grpo_train | metrics_logged mode=train step=1640

                                                       
{'loss': 0.0239, 'grad_norm': 5.066521167755127, 'learning_rate': 5.033333333333333e-06, 'num_tokens': 3688136.0, 'completions/mean_length': 69.75, 'completions/min_length': 67.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9961869716644287, 'rewards/meter/std': 0.0016419171588495374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961869716644287, 'rewards/total_composite/std': 0.0016419171588495374, 'reward': 0.9961869716644287, 'reward_std': 0.0016419151797890663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036743052303791046, 'sampling/sampling_logp_difference/max': 0.8399066925048828, 'sampling/importance_sampling_ratio/min': 0.4317508041858673, 'sampling/importance_sampling_ratio/mean': 1.0081746578216553, 'sampling/importance_sampling_ratio/max': 1.939770221710205, 'entropy': 0.2406612578779459, 'clip_ratio/low_mean': 0.010499144555069506, 'clip_ratio/low_min': 0.010499144555069506, 'clip_ratio/high_mean': 0.012254902278073132, 'clip_ratio/high_max': 0.012254902278073132, 'clip_ratio/region_mean': 0.022754046833142638, 'reward_total_mean': 0.9961869716644287, 'reward_meter_mean': 0.9961869716644287, 'reward_meter_std': 0.0016419171588495374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961869716644287, 'reward_total_composite_std': 0.0016419171588495374, 'epoch': 0.07}

 50%|████▉     | 1640/3300 [1:57:12<2:17:46,  4.98s/it]INFO 04-12 00:34:03 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1641/3300 [1:57:16<2:15:50,  4.91s/it]2026-04-12 00:34:07,823 | INFO | train_grpo_train | metrics_logged mode=train step=1641

                                                       
{'loss': -0.0356, 'grad_norm': 6.707535266876221, 'learning_rate': 5.030303030303031e-06, 'num_tokens': 3690011.0, 'completions/mean_length': 69.375, 'completions/min_length': 61.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.375, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.7832512855529785, 'rewards/meter/std': 0.33663251996040344, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7832512855529785, 'rewards/total_composite/std': 0.33663251996040344, 'reward': 0.7832512855529785, 'reward_std': 0.33663249015808105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09360064566135406, 'sampling/sampling_logp_difference/max': 1.031529426574707, 'sampling/importance_sampling_ratio/min': 0.35646137595176697, 'sampling/importance_sampling_ratio/mean': 1.0248838663101196, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9592136070132256, 'clip_ratio/low_mean': 0.01592771988362074, 'clip_ratio/low_min': 0.01592771988362074, 'clip_ratio/high_mean': 0.047490152064710855, 'clip_ratio/high_max': 0.047490152064710855, 'clip_ratio/region_mean': 0.0634178719483316, 'reward_total_mean': 0.7832512855529785, 'reward_meter_mean': 0.7832512855529785, 'reward_meter_std': 0.33663251996040344, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7832512855529785, 'reward_total_composite_std': 0.33663251996040344, 'epoch': 0.07}

 50%|████▉     | 1641/3300 [1:57:16<2:15:50,  4.91s/it]INFO 04-12 00:34:08 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1642/3300 [1:57:21<2:11:23,  4.75s/it]2026-04-12 00:34:12,210 | INFO | train_grpo_train | metrics_logged mode=train step=1642

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.027272727272728e-06, 'num_tokens': 3691659.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.993520200252533, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.993520200252533, 'rewards/total_composite/std': 0.0, 'reward': 0.993520200252533, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005443372065201402, 'sampling/sampling_logp_difference/max': 0.014515344053506851, 'sampling/importance_sampling_ratio/min': 0.9973773956298828, 'sampling/importance_sampling_ratio/mean': 1.0005112886428833, 'sampling/importance_sampling_ratio/max': 1.0146211385726929, 'entropy': 0.0062332607340067625, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.993520200252533, 'reward_meter_mean': 0.993520200252533, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.993520200252533, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 50%|████▉     | 1642/3300 [1:57:21<2:11:23,  4.75s/it]INFO 04-12 00:34:12 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1643/3300 [1:57:27<2:26:04,  5.29s/it]2026-04-12 00:34:18,750 | INFO | train_grpo_train | metrics_logged mode=train step=1643

                                                       
{'loss': 0.0248, 'grad_norm': 3.824038028717041, 'learning_rate': 5.024242424242425e-06, 'num_tokens': 3695039.0, 'completions/mean_length': 203.5, 'completions/min_length': 193.0, 'completions/max_length': 227.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 203.5, 'completions/min_terminated_length': 193.0, 'completions/max_terminated_length': 227.0, 'rewards/meter/mean': 0.9925307035446167, 'rewards/meter/std': 0.010234549641609192, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9925307035446167, 'rewards/total_composite/std': 0.010234549641609192, 'reward': 0.9925307035446167, 'reward_std': 0.010234540328383446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07599979639053345, 'sampling/sampling_logp_difference/max': 1.208749771118164, 'sampling/importance_sampling_ratio/min': 0.29857030510902405, 'sampling/importance_sampling_ratio/mean': 1.0239778757095337, 'sampling/importance_sampling_ratio/max': 1.9295165538787842, 'entropy': 0.8424399569630623, 'clip_ratio/low_mean': 0.009563126135617495, 'clip_ratio/low_min': 0.009563126135617495, 'clip_ratio/high_mean': 0.042821857146918774, 'clip_ratio/high_max': 0.042821857146918774, 'clip_ratio/region_mean': 0.05238498328253627, 'reward_total_mean': 0.9925307035446167, 'reward_meter_mean': 0.9925307035446167, 'reward_meter_std': 0.010234549641609192, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9925307035446167, 'reward_total_composite_std': 0.010234549641609192, 'epoch': 0.07}

 50%|████▉     | 1643/3300 [1:57:27<2:26:04,  5.29s/it]INFO 04-12 00:34:18 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1644/3300 [1:57:33<2:30:58,  5.47s/it]2026-04-12 00:34:24,641 | INFO | train_grpo_train | metrics_logged mode=train step=1644

                                                       
{'loss': -0.0099, 'grad_norm': 3.962242603302002, 'learning_rate': 5.021212121212121e-06, 'num_tokens': 3697742.0, 'completions/mean_length': 146.875, 'completions/min_length': 142.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 146.875, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9914488792419434, 'rewards/meter/std': 0.019252995029091835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9379702806472778, 'rewards/total_composite/std': 0.07060353457927704, 'reward': 0.9379702806472778, 'reward_std': 0.07060354202985764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0685059130191803, 'sampling/sampling_logp_difference/max': 1.3913602828979492, 'sampling/importance_sampling_ratio/min': 0.24873672425746918, 'sampling/importance_sampling_ratio/mean': 1.0191352367401123, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6761485747992992, 'clip_ratio/low_mean': 0.016500307247042656, 'clip_ratio/low_min': 0.016500307247042656, 'clip_ratio/high_mean': 0.037598957773298025, 'clip_ratio/high_max': 0.037598957773298025, 'clip_ratio/region_mean': 0.05409926502034068, 'reward_total_mean': 0.9379702806472778, 'reward_meter_mean': 0.9914488792419434, 'reward_meter_std': 0.019252995029091835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9379702806472778, 'reward_total_composite_std': 0.07060353457927704, 'epoch': 0.07}

 50%|████▉     | 1644/3300 [1:57:33<2:30:58,  5.47s/it]INFO 04-12 00:34:24 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1645/3300 [1:57:38<2:24:48,  5.25s/it]2026-04-12 00:34:29,372 | INFO | train_grpo_train | metrics_logged mode=train step=1645

                                                       
{'loss': -0.0102, 'grad_norm': 7.402265548706055, 'learning_rate': 5.0181818181818186e-06, 'num_tokens': 3699539.0, 'completions/mean_length': 69.625, 'completions/min_length': 62.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.6198569536209106, 'rewards/meter/std': 0.3495922088623047, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6198569536209106, 'rewards/total_composite/std': 0.3495922088623047, 'reward': 0.6198569536209106, 'reward_std': 0.3495922088623047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10739893466234207, 'sampling/sampling_logp_difference/max': 1.626225471496582, 'sampling/importance_sampling_ratio/min': 0.1966705024242401, 'sampling/importance_sampling_ratio/mean': 1.0240029096603394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.080611765384674, 'clip_ratio/low_mean': 0.022980431327596307, 'clip_ratio/low_min': 0.022980431327596307, 'clip_ratio/high_mean': 0.053518022410571575, 'clip_ratio/high_max': 0.053518022410571575, 'clip_ratio/region_mean': 0.07649845373816788, 'reward_total_mean': 0.6198569536209106, 'reward_meter_mean': 0.6198569536209106, 'reward_meter_std': 0.3495922088623047, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6198569536209106, 'reward_total_composite_std': 0.3495922088623047, 'epoch': 0.07}

 50%|████▉     | 1645/3300 [1:57:38<2:24:48,  5.25s/it]INFO 04-12 00:34:29 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1646/3300 [1:57:42<2:17:15,  4.98s/it]2026-04-12 00:34:33,720 | INFO | train_grpo_train | metrics_logged mode=train step=1646

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.015151515151515e-06, 'num_tokens': 3701219.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.993520200252533, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.993520200252533, 'rewards/total_composite/std': 0.0, 'reward': 0.993520200252533, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0016902285860851407, 'sampling/sampling_logp_difference/max': 0.07047711312770844, 'sampling/importance_sampling_ratio/min': 0.9319490194320679, 'sampling/importance_sampling_ratio/mean': 1.0011212825775146, 'sampling/importance_sampling_ratio/max': 1.0615943670272827, 'entropy': 0.010349510470405221, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.993520200252533, 'reward_meter_mean': 0.993520200252533, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.993520200252533, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 50%|████▉     | 1646/3300 [1:57:42<2:17:15,  4.98s/it]INFO 04-12 00:34:33 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1647/3300 [1:57:47<2:19:18,  5.06s/it]2026-04-12 00:34:38,957 | INFO | train_grpo_train | metrics_logged mode=train step=1647

                                                       
{'loss': 0.0056, 'grad_norm': 4.599454879760742, 'learning_rate': 5.012121212121212e-06, 'num_tokens': 3703551.0, 'completions/mean_length': 115.5, 'completions/min_length': 110.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.5, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.989443302154541, 'rewards/meter/std': 0.009337217546999454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.989443302154541, 'rewards/total_composite/std': 0.009337217546999454, 'reward': 0.989443302154541, 'reward_std': 0.009337219409644604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07534082978963852, 'sampling/sampling_logp_difference/max': 1.2979240417480469, 'sampling/importance_sampling_ratio/min': 0.2730981409549713, 'sampling/importance_sampling_ratio/mean': 1.0189144611358643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7444454878568649, 'clip_ratio/low_mean': 0.02626075316220522, 'clip_ratio/low_min': 0.02626075316220522, 'clip_ratio/high_mean': 0.025925687979906797, 'clip_ratio/high_max': 0.025925687979906797, 'clip_ratio/region_mean': 0.05218644114211202, 'reward_total_mean': 0.989443302154541, 'reward_meter_mean': 0.989443302154541, 'reward_meter_std': 0.009337217546999454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.989443302154541, 'reward_total_composite_std': 0.009337217546999454, 'epoch': 0.07}

 50%|████▉     | 1647/3300 [1:57:47<2:19:18,  5.06s/it]INFO 04-12 00:34:39 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1648/3300 [1:57:52<2:18:57,  5.05s/it]2026-04-12 00:34:43,982 | INFO | train_grpo_train | metrics_logged mode=train step=1648

                                                       
{'loss': 0.0453, 'grad_norm': 5.8860039710998535, 'learning_rate': 5.009090909090909e-06, 'num_tokens': 3705642.0, 'completions/mean_length': 103.375, 'completions/min_length': 97.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.375, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.877294659614563, 'rewards/meter/std': 0.19824109971523285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.877294659614563, 'rewards/total_composite/std': 0.19824109971523285, 'reward': 0.877294659614563, 'reward_std': 0.19824109971523285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10711432248353958, 'sampling/sampling_logp_difference/max': 1.7170472145080566, 'sampling/importance_sampling_ratio/min': 0.1795956790447235, 'sampling/importance_sampling_ratio/mean': 1.031799077987671, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.183668315410614, 'clip_ratio/low_mean': 0.01907266629859805, 'clip_ratio/low_min': 0.01907266629859805, 'clip_ratio/high_mean': 0.06016605906188488, 'clip_ratio/high_max': 0.06016605906188488, 'clip_ratio/region_mean': 0.07923872536048293, 'reward_total_mean': 0.877294659614563, 'reward_meter_mean': 0.877294659614563, 'reward_meter_std': 0.19824109971523285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.877294659614563, 'reward_total_composite_std': 0.19824109971523285, 'epoch': 0.07}

 50%|████▉     | 1648/3300 [1:57:52<2:18:57,  5.05s/it]INFO 04-12 00:34:44 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1649/3300 [1:57:57<2:11:03,  4.76s/it]2026-04-12 00:34:48,078 | INFO | train_grpo_train | metrics_logged mode=train step=1649

                                                       
{'loss': -0.0025, 'grad_norm': 5.83959436416626, 'learning_rate': 5.006060606060607e-06, 'num_tokens': 3707194.0, 'completions/mean_length': 32.0, 'completions/min_length': 31.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9442778825759888, 'rewards/meter/std': 0.059931062161922455, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9442778825759888, 'rewards/total_composite/std': 0.059931062161922455, 'reward': 0.9442778825759888, 'reward_std': 0.059931062161922455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05596788227558136, 'sampling/sampling_logp_difference/max': 1.9147157669067383, 'sampling/importance_sampling_ratio/min': 0.14738371968269348, 'sampling/importance_sampling_ratio/mean': 1.007331371307373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4391027558594942, 'clip_ratio/low_mean': 0.008064515888690948, 'clip_ratio/low_min': 0.008064515888690948, 'clip_ratio/high_mean': 0.03531280532479286, 'clip_ratio/high_max': 0.03531280532479286, 'clip_ratio/region_mean': 0.04337732121348381, 'reward_total_mean': 0.9442778825759888, 'reward_meter_mean': 0.9442778825759888, 'reward_meter_std': 0.059931062161922455, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9442778825759888, 'reward_total_composite_std': 0.059931062161922455, 'epoch': 0.07}

 50%|████▉     | 1649/3300 [1:57:57<2:11:03,  4.76s/it]INFO 04-12 00:34:48 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1650/3300 [1:58:01<2:08:48,  4.68s/it]2026-04-12 00:34:52,587 | INFO | train_grpo_train | metrics_logged mode=train step=1650

                                                       
{'loss': -0.0102, 'grad_norm': 4.691352367401123, 'learning_rate': 5.003030303030303e-06, 'num_tokens': 3708819.0, 'completions/mean_length': 32.125, 'completions/min_length': 31.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9763838648796082, 'rewards/meter/std': 0.04593488201498985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9763838648796082, 'rewards/total_composite/std': 0.04593488201498985, 'reward': 0.9763838648796082, 'reward_std': 0.04593489319086075, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05742637813091278, 'sampling/sampling_logp_difference/max': 1.3233098983764648, 'sampling/importance_sampling_ratio/min': 0.2662525773048401, 'sampling/importance_sampling_ratio/mean': 1.0103769302368164, 'sampling/importance_sampling_ratio/max': 1.4046815633773804, 'entropy': 0.44733541272580624, 'clip_ratio/low_mean': 0.004032257944345474, 'clip_ratio/low_min': 0.004032257944345474, 'clip_ratio/high_mean': 0.0337981628254056, 'clip_ratio/high_max': 0.0337981628254056, 'clip_ratio/region_mean': 0.03783042076975107, 'reward_total_mean': 0.9763838648796082, 'reward_meter_mean': 0.9763838648796082, 'reward_meter_std': 0.04593488201498985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9763838648796082, 'reward_total_composite_std': 0.04593488201498985, 'epoch': 0.07}

 50%|█████     | 1650/3300 [1:58:01<2:08:48,  4.68s/it]INFO 04-12 00:34:52 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:33,  3.03s/it]

 23%|██▎       | 3/13 [00:12<00:42,  4.29s/it]

 31%|███       | 4/13 [00:17<00:43,  4.83s/it]

 38%|███▊      | 5/13 [00:23<00:40,  5.10s/it]

 46%|████▌     | 6/13 [00:29<00:37,  5.38s/it]

 54%|█████▍    | 7/13 [00:34<00:31,  5.24s/it]

 62%|██████▏   | 8/13 [00:41<00:29,  5.89s/it]

 69%|██████▉   | 9/13 [00:48<00:24,  6.14s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.71s/it]

 85%|████████▍ | 11/13 [01:00<00:12,  6.33s/it]

 92%|█████████▏| 12/13 [01:08<00:06,  6.59s/it]

100%|██████████| 13/13 [01:12<00:00,  6.10s/it]2026-04-12 00:36:11,915 | INFO | train_grpo_train | metrics_logged mode=eval step=1650

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 79.2708, 'eval_samples_per_second': 1.312, 'eval_steps_per_second': 0.164, 'eval_num_tokens': 3708819.0, 'eval_completions/mean_length': 212.91346153846155, 'eval_completions/min_length': 62.0, 'eval_completions/max_length': 409.3076923076923, 'eval_completions/clipped_ratio': 0.028846153846153848, 'eval_completions/mean_terminated_length': 204.67307927058295, 'eval_completions/min_terminated_length': 62.0, 'eval_completions/max_terminated_length': 372.9230769230769, 'eval_rewards/meter/mean': 0.7244124962733343, 'eval_rewards/meter/std': 0.36815990120745623, 'eval_rewards/count_adherence/mean': 0.9108829039793748, 'eval_rewards/count_adherence/std': 0.11557848694232795, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.8433380264502305, 'eval_rewards/repeat_penalty/std': 0.15947996452450752, 'eval_rewards/total_composite/mean': 0.5518424786054171, 'eval_rewards/total_composite/std': 0.34955332485529095, 'eval_reward': 0.5518424786054171, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.035442069316139586, 'eval_sampling/sampling_logp_difference/max': 1.0809908371705275, 'eval_sampling/importance_sampling_ratio/min': 0.3488806807077848, 'eval_sampling/importance_sampling_ratio/mean': 1.009941733800448, 'eval_sampling/importance_sampling_ratio/max': 1.5815464166494517, 'eval_entropy': 0.4149218121400246, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5518424786054171, 'eval_reward_meter_mean': 0.7244124962733343, 'eval_reward_meter_std': 0.36815990120745623, 'eval_reward_count_adherence_mean': 0.9108829039793748, 'eval_reward_count_adherence_std': 0.11557848694232795, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.8433380264502305, 'eval_reward_repeat_penalty_std': 0.15947996452450752, 'eval_reward_total_composite_mean': 0.5518424786054171, 'eval_reward_total_composite_std': 0.34955332485529095, 'epoch': 0.07}

 50%|█████     | 1650/3300 [1:59:20<2:08:48,  4.68s/it]

100%|██████████| 13/13 [01:13<00:00,  6.10s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 50%|█████     | 1651/3300 [1:59:27<13:22:13, 29.19s/it]2026-04-12 00:36:18,953 | INFO | train_grpo_train | metrics_logged mode=train step=1651

                                                        
{'loss': 0.0001, 'grad_norm': 0.5019415616989136, 'learning_rate': 5e-06, 'num_tokens': 3710592.0, 'completions/mean_length': 67.625, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9974769949913025, 'rewards/meter/std': 0.001003190758638084, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974769949913025, 'rewards/total_composite/std': 0.001003190758638084, 'reward': 0.9974769949913025, 'reward_std': 0.001003190758638084, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009622362442314625, 'sampling/sampling_logp_difference/max': 0.6310257911682129, 'sampling/importance_sampling_ratio/min': 0.5320457816123962, 'sampling/importance_sampling_ratio/mean': 1.0023994445800781, 'sampling/importance_sampling_ratio/max': 1.4636121988296509, 'entropy': 0.06585087720304728, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.005569578497670591, 'clip_ratio/high_max': 0.005569578497670591, 'clip_ratio/region_mean': 0.007407813798636198, 'reward_total_mean': 0.9974769949913025, 'reward_meter_mean': 0.9974769949913025, 'reward_meter_std': 0.001003190758638084, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974769949913025, 'reward_total_composite_std': 0.001003190758638084, 'epoch': 0.07}

 50%|█████     | 1651/3300 [1:59:27<13:22:13, 29.19s/it]INFO 04-12 00:36:19 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1652/3300 [1:59:32<10:01:24, 21.90s/it]2026-04-12 00:36:23,827 | INFO | train_grpo_train | metrics_logged mode=train step=1652

                                                        
{'loss': -0.0038, 'grad_norm': 1.1626874208450317, 'learning_rate': 4.996969696969698e-06, 'num_tokens': 3712383.0, 'completions/mean_length': 67.875, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9978716373443604, 'rewards/meter/std': 7.145507697714493e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978716373443604, 'rewards/total_composite/std': 7.145507697714493e-05, 'reward': 0.9978716373443604, 'reward_std': 7.145912968553603e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011869369074702263, 'sampling/sampling_logp_difference/max': 0.9046740531921387, 'sampling/importance_sampling_ratio/min': 0.4046737849712372, 'sampling/importance_sampling_ratio/mean': 1.0000650882720947, 'sampling/importance_sampling_ratio/max': 1.2595540285110474, 'entropy': 0.07074726792052388, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007299659075215459, 'clip_ratio/high_max': 0.007299659075215459, 'clip_ratio/region_mean': 0.007299659075215459, 'reward_total_mean': 0.9978716373443604, 'reward_meter_mean': 0.9978716373443604, 'reward_meter_std': 7.145507697714493e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978716373443604, 'reward_total_composite_std': 7.145507697714493e-05, 'epoch': 0.07}

 50%|█████     | 1652/3300 [1:59:32<10:01:24, 21.90s/it]INFO 04-12 00:36:24 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1653/3300 [1:59:43<8:28:03, 18.51s/it] 2026-04-12 00:36:34,429 | INFO | train_grpo_train | metrics_logged mode=train step=1653

                                                       
{'loss': -0.3504, 'grad_norm': 1.3737280368804932, 'learning_rate': 4.993939393939394e-06, 'num_tokens': 3716159.0, 'completions/mean_length': 382.0, 'completions/min_length': 321.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 338.66668701171875, 'completions/min_terminated_length': 321.0, 'completions/max_terminated_length': 361.0, 'rewards/meter/mean': 0.8152369260787964, 'rewards/meter/std': 0.1942194551229477, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.9428104758262634, 'rewards/repeat_penalty/std': 0.05349903926253319, 'rewards/total_composite/mean': 0.5758238434791565, 'rewards/total_composite/std': 0.36915671825408936, 'reward': 0.5758238434791565, 'reward_std': 0.36915671825408936, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08613347262144089, 'sampling/sampling_logp_difference/max': 2.176515579223633, 'sampling/importance_sampling_ratio/min': 0.11343610286712646, 'sampling/importance_sampling_ratio/mean': 1.0216100215911865, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7133785486221313, 'clip_ratio/low_mean': 0.007788162212818861, 'clip_ratio/low_min': 0.007788162212818861, 'clip_ratio/high_mean': 0.03797963773831725, 'clip_ratio/high_max': 0.03797963773831725, 'clip_ratio/region_mean': 0.04576779995113611, 'reward_total_mean': 0.5758238434791565, 'reward_meter_mean': 0.8152369260787964, 'reward_meter_std': 0.1942194551229477, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.9428104758262634, 'reward_repeat_penalty_std': 0.05349903926253319, 'reward_total_composite_mean': 0.5758238434791565, 'reward_total_composite_std': 0.36915671825408936, 'epoch': 0.07}

 50%|█████     | 1653/3300 [1:59:43<8:28:03, 18.51s/it]INFO 04-12 00:36:34 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1654/3300 [1:59:48<6:35:33, 14.42s/it]2026-04-12 00:36:39,336 | INFO | train_grpo_train | metrics_logged mode=train step=1654

                                                       
{'loss': -0.0232, 'grad_norm': 3.5595715045928955, 'learning_rate': 4.990909090909091e-06, 'num_tokens': 3718004.0, 'completions/mean_length': 63.625, 'completions/min_length': 62.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9683968424797058, 'rewards/meter/std': 0.024647340178489685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9683968424797058, 'rewards/total_composite/std': 0.024647340178489685, 'reward': 0.9683968424797058, 'reward_std': 0.024647342041134834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03070976585149765, 'sampling/sampling_logp_difference/max': 5.06722354888916, 'sampling/importance_sampling_ratio/min': 0.00629988731816411, 'sampling/importance_sampling_ratio/mean': 0.9969826936721802, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0451802066527307, 'clip_ratio/low_mean': 0.010080644860863686, 'clip_ratio/low_min': 0.010080644860863686, 'clip_ratio/high_mean': 0.00562528264708817, 'clip_ratio/high_max': 0.00562528264708817, 'clip_ratio/region_mean': 0.015705927507951856, 'reward_total_mean': 0.9683968424797058, 'reward_meter_mean': 0.9683968424797058, 'reward_meter_std': 0.024647340178489685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9683968424797058, 'reward_total_composite_std': 0.024647340178489685, 'epoch': 0.07}

 50%|█████     | 1654/3300 [1:59:48<6:35:33, 14.42s/it]INFO 04-12 00:36:39 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1655/3300 [1:59:53<5:16:52, 11.56s/it]2026-04-12 00:36:44,220 | INFO | train_grpo_train | metrics_logged mode=train step=1655

                                                       
{'loss': -0.0142, 'grad_norm': 5.8942670822143555, 'learning_rate': 4.987878787878789e-06, 'num_tokens': 3719438.0, 'completions/mean_length': 33.25, 'completions/min_length': 32.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9977890849113464, 'rewards/meter/std': 0.002695126226171851, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977890849113464, 'rewards/total_composite/std': 0.002695126226171851, 'reward': 0.9977890849113464, 'reward_std': 0.002695116912946105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0360952690243721, 'sampling/sampling_logp_difference/max': 4.130129814147949, 'sampling/importance_sampling_ratio/min': 0.016080791130661964, 'sampling/importance_sampling_ratio/mean': 0.9885497093200684, 'sampling/importance_sampling_ratio/max': 1.4004013538360596, 'entropy': 0.03417817200534046, 'clip_ratio/low_mean': 0.0078125, 'clip_ratio/low_min': 0.0078125, 'clip_ratio/high_mean': 0.014928699005395174, 'clip_ratio/high_max': 0.014928699005395174, 'clip_ratio/region_mean': 0.022741199005395174, 'reward_total_mean': 0.9977890849113464, 'reward_meter_mean': 0.9977890849113464, 'reward_meter_std': 0.002695126226171851, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977890849113464, 'reward_total_composite_std': 0.002695126226171851, 'epoch': 0.07}

 50%|█████     | 1655/3300 [1:59:53<5:16:52, 11.56s/it]INFO 04-12 00:36:44 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1656/3300 [2:00:03<5:04:10, 11.10s/it]2026-04-12 00:36:54,226 | INFO | train_grpo_train | metrics_logged mode=train step=1656

                                                       
{'loss': -0.2605, 'grad_norm': 1.0378773212432861, 'learning_rate': 4.984848484848485e-06, 'num_tokens': 3722366.0, 'completions/mean_length': 235.0, 'completions/min_length': 182.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 195.42857360839844, 'completions/min_terminated_length': 182.0, 'completions/max_terminated_length': 214.0, 'rewards/meter/mean': 0.8787662982940674, 'rewards/meter/std': 0.3354223370552063, 'rewards/count_adherence/mean': 0.925000011920929, 'rewards/count_adherence/std': 0.2121320217847824, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9861111044883728, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.8588160872459412, 'rewards/total_composite/std': 0.34913378953933716, 'reward': 0.8588160872459412, 'reward_std': 0.34913378953933716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0746423751115799, 'sampling/sampling_logp_difference/max': 1.7693357467651367, 'sampling/importance_sampling_ratio/min': 0.17044617235660553, 'sampling/importance_sampling_ratio/mean': 1.0213563442230225, 'sampling/importance_sampling_ratio/max': 1.8245017528533936, 'entropy': 0.6871326193213463, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.046408120542764664, 'clip_ratio/high_max': 0.046408120542764664, 'clip_ratio/region_mean': 0.046408120542764664, 'reward_total_mean': 0.8588160872459412, 'reward_meter_mean': 0.8787662982940674, 'reward_meter_std': 0.3354223370552063, 'reward_count_adherence_mean': 0.925000011920929, 'reward_count_adherence_std': 0.2121320217847824, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9861111044883728, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.8588160872459412, 'reward_total_composite_std': 0.34913378953933716, 'epoch': 0.07}

 50%|█████     | 1656/3300 [2:00:03<5:04:10, 11.10s/it]INFO 04-12 00:36:54 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1657/3300 [2:00:07<4:12:14,  9.21s/it]2026-04-12 00:36:59,038 | INFO | train_grpo_train | metrics_logged mode=train step=1657

                                                       
{'loss': -0.0037, 'grad_norm': 6.6080169677734375, 'learning_rate': 4.981818181818182e-06, 'num_tokens': 3724136.0, 'completions/mean_length': 63.25, 'completions/min_length': 62.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.25, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9078664779663086, 'rewards/meter/std': 0.15767432749271393, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9078664779663086, 'rewards/total_composite/std': 0.15767432749271393, 'reward': 0.9078664779663086, 'reward_std': 0.15767431259155273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043513912707567215, 'sampling/sampling_logp_difference/max': 1.9472687244415283, 'sampling/importance_sampling_ratio/min': 0.14266319572925568, 'sampling/importance_sampling_ratio/mean': 0.998699963092804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3440733216702938, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.021861399058252573, 'clip_ratio/high_max': 0.021861399058252573, 'clip_ratio/region_mean': 0.021861399058252573, 'reward_total_mean': 0.9078664779663086, 'reward_meter_mean': 0.9078664779663086, 'reward_meter_std': 0.15767432749271393, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9078664779663086, 'reward_total_composite_std': 0.15767432749271393, 'epoch': 0.07}

 50%|█████     | 1657/3300 [2:00:08<4:12:14,  9.21s/it]INFO 04-12 00:36:59 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1658/3300 [2:00:12<3:36:10,  7.90s/it]2026-04-12 00:37:03,875 | INFO | train_grpo_train | metrics_logged mode=train step=1658

                                                       
{'loss': 0.0035, 'grad_norm': 2.866943597793579, 'learning_rate': 4.978787878787879e-06, 'num_tokens': 3726020.0, 'completions/mean_length': 64.5, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9981748461723328, 'rewards/meter/std': 7.74198560975492e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981748461723328, 'rewards/total_composite/std': 7.74198560975492e-05, 'reward': 0.9981748461723328, 'reward_std': 7.74198560975492e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007607936859130859, 'sampling/sampling_logp_difference/max': 1.5065312385559082, 'sampling/importance_sampling_ratio/min': 0.22167760133743286, 'sampling/importance_sampling_ratio/mean': 0.9998639822006226, 'sampling/importance_sampling_ratio/max': 1.3977038860321045, 'entropy': 0.02129516121931374, 'clip_ratio/low_mean': 0.001923076924867928, 'clip_ratio/low_min': 0.001923076924867928, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.003876201924867928, 'reward_total_mean': 0.9981748461723328, 'reward_meter_mean': 0.9981748461723328, 'reward_meter_std': 7.74198560975492e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981748461723328, 'reward_total_composite_std': 7.74198560975492e-05, 'epoch': 0.07}

 50%|█████     | 1658/3300 [2:00:12<3:36:10,  7.90s/it]INFO 04-12 00:37:04 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1659/3300 [2:00:17<3:10:45,  6.97s/it]2026-04-12 00:37:08,691 | INFO | train_grpo_train | metrics_logged mode=train step=1659

                                                       
{'loss': 0.0129, 'grad_norm': 5.0424885749816895, 'learning_rate': 4.975757575757576e-06, 'num_tokens': 3727828.0, 'completions/mean_length': 68.0, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9952208995819092, 'rewards/meter/std': 0.004935278091579676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952208995819092, 'rewards/total_composite/std': 0.004935278091579676, 'reward': 0.9952208995819092, 'reward_std': 0.004935278091579676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009829939343035221, 'sampling/sampling_logp_difference/max': 0.8581647872924805, 'sampling/importance_sampling_ratio/min': 0.42393937706947327, 'sampling/importance_sampling_ratio/mean': 1.0028293132781982, 'sampling/importance_sampling_ratio/max': 1.3186924457550049, 'entropy': 0.06244664313271642, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005542142200283706, 'clip_ratio/high_max': 0.005542142200283706, 'clip_ratio/region_mean': 0.005542142200283706, 'reward_total_mean': 0.9952208995819092, 'reward_meter_mean': 0.9952208995819092, 'reward_meter_std': 0.004935278091579676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952208995819092, 'reward_total_composite_std': 0.004935278091579676, 'epoch': 0.07}

 50%|█████     | 1659/3300 [2:00:17<3:10:45,  6.97s/it]INFO 04-12 00:37:08 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1660/3300 [2:00:23<2:59:31,  6.57s/it]2026-04-12 00:37:14,304 | INFO | train_grpo_train | metrics_logged mode=train step=1660

                                                       
{'loss': -0.0015, 'grad_norm': 0.4592190384864807, 'learning_rate': 4.972727272727273e-06, 'num_tokens': 3730317.0, 'completions/mean_length': 139.125, 'completions/min_length': 139.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 139.125, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9948010444641113, 'rewards/meter/std': 0.004489070735871792, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7461007833480835, 'rewards/total_composite/std': 0.003366807708516717, 'reward': 0.7461007833480835, 'reward_std': 0.0033668053802102804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004387721884995699, 'sampling/sampling_logp_difference/max': 1.0749011039733887, 'sampling/importance_sampling_ratio/min': 0.34133151173591614, 'sampling/importance_sampling_ratio/mean': 0.9996078014373779, 'sampling/importance_sampling_ratio/max': 1.1272794008255005, 'entropy': 0.016455981065519154, 'clip_ratio/low_mean': 0.0008992805960588157, 'clip_ratio/low_min': 0.0008992805960588157, 'clip_ratio/high_mean': 0.0017985611921176314, 'clip_ratio/high_max': 0.0017985611921176314, 'clip_ratio/region_mean': 0.002697841788176447, 'reward_total_mean': 0.7461007833480835, 'reward_meter_mean': 0.9948010444641113, 'reward_meter_std': 0.004489070735871792, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7461007833480835, 'reward_total_composite_std': 0.003366807708516717, 'epoch': 0.07}

 50%|█████     | 1660/3300 [2:00:23<2:59:31,  6.57s/it]INFO 04-12 00:37:14 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1661/3300 [2:00:28<2:44:58,  6.04s/it]2026-04-12 00:37:19,117 | INFO | train_grpo_train | metrics_logged mode=train step=1661

                                                       
{'loss': 0.0154, 'grad_norm': 5.378631114959717, 'learning_rate': 4.9696969696969696e-06, 'num_tokens': 3732140.0, 'completions/mean_length': 58.875, 'completions/min_length': 58.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9912177324295044, 'rewards/meter/std': 0.010196500457823277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9912177324295044, 'rewards/total_composite/std': 0.010196500457823277, 'reward': 0.9912177324295044, 'reward_std': 0.010196508839726448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015276388265192509, 'sampling/sampling_logp_difference/max': 0.687659740447998, 'sampling/importance_sampling_ratio/min': 0.5027512907981873, 'sampling/importance_sampling_ratio/mean': 1.0054762363433838, 'sampling/importance_sampling_ratio/max': 1.3040835857391357, 'entropy': 0.14373003505170345, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.008511104620993137, 'clip_ratio/high_max': 0.008511104620993137, 'clip_ratio/region_mean': 0.008511104620993137, 'reward_total_mean': 0.9912177324295044, 'reward_meter_mean': 0.9912177324295044, 'reward_meter_std': 0.010196500457823277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9912177324295044, 'reward_total_composite_std': 0.010196500457823277, 'epoch': 0.07}

 50%|█████     | 1661/3300 [2:00:28<2:44:58,  6.04s/it]INFO 04-12 00:37:19 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1662/3300 [2:00:32<2:33:58,  5.64s/it]2026-04-12 00:37:23,818 | INFO | train_grpo_train | metrics_logged mode=train step=1662

                                                       
{'loss': 0.0131, 'grad_norm': 3.579711675643921, 'learning_rate': 4.966666666666667e-06, 'num_tokens': 3733822.0, 'completions/mean_length': 59.25, 'completions/min_length': 58.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9940330982208252, 'rewards/meter/std': 0.002611657604575157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9940330982208252, 'rewards/total_composite/std': 0.002611657604575157, 'reward': 0.9940330982208252, 'reward_std': 0.002611662959679961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015829751268029213, 'sampling/sampling_logp_difference/max': 0.7537941932678223, 'sampling/importance_sampling_ratio/min': 0.4705776870250702, 'sampling/importance_sampling_ratio/mean': 1.0035570859909058, 'sampling/importance_sampling_ratio/max': 1.5864135026931763, 'entropy': 0.1442255014553666, 'clip_ratio/low_mean': 0.008064515888690948, 'clip_ratio/low_min': 0.008064515888690948, 'clip_ratio/high_mean': 0.008474576286971569, 'clip_ratio/high_max': 0.008474576286971569, 'clip_ratio/region_mean': 0.016539092175662518, 'reward_total_mean': 0.9940330982208252, 'reward_meter_mean': 0.9940330982208252, 'reward_meter_std': 0.002611657604575157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9940330982208252, 'reward_total_composite_std': 0.002611657604575157, 'epoch': 0.07}

 50%|█████     | 1662/3300 [2:00:32<2:33:58,  5.64s/it]INFO 04-12 00:37:24 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1663/3300 [2:00:37<2:28:17,  5.44s/it]2026-04-12 00:37:28,774 | INFO | train_grpo_train | metrics_logged mode=train step=1663

                                                       
{'loss': 0.0149, 'grad_norm': 4.979900360107422, 'learning_rate': 4.963636363636364e-06, 'num_tokens': 3735532.0, 'completions/mean_length': 72.75, 'completions/min_length': 69.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9954915046691895, 'rewards/meter/std': 0.002510088263079524, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9954915046691895, 'rewards/total_composite/std': 0.002510088263079524, 'reward': 0.9954915046691895, 'reward_std': 0.002510099671781063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06602619588375092, 'sampling/sampling_logp_difference/max': 1.1701393127441406, 'sampling/importance_sampling_ratio/min': 0.31032371520996094, 'sampling/importance_sampling_ratio/mean': 1.0033316612243652, 'sampling/importance_sampling_ratio/max': 1.7458137273788452, 'entropy': 0.49573158100247383, 'clip_ratio/low_mean': 0.010420631850138307, 'clip_ratio/low_min': 0.010420631850138307, 'clip_ratio/high_mean': 0.03229976072907448, 'clip_ratio/high_max': 0.03229976072907448, 'clip_ratio/region_mean': 0.042720392579212785, 'reward_total_mean': 0.9954915046691895, 'reward_meter_mean': 0.9954915046691895, 'reward_meter_std': 0.002510088263079524, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9954915046691895, 'reward_total_composite_std': 0.002510088263079524, 'epoch': 0.07}

 50%|█████     | 1663/3300 [2:00:37<2:28:17,  5.44s/it]INFO 04-12 00:37:29 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1664/3300 [2:00:43<2:27:04,  5.39s/it]2026-04-12 00:37:34,070 | INFO | train_grpo_train | metrics_logged mode=train step=1664

                                                       
{'loss': -0.0026, 'grad_norm': 3.2529075145721436, 'learning_rate': 4.9606060606060605e-06, 'num_tokens': 3737762.0, 'completions/mean_length': 109.75, 'completions/min_length': 102.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.75, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9983501434326172, 'rewards/meter/std': 0.000316560355713591, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9983501434326172, 'rewards/total_composite/std': 0.000316560355713591, 'reward': 0.9983501434326172, 'reward_std': 0.0003165603266097605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05091632902622223, 'sampling/sampling_logp_difference/max': 1.1261930465698242, 'sampling/importance_sampling_ratio/min': 0.32426539063453674, 'sampling/importance_sampling_ratio/mean': 1.0035275220870972, 'sampling/importance_sampling_ratio/max': 1.7752270698547363, 'entropy': 0.3692820519208908, 'clip_ratio/low_mean': 0.004454409005120397, 'clip_ratio/low_min': 0.004454409005120397, 'clip_ratio/high_mean': 0.028268360998481512, 'clip_ratio/high_max': 0.028268360998481512, 'clip_ratio/region_mean': 0.03272277000360191, 'reward_total_mean': 0.9983501434326172, 'reward_meter_mean': 0.9983501434326172, 'reward_meter_std': 0.000316560355713591, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9983501434326172, 'reward_total_composite_std': 0.000316560355713591, 'epoch': 0.07}

 50%|█████     | 1664/3300 [2:00:43<2:27:04,  5.39s/it]INFO 04-12 00:37:34 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1665/3300 [2:00:51<2:51:17,  6.29s/it]2026-04-12 00:37:42,436 | INFO | train_grpo_train | metrics_logged mode=train step=1665

                                                       
{'loss': 0.0217, 'grad_norm': 2.9181759357452393, 'learning_rate': 4.957575757575758e-06, 'num_tokens': 3742119.0, 'completions/mean_length': 340.625, 'completions/min_length': 320.0, 'completions/max_length': 366.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 340.625, 'completions/min_terminated_length': 320.0, 'completions/max_terminated_length': 366.0, 'rewards/meter/mean': 0.9840777516365051, 'rewards/meter/std': 0.037384092807769775, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.7647058963775635, 'rewards/repeat_penalty/std': 0.10428297519683838, 'rewards/total_composite/mean': 0.5716930627822876, 'rewards/total_composite/std': 0.2410053312778473, 'reward': 0.5716930627822876, 'reward_std': 0.2410053312778473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05633535981178284, 'sampling/sampling_logp_difference/max': 10.131498336791992, 'sampling/importance_sampling_ratio/min': 3.980578185291961e-05, 'sampling/importance_sampling_ratio/mean': 1.0077635049819946, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3913237787783146, 'clip_ratio/low_mean': 0.004957506898790598, 'clip_ratio/low_min': 0.004957506898790598, 'clip_ratio/high_mean': 0.029565565433586016, 'clip_ratio/high_max': 0.029565565433586016, 'clip_ratio/region_mean': 0.034523072332376614, 'reward_total_mean': 0.5716930627822876, 'reward_meter_mean': 0.9840777516365051, 'reward_meter_std': 0.037384092807769775, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.7647058963775635, 'reward_repeat_penalty_std': 0.10428297519683838, 'reward_total_composite_mean': 0.5716930627822876, 'reward_total_composite_std': 0.2410053312778473, 'epoch': 0.07}

 50%|█████     | 1665/3300 [2:00:51<2:51:17,  6.29s/it]INFO 04-12 00:37:42 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1666/3300 [2:00:59<3:02:40,  6.71s/it]2026-04-12 00:37:50,127 | INFO | train_grpo_train | metrics_logged mode=train step=1666

                                                       
{'loss': -0.0177, 'grad_norm': 1.908144474029541, 'learning_rate': 4.954545454545455e-06, 'num_tokens': 3746080.0, 'completions/mean_length': 302.125, 'completions/min_length': 285.0, 'completions/max_length': 314.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 302.125, 'completions/min_terminated_length': 285.0, 'completions/max_terminated_length': 314.0, 'rewards/meter/mean': 0.9341138005256653, 'rewards/meter/std': 0.05671977251768112, 'rewards/count_adherence/mean': 0.8181818127632141, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.08169000595808029, 'rewards/total_composite/mean': 0.5741910934448242, 'rewards/total_composite/std': 0.07986098527908325, 'reward': 0.5741910934448242, 'reward_std': 0.07986099272966385, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03923000767827034, 'sampling/sampling_logp_difference/max': 2.483020067214966, 'sampling/importance_sampling_ratio/min': 0.08349069207906723, 'sampling/importance_sampling_ratio/mean': 1.0056406259536743, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2841501086950302, 'clip_ratio/low_mean': 0.01092962856637314, 'clip_ratio/low_min': 0.01092962856637314, 'clip_ratio/high_mean': 0.01824606256559491, 'clip_ratio/high_max': 0.01824606256559491, 'clip_ratio/region_mean': 0.02917569113196805, 'reward_total_mean': 0.5741910934448242, 'reward_meter_mean': 0.9341138005256653, 'reward_meter_std': 0.05671977251768112, 'reward_count_adherence_mean': 0.8181818127632141, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.08169000595808029, 'reward_total_composite_mean': 0.5741910934448242, 'reward_total_composite_std': 0.07986098527908325, 'epoch': 0.07}

 50%|█████     | 1666/3300 [2:00:59<3:02:40,  6.71s/it]INFO 04-12 00:37:50 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1667/3300 [2:01:04<2:49:11,  6.22s/it]2026-04-12 00:37:55,200 | INFO | train_grpo_train | metrics_logged mode=train step=1667

                                                       
{'loss': 0.0056, 'grad_norm': 2.4326391220092773, 'learning_rate': 4.951515151515152e-06, 'num_tokens': 3748310.0, 'completions/mean_length': 101.75, 'completions/min_length': 101.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.75, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9970037937164307, 'rewards/meter/std': 0.0013122691307216883, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7976030111312866, 'rewards/total_composite/std': 0.0010498042684048414, 'reward': 0.7976030111312866, 'reward_std': 0.0010498074116185308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010477752424776554, 'sampling/sampling_logp_difference/max': 1.087967872619629, 'sampling/importance_sampling_ratio/min': 0.33690041303634644, 'sampling/importance_sampling_ratio/mean': 1.0012123584747314, 'sampling/importance_sampling_ratio/max': 1.46913480758667, 'entropy': 0.08447205042466521, 'clip_ratio/low_mean': 0.0024390824837610126, 'clip_ratio/low_min': 0.0024390824837610126, 'clip_ratio/high_mean': 0.004938361467793584, 'clip_ratio/high_max': 0.004938361467793584, 'clip_ratio/region_mean': 0.007377443951554596, 'reward_total_mean': 0.7976030111312866, 'reward_meter_mean': 0.9970037937164307, 'reward_meter_std': 0.0013122691307216883, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7976030111312866, 'reward_total_composite_std': 0.0010498042684048414, 'epoch': 0.07}

 51%|█████     | 1667/3300 [2:01:04<2:49:11,  6.22s/it]INFO 04-12 00:37:55 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1668/3300 [2:01:09<2:41:21,  5.93s/it]2026-04-12 00:38:00,486 | INFO | train_grpo_train | metrics_logged mode=train step=1668

                                                       
{'loss': -0.0005, 'grad_norm': 0.08446517586708069, 'learning_rate': 4.9484848484848495e-06, 'num_tokens': 3750844.0, 'completions/mean_length': 133.75, 'completions/min_length': 133.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.75, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9979548454284668, 'rewards/meter/std': 2.2615500711253844e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7142857313156128, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7128248810768127, 'rewards/total_composite/std': 1.6168985894182697e-05, 'reward': 0.7128248810768127, 'reward_std': 1.6179135855054483e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004637876525521278, 'sampling/sampling_logp_difference/max': 1.0087156295776367, 'sampling/importance_sampling_ratio/min': 0.36468705534935, 'sampling/importance_sampling_ratio/mean': 1.001842975616455, 'sampling/importance_sampling_ratio/max': 1.2750511169433594, 'entropy': 0.027194248279556632, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7128248810768127, 'reward_meter_mean': 0.9979548454284668, 'reward_meter_std': 2.2615500711253844e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7142857313156128, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7128248810768127, 'reward_total_composite_std': 1.6168985894182697e-05, 'epoch': 0.07}

 51%|█████     | 1668/3300 [2:01:09<2:41:21,  5.93s/it]INFO 04-12 00:38:00 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1669/3300 [2:01:14<2:31:05,  5.56s/it]2026-04-12 00:38:05,155 | INFO | train_grpo_train | metrics_logged mode=train step=1669

                                                       
{'loss': -0.0009, 'grad_norm': 0.0685642808675766, 'learning_rate': 4.945454545454546e-06, 'num_tokens': 3752596.0, 'completions/mean_length': 68.0, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9978952407836914, 'rewards/meter/std': 1.1280608305241913e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978952407836914, 'rewards/total_composite/std': 1.1280608305241913e-05, 'reward': 0.9978952407836914, 'reward_std': 1.128060739574721e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004578685853630304, 'sampling/sampling_logp_difference/max': 0.41336488723754883, 'sampling/importance_sampling_ratio/min': 0.6614209413528442, 'sampling/importance_sampling_ratio/mean': 1.0022841691970825, 'sampling/importance_sampling_ratio/max': 1.0789562463760376, 'entropy': 0.03746737586334348, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018115942366421223, 'clip_ratio/high_max': 0.0018115942366421223, 'clip_ratio/region_mean': 0.0018115942366421223, 'reward_total_mean': 0.9978952407836914, 'reward_meter_mean': 0.9978952407836914, 'reward_meter_std': 1.1280608305241913e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978952407836914, 'reward_total_composite_std': 1.1280608305241913e-05, 'epoch': 0.07}

 51%|█████     | 1669/3300 [2:01:14<2:31:05,  5.56s/it]INFO 04-12 00:38:05 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1670/3300 [2:01:18<2:24:27,  5.32s/it]2026-04-12 00:38:09,920 | INFO | train_grpo_train | metrics_logged mode=train step=1670

                                                       
{'loss': 0.0003, 'grad_norm': 3.68241024017334, 'learning_rate': 4.942424242424243e-06, 'num_tokens': 3754401.0, 'completions/mean_length': 63.625, 'completions/min_length': 62.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.986209511756897, 'rewards/meter/std': 0.004416565876454115, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.986209511756897, 'rewards/total_composite/std': 0.004416565876454115, 'reward': 0.986209511756897, 'reward_std': 0.004416569136083126, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029973594471812248, 'sampling/sampling_logp_difference/max': 1.5556812286376953, 'sampling/importance_sampling_ratio/min': 0.2110455483198166, 'sampling/importance_sampling_ratio/mean': 1.000848650932312, 'sampling/importance_sampling_ratio/max': 1.363709807395935, 'entropy': 0.1983039677143097, 'clip_ratio/low_mean': 0.004032257944345474, 'clip_ratio/low_min': 0.004032257944345474, 'clip_ratio/high_mean': 0.01764108322095126, 'clip_ratio/high_max': 0.01764108322095126, 'clip_ratio/region_mean': 0.021673341165296733, 'reward_total_mean': 0.986209511756897, 'reward_meter_mean': 0.986209511756897, 'reward_meter_std': 0.004416565876454115, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.986209511756897, 'reward_total_composite_std': 0.004416565876454115, 'epoch': 0.07}

 51%|█████     | 1670/3300 [2:01:18<2:24:27,  5.32s/it]INFO 04-12 00:38:10 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1671/3300 [2:01:23<2:22:16,  5.24s/it]2026-04-12 00:38:14,972 | INFO | train_grpo_train | metrics_logged mode=train step=1671

                                                       
{'loss': -0.0002, 'grad_norm': 1.1145938634872437, 'learning_rate': 4.93939393939394e-06, 'num_tokens': 3756527.0, 'completions/mean_length': 100.75, 'completions/min_length': 100.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.75, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9979295134544373, 'rewards/meter/std': 2.0676665371865965e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8250000476837158, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8232920169830322, 'rewards/total_composite/std': 0.07056734710931778, 'reward': 0.8232920169830322, 'reward_std': 0.07056733220815659, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0027057162951678038, 'sampling/sampling_logp_difference/max': 0.21077418327331543, 'sampling/importance_sampling_ratio/min': 0.809956967830658, 'sampling/importance_sampling_ratio/mean': 1.0012245178222656, 'sampling/importance_sampling_ratio/max': 1.0940825939178467, 'entropy': 0.022657749010249972, 'clip_ratio/low_mean': 0.0012499999720603228, 'clip_ratio/low_min': 0.0012499999720603228, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0012499999720603228, 'reward_total_mean': 0.8232920169830322, 'reward_meter_mean': 0.9979295134544373, 'reward_meter_std': 2.0676665371865965e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8250000476837158, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8232920169830322, 'reward_total_composite_std': 0.07056734710931778, 'epoch': 0.07}

 51%|█████     | 1671/3300 [2:01:23<2:22:16,  5.24s/it]INFO 04-12 00:38:15 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1672/3300 [2:01:28<2:14:34,  4.96s/it]2026-04-12 00:38:19,299 | INFO | train_grpo_train | metrics_logged mode=train step=1672

                                                       
{'loss': -0.0112, 'grad_norm': 4.178384304046631, 'learning_rate': 4.936363636363637e-06, 'num_tokens': 3758078.0, 'completions/mean_length': 32.875, 'completions/min_length': 32.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9963486194610596, 'rewards/meter/std': 0.006714941468089819, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9963486194610596, 'rewards/total_composite/std': 0.006714941468089819, 'reward': 0.9963486194610596, 'reward_std': 0.006714944262057543, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003993004094809294, 'sampling/sampling_logp_difference/max': 0.5090913772583008, 'sampling/importance_sampling_ratio/min': 0.6010414361953735, 'sampling/importance_sampling_ratio/mean': 1.000542163848877, 'sampling/importance_sampling_ratio/max': 1.1358792781829834, 'entropy': 0.02497886586934328, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9963486194610596, 'reward_meter_mean': 0.9963486194610596, 'reward_meter_std': 0.006714941468089819, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9963486194610596, 'reward_total_composite_std': 0.006714941468089819, 'epoch': 0.07}

 51%|█████     | 1672/3300 [2:01:28<2:14:34,  4.96s/it]INFO 04-12 00:38:19 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1673/3300 [2:01:33<2:17:18,  5.06s/it]2026-04-12 00:38:24,588 | INFO | train_grpo_train | metrics_logged mode=train step=1673

                                                       
{'loss': 0.0159, 'grad_norm': 5.359276294708252, 'learning_rate': 4.933333333333334e-06, 'num_tokens': 3760034.0, 'completions/mean_length': 73.5, 'completions/min_length': 68.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.995732307434082, 'rewards/meter/std': 0.00316053768619895, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.995732307434082, 'rewards/total_composite/std': 0.00316053768619895, 'reward': 0.995732307434082, 'reward_std': 0.0031605414114892483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05547928065061569, 'sampling/sampling_logp_difference/max': 1.5614430904388428, 'sampling/importance_sampling_ratio/min': 0.2098330408334732, 'sampling/importance_sampling_ratio/mean': 1.0138006210327148, 'sampling/importance_sampling_ratio/max': 1.7035573720932007, 'entropy': 0.47178683429956436, 'clip_ratio/low_mean': 0.008515364723280072, 'clip_ratio/low_min': 0.008515364723280072, 'clip_ratio/high_mean': 0.0351892322069034, 'clip_ratio/high_max': 0.0351892322069034, 'clip_ratio/region_mean': 0.04370459693018347, 'reward_total_mean': 0.995732307434082, 'reward_meter_mean': 0.995732307434082, 'reward_meter_std': 0.00316053768619895, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.995732307434082, 'reward_total_composite_std': 0.00316053768619895, 'epoch': 0.07}

 51%|█████     | 1673/3300 [2:01:33<2:17:18,  5.06s/it]INFO 04-12 00:38:24 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1674/3300 [2:01:38<2:12:49,  4.90s/it]2026-04-12 00:38:29,109 | INFO | train_grpo_train | metrics_logged mode=train step=1674

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.9303030303030305e-06, 'num_tokens': 3761666.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.993520200252533, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.993520200252533, 'rewards/total_composite/std': 0.0, 'reward': 0.993520200252533, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0007125565898604691, 'sampling/sampling_logp_difference/max': 0.018451236188411713, 'sampling/importance_sampling_ratio/min': 0.9973061680793762, 'sampling/importance_sampling_ratio/mean': 1.0006948709487915, 'sampling/importance_sampling_ratio/max': 1.0186225175857544, 'entropy': 0.005282787780743092, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.993520200252533, 'reward_meter_mean': 0.993520200252533, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.993520200252533, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 51%|█████     | 1674/3300 [2:01:38<2:12:49,  4.90s/it]INFO 04-12 00:38:29 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1675/3300 [2:01:42<2:12:28,  4.89s/it]2026-04-12 00:38:33,974 | INFO | train_grpo_train | metrics_logged mode=train step=1675

                                                       
{'loss': 0.0014, 'grad_norm': 0.6250240802764893, 'learning_rate': 4.927272727272728e-06, 'num_tokens': 3763698.0, 'completions/mean_length': 86.0, 'completions/min_length': 86.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.0, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9938453435897827, 'rewards/meter/std': 0.0002583391033113003, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7950762510299683, 'rewards/total_composite/std': 0.00020666708587668836, 'reward': 0.7950762510299683, 'reward_std': 0.00020667610806412995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0035054178442806005, 'sampling/sampling_logp_difference/max': 1.0392670631408691, 'sampling/importance_sampling_ratio/min': 0.9702885746955872, 'sampling/importance_sampling_ratio/mean': 1.0030723810195923, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.017064974061213434, 'clip_ratio/low_mean': 0.0014534883666783571, 'clip_ratio/low_min': 0.0014534883666783571, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0014534883666783571, 'reward_total_mean': 0.7950762510299683, 'reward_meter_mean': 0.9938453435897827, 'reward_meter_std': 0.0002583391033113003, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7950762510299683, 'reward_total_composite_std': 0.00020666708587668836, 'epoch': 0.07}

 51%|█████     | 1675/3300 [2:01:42<2:12:28,  4.89s/it]INFO 04-12 00:38:34 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1676/3300 [2:01:48<2:14:30,  4.97s/it]2026-04-12 00:38:39,129 | INFO | train_grpo_train | metrics_logged mode=train step=1676

                                                       
{'loss': 0.0003, 'grad_norm': 1.5943388938903809, 'learning_rate': 4.924242424242425e-06, 'num_tokens': 3765698.0, 'completions/mean_length': 101.0, 'completions/min_length': 100.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.0, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9978747963905334, 'rewards/meter/std': 0.0001633708452573046, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8731323480606079, 'rewards/total_composite/std': 0.10320712625980377, 'reward': 0.8731323480606079, 'reward_std': 0.10320711135864258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008153069764375687, 'sampling/sampling_logp_difference/max': 1.007601261138916, 'sampling/importance_sampling_ratio/min': 0.3650937080383301, 'sampling/importance_sampling_ratio/mean': 1.0001577138900757, 'sampling/importance_sampling_ratio/max': 1.161500334739685, 'entropy': 0.04971472639590502, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.00492622796446085, 'clip_ratio/high_max': 0.00492622796446085, 'clip_ratio/region_mean': 0.00492622796446085, 'reward_total_mean': 0.8731323480606079, 'reward_meter_mean': 0.9978747963905334, 'reward_meter_std': 0.0001633708452573046, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8731323480606079, 'reward_total_composite_std': 0.10320712625980377, 'epoch': 0.07}

 51%|█████     | 1676/3300 [2:01:48<2:14:30,  4.97s/it]INFO 04-12 00:38:39 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1677/3300 [2:01:52<2:11:02,  4.84s/it]2026-04-12 00:38:43,677 | INFO | train_grpo_train | metrics_logged mode=train step=1677

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.9212121212121214e-06, 'num_tokens': 3767450.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9982472658157349, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982472658157349, 'rewards/total_composite/std': 0.0, 'reward': 0.9982472658157349, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0014719502069056034, 'sampling/sampling_logp_difference/max': 0.2051541954278946, 'sampling/importance_sampling_ratio/min': 0.8145217299461365, 'sampling/importance_sampling_ratio/mean': 1.0002412796020508, 'sampling/importance_sampling_ratio/max': 1.0495675802230835, 'entropy': 0.00934242527000606, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9982472658157349, 'reward_meter_mean': 0.9982472658157349, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982472658157349, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 51%|█████     | 1677/3300 [2:01:52<2:11:02,  4.84s/it]INFO 04-12 00:38:43 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1678/3300 [2:01:57<2:12:56,  4.92s/it]2026-04-12 00:38:48,767 | INFO | train_grpo_train | metrics_logged mode=train step=1678

                                                       
{'loss': 0.01, 'grad_norm': 1.9516417980194092, 'learning_rate': 4.918181818181819e-06, 'num_tokens': 3769712.0, 'completions/mean_length': 99.75, 'completions/min_length': 98.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.75, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9527222514152527, 'rewards/meter/std': 0.0017279081512242556, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7857142686843872, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.748537540435791, 'rewards/total_composite/std': 0.0724104791879654, 'reward': 0.748537540435791, 'reward_std': 0.0724104791879654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01071135699748993, 'sampling/sampling_logp_difference/max': 1.6311063766479492, 'sampling/importance_sampling_ratio/min': 0.19571292400360107, 'sampling/importance_sampling_ratio/mean': 0.999224066734314, 'sampling/importance_sampling_ratio/max': 1.2951877117156982, 'entropy': 0.04197088209912181, 'clip_ratio/low_mean': 0.0012376237427815795, 'clip_ratio/low_min': 0.0012376237427815795, 'clip_ratio/high_mean': 0.0037499999161809683, 'clip_ratio/high_max': 0.0037499999161809683, 'clip_ratio/region_mean': 0.004987623658962548, 'reward_total_mean': 0.748537540435791, 'reward_meter_mean': 0.9527222514152527, 'reward_meter_std': 0.0017279081512242556, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7857142686843872, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.748537540435791, 'reward_total_composite_std': 0.0724104791879654, 'epoch': 0.07}

 51%|█████     | 1678/3300 [2:01:57<2:12:56,  4.92s/it]INFO 04-12 00:38:49 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1679/3300 [2:02:04<2:26:29,  5.42s/it]2026-04-12 00:38:55,369 | INFO | train_grpo_train | metrics_logged mode=train step=1679

                                                       
{'loss': -0.006, 'grad_norm': 1.1392688751220703, 'learning_rate': 4.915151515151516e-06, 'num_tokens': 3772903.0, 'completions/mean_length': 216.875, 'completions/min_length': 214.0, 'completions/max_length': 224.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 216.875, 'completions/min_terminated_length': 214.0, 'completions/max_terminated_length': 224.0, 'rewards/meter/mean': 0.9947628974914551, 'rewards/meter/std': 0.0007815666613169014, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6442307829856873, 'rewards/repeat_penalty/std': 0.039811473339796066, 'rewards/total_composite/mean': 0.5340501070022583, 'rewards/total_composite/std': 0.03305770829319954, 'reward': 0.5340501070022583, 'reward_std': 0.03305771201848984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009834333322942257, 'sampling/sampling_logp_difference/max': 1.1391210556030273, 'sampling/importance_sampling_ratio/min': 0.32010024785995483, 'sampling/importance_sampling_ratio/mean': 1.0005079507827759, 'sampling/importance_sampling_ratio/max': 1.3776015043258667, 'entropy': 0.052406568080186844, 'clip_ratio/low_mean': 0.0011655074777081609, 'clip_ratio/low_min': 0.0011655074777081609, 'clip_ratio/high_mean': 0.003399575361981988, 'clip_ratio/high_max': 0.003399575361981988, 'clip_ratio/region_mean': 0.004565082839690149, 'reward_total_mean': 0.5340501070022583, 'reward_meter_mean': 0.9947628974914551, 'reward_meter_std': 0.0007815666613169014, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6442307829856873, 'reward_repeat_penalty_std': 0.039811473339796066, 'reward_total_composite_mean': 0.5340501070022583, 'reward_total_composite_std': 0.03305770829319954, 'epoch': 0.07}

 51%|█████     | 1679/3300 [2:02:04<2:26:29,  5.42s/it]INFO 04-12 00:38:55 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1680/3300 [2:02:13<2:58:40,  6.62s/it]2026-04-12 00:39:04,775 | INFO | train_grpo_train | metrics_logged mode=train step=1680

                                                       
{'loss': -0.0168, 'grad_norm': 2.291543960571289, 'learning_rate': 4.912121212121212e-06, 'num_tokens': 3777950.0, 'completions/mean_length': 401.875, 'completions/min_length': 373.0, 'completions/max_length': 455.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 401.875, 'completions/min_terminated_length': 373.0, 'completions/max_terminated_length': 455.0, 'rewards/meter/mean': 0.9955718517303467, 'rewards/meter/std': 0.004510289058089256, 'rewards/count_adherence/mean': 0.7232142686843872, 'rewards/count_adherence/std': 0.02525380253791809, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8195801973342896, 'rewards/repeat_penalty/std': 0.1243370845913887, 'rewards/total_composite/mean': 0.5898604393005371, 'rewards/total_composite/std': 0.08871299028396606, 'reward': 0.5898604393005371, 'reward_std': 0.08871299028396606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0480833500623703, 'sampling/sampling_logp_difference/max': 1.780477523803711, 'sampling/importance_sampling_ratio/min': 0.17153088748455048, 'sampling/importance_sampling_ratio/mean': 1.0117672681808472, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3997117578983307, 'clip_ratio/low_mean': 0.010193796595558524, 'clip_ratio/low_min': 0.010193796595558524, 'clip_ratio/high_mean': 0.0216394632589072, 'clip_ratio/high_max': 0.0216394632589072, 'clip_ratio/region_mean': 0.03183325985446572, 'reward_total_mean': 0.5898604393005371, 'reward_meter_mean': 0.9955718517303467, 'reward_meter_std': 0.004510289058089256, 'reward_count_adherence_mean': 0.7232142686843872, 'reward_count_adherence_std': 0.02525380253791809, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8195801973342896, 'reward_repeat_penalty_std': 0.1243370845913887, 'reward_total_composite_mean': 0.5898604393005371, 'reward_total_composite_std': 0.08871299028396606, 'epoch': 0.07}

 51%|█████     | 1680/3300 [2:02:13<2:58:40,  6.62s/it]INFO 04-12 00:39:05 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1681/3300 [2:02:17<2:37:48,  5.85s/it]2026-04-12 00:39:08,830 | INFO | train_grpo_train | metrics_logged mode=train step=1681

                                                       
{'loss': -0.0134, 'grad_norm': 4.46686315536499, 'learning_rate': 4.90909090909091e-06, 'num_tokens': 3779464.0, 'completions/mean_length': 30.25, 'completions/min_length': 29.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.9945183992385864, 'rewards/meter/std': 0.000778909248765558, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9945183992385864, 'rewards/total_composite/std': 0.000778909248765558, 'reward': 0.9945183992385864, 'reward_std': 0.0007789283408783376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0202788058668375, 'sampling/sampling_logp_difference/max': 0.6706576347351074, 'sampling/importance_sampling_ratio/min': 0.5113722085952759, 'sampling/importance_sampling_ratio/mean': 0.9981259703636169, 'sampling/importance_sampling_ratio/max': 1.514021873474121, 'entropy': 0.1191922826692462, 'clip_ratio/low_mean': 0.008477011695504189, 'clip_ratio/low_min': 0.008477011695504189, 'clip_ratio/high_mean': 0.02016128972172737, 'clip_ratio/high_max': 0.02016128972172737, 'clip_ratio/region_mean': 0.02863830141723156, 'reward_total_mean': 0.9945183992385864, 'reward_meter_mean': 0.9945183992385864, 'reward_meter_std': 0.000778909248765558, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9945183992385864, 'reward_total_composite_std': 0.000778909248765558, 'epoch': 0.07}

 51%|█████     | 1681/3300 [2:02:17<2:37:48,  5.85s/it]INFO 04-12 00:39:09 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1682/3300 [2:02:22<2:27:30,  5.47s/it]2026-04-12 00:39:13,415 | INFO | train_grpo_train | metrics_logged mode=train step=1682

                                                       
{'loss': -0.0204, 'grad_norm': 7.099659442901611, 'learning_rate': 4.906060606060606e-06, 'num_tokens': 3781340.0, 'completions/mean_length': 72.5, 'completions/min_length': 67.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.5, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9961867928504944, 'rewards/meter/std': 0.0030412215273827314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961867928504944, 'rewards/total_composite/std': 0.0030412215273827314, 'reward': 0.9961867928504944, 'reward_std': 0.003041210351511836, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06393711268901825, 'sampling/sampling_logp_difference/max': 1.4125175476074219, 'sampling/importance_sampling_ratio/min': 0.24352942407131195, 'sampling/importance_sampling_ratio/mean': 1.0079271793365479, 'sampling/importance_sampling_ratio/max': 1.853277325630188, 'entropy': 0.46719325333833694, 'clip_ratio/low_mean': 0.007042253389954567, 'clip_ratio/low_min': 0.007042253389954567, 'clip_ratio/high_mean': 0.0440298137255013, 'clip_ratio/high_max': 0.0440298137255013, 'clip_ratio/region_mean': 0.051072067115455866, 'reward_total_mean': 0.9961867928504944, 'reward_meter_mean': 0.9961867928504944, 'reward_meter_std': 0.0030412215273827314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961867928504944, 'reward_total_composite_std': 0.0030412215273827314, 'epoch': 0.07}

 51%|█████     | 1682/3300 [2:02:22<2:27:30,  5.47s/it]INFO 04-12 00:39:13 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1683/3300 [2:02:26<2:18:44,  5.15s/it]2026-04-12 00:39:17,816 | INFO | train_grpo_train | metrics_logged mode=train step=1683

                                                       
{'loss': -0.0016, 'grad_norm': 0.45891278982162476, 'learning_rate': 4.903030303030303e-06, 'num_tokens': 3783091.0, 'completions/mean_length': 63.875, 'completions/min_length': 63.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9982389211654663, 'rewards/meter/std': 2.360223516006954e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982389211654663, 'rewards/total_composite/std': 2.360223516006954e-05, 'reward': 0.9982389211654663, 'reward_std': 2.360223516006954e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004287322983145714, 'sampling/sampling_logp_difference/max': 0.9284515380859375, 'sampling/importance_sampling_ratio/min': 0.3951651453971863, 'sampling/importance_sampling_ratio/mean': 1.000044345855713, 'sampling/importance_sampling_ratio/max': 1.0655800104141235, 'entropy': 0.01849041902460158, 'clip_ratio/low_mean': 0.0019841270986944437, 'clip_ratio/low_min': 0.0019841270986944437, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0019841270986944437, 'reward_total_mean': 0.9982389211654663, 'reward_meter_mean': 0.9982389211654663, 'reward_meter_std': 2.360223516006954e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982389211654663, 'reward_total_composite_std': 2.360223516006954e-05, 'epoch': 0.07}

 51%|█████     | 1683/3300 [2:02:26<2:18:44,  5.15s/it]INFO 04-12 00:39:18 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1684/3300 [2:02:32<2:24:03,  5.35s/it]2026-04-12 00:39:23,628 | INFO | train_grpo_train | metrics_logged mode=train step=1684

                                                       
{'loss': 0.0107, 'grad_norm': 3.898653507232666, 'learning_rate': 4.9000000000000005e-06, 'num_tokens': 3785743.0, 'completions/mean_length': 159.5, 'completions/min_length': 148.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.5, 'completions/min_terminated_length': 148.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.9947781562805176, 'rewards/meter/std': 0.007610421162098646, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9769483804702759, 'rewards/total_composite/std': 0.049505408853292465, 'reward': 0.9769483804702759, 'reward_std': 0.049505408853292465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059815503656864166, 'sampling/sampling_logp_difference/max': 1.4108057022094727, 'sampling/importance_sampling_ratio/min': 0.24394665658473969, 'sampling/importance_sampling_ratio/mean': 1.0116267204284668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5180013440549374, 'clip_ratio/low_mean': 0.007966701406985521, 'clip_ratio/low_min': 0.007966701406985521, 'clip_ratio/high_mean': 0.024855362717062235, 'clip_ratio/high_max': 0.024855362717062235, 'clip_ratio/region_mean': 0.032822064124047756, 'reward_total_mean': 0.9769483804702759, 'reward_meter_mean': 0.9947781562805176, 'reward_meter_std': 0.007610421162098646, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9769483804702759, 'reward_total_composite_std': 0.049505408853292465, 'epoch': 0.07}

 51%|█████     | 1684/3300 [2:02:32<2:24:03,  5.35s/it]INFO 04-12 00:39:23 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1685/3300 [2:02:37<2:17:30,  5.11s/it]2026-04-12 00:39:28,182 | INFO | train_grpo_train | metrics_logged mode=train step=1685

                                                       
{'loss': -0.0028, 'grad_norm': 2.095104932785034, 'learning_rate': 4.896969696969697e-06, 'num_tokens': 3787673.0, 'completions/mean_length': 64.25, 'completions/min_length': 64.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9981728792190552, 'rewards/meter/std': 0.0002840534143615514, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981728792190552, 'rewards/total_composite/std': 0.0002840534143615514, 'reward': 0.9981728792190552, 'reward_std': 0.00028403810574673116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011698653921484947, 'sampling/sampling_logp_difference/max': 1.6340885162353516, 'sampling/importance_sampling_ratio/min': 0.1951301395893097, 'sampling/importance_sampling_ratio/mean': 0.9983230829238892, 'sampling/importance_sampling_ratio/max': 1.5041022300720215, 'entropy': 0.04154683533124626, 'clip_ratio/low_mean': 0.005859375, 'clip_ratio/low_min': 0.005859375, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.009647253900766373, 'reward_total_mean': 0.9981728792190552, 'reward_meter_mean': 0.9981728792190552, 'reward_meter_std': 0.0002840534143615514, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981728792190552, 'reward_total_composite_std': 0.0002840534143615514, 'epoch': 0.07}

 51%|█████     | 1685/3300 [2:02:37<2:17:30,  5.11s/it]INFO 04-12 00:39:28 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1686/3300 [2:02:41<2:13:09,  4.95s/it]2026-04-12 00:39:32,757 | INFO | train_grpo_train | metrics_logged mode=train step=1686

                                                       
{'loss': -0.0182, 'grad_norm': 4.037059783935547, 'learning_rate': 4.893939393939394e-06, 'num_tokens': 3789623.0, 'completions/mean_length': 71.75, 'completions/min_length': 68.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.75, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.997489333152771, 'rewards/meter/std': 0.0010296773398295045, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997489333152771, 'rewards/total_composite/std': 0.0010296773398295045, 'reward': 0.997489333152771, 'reward_std': 0.0010296726832166314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04423650726675987, 'sampling/sampling_logp_difference/max': 1.0650873184204102, 'sampling/importance_sampling_ratio/min': 0.3446977436542511, 'sampling/importance_sampling_ratio/mean': 1.0071872472763062, 'sampling/importance_sampling_ratio/max': 1.5631378889083862, 'entropy': 0.4106294997036457, 'clip_ratio/low_mean': 0.017920455895364285, 'clip_ratio/low_min': 0.017920455895364285, 'clip_ratio/high_mean': 0.02395429043099284, 'clip_ratio/high_max': 0.02395429043099284, 'clip_ratio/region_mean': 0.041874746326357126, 'reward_total_mean': 0.997489333152771, 'reward_meter_mean': 0.997489333152771, 'reward_meter_std': 0.0010296773398295045, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997489333152771, 'reward_total_composite_std': 0.0010296773398295045, 'epoch': 0.07}

 51%|█████     | 1686/3300 [2:02:41<2:13:09,  4.95s/it]INFO 04-12 00:39:32 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1687/3300 [2:02:45<2:06:00,  4.69s/it]2026-04-12 00:39:36,832 | INFO | train_grpo_train | metrics_logged mode=train step=1687

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.8909090909090914e-06, 'num_tokens': 3791023.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9977574944496155, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977574944496155, 'rewards/total_composite/std': 0.0, 'reward': 0.9977574944496155, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0031034064013510942, 'sampling/sampling_logp_difference/max': 0.04160819947719574, 'sampling/importance_sampling_ratio/min': 0.9605222940444946, 'sampling/importance_sampling_ratio/mean': 1.0023516416549683, 'sampling/importance_sampling_ratio/max': 1.0424859523773193, 'entropy': 0.02776301186531782, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9977574944496155, 'reward_meter_mean': 0.9977574944496155, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977574944496155, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 51%|█████     | 1687/3300 [2:02:45<2:06:00,  4.69s/it]INFO 04-12 00:39:37 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1688/3300 [2:02:49<2:01:39,  4.53s/it]2026-04-12 00:39:40,989 | INFO | train_grpo_train | metrics_logged mode=train step=1688

                                                       
{'loss': 0.0003, 'grad_norm': 4.740636348724365, 'learning_rate': 4.887878787878788e-06, 'num_tokens': 3792659.0, 'completions/mean_length': 35.5, 'completions/min_length': 35.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9980167150497437, 'rewards/meter/std': 0.00040671354508958757, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980167150497437, 'rewards/total_composite/std': 0.00040671354508958757, 'reward': 0.9980167150497437, 'reward_std': 0.0004067234694957733, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021727347746491432, 'sampling/sampling_logp_difference/max': 0.9272527694702148, 'sampling/importance_sampling_ratio/min': 0.3956391513347626, 'sampling/importance_sampling_ratio/mean': 0.9994746446609497, 'sampling/importance_sampling_ratio/max': 1.5504966974258423, 'entropy': 0.12218342442065477, 'clip_ratio/low_mean': 0.0070436508394777775, 'clip_ratio/low_min': 0.0070436508394777775, 'clip_ratio/high_mean': 0.0035714285913854837, 'clip_ratio/high_max': 0.0035714285913854837, 'clip_ratio/region_mean': 0.010615079430863261, 'reward_total_mean': 0.9980167150497437, 'reward_meter_mean': 0.9980167150497437, 'reward_meter_std': 0.00040671354508958757, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980167150497437, 'reward_total_composite_std': 0.00040671354508958757, 'epoch': 0.07}

 51%|█████     | 1688/3300 [2:02:49<2:01:39,  4.53s/it]INFO 04-12 00:39:41 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1689/3300 [2:02:54<1:59:49,  4.46s/it]2026-04-12 00:39:45,297 | INFO | train_grpo_train | metrics_logged mode=train step=1689

                                                       
{'loss': 0.0024, 'grad_norm': 0.21065974235534668, 'learning_rate': 4.884848484848485e-06, 'num_tokens': 3794387.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.998286783695221, 'rewards/meter/std': 0.00011179452121723443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998286783695221, 'rewards/total_composite/std': 0.00011179452121723443, 'reward': 0.998286783695221, 'reward_std': 0.00011178548447787762, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0037898647133260965, 'sampling/sampling_logp_difference/max': 1.470503807067871, 'sampling/importance_sampling_ratio/min': 0.22980967164039612, 'sampling/importance_sampling_ratio/mean': 0.9991194009780884, 'sampling/importance_sampling_ratio/max': 1.028564214706421, 'entropy': 0.008520788454916328, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.998286783695221, 'reward_meter_mean': 0.998286783695221, 'reward_meter_std': 0.00011179452121723443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998286783695221, 'reward_total_composite_std': 0.00011179452121723443, 'epoch': 0.07}

 51%|█████     | 1689/3300 [2:02:54<1:59:49,  4.46s/it]INFO 04-12 00:39:45 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1690/3300 [2:02:58<1:57:15,  4.37s/it]2026-04-12 00:39:49,451 | INFO | train_grpo_train | metrics_logged mode=train step=1690

                                                       
{'loss': 0.0367, 'grad_norm': 7.852858066558838, 'learning_rate': 4.881818181818182e-06, 'num_tokens': 3795915.0, 'completions/mean_length': 36.0, 'completions/min_length': 34.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.934417724609375, 'rewards/meter/std': 0.06754057109355927, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.934417724609375, 'rewards/total_composite/std': 0.06754057109355927, 'reward': 0.934417724609375, 'reward_std': 0.06754057854413986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05185196176171303, 'sampling/sampling_logp_difference/max': 1.1850357055664062, 'sampling/importance_sampling_ratio/min': 0.3057352602481842, 'sampling/importance_sampling_ratio/mean': 1.0148731470108032, 'sampling/importance_sampling_ratio/max': 1.7753301858901978, 'entropy': 0.3176651671528816, 'clip_ratio/low_mean': 0.003289473708719015, 'clip_ratio/low_min': 0.003289473708719015, 'clip_ratio/high_mean': 0.04497848777100444, 'clip_ratio/high_max': 0.04497848777100444, 'clip_ratio/region_mean': 0.048267961479723454, 'reward_total_mean': 0.934417724609375, 'reward_meter_mean': 0.934417724609375, 'reward_meter_std': 0.06754057109355927, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.934417724609375, 'reward_total_composite_std': 0.06754057109355927, 'epoch': 0.07}

 51%|█████     | 1690/3300 [2:02:58<1:57:15,  4.37s/it]INFO 04-12 00:39:49 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1691/3300 [2:03:03<2:05:38,  4.69s/it]2026-04-12 00:39:54,873 | INFO | train_grpo_train | metrics_logged mode=train step=1691

                                                       
{'loss': -0.0057, 'grad_norm': 3.046787738800049, 'learning_rate': 4.878787878787879e-06, 'num_tokens': 3798615.0, 'completions/mean_length': 140.5, 'completions/min_length': 136.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.5, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.9408473372459412, 'rewards/meter/std': 0.04903604835271835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.09155284613370895, 'rewards/total_composite/mean': 0.824957013130188, 'rewards/total_composite/std': 0.11124026030302048, 'reward': 0.824957013130188, 'reward_std': 0.11124025285243988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04019925743341446, 'sampling/sampling_logp_difference/max': 1.1345267295837402, 'sampling/importance_sampling_ratio/min': 0.32157430052757263, 'sampling/importance_sampling_ratio/mean': 1.012131690979004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28699430637061596, 'clip_ratio/low_mean': 0.014275046763941646, 'clip_ratio/low_min': 0.014275046763941646, 'clip_ratio/high_mean': 0.025541604205500335, 'clip_ratio/high_max': 0.025541604205500335, 'clip_ratio/region_mean': 0.03981665096944198, 'reward_total_mean': 0.824957013130188, 'reward_meter_mean': 0.9408473372459412, 'reward_meter_std': 0.04903604835271835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.09155284613370895, 'reward_total_composite_mean': 0.824957013130188, 'reward_total_composite_std': 0.11124026030302048, 'epoch': 0.07}

 51%|█████     | 1691/3300 [2:03:03<2:05:38,  4.69s/it]INFO 04-12 00:39:55 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1692/3300 [2:03:09<2:12:12,  4.93s/it]2026-04-12 00:40:00,382 | INFO | train_grpo_train | metrics_logged mode=train step=1692

                                                       
{'loss': -0.0153, 'grad_norm': 2.5923614501953125, 'learning_rate': 4.875757575757576e-06, 'num_tokens': 3801374.0, 'completions/mean_length': 148.875, 'completions/min_length': 142.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.875, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9931907653808594, 'rewards/meter/std': 0.0014422648819163442, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6305556297302246, 'rewards/repeat_penalty/std': 0.0602024681866169, 'rewards/total_composite/mean': 0.6263105869293213, 'rewards/total_composite/std': 0.060353342443704605, 'reward': 0.6263105869293213, 'reward_std': 0.06035333871841431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009429208934307098, 'sampling/sampling_logp_difference/max': 1.152205228805542, 'sampling/importance_sampling_ratio/min': 0.3159392774105072, 'sampling/importance_sampling_ratio/mean': 1.001092791557312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03733880212530494, 'clip_ratio/low_mean': 0.005935311957728118, 'clip_ratio/low_min': 0.005935311957728118, 'clip_ratio/high_mean': 0.006003972492180765, 'clip_ratio/high_max': 0.006003972492180765, 'clip_ratio/region_mean': 0.011939284449908882, 'reward_total_mean': 0.6263105869293213, 'reward_meter_mean': 0.9931907653808594, 'reward_meter_std': 0.0014422648819163442, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6305556297302246, 'reward_repeat_penalty_std': 0.0602024681866169, 'reward_total_composite_mean': 0.6263105869293213, 'reward_total_composite_std': 0.060353342443704605, 'epoch': 0.07}

 51%|█████▏    | 1692/3300 [2:03:09<2:12:12,  4.93s/it]INFO 04-12 00:40:00 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1693/3300 [2:03:13<2:08:47,  4.81s/it]2026-04-12 00:40:04,901 | INFO | train_grpo_train | metrics_logged mode=train step=1693

                                                       
{'loss': 0.013, 'grad_norm': 5.695197105407715, 'learning_rate': 4.872727272727273e-06, 'num_tokens': 3803064.0, 'completions/mean_length': 67.25, 'completions/min_length': 62.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7432924509048462, 'rewards/meter/std': 0.2945479154586792, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7432924509048462, 'rewards/total_composite/std': 0.2945479154586792, 'reward': 0.7432924509048462, 'reward_std': 0.2945479154586792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05113214999437332, 'sampling/sampling_logp_difference/max': 1.1394152641296387, 'sampling/importance_sampling_ratio/min': 0.3200061023235321, 'sampling/importance_sampling_ratio/mean': 1.0079994201660156, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40255092084407806, 'clip_ratio/low_mean': 0.009146409342065454, 'clip_ratio/low_min': 0.009146409342065454, 'clip_ratio/high_mean': 0.02981708408333361, 'clip_ratio/high_max': 0.02981708408333361, 'clip_ratio/region_mean': 0.038963493425399065, 'reward_total_mean': 0.7432924509048462, 'reward_meter_mean': 0.7432924509048462, 'reward_meter_std': 0.2945479154586792, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7432924509048462, 'reward_total_composite_std': 0.2945479154586792, 'epoch': 0.07}

 51%|█████▏    | 1693/3300 [2:03:13<2:08:47,  4.81s/it]INFO 04-12 00:40:05 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1694/3300 [2:03:19<2:16:48,  5.11s/it]2026-04-12 00:40:10,720 | INFO | train_grpo_train | metrics_logged mode=train step=1694

                                                       
{'loss': 0.0364, 'grad_norm': 2.8474440574645996, 'learning_rate': 4.8696969696969705e-06, 'num_tokens': 3805883.0, 'completions/mean_length': 169.375, 'completions/min_length': 158.0, 'completions/max_length': 182.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 169.375, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.6835033893585205, 'rewards/meter/std': 0.2869633436203003, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9041666388511658, 'rewards/repeat_penalty/std': 0.07100624591112137, 'rewards/total_composite/mean': 0.6048145294189453, 'rewards/total_composite/std': 0.23514944314956665, 'reward': 0.6048145294189453, 'reward_std': 0.23514942824840546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03621907904744148, 'sampling/sampling_logp_difference/max': 0.9100509285926819, 'sampling/importance_sampling_ratio/min': 0.40250372886657715, 'sampling/importance_sampling_ratio/mean': 1.00909423828125, 'sampling/importance_sampling_ratio/max': 1.8559550046920776, 'entropy': 0.3172535989433527, 'clip_ratio/low_mean': 0.010541909374296665, 'clip_ratio/low_min': 0.010541909374296665, 'clip_ratio/high_mean': 0.020258112344890833, 'clip_ratio/high_max': 0.020258112344890833, 'clip_ratio/region_mean': 0.030800021719187498, 'reward_total_mean': 0.6048145294189453, 'reward_meter_mean': 0.6835033893585205, 'reward_meter_std': 0.2869633436203003, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9041666388511658, 'reward_repeat_penalty_std': 0.07100624591112137, 'reward_total_composite_mean': 0.6048145294189453, 'reward_total_composite_std': 0.23514944314956665, 'epoch': 0.07}

 51%|█████▏    | 1694/3300 [2:03:19<2:16:48,  5.11s/it]INFO 04-12 00:40:10 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1695/3300 [2:03:24<2:11:19,  4.91s/it]2026-04-12 00:40:15,163 | INFO | train_grpo_train | metrics_logged mode=train step=1695

                                                       
{'loss': -0.0183, 'grad_norm': 2.8666982650756836, 'learning_rate': 4.866666666666667e-06, 'num_tokens': 3807608.0, 'completions/mean_length': 60.625, 'completions/min_length': 58.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8901118040084839, 'rewards/meter/std': 0.2941354215145111, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8901118040084839, 'rewards/total_composite/std': 0.2941354215145111, 'reward': 0.8901118040084839, 'reward_std': 0.2941353917121887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02997700497508049, 'sampling/sampling_logp_difference/max': 1.047898292541504, 'sampling/importance_sampling_ratio/min': 0.35067400336265564, 'sampling/importance_sampling_ratio/mean': 0.9950342178344727, 'sampling/importance_sampling_ratio/max': 1.2877650260925293, 'entropy': 0.17649692203849554, 'clip_ratio/low_mean': 0.006465517450124025, 'clip_ratio/low_min': 0.006465517450124025, 'clip_ratio/high_mean': 0.03078524977900088, 'clip_ratio/high_max': 0.03078524977900088, 'clip_ratio/region_mean': 0.037250767229124904, 'reward_total_mean': 0.8901118040084839, 'reward_meter_mean': 0.8901118040084839, 'reward_meter_std': 0.2941354215145111, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8901118040084839, 'reward_total_composite_std': 0.2941354215145111, 'epoch': 0.07}

 51%|█████▏    | 1695/3300 [2:03:24<2:11:19,  4.91s/it]INFO 04-12 00:40:15 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1696/3300 [2:03:29<2:12:21,  4.95s/it]2026-04-12 00:40:20,213 | INFO | train_grpo_train | metrics_logged mode=train step=1696

                                                       
{'loss': -0.0017, 'grad_norm': 3.8953700065612793, 'learning_rate': 4.863636363636364e-06, 'num_tokens': 3809772.0, 'completions/mean_length': 106.5, 'completions/min_length': 103.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.5, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9039969444274902, 'rewards/meter/std': 0.09514570236206055, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9039969444274902, 'rewards/total_composite/std': 0.09514570236206055, 'reward': 0.9039969444274902, 'reward_std': 0.09514573216438293, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04417286068201065, 'sampling/sampling_logp_difference/max': 1.5998225212097168, 'sampling/importance_sampling_ratio/min': 0.20193234086036682, 'sampling/importance_sampling_ratio/mean': 0.9994289875030518, 'sampling/importance_sampling_ratio/max': 1.748896598815918, 'entropy': 0.3774713296443224, 'clip_ratio/low_mean': 0.010750595480203629, 'clip_ratio/low_min': 0.010750595480203629, 'clip_ratio/high_mean': 0.03136626980267465, 'clip_ratio/high_max': 0.03136626980267465, 'clip_ratio/region_mean': 0.04211686528287828, 'reward_total_mean': 0.9039969444274902, 'reward_meter_mean': 0.9039969444274902, 'reward_meter_std': 0.09514570236206055, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9039969444274902, 'reward_total_composite_std': 0.09514570236206055, 'epoch': 0.07}

 51%|█████▏    | 1696/3300 [2:03:29<2:12:21,  4.95s/it]INFO 04-12 00:40:20 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1697/3300 [2:03:33<2:09:35,  4.85s/it]2026-04-12 00:40:24,826 | INFO | train_grpo_train | metrics_logged mode=train step=1697

                                                       
{'loss': 0.0098, 'grad_norm': 12.74048137664795, 'learning_rate': 4.8606060606060615e-06, 'num_tokens': 3811557.0, 'completions/mean_length': 69.125, 'completions/min_length': 64.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7374060153961182, 'rewards/meter/std': 0.2923416793346405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7374060153961182, 'rewards/total_composite/std': 0.2923416793346405, 'reward': 0.7374060153961182, 'reward_std': 0.2923416793346405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051393892616033554, 'sampling/sampling_logp_difference/max': 0.840245246887207, 'sampling/importance_sampling_ratio/min': 0.43160465359687805, 'sampling/importance_sampling_ratio/mean': 1.0157667398452759, 'sampling/importance_sampling_ratio/max': 1.9715321063995361, 'entropy': 0.3504845257848501, 'clip_ratio/low_mean': 0.011194029822945595, 'clip_ratio/low_min': 0.011194029822945595, 'clip_ratio/high_mean': 0.026841548271477222, 'clip_ratio/high_max': 0.026841548271477222, 'clip_ratio/region_mean': 0.03803557809442282, 'reward_total_mean': 0.7374060153961182, 'reward_meter_mean': 0.7374060153961182, 'reward_meter_std': 0.2923416793346405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7374060153961182, 'reward_total_composite_std': 0.2923416793346405, 'epoch': 0.07}

 51%|█████▏    | 1697/3300 [2:03:33<2:09:35,  4.85s/it]INFO 04-12 00:40:25 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1698/3300 [2:03:43<2:47:07,  6.26s/it]2026-04-12 00:40:34,369 | INFO | train_grpo_train | metrics_logged mode=train step=1698

                                                       
{'loss': 0.0122, 'grad_norm': 2.547250747680664, 'learning_rate': 4.857575757575758e-06, 'num_tokens': 3817138.0, 'completions/mean_length': 456.625, 'completions/min_length': 442.0, 'completions/max_length': 476.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 456.625, 'completions/min_terminated_length': 442.0, 'completions/max_terminated_length': 476.0, 'rewards/meter/mean': 0.9678418636322021, 'rewards/meter/std': 0.03973681107163429, 'rewards/count_adherence/mean': 0.7678571343421936, 'rewards/count_adherence/std': 0.033064987510442734, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9239448308944702, 'rewards/repeat_penalty/std': 0.04794442281126976, 'rewards/total_composite/mean': 0.6861064434051514, 'rewards/total_composite/std': 0.04557538405060768, 'reward': 0.6861064434051514, 'reward_std': 0.04557539150118828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.055158257484436035, 'sampling/sampling_logp_difference/max': 2.418172836303711, 'sampling/importance_sampling_ratio/min': 0.08908424526453018, 'sampling/importance_sampling_ratio/mean': 1.0073450803756714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4843989312648773, 'clip_ratio/low_mean': 0.01929042232222855, 'clip_ratio/low_min': 0.01929042232222855, 'clip_ratio/high_mean': 0.023344212910160422, 'clip_ratio/high_max': 0.023344212910160422, 'clip_ratio/region_mean': 0.04263463523238897, 'reward_total_mean': 0.6861064434051514, 'reward_meter_mean': 0.9678418636322021, 'reward_meter_std': 0.03973681107163429, 'reward_count_adherence_mean': 0.7678571343421936, 'reward_count_adherence_std': 0.033064987510442734, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9239448308944702, 'reward_repeat_penalty_std': 0.04794442281126976, 'reward_total_composite_mean': 0.6861064434051514, 'reward_total_composite_std': 0.04557538405060768, 'epoch': 0.07}

 51%|█████▏    | 1698/3300 [2:03:43<2:47:07,  6.26s/it]INFO 04-12 00:40:34 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1699/3300 [2:03:47<2:33:51,  5.77s/it]2026-04-12 00:40:38,984 | INFO | train_grpo_train | metrics_logged mode=train step=1699

                                                       
{'loss': -0.0066, 'grad_norm': 3.725172758102417, 'learning_rate': 4.854545454545455e-06, 'num_tokens': 3818997.0, 'completions/mean_length': 72.375, 'completions/min_length': 70.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9986300468444824, 'rewards/meter/std': 0.0003518488083500415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9986300468444824, 'rewards/total_composite/std': 0.0003518488083500415, 'reward': 0.9986300468444824, 'reward_std': 0.00035184345324523747, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03500673174858093, 'sampling/sampling_logp_difference/max': 1.481827974319458, 'sampling/importance_sampling_ratio/min': 0.22722196578979492, 'sampling/importance_sampling_ratio/mean': 1.0039602518081665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23103010654449463, 'clip_ratio/low_mean': 0.010468951310031116, 'clip_ratio/low_min': 0.010468951310031116, 'clip_ratio/high_mean': 0.01543669670354575, 'clip_ratio/high_max': 0.01543669670354575, 'clip_ratio/region_mean': 0.025905648013576865, 'reward_total_mean': 0.9986300468444824, 'reward_meter_mean': 0.9986300468444824, 'reward_meter_std': 0.0003518488083500415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9986300468444824, 'reward_total_composite_std': 0.0003518488083500415, 'epoch': 0.07}

 51%|█████▏    | 1699/3300 [2:03:47<2:33:51,  5.77s/it]INFO 04-12 00:40:39 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1700/3300 [2:03:53<2:33:21,  5.75s/it]2026-04-12 00:40:44,699 | INFO | train_grpo_train | metrics_logged mode=train step=1700

                                                       
{'loss': 0.0025, 'grad_norm': 1.9409257173538208, 'learning_rate': 4.851515151515152e-06, 'num_tokens': 3822000.0, 'completions/mean_length': 171.375, 'completions/min_length': 170.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.375, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.9947851896286011, 'rewards/meter/std': 0.0019038140308111906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6875, 'rewards/repeat_penalty/std': 0.035355325788259506, 'rewards/total_composite/mean': 0.6838877201080322, 'rewards/total_composite/std': 0.034654706716537476, 'reward': 0.6838877201080322, 'reward_std': 0.034654706716537476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014711027964949608, 'sampling/sampling_logp_difference/max': 6.82924747467041, 'sampling/importance_sampling_ratio/min': 0.0010816717986017466, 'sampling/importance_sampling_ratio/mean': 0.9983268976211548, 'sampling/importance_sampling_ratio/max': 1.4856215715408325, 'entropy': 0.04710668884217739, 'clip_ratio/low_mean': 0.0007267441833391786, 'clip_ratio/low_min': 0.0007267441833391786, 'clip_ratio/high_mean': 0.008750928391236812, 'clip_ratio/high_max': 0.008750928391236812, 'clip_ratio/region_mean': 0.00947767257457599, 'reward_total_mean': 0.6838877201080322, 'reward_meter_mean': 0.9947851896286011, 'reward_meter_std': 0.0019038140308111906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6875, 'reward_repeat_penalty_std': 0.035355325788259506, 'reward_total_composite_mean': 0.6838877201080322, 'reward_total_composite_std': 0.034654706716537476, 'epoch': 0.07}

 52%|█████▏    | 1700/3300 [2:03:53<2:33:21,  5.75s/it]INFO 04-12 00:40:44 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:30,  2.78s/it]

 23%|██▎       | 3/13 [00:09<00:31,  3.11s/it]

 31%|███       | 4/13 [00:14<00:34,  3.84s/it]

 38%|███▊      | 5/13 [00:19<00:34,  4.37s/it]

 46%|████▌     | 6/13 [00:25<00:34,  4.97s/it]

 54%|█████▍    | 7/13 [00:30<00:29,  4.97s/it]

 62%|██████▏   | 8/13 [00:37<00:27,  5.56s/it]

 69%|██████▉   | 9/13 [00:44<00:23,  5.99s/it]

 77%|███████▋  | 10/13 [00:49<00:16,  5.61s/it]

 85%|████████▍ | 11/13 [00:55<00:11,  5.76s/it]

 92%|█████████▏| 12/13 [01:01<00:05,  5.76s/it]

100%|██████████| 13/13 [01:06<00:00,  5.62s/it]2026-04-12 00:41:57,403 | INFO | train_grpo_train | metrics_logged mode=eval step=1700

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 72.6468, 'eval_samples_per_second': 1.432, 'eval_steps_per_second': 0.179, 'eval_num_tokens': 3822000.0, 'eval_completions/mean_length': 213.16346153846155, 'eval_completions/min_length': 64.15384615384616, 'eval_completions/max_length': 386.46153846153845, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 213.16346153846155, 'eval_completions/min_terminated_length': 64.15384615384616, 'eval_completions/max_terminated_length': 386.46153846153845, 'eval_rewards/meter/mean': 0.7185193575345553, 'eval_rewards/meter/std': 0.4064074754714966, 'eval_rewards/count_adherence/mean': 0.9180220273824838, 'eval_rewards/count_adherence/std': 0.10367171552318794, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.7850257295828599, 'eval_rewards/repeat_penalty/std': 0.15948278972735772, 'eval_rewards/total_composite/mean': 0.5148563522558945, 'eval_rewards/total_composite/std': 0.3435696088350736, 'eval_reward': 0.5148563522558945, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.020774336352657814, 'eval_sampling/sampling_logp_difference/max': 1.5070369427020733, 'eval_sampling/importance_sampling_ratio/min': 0.3100252260382359, 'eval_sampling/importance_sampling_ratio/mean': 1.0051932701697717, 'eval_sampling/importance_sampling_ratio/max': 1.4555618029374342, 'eval_entropy': 0.20546403641884142, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5148563522558945, 'eval_reward_meter_mean': 0.7185193575345553, 'eval_reward_meter_std': 0.4064074754714966, 'eval_reward_count_adherence_mean': 0.9180220273824838, 'eval_reward_count_adherence_std': 0.10367171552318794, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.7850257295828599, 'eval_reward_repeat_penalty_std': 0.15948278972735772, 'eval_reward_total_composite_mean': 0.5148563522558945, 'eval_reward_total_composite_std': 0.3435696088350736, 'epoch': 0.07}

 52%|█████▏    | 1700/3300 [2:05:06<2:33:21,  5.75s/it]

100%|██████████| 13/13 [01:06<00:00,  5.62s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 52%|█████▏    | 1701/3300 [2:05:15<12:41:02, 28.56s/it]2026-04-12 00:42:06,476 | INFO | train_grpo_train | metrics_logged mode=train step=1701

                                                        
{'loss': 0.0094, 'grad_norm': 2.8516652584075928, 'learning_rate': 4.848484848484849e-06, 'num_tokens': 3825370.0, 'completions/mean_length': 247.25, 'completions/min_length': 228.0, 'completions/max_length': 266.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 247.25, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 266.0, 'rewards/meter/mean': 0.8367831110954285, 'rewards/meter/std': 0.21646961569786072, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7960164546966553, 'rewards/repeat_penalty/std': 0.12943987548351288, 'rewards/total_composite/mean': 0.6609899997711182, 'rewards/total_composite/std': 0.1882331371307373, 'reward': 0.6609899997711182, 'reward_std': 0.1882331222295761, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04899187758564949, 'sampling/sampling_logp_difference/max': 11.93899917602539, 'sampling/importance_sampling_ratio/min': 6.5306817305099685e-06, 'sampling/importance_sampling_ratio/mean': 1.0024902820587158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2732624653726816, 'clip_ratio/low_mean': 0.008105989079922438, 'clip_ratio/low_min': 0.008105989079922438, 'clip_ratio/high_mean': 0.02072087489068508, 'clip_ratio/high_max': 0.02072087489068508, 'clip_ratio/region_mean': 0.02882686397060752, 'reward_total_mean': 0.6609899997711182, 'reward_meter_mean': 0.8367831110954285, 'reward_meter_std': 0.21646961569786072, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7960164546966553, 'reward_repeat_penalty_std': 0.12943987548351288, 'reward_total_composite_mean': 0.6609899997711182, 'reward_total_composite_std': 0.1882331371307373, 'epoch': 0.07}

 52%|█████▏    | 1701/3300 [2:05:15<12:41:02, 28.56s/it]INFO 04-12 00:42:06 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1702/3300 [2:05:20<9:33:33, 21.54s/it] 2026-04-12 00:42:11,625 | INFO | train_grpo_train | metrics_logged mode=train step=1702

                                                       
{'loss': 0.013, 'grad_norm': 3.22273325920105, 'learning_rate': 4.845454545454546e-06, 'num_tokens': 3827574.0, 'completions/mean_length': 108.5, 'completions/min_length': 102.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.5, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.9983426332473755, 'rewards/meter/std': 0.000583991059102118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8985103964805603, 'rewards/total_composite/std': 0.10674867779016495, 'reward': 0.8985103964805603, 'reward_std': 0.10674867779016495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03188902884721756, 'sampling/sampling_logp_difference/max': 1.428539752960205, 'sampling/importance_sampling_ratio/min': 0.2396586388349533, 'sampling/importance_sampling_ratio/mean': 1.0079683065414429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24685455113649368, 'clip_ratio/low_mean': 0.011560323182493448, 'clip_ratio/low_min': 0.011560323182493448, 'clip_ratio/high_mean': 0.012898168293759227, 'clip_ratio/high_max': 0.012898168293759227, 'clip_ratio/region_mean': 0.024458491476252675, 'reward_total_mean': 0.8985103964805603, 'reward_meter_mean': 0.9983426332473755, 'reward_meter_std': 0.000583991059102118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8985103964805603, 'reward_total_composite_std': 0.10674867779016495, 'epoch': 0.07}

 52%|█████▏    | 1702/3300 [2:05:20<9:33:33, 21.54s/it]INFO 04-12 00:42:11 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1703/3300 [2:05:28<7:45:16, 17.48s/it]2026-04-12 00:42:19,642 | INFO | train_grpo_train | metrics_logged mode=train step=1703

                                                       
{'loss': 0.0086, 'grad_norm': 1.9336153268814087, 'learning_rate': 4.842424242424243e-06, 'num_tokens': 3832143.0, 'completions/mean_length': 328.125, 'completions/min_length': 316.0, 'completions/max_length': 352.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 328.125, 'completions/min_terminated_length': 316.0, 'completions/max_terminated_length': 352.0, 'rewards/meter/mean': 0.9899157285690308, 'rewards/meter/std': 0.008645469322800636, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9504464268684387, 'rewards/repeat_penalty/std': 0.030713941901922226, 'rewards/total_composite/mean': 0.8237563371658325, 'rewards/total_composite/std': 0.054851170629262924, 'reward': 0.8237563371658325, 'reward_std': 0.054851170629262924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05240955203771591, 'sampling/sampling_logp_difference/max': 2.064100742340088, 'sampling/importance_sampling_ratio/min': 0.12693238258361816, 'sampling/importance_sampling_ratio/mean': 1.0071643590927124, 'sampling/importance_sampling_ratio/max': 1.909447431564331, 'entropy': 0.4287557154893875, 'clip_ratio/low_mean': 0.017736590933054686, 'clip_ratio/low_min': 0.017736590933054686, 'clip_ratio/high_mean': 0.019598021870478988, 'clip_ratio/high_max': 0.019598021870478988, 'clip_ratio/region_mean': 0.03733461280353367, 'reward_total_mean': 0.8237563371658325, 'reward_meter_mean': 0.9899157285690308, 'reward_meter_std': 0.008645469322800636, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9504464268684387, 'reward_repeat_penalty_std': 0.030713941901922226, 'reward_total_composite_mean': 0.8237563371658325, 'reward_total_composite_std': 0.054851170629262924, 'epoch': 0.07}

 52%|█████▏    | 1703/3300 [2:05:28<7:45:16, 17.48s/it]INFO 04-12 00:42:19 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1704/3300 [2:05:34<6:10:45, 13.94s/it]2026-04-12 00:42:25,317 | INFO | train_grpo_train | metrics_logged mode=train step=1704

                                                       
{'loss': 0.0231, 'grad_norm': 4.468354225158691, 'learning_rate': 4.83939393939394e-06, 'num_tokens': 3834840.0, 'completions/mean_length': 143.125, 'completions/min_length': 140.0, 'completions/max_length': 151.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.125, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.9751874804496765, 'rewards/meter/std': 0.05766330659389496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7313905954360962, 'rewards/total_composite/std': 0.04324747994542122, 'reward': 0.7313905954360962, 'reward_std': 0.04324747994542122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013407315127551556, 'sampling/sampling_logp_difference/max': 1.5817131996154785, 'sampling/importance_sampling_ratio/min': 0.20562252402305603, 'sampling/importance_sampling_ratio/mean': 1.0032126903533936, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05452722427435219, 'clip_ratio/low_mean': 0.002483443822711706, 'clip_ratio/low_min': 0.002483443822711706, 'clip_ratio/high_mean': 0.008767852385062724, 'clip_ratio/high_max': 0.008767852385062724, 'clip_ratio/region_mean': 0.01125129620777443, 'reward_total_mean': 0.7313905954360962, 'reward_meter_mean': 0.9751874804496765, 'reward_meter_std': 0.05766330659389496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7313905954360962, 'reward_total_composite_std': 0.04324747994542122, 'epoch': 0.07}

 52%|█████▏    | 1704/3300 [2:05:34<6:10:45, 13.94s/it]INFO 04-12 00:42:25 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1705/3300 [2:05:38<4:56:35, 11.16s/it]2026-04-12 00:42:29,985 | INFO | train_grpo_train | metrics_logged mode=train step=1705

                                                       
{'loss': 0.0492, 'grad_norm': 9.636635780334473, 'learning_rate': 4.836363636363637e-06, 'num_tokens': 3836711.0, 'completions/mean_length': 75.875, 'completions/min_length': 71.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9914894104003906, 'rewards/meter/std': 0.009747824631631374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9914894104003906, 'rewards/total_composite/std': 0.009747824631631374, 'reward': 0.9914894104003906, 'reward_std': 0.009747819975018501, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07412047684192657, 'sampling/sampling_logp_difference/max': 1.237552523612976, 'sampling/importance_sampling_ratio/min': 0.290093332529068, 'sampling/importance_sampling_ratio/mean': 1.0037219524383545, 'sampling/importance_sampling_ratio/max': 1.6591025590896606, 'entropy': 0.66333282366395, 'clip_ratio/low_mean': 0.02736378228291869, 'clip_ratio/low_min': 0.02736378228291869, 'clip_ratio/high_mean': 0.033178919460624456, 'clip_ratio/high_max': 0.033178919460624456, 'clip_ratio/region_mean': 0.06054270174354315, 'reward_total_mean': 0.9914894104003906, 'reward_meter_mean': 0.9914894104003906, 'reward_meter_std': 0.009747824631631374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9914894104003906, 'reward_total_composite_std': 0.009747824631631374, 'epoch': 0.07}

 52%|█████▏    | 1705/3300 [2:05:38<4:56:35, 11.16s/it]INFO 04-12 00:42:30 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1706/3300 [2:05:46<4:27:18, 10.06s/it]2026-04-12 00:42:37,488 | INFO | train_grpo_train | metrics_logged mode=train step=1706

                                                       
{'loss': -0.016, 'grad_norm': 1.8596045970916748, 'learning_rate': 4.833333333333333e-06, 'num_tokens': 3840705.0, 'completions/mean_length': 302.25, 'completions/min_length': 279.0, 'completions/max_length': 322.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 302.25, 'completions/min_terminated_length': 279.0, 'completions/max_terminated_length': 322.0, 'rewards/meter/mean': 0.998089075088501, 'rewards/meter/std': 0.0013965349644422531, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.604687511920929, 'rewards/repeat_penalty/std': 0.15109451115131378, 'rewards/total_composite/mean': 0.6033692955970764, 'rewards/total_composite/std': 0.14974084496498108, 'reward': 0.6033692955970764, 'reward_std': 0.1497408151626587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022071924060583115, 'sampling/sampling_logp_difference/max': 7.479898929595947, 'sampling/importance_sampling_ratio/min': 0.0005643144831992686, 'sampling/importance_sampling_ratio/mean': 1.0016717910766602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13442990742623806, 'clip_ratio/low_mean': 0.007105766620952636, 'clip_ratio/low_min': 0.007105766620952636, 'clip_ratio/high_mean': 0.0056348692160099745, 'clip_ratio/high_max': 0.0056348692160099745, 'clip_ratio/region_mean': 0.01274063583696261, 'reward_total_mean': 0.6033692955970764, 'reward_meter_mean': 0.998089075088501, 'reward_meter_std': 0.0013965349644422531, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.604687511920929, 'reward_repeat_penalty_std': 0.15109451115131378, 'reward_total_composite_mean': 0.6033692955970764, 'reward_total_composite_std': 0.14974084496498108, 'epoch': 0.07}

 52%|█████▏    | 1706/3300 [2:05:46<4:27:18, 10.06s/it]INFO 04-12 00:42:37 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1707/3300 [2:05:50<3:42:07,  8.37s/it]2026-04-12 00:42:41,896 | INFO | train_grpo_train | metrics_logged mode=train step=1707

                                                       
{'loss': 0.0035, 'grad_norm': 4.459820747375488, 'learning_rate': 4.830303030303031e-06, 'num_tokens': 3842546.0, 'completions/mean_length': 61.125, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9614025354385376, 'rewards/meter/std': 0.046098750084638596, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9614025354385376, 'rewards/total_composite/std': 0.046098750084638596, 'reward': 0.9614025354385376, 'reward_std': 0.046098742634058, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0242039505392313, 'sampling/sampling_logp_difference/max': 0.7878568172454834, 'sampling/importance_sampling_ratio/min': 0.5145033597946167, 'sampling/importance_sampling_ratio/mean': 1.005699872970581, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1731706503778696, 'clip_ratio/low_mean': 0.004166666883975267, 'clip_ratio/low_min': 0.004166666883975267, 'clip_ratio/high_mean': 0.014213158516213298, 'clip_ratio/high_max': 0.014213158516213298, 'clip_ratio/region_mean': 0.018379825400188565, 'reward_total_mean': 0.9614025354385376, 'reward_meter_mean': 0.9614025354385376, 'reward_meter_std': 0.046098750084638596, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9614025354385376, 'reward_total_composite_std': 0.046098750084638596, 'epoch': 0.07}

 52%|█████▏    | 1707/3300 [2:05:50<3:42:07,  8.37s/it]INFO 04-12 00:42:42 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1708/3300 [2:05:55<3:12:21,  7.25s/it]2026-04-12 00:42:46,545 | INFO | train_grpo_train | metrics_logged mode=train step=1708

                                                       
{'loss': -0.0028, 'grad_norm': 4.370044231414795, 'learning_rate': 4.827272727272728e-06, 'num_tokens': 3844472.0, 'completions/mean_length': 75.75, 'completions/min_length': 72.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9969611167907715, 'rewards/meter/std': 0.0019538968335837126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969611167907715, 'rewards/total_composite/std': 0.0019538968335837126, 'reward': 0.9969611167907715, 'reward_std': 0.001953905913978815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048442739993333817, 'sampling/sampling_logp_difference/max': 1.1050148010253906, 'sampling/importance_sampling_ratio/min': 0.33120596408843994, 'sampling/importance_sampling_ratio/mean': 1.0110034942626953, 'sampling/importance_sampling_ratio/max': 1.6920305490493774, 'entropy': 0.360767001286149, 'clip_ratio/low_mean': 0.011602062964811921, 'clip_ratio/low_min': 0.011602062964811921, 'clip_ratio/high_mean': 0.019748708698898554, 'clip_ratio/high_max': 0.019748708698898554, 'clip_ratio/region_mean': 0.031350771663710475, 'reward_total_mean': 0.9969611167907715, 'reward_meter_mean': 0.9969611167907715, 'reward_meter_std': 0.0019538968335837126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969611167907715, 'reward_total_composite_std': 0.0019538968335837126, 'epoch': 0.07}

 52%|█████▏    | 1708/3300 [2:05:55<3:12:21,  7.25s/it]INFO 04-12 00:42:46 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1709/3300 [2:06:00<2:52:29,  6.51s/it]2026-04-12 00:42:51,314 | INFO | train_grpo_train | metrics_logged mode=train step=1709

                                                       
{'loss': 0.0066, 'grad_norm': 7.5510640144348145, 'learning_rate': 4.824242424242424e-06, 'num_tokens': 3846403.0, 'completions/mean_length': 79.375, 'completions/min_length': 71.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.375, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9834517240524292, 'rewards/meter/std': 0.016320018097758293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9834517240524292, 'rewards/total_composite/std': 0.016320018097758293, 'reward': 0.9834517240524292, 'reward_std': 0.016320008784532547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06894521415233612, 'sampling/sampling_logp_difference/max': 1.456099510192871, 'sampling/importance_sampling_ratio/min': 0.2331438809633255, 'sampling/importance_sampling_ratio/mean': 1.019029974937439, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.663462370634079, 'clip_ratio/low_mean': 0.029091242235153913, 'clip_ratio/low_min': 0.029091242235153913, 'clip_ratio/high_mean': 0.029255983070470393, 'clip_ratio/high_max': 0.029255983070470393, 'clip_ratio/region_mean': 0.058347225305624306, 'reward_total_mean': 0.9834517240524292, 'reward_meter_mean': 0.9834517240524292, 'reward_meter_std': 0.016320018097758293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9834517240524292, 'reward_total_composite_std': 0.016320018097758293, 'epoch': 0.07}

 52%|█████▏    | 1709/3300 [2:06:00<2:52:29,  6.51s/it]INFO 04-12 00:42:51 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1710/3300 [2:06:05<2:40:21,  6.05s/it]2026-04-12 00:42:56,306 | INFO | train_grpo_train | metrics_logged mode=train step=1710

                                                       
{'loss': 0.0194, 'grad_norm': 2.7458198070526123, 'learning_rate': 4.8212121212121215e-06, 'num_tokens': 3848628.0, 'completions/mean_length': 105.125, 'completions/min_length': 102.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.125, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9941129684448242, 'rewards/meter/std': 0.0036593328695744276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9941129684448242, 'rewards/total_composite/std': 0.0036593328695744276, 'reward': 0.9941129684448242, 'reward_std': 0.0036593314725905657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020809238776564598, 'sampling/sampling_logp_difference/max': 1.1916732788085938, 'sampling/importance_sampling_ratio/min': 0.3037126362323761, 'sampling/importance_sampling_ratio/mean': 1.0030418634414673, 'sampling/importance_sampling_ratio/max': 1.5832405090332031, 'entropy': 0.17343183141201735, 'clip_ratio/low_mean': 0.01390780950896442, 'clip_ratio/low_min': 0.01390780950896442, 'clip_ratio/high_mean': 0.008437702199444175, 'clip_ratio/high_max': 0.008437702199444175, 'clip_ratio/region_mean': 0.022345511708408594, 'reward_total_mean': 0.9941129684448242, 'reward_meter_mean': 0.9941129684448242, 'reward_meter_std': 0.0036593328695744276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9941129684448242, 'reward_total_composite_std': 0.0036593328695744276, 'epoch': 0.07}

 52%|█████▏    | 1710/3300 [2:06:05<2:40:21,  6.05s/it]INFO 04-12 00:42:56 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1711/3300 [2:06:09<2:27:30,  5.57s/it]2026-04-12 00:43:00,756 | INFO | train_grpo_train | metrics_logged mode=train step=1711

                                                       
{'loss': 0.0035, 'grad_norm': 6.050206661224365, 'learning_rate': 4.818181818181819e-06, 'num_tokens': 3850396.0, 'completions/mean_length': 62.0, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.992821455001831, 'rewards/meter/std': 0.0031417824793606997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.992821455001831, 'rewards/total_composite/std': 0.0031417824793606997, 'reward': 0.992821455001831, 'reward_std': 0.00314177293330431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028909733518958092, 'sampling/sampling_logp_difference/max': 0.8604459762573242, 'sampling/importance_sampling_ratio/min': 0.4229734241962433, 'sampling/importance_sampling_ratio/mean': 1.0075486898422241, 'sampling/importance_sampling_ratio/max': 1.8755117654800415, 'entropy': 0.21017488650977612, 'clip_ratio/low_mean': 0.010147797176614404, 'clip_ratio/low_min': 0.010147797176614404, 'clip_ratio/high_mean': 0.010149948066100478, 'clip_ratio/high_max': 0.010149948066100478, 'clip_ratio/region_mean': 0.020297745242714882, 'reward_total_mean': 0.992821455001831, 'reward_meter_mean': 0.992821455001831, 'reward_meter_std': 0.0031417824793606997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.992821455001831, 'reward_total_composite_std': 0.0031417824793606997, 'epoch': 0.07}

 52%|█████▏    | 1711/3300 [2:06:09<2:27:30,  5.57s/it]INFO 04-12 00:43:00 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1712/3300 [2:06:13<2:16:06,  5.14s/it]2026-04-12 00:43:04,895 | INFO | train_grpo_train | metrics_logged mode=train step=1712

                                                       
{'loss': 0.0252, 'grad_norm': 9.417886734008789, 'learning_rate': 4.815151515151515e-06, 'num_tokens': 3851781.0, 'completions/mean_length': 31.125, 'completions/min_length': 30.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.99363112449646, 'rewards/meter/std': 0.0030946105252951384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99363112449646, 'rewards/total_composite/std': 0.0030946105252951384, 'reward': 0.99363112449646, 'reward_std': 0.0030946091283112764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01712997443974018, 'sampling/sampling_logp_difference/max': 0.6930625438690186, 'sampling/importance_sampling_ratio/min': 0.5000423192977905, 'sampling/importance_sampling_ratio/mean': 1.0098103284835815, 'sampling/importance_sampling_ratio/max': 1.6804507970809937, 'entropy': 0.138029710855335, 'clip_ratio/low_mean': 0.011742424685508013, 'clip_ratio/low_min': 0.011742424685508013, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.011742424685508013, 'reward_total_mean': 0.99363112449646, 'reward_meter_mean': 0.99363112449646, 'reward_meter_std': 0.0030946105252951384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99363112449646, 'reward_total_composite_std': 0.0030946105252951384, 'epoch': 0.07}

 52%|█████▏    | 1712/3300 [2:06:13<2:16:06,  5.14s/it]INFO 04-12 00:43:05 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1713/3300 [2:06:18<2:11:50,  4.98s/it]2026-04-12 00:43:09,513 | INFO | train_grpo_train | metrics_logged mode=train step=1713

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.8121212121212125e-06, 'num_tokens': 3853917.0, 'completions/mean_length': 86.0, 'completions/min_length': 86.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.0, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9939366579055786, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7951493263244629, 'rewards/total_composite/std': 0.0, 'reward': 0.7951493263244629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00075820047641173, 'sampling/sampling_logp_difference/max': 0.02473144233226776, 'sampling/importance_sampling_ratio/min': 0.975571870803833, 'sampling/importance_sampling_ratio/mean': 1.0006271600723267, 'sampling/importance_sampling_ratio/max': 1.0226460695266724, 'entropy': 0.007212614000309259, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7951493263244629, 'reward_meter_mean': 0.9939366579055786, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7951493263244629, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 52%|█████▏    | 1713/3300 [2:06:18<2:11:50,  4.98s/it]INFO 04-12 00:43:09 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1714/3300 [2:06:22<2:06:26,  4.78s/it]2026-04-12 00:43:13,830 | INFO | train_grpo_train | metrics_logged mode=train step=1714

                                                       
{'loss': 0.0002, 'grad_norm': 0.25375768542289734, 'learning_rate': 4.80909090909091e-06, 'num_tokens': 3855829.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9982520937919617, 'rewards/meter/std': 1.3676654816663358e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982520937919617, 'rewards/total_composite/std': 1.3676654816663358e-05, 'reward': 0.9982520937919617, 'reward_std': 1.3667625353264157e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0014053680934011936, 'sampling/sampling_logp_difference/max': 0.2431938648223877, 'sampling/importance_sampling_ratio/min': 0.7841194868087769, 'sampling/importance_sampling_ratio/mean': 0.9998658895492554, 'sampling/importance_sampling_ratio/max': 1.0173074007034302, 'entropy': 0.007174990780185908, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9982520937919617, 'reward_meter_mean': 0.9982520937919617, 'reward_meter_std': 1.3676654816663358e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982520937919617, 'reward_total_composite_std': 1.3676654816663358e-05, 'epoch': 0.07}

 52%|█████▏    | 1714/3300 [2:06:22<2:06:26,  4.78s/it]INFO 04-12 00:43:14 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1715/3300 [2:06:27<2:03:44,  4.68s/it]2026-04-12 00:43:18,285 | INFO | train_grpo_train | metrics_logged mode=train step=1715

                                                       
{'loss': 0.0006, 'grad_norm': 0.6869344115257263, 'learning_rate': 4.806060606060606e-06, 'num_tokens': 3857509.0, 'completions/mean_length': 50.0, 'completions/min_length': 50.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9374866485595703, 'rewards/meter/std': 3.270595334470272e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9374866485595703, 'rewards/total_composite/std': 3.270595334470272e-05, 'reward': 0.9374866485595703, 'reward_std': 3.270595334470272e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006228248123079538, 'sampling/sampling_logp_difference/max': 0.5891187191009521, 'sampling/importance_sampling_ratio/min': 0.5548160076141357, 'sampling/importance_sampling_ratio/mean': 1.0022356510162354, 'sampling/importance_sampling_ratio/max': 1.3050650358200073, 'entropy': 0.04075565282255411, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0024999999441206455, 'clip_ratio/high_max': 0.0024999999441206455, 'clip_ratio/region_mean': 0.0024999999441206455, 'reward_total_mean': 0.9374866485595703, 'reward_meter_mean': 0.9374866485595703, 'reward_meter_std': 3.270595334470272e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9374866485595703, 'reward_total_composite_std': 3.270595334470272e-05, 'epoch': 0.07}

 52%|█████▏    | 1715/3300 [2:06:27<2:03:44,  4.68s/it]INFO 04-12 00:43:18 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1716/3300 [2:06:31<2:03:25,  4.68s/it]2026-04-12 00:43:22,945 | INFO | train_grpo_train | metrics_logged mode=train step=1716

                                                       
{'loss': -0.0025, 'grad_norm': 1.8621673583984375, 'learning_rate': 4.803030303030303e-06, 'num_tokens': 3859301.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9931608438491821, 'rewards/meter/std': 0.0010164134437218308, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9931608438491821, 'rewards/total_composite/std': 0.0010164134437218308, 'reward': 0.9931608438491821, 'reward_std': 0.001016413327306509, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003477482357993722, 'sampling/sampling_logp_difference/max': 1.2498140335083008, 'sampling/importance_sampling_ratio/min': 0.2865580916404724, 'sampling/importance_sampling_ratio/mean': 0.9991500377655029, 'sampling/importance_sampling_ratio/max': 1.0107558965682983, 'entropy': 0.007712913560681045, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9931608438491821, 'reward_meter_mean': 0.9931608438491821, 'reward_meter_std': 0.0010164134437218308, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9931608438491821, 'reward_total_composite_std': 0.0010164134437218308, 'epoch': 0.07}

 52%|█████▏    | 1716/3300 [2:06:31<2:03:25,  4.68s/it]INFO 04-12 00:43:23 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1717/3300 [2:06:41<2:39:14,  6.04s/it]2026-04-12 00:43:32,147 | INFO | train_grpo_train | metrics_logged mode=train step=1717

                                                       
{'loss': 0.0087, 'grad_norm': 2.0890252590179443, 'learning_rate': 4.800000000000001e-06, 'num_tokens': 3864531.0, 'completions/mean_length': 403.75, 'completions/min_length': 377.0, 'completions/max_length': 424.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 403.75, 'completions/min_terminated_length': 377.0, 'completions/max_terminated_length': 424.0, 'rewards/meter/mean': 0.9169327616691589, 'rewards/meter/std': 0.19917452335357666, 'rewards/count_adherence/mean': 0.6583333015441895, 'rewards/count_adherence/std': 0.0235702246427536, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9466373920440674, 'rewards/repeat_penalty/std': 0.04092821478843689, 'rewards/total_composite/mean': 0.49961453676223755, 'rewards/total_composite/std': 0.24736955761909485, 'reward': 0.49961453676223755, 'reward_std': 0.24736955761909485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049914389848709106, 'sampling/sampling_logp_difference/max': 1.5311355590820312, 'sampling/importance_sampling_ratio/min': 0.2162899225950241, 'sampling/importance_sampling_ratio/mean': 1.0099897384643555, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4459982290863991, 'clip_ratio/low_mean': 0.007035988033749163, 'clip_ratio/low_min': 0.007035988033749163, 'clip_ratio/high_mean': 0.033190221060067415, 'clip_ratio/high_max': 0.033190221060067415, 'clip_ratio/region_mean': 0.04022620909381658, 'reward_total_mean': 0.49961453676223755, 'reward_meter_mean': 0.9169327616691589, 'reward_meter_std': 0.19917452335357666, 'reward_count_adherence_mean': 0.6583333015441895, 'reward_count_adherence_std': 0.0235702246427536, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9466373920440674, 'reward_repeat_penalty_std': 0.04092821478843689, 'reward_total_composite_mean': 0.49961453676223755, 'reward_total_composite_std': 0.24736955761909485, 'epoch': 0.07}

 52%|█████▏    | 1717/3300 [2:06:41<2:39:14,  6.04s/it]INFO 04-12 00:43:32 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1718/3300 [2:06:45<2:28:08,  5.62s/it]2026-04-12 00:43:36,800 | INFO | train_grpo_train | metrics_logged mode=train step=1718

                                                       
{'loss': 0.0003, 'grad_norm': 0.21243786811828613, 'learning_rate': 4.796969696969697e-06, 'num_tokens': 3866227.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9982569217681885, 'rewards/meter/std': 1.7906948414747603e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982569217681885, 'rewards/total_composite/std': 1.7906948414747603e-05, 'reward': 0.9982569217681885, 'reward_std': 1.7897747966344468e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00201884051784873, 'sampling/sampling_logp_difference/max': 0.3060312271118164, 'sampling/importance_sampling_ratio/min': 0.7363636493682861, 'sampling/importance_sampling_ratio/mean': 0.9998619556427002, 'sampling/importance_sampling_ratio/max': 1.103933572769165, 'entropy': 0.009457326959818602, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.001953125, 'reward_total_mean': 0.9982569217681885, 'reward_meter_mean': 0.9982569217681885, 'reward_meter_std': 1.7906948414747603e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982569217681885, 'reward_total_composite_std': 1.7906948414747603e-05, 'epoch': 0.07}

 52%|█████▏    | 1718/3300 [2:06:45<2:28:08,  5.62s/it]INFO 04-12 00:43:37 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1719/3300 [2:06:53<2:44:08,  6.23s/it]2026-04-12 00:43:44,458 | INFO | train_grpo_train | metrics_logged mode=train step=1719

                                                       
{'loss': -0.0103, 'grad_norm': 2.303537368774414, 'learning_rate': 4.793939393939394e-06, 'num_tokens': 3870642.0, 'completions/mean_length': 282.875, 'completions/min_length': 273.0, 'completions/max_length': 292.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 282.875, 'completions/min_terminated_length': 273.0, 'completions/max_terminated_length': 292.0, 'rewards/meter/mean': 0.9829399585723877, 'rewards/meter/std': 0.022622600197792053, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.942307710647583, 'rewards/repeat_penalty/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.9256449937820435, 'rewards/total_composite/std': 0.04545941948890686, 'reward': 0.9256449937820435, 'reward_std': 0.04545941576361656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06017490476369858, 'sampling/sampling_logp_difference/max': 4.142031669616699, 'sampling/importance_sampling_ratio/min': 0.015890534967184067, 'sampling/importance_sampling_ratio/mean': 1.0142277479171753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5307335741817951, 'clip_ratio/low_mean': 0.021380615420639515, 'clip_ratio/low_min': 0.021380615420639515, 'clip_ratio/high_mean': 0.02710696868598461, 'clip_ratio/high_max': 0.02710696868598461, 'clip_ratio/region_mean': 0.048487584106624126, 'reward_total_mean': 0.9256449937820435, 'reward_meter_mean': 0.9829399585723877, 'reward_meter_std': 0.022622600197792053, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.942307710647583, 'reward_repeat_penalty_std': 0.054392825812101364, 'reward_total_composite_mean': 0.9256449937820435, 'reward_total_composite_std': 0.04545941948890686, 'epoch': 0.07}

 52%|█████▏    | 1719/3300 [2:06:53<2:44:08,  6.23s/it]INFO 04-12 00:43:44 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1720/3300 [2:07:02<3:06:13,  7.07s/it]2026-04-12 00:43:53,491 | INFO | train_grpo_train | metrics_logged mode=train step=1720

                                                       
{'loss': -0.0169, 'grad_norm': 2.2101805210113525, 'learning_rate': 4.790909090909091e-06, 'num_tokens': 3875535.0, 'completions/mean_length': 400.625, 'completions/min_length': 368.0, 'completions/max_length': 418.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 400.625, 'completions/min_terminated_length': 368.0, 'completions/max_terminated_length': 418.0, 'rewards/meter/mean': 0.9720591306686401, 'rewards/meter/std': 0.04081534594297409, 'rewards/count_adherence/mean': 0.6499999761581421, 'rewards/count_adherence/std': 0.030860668048262596, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9199131727218628, 'rewards/repeat_penalty/std': 0.08860698342323303, 'rewards/total_composite/mean': 0.5811381340026855, 'rewards/total_composite/std': 0.0678037479519844, 'reward': 0.5811381340026855, 'reward_std': 0.06780374050140381, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05924157425761223, 'sampling/sampling_logp_difference/max': 1.3779230117797852, 'sampling/importance_sampling_ratio/min': 0.2521016299724579, 'sampling/importance_sampling_ratio/mean': 1.0108215808868408, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5437066555023193, 'clip_ratio/low_mean': 0.01499476860044524, 'clip_ratio/low_min': 0.01499476860044524, 'clip_ratio/high_mean': 0.02708326978608966, 'clip_ratio/high_max': 0.02708326978608966, 'clip_ratio/region_mean': 0.0420780383865349, 'reward_total_mean': 0.5811381340026855, 'reward_meter_mean': 0.9720591306686401, 'reward_meter_std': 0.04081534594297409, 'reward_count_adherence_mean': 0.6499999761581421, 'reward_count_adherence_std': 0.030860668048262596, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9199131727218628, 'reward_repeat_penalty_std': 0.08860698342323303, 'reward_total_composite_mean': 0.5811381340026855, 'reward_total_composite_std': 0.0678037479519844, 'epoch': 0.07}

 52%|█████▏    | 1720/3300 [2:07:02<3:06:13,  7.07s/it]INFO 04-12 00:43:53 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1721/3300 [2:07:07<2:48:51,  6.42s/it]2026-04-12 00:43:58,373 | INFO | train_grpo_train | metrics_logged mode=train step=1721

                                                       
{'loss': 0.0135, 'grad_norm': 5.097805023193359, 'learning_rate': 4.787878787878788e-06, 'num_tokens': 3877383.0, 'completions/mean_length': 72.0, 'completions/min_length': 70.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9644453525543213, 'rewards/meter/std': 0.02713916078209877, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9644453525543213, 'rewards/total_composite/std': 0.02713916078209877, 'reward': 0.9644453525543213, 'reward_std': 0.027139168232679367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0413968488574028, 'sampling/sampling_logp_difference/max': 1.1499383449554443, 'sampling/importance_sampling_ratio/min': 0.31665629148483276, 'sampling/importance_sampling_ratio/mean': 1.0127274990081787, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3467879444360733, 'clip_ratio/low_mean': 0.013723055250011384, 'clip_ratio/low_min': 0.013723055250011384, 'clip_ratio/high_mean': 0.02464998373761773, 'clip_ratio/high_max': 0.02464998373761773, 'clip_ratio/region_mean': 0.038373038987629116, 'reward_total_mean': 0.9644453525543213, 'reward_meter_mean': 0.9644453525543213, 'reward_meter_std': 0.02713916078209877, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9644453525543213, 'reward_total_composite_std': 0.02713916078209877, 'epoch': 0.07}

 52%|█████▏    | 1721/3300 [2:07:07<2:48:51,  6.42s/it]INFO 04-12 00:43:58 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1722/3300 [2:07:15<3:01:07,  6.89s/it]2026-04-12 00:44:06,359 | INFO | train_grpo_train | metrics_logged mode=train step=1722

                                                       
{'loss': 0.0331, 'grad_norm': 2.7361412048339844, 'learning_rate': 4.784848484848485e-06, 'num_tokens': 3881642.0, 'completions/mean_length': 319.375, 'completions/min_length': 307.0, 'completions/max_length': 340.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 319.375, 'completions/min_terminated_length': 307.0, 'completions/max_terminated_length': 340.0, 'rewards/meter/mean': 0.7120033502578735, 'rewards/meter/std': 0.35525479912757874, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666865348816, 'rewards/repeat_penalty/std': 0.1321374922990799, 'rewards/total_composite/mean': 0.6435312032699585, 'rewards/total_composite/std': 0.32674068212509155, 'reward': 0.6435312032699585, 'reward_std': 0.32674068212509155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06873981654644012, 'sampling/sampling_logp_difference/max': 2.4230339527130127, 'sampling/importance_sampling_ratio/min': 0.17653696238994598, 'sampling/importance_sampling_ratio/mean': 1.011804461479187, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.594971913844347, 'clip_ratio/low_mean': 0.012550130486488342, 'clip_ratio/low_min': 0.012550130486488342, 'clip_ratio/high_mean': 0.04609629465267062, 'clip_ratio/high_max': 0.04609629465267062, 'clip_ratio/region_mean': 0.058646425139158964, 'reward_total_mean': 0.6435312032699585, 'reward_meter_mean': 0.7120033502578735, 'reward_meter_std': 0.35525479912757874, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666865348816, 'reward_repeat_penalty_std': 0.1321374922990799, 'reward_total_composite_mean': 0.6435312032699585, 'reward_total_composite_std': 0.32674068212509155, 'epoch': 0.07}

 52%|█████▏    | 1722/3300 [2:07:15<3:01:07,  6.89s/it]INFO 04-12 00:44:06 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1723/3300 [2:07:19<2:41:29,  6.14s/it]2026-04-12 00:44:10,769 | INFO | train_grpo_train | metrics_logged mode=train step=1723

                                                       
{'loss': -0.0342, 'grad_norm': 6.818909645080566, 'learning_rate': 4.7818181818181825e-06, 'num_tokens': 3883271.0, 'completions/mean_length': 38.625, 'completions/min_length': 35.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9971029758453369, 'rewards/meter/std': 0.0023624140303581953, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971029758453369, 'rewards/total_composite/std': 0.0023624140303581953, 'reward': 0.9971029758453369, 'reward_std': 0.002362414263188839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05782073736190796, 'sampling/sampling_logp_difference/max': 1.0367670059204102, 'sampling/importance_sampling_ratio/min': 0.354599267244339, 'sampling/importance_sampling_ratio/mean': 1.012098789215088, 'sampling/importance_sampling_ratio/max': 1.597930669784546, 'entropy': 0.50562334805727, 'clip_ratio/low_mean': 0.017084942432120442, 'clip_ratio/low_min': 0.017084942432120442, 'clip_ratio/high_mean': 0.028194751124829054, 'clip_ratio/high_max': 0.028194751124829054, 'clip_ratio/region_mean': 0.045279693556949496, 'reward_total_mean': 0.9971029758453369, 'reward_meter_mean': 0.9971029758453369, 'reward_meter_std': 0.0023624140303581953, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971029758453369, 'reward_total_composite_std': 0.0023624140303581953, 'epoch': 0.07}

 52%|█████▏    | 1723/3300 [2:07:19<2:41:29,  6.14s/it]INFO 04-12 00:44:11 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1724/3300 [2:07:25<2:36:52,  5.97s/it]2026-04-12 00:44:16,343 | INFO | train_grpo_train | metrics_logged mode=train step=1724

                                                       
{'loss': 0.0002, 'grad_norm': 0.12045075744390488, 'learning_rate': 4.77878787878788e-06, 'num_tokens': 3885143.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9982585310935974, 'rewards/meter/std': 4.937959602102637e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982585310935974, 'rewards/total_composite/std': 4.937959602102637e-05, 'reward': 0.9982585310935974, 'reward_std': 4.9387075705453753e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0029545121360570192, 'sampling/sampling_logp_difference/max': 0.4872865676879883, 'sampling/importance_sampling_ratio/min': 0.6142910122871399, 'sampling/importance_sampling_ratio/mean': 1.001138687133789, 'sampling/importance_sampling_ratio/max': 1.6037458181381226, 'entropy': 0.01320492452941835, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.9982585310935974, 'reward_meter_mean': 0.9982585310935974, 'reward_meter_std': 4.937959602102637e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982585310935974, 'reward_total_composite_std': 4.937959602102637e-05, 'epoch': 0.07}

 52%|█████▏    | 1724/3300 [2:07:25<2:36:52,  5.97s/it]INFO 04-12 00:44:16 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1725/3300 [2:07:29<2:26:06,  5.57s/it]2026-04-12 00:44:20,951 | INFO | train_grpo_train | metrics_logged mode=train step=1725

                                                       
{'loss': -0.0117, 'grad_norm': 3.2427947521209717, 'learning_rate': 4.775757575757576e-06, 'num_tokens': 3886827.0, 'completions/mean_length': 62.5, 'completions/min_length': 59.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9951238632202148, 'rewards/meter/std': 0.0010399873135611415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9951238632202148, 'rewards/total_composite/std': 0.0010399873135611415, 'reward': 0.9951238632202148, 'reward_std': 0.0010399814927950501, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009362027049064636, 'sampling/sampling_logp_difference/max': 0.8125953674316406, 'sampling/importance_sampling_ratio/min': 0.4437049925327301, 'sampling/importance_sampling_ratio/mean': 1.002244234085083, 'sampling/importance_sampling_ratio/max': 1.5123201608657837, 'entropy': 0.06342153111472726, 'clip_ratio/low_mean': 0.0042372881434857845, 'clip_ratio/low_min': 0.0042372881434857845, 'clip_ratio/high_mean': 0.0019841270986944437, 'clip_ratio/high_max': 0.0019841270986944437, 'clip_ratio/region_mean': 0.006221415242180228, 'reward_total_mean': 0.9951238632202148, 'reward_meter_mean': 0.9951238632202148, 'reward_meter_std': 0.0010399873135611415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9951238632202148, 'reward_total_composite_std': 0.0010399873135611415, 'epoch': 0.07}

 52%|█████▏    | 1725/3300 [2:07:29<2:26:06,  5.57s/it]INFO 04-12 00:44:21 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1726/3300 [2:07:35<2:25:34,  5.55s/it]2026-04-12 00:44:26,467 | INFO | train_grpo_train | metrics_logged mode=train step=1726

                                                       
{'loss': 0.0135, 'grad_norm': 3.5299017429351807, 'learning_rate': 4.772727272727273e-06, 'num_tokens': 3889004.0, 'completions/mean_length': 130.125, 'completions/min_length': 125.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.125, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.5966504812240601, 'rewards/meter/std': 0.2748834788799286, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8214285373687744, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.48184216022491455, 'rewards/total_composite/std': 0.2112974375486374, 'reward': 0.48184216022491455, 'reward_std': 0.2112974226474762, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03716103360056877, 'sampling/sampling_logp_difference/max': 3.358452320098877, 'sampling/importance_sampling_ratio/min': 0.034789059311151505, 'sampling/importance_sampling_ratio/mean': 1.0085442066192627, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.264286732301116, 'clip_ratio/low_mean': 0.011303507490083575, 'clip_ratio/low_min': 0.011303507490083575, 'clip_ratio/high_mean': 0.009680354851298034, 'clip_ratio/high_max': 0.009680354851298034, 'clip_ratio/region_mean': 0.02098386234138161, 'reward_total_mean': 0.48184216022491455, 'reward_meter_mean': 0.5966504812240601, 'reward_meter_std': 0.2748834788799286, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8214285373687744, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.48184216022491455, 'reward_total_composite_std': 0.2112974375486374, 'epoch': 0.07}

 52%|█████▏    | 1726/3300 [2:07:35<2:25:34,  5.55s/it]INFO 04-12 00:44:26 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1727/3300 [2:07:40<2:19:09,  5.31s/it]2026-04-12 00:44:31,217 | INFO | train_grpo_train | metrics_logged mode=train step=1727

                                                       
{'loss': -0.0028, 'grad_norm': 1.0005321502685547, 'learning_rate': 4.769696969696971e-06, 'num_tokens': 3890696.0, 'completions/mean_length': 68.5, 'completions/min_length': 67.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9954667687416077, 'rewards/meter/std': 0.004938251804560423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9954667687416077, 'rewards/total_composite/std': 0.004938251804560423, 'reward': 0.9954667687416077, 'reward_std': 0.004938257858157158, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01310752984136343, 'sampling/sampling_logp_difference/max': 0.6610031127929688, 'sampling/importance_sampling_ratio/min': 0.5163331627845764, 'sampling/importance_sampling_ratio/mean': 1.0034555196762085, 'sampling/importance_sampling_ratio/max': 1.4678758382797241, 'entropy': 0.11760625522583723, 'clip_ratio/low_mean': 0.005514706019312143, 'clip_ratio/low_min': 0.005514706019312143, 'clip_ratio/high_mean': 0.011004327097907662, 'clip_ratio/high_max': 0.011004327097907662, 'clip_ratio/region_mean': 0.016519033117219806, 'reward_total_mean': 0.9954667687416077, 'reward_meter_mean': 0.9954667687416077, 'reward_meter_std': 0.004938251804560423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9954667687416077, 'reward_total_composite_std': 0.004938251804560423, 'epoch': 0.07}

 52%|█████▏    | 1727/3300 [2:07:40<2:19:09,  5.31s/it]INFO 04-12 00:44:31 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1728/3300 [2:07:45<2:16:37,  5.21s/it]2026-04-12 00:44:36,252 | INFO | train_grpo_train | metrics_logged mode=train step=1728

                                                       
{'loss': 0.0319, 'grad_norm': 5.639673709869385, 'learning_rate': 4.766666666666667e-06, 'num_tokens': 3892529.0, 'completions/mean_length': 82.125, 'completions/min_length': 72.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.125, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9778019189834595, 'rewards/meter/std': 0.024416539818048477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9778019189834595, 'rewards/total_composite/std': 0.024416539818048477, 'reward': 0.9778019189834595, 'reward_std': 0.024416528642177582, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.055459219962358475, 'sampling/sampling_logp_difference/max': 1.5115740299224854, 'sampling/importance_sampling_ratio/min': 0.31155264377593994, 'sampling/importance_sampling_ratio/mean': 1.0134650468826294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5294598899781704, 'clip_ratio/low_mean': 0.005952381179668009, 'clip_ratio/low_min': 0.005952381179668009, 'clip_ratio/high_mean': 0.041200052946805954, 'clip_ratio/high_max': 0.041200052946805954, 'clip_ratio/region_mean': 0.04715243412647396, 'reward_total_mean': 0.9778019189834595, 'reward_meter_mean': 0.9778019189834595, 'reward_meter_std': 0.024416539818048477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9778019189834595, 'reward_total_composite_std': 0.024416539818048477, 'epoch': 0.07}

 52%|█████▏    | 1728/3300 [2:07:45<2:16:37,  5.21s/it]INFO 04-12 00:44:36 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1729/3300 [2:07:52<2:31:46,  5.80s/it]2026-04-12 00:44:43,412 | INFO | train_grpo_train | metrics_logged mode=train step=1729

                                                       
{'loss': 0.0142, 'grad_norm': 2.5409841537475586, 'learning_rate': 4.763636363636364e-06, 'num_tokens': 3895339.0, 'completions/mean_length': 150.25, 'completions/min_length': 135.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.25, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9943567514419556, 'rewards/meter/std': 0.0023485159035772085, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8415178656578064, 'rewards/repeat_penalty/std': 0.056821081787347794, 'rewards/total_composite/mean': 0.836736798286438, 'rewards/total_composite/std': 0.05605412647128105, 'reward': 0.836736798286438, 'reward_std': 0.05605413764715195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016792375594377518, 'sampling/sampling_logp_difference/max': 1.8594321012496948, 'sampling/importance_sampling_ratio/min': 0.15576106309890747, 'sampling/importance_sampling_ratio/mean': 1.0042328834533691, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09930440969765186, 'clip_ratio/low_mean': 0.0008223684271797538, 'clip_ratio/low_min': 0.0008223684271797538, 'clip_ratio/high_mean': 0.01018475356977433, 'clip_ratio/high_max': 0.01018475356977433, 'clip_ratio/region_mean': 0.011007121996954083, 'reward_total_mean': 0.836736798286438, 'reward_meter_mean': 0.9943567514419556, 'reward_meter_std': 0.0023485159035772085, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8415178656578064, 'reward_repeat_penalty_std': 0.056821081787347794, 'reward_total_composite_mean': 0.836736798286438, 'reward_total_composite_std': 0.05605412647128105, 'epoch': 0.07}

 52%|█████▏    | 1729/3300 [2:07:52<2:31:46,  5.80s/it]INFO 04-12 00:44:43 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1730/3300 [2:07:59<2:42:08,  6.20s/it]2026-04-12 00:44:50,542 | INFO | train_grpo_train | metrics_logged mode=train step=1730

                                                       
{'loss': 0.0161, 'grad_norm': 4.371915340423584, 'learning_rate': 4.760606060606061e-06, 'num_tokens': 3897799.0, 'completions/mean_length': 131.5, 'completions/min_length': 126.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.5, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9530513882637024, 'rewards/meter/std': 0.020072180777788162, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6852272748947144, 'rewards/repeat_penalty/std': 0.10432641208171844, 'rewards/total_composite/mean': 0.6233780384063721, 'rewards/total_composite/std': 0.12500165402889252, 'reward': 0.6233780384063721, 'reward_std': 0.12500165402889252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030718868598341942, 'sampling/sampling_logp_difference/max': 1.7257447242736816, 'sampling/importance_sampling_ratio/min': 0.17804041504859924, 'sampling/importance_sampling_ratio/mean': 1.0002104043960571, 'sampling/importance_sampling_ratio/max': 1.8074078559875488, 'entropy': 0.14938442967832088, 'clip_ratio/low_mean': 0.013116939691826701, 'clip_ratio/low_min': 0.013116939691826701, 'clip_ratio/high_mean': 0.024344916339032352, 'clip_ratio/high_max': 0.024344916339032352, 'clip_ratio/region_mean': 0.03746185603085905, 'reward_total_mean': 0.6233780384063721, 'reward_meter_mean': 0.9530513882637024, 'reward_meter_std': 0.020072180777788162, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6852272748947144, 'reward_repeat_penalty_std': 0.10432641208171844, 'reward_total_composite_mean': 0.6233780384063721, 'reward_total_composite_std': 0.12500165402889252, 'epoch': 0.07}

 52%|█████▏    | 1730/3300 [2:07:59<2:42:08,  6.20s/it]INFO 04-12 00:44:50 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1731/3300 [2:08:05<2:42:40,  6.22s/it]2026-04-12 00:44:56,820 | INFO | train_grpo_train | metrics_logged mode=train step=1731

                                                       
{'loss': -0.0001, 'grad_norm': 5.360625267028809, 'learning_rate': 4.757575757575758e-06, 'num_tokens': 3899601.0, 'completions/mean_length': 68.25, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.996942937374115, 'rewards/meter/std': 0.0017319588223472238, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.996942937374115, 'rewards/total_composite/std': 0.0017319588223472238, 'reward': 0.996942937374115, 'reward_std': 0.0017319645266979933, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017641214653849602, 'sampling/sampling_logp_difference/max': 1.565748691558838, 'sampling/importance_sampling_ratio/min': 0.20893153548240662, 'sampling/importance_sampling_ratio/mean': 1.0010026693344116, 'sampling/importance_sampling_ratio/max': 1.5452033281326294, 'entropy': 0.08362119551748037, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.007273018010891974, 'clip_ratio/high_max': 0.007273018010891974, 'clip_ratio/region_mean': 0.010949488612823188, 'reward_total_mean': 0.996942937374115, 'reward_meter_mean': 0.996942937374115, 'reward_meter_std': 0.0017319588223472238, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.996942937374115, 'reward_total_composite_std': 0.0017319588223472238, 'epoch': 0.07}

 52%|█████▏    | 1731/3300 [2:08:05<2:42:40,  6.22s/it]INFO 04-12 00:44:57 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1732/3300 [2:08:10<2:32:28,  5.83s/it]2026-04-12 00:45:01,705 | INFO | train_grpo_train | metrics_logged mode=train step=1732

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.754545454545455e-06, 'num_tokens': 3901161.0, 'completions/mean_length': 30.0, 'completions/min_length': 30.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.992271900177002, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.992271900177002, 'rewards/total_composite/std': 0.0, 'reward': 0.992271900177002, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0011375559261068702, 'sampling/sampling_logp_difference/max': 0.07951271533966064, 'sampling/importance_sampling_ratio/min': 0.9235662817955017, 'sampling/importance_sampling_ratio/mean': 1.0000195503234863, 'sampling/importance_sampling_ratio/max': 1.0209813117980957, 'entropy': 0.008589746314100921, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.992271900177002, 'reward_meter_mean': 0.992271900177002, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.992271900177002, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 52%|█████▏    | 1732/3300 [2:08:10<2:32:28,  5.83s/it]INFO 04-12 00:45:01 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1733/3300 [2:08:15<2:21:33,  5.42s/it]2026-04-12 00:45:06,160 | INFO | train_grpo_train | metrics_logged mode=train step=1733

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.751515151515152e-06, 'num_tokens': 3902961.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.993520200252533, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.993520200252533, 'rewards/total_composite/std': 0.0, 'reward': 0.993520200252533, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0015307251596823335, 'sampling/sampling_logp_difference/max': 0.11124386638402939, 'sampling/importance_sampling_ratio/min': 0.8947205543518066, 'sampling/importance_sampling_ratio/mean': 1.000504970550537, 'sampling/importance_sampling_ratio/max': 1.0491514205932617, 'entropy': 0.012850302853621542, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.993520200252533, 'reward_meter_mean': 0.993520200252533, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.993520200252533, 'reward_total_composite_std': 0.0, 'epoch': 0.07}

 53%|█████▎    | 1733/3300 [2:08:15<2:21:33,  5.42s/it]INFO 04-12 00:45:06 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1734/3300 [2:08:19<2:13:38,  5.12s/it]2026-04-12 00:45:10,572 | INFO | train_grpo_train | metrics_logged mode=train step=1734

                                                       
{'loss': -0.0114, 'grad_norm': 2.5078203678131104, 'learning_rate': 4.748484848484849e-06, 'num_tokens': 3904691.0, 'completions/mean_length': 62.25, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9855538606643677, 'rewards/meter/std': 0.027445603162050247, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9855538606643677, 'rewards/total_composite/std': 0.027445603162050247, 'reward': 0.9855538606643677, 'reward_std': 0.0274455975741148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021761858835816383, 'sampling/sampling_logp_difference/max': 1.1243009567260742, 'sampling/importance_sampling_ratio/min': 0.3248794972896576, 'sampling/importance_sampling_ratio/mean': 0.9950600266456604, 'sampling/importance_sampling_ratio/max': 1.428898811340332, 'entropy': 0.08833573758602142, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.01626984216272831, 'clip_ratio/high_max': 0.01626984216272831, 'clip_ratio/region_mean': 0.018353175604715943, 'reward_total_mean': 0.9855538606643677, 'reward_meter_mean': 0.9855538606643677, 'reward_meter_std': 0.027445603162050247, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9855538606643677, 'reward_total_composite_std': 0.027445603162050247, 'epoch': 0.07}

 53%|█████▎    | 1734/3300 [2:08:19<2:13:38,  5.12s/it]INFO 04-12 00:45:10 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1735/3300 [2:08:23<2:08:02,  4.91s/it]2026-04-12 00:45:14,992 | INFO | train_grpo_train | metrics_logged mode=train step=1735

                                                       
{'loss': -0.0859, 'grad_norm': 3.9663925170898438, 'learning_rate': 4.745454545454546e-06, 'num_tokens': 3906499.0, 'completions/mean_length': 56.0, 'completions/min_length': 42.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9921239614486694, 'rewards/meter/std': 0.0039491597563028336, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9921239614486694, 'rewards/total_composite/std': 0.0039491597563028336, 'reward': 0.9921239614486694, 'reward_std': 0.0039491597563028336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006989335175603628, 'sampling/sampling_logp_difference/max': 0.8487052917480469, 'sampling/importance_sampling_ratio/min': 0.42796868085861206, 'sampling/importance_sampling_ratio/mean': 1.000370740890503, 'sampling/importance_sampling_ratio/max': 1.7702317237854004, 'entropy': 0.037228189525194466, 'clip_ratio/low_mean': 0.008928571827709675, 'clip_ratio/low_min': 0.008928571827709675, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.008928571827709675, 'reward_total_mean': 0.9921239614486694, 'reward_meter_mean': 0.9921239614486694, 'reward_meter_std': 0.0039491597563028336, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9921239614486694, 'reward_total_composite_std': 0.0039491597563028336, 'epoch': 0.07}

 53%|█████▎    | 1735/3300 [2:08:23<2:08:02,  4.91s/it]INFO 04-12 00:45:15 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1736/3300 [2:08:28<2:05:26,  4.81s/it]2026-04-12 00:45:19,576 | INFO | train_grpo_train | metrics_logged mode=train step=1736

                                                       
{'loss': 0.03, 'grad_norm': 5.453359127044678, 'learning_rate': 4.7424242424242426e-06, 'num_tokens': 3908212.0, 'completions/mean_length': 63.125, 'completions/min_length': 60.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.125, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8737130761146545, 'rewards/meter/std': 0.3438531160354614, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8737130761146545, 'rewards/total_composite/std': 0.3438531160354614, 'reward': 0.8737130761146545, 'reward_std': 0.34385308623313904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01911444216966629, 'sampling/sampling_logp_difference/max': 0.978426456451416, 'sampling/importance_sampling_ratio/min': 0.3759021461009979, 'sampling/importance_sampling_ratio/mean': 1.003002643585205, 'sampling/importance_sampling_ratio/max': 1.8557262420654297, 'entropy': 0.12713485257700086, 'clip_ratio/low_mean': 0.0037313431967049837, 'clip_ratio/low_min': 0.0037313431967049837, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0037313431967049837, 'reward_total_mean': 0.8737130761146545, 'reward_meter_mean': 0.8737130761146545, 'reward_meter_std': 0.3438531160354614, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8737130761146545, 'reward_total_composite_std': 0.3438531160354614, 'epoch': 0.07}

 53%|█████▎    | 1736/3300 [2:08:28<2:05:26,  4.81s/it]INFO 04-12 00:45:19 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1737/3300 [2:08:33<2:09:15,  4.96s/it]2026-04-12 00:45:24,899 | INFO | train_grpo_train | metrics_logged mode=train step=1737

                                                       
{'loss': -0.0033, 'grad_norm': 4.734118461608887, 'learning_rate': 4.73939393939394e-06, 'num_tokens': 3910546.0, 'completions/mean_length': 129.75, 'completions/min_length': 126.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.75, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.8132379651069641, 'rewards/meter/std': 0.2665981948375702, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8392857313156128, 'rewards/repeat_penalty/std': 0.09155284613370895, 'rewards/total_composite/mean': 0.6668680310249329, 'rewards/total_composite/std': 0.21354439854621887, 'reward': 0.6668680310249329, 'reward_std': 0.21354439854621887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02951814979314804, 'sampling/sampling_logp_difference/max': 1.1711840629577637, 'sampling/importance_sampling_ratio/min': 0.3099996745586395, 'sampling/importance_sampling_ratio/mean': 1.0038542747497559, 'sampling/importance_sampling_ratio/max': 1.6680244207382202, 'entropy': 0.18008941877633333, 'clip_ratio/low_mean': 0.008928571944124997, 'clip_ratio/low_min': 0.008928571944124997, 'clip_ratio/high_mean': 0.010484297177754343, 'clip_ratio/high_max': 0.010484297177754343, 'clip_ratio/region_mean': 0.01941286912187934, 'reward_total_mean': 0.6668680310249329, 'reward_meter_mean': 0.8132379651069641, 'reward_meter_std': 0.2665981948375702, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8392857313156128, 'reward_repeat_penalty_std': 0.09155284613370895, 'reward_total_composite_mean': 0.6668680310249329, 'reward_total_composite_std': 0.21354439854621887, 'epoch': 0.07}

 53%|█████▎    | 1737/3300 [2:08:33<2:09:15,  4.96s/it]INFO 04-12 00:45:25 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1738/3300 [2:08:39<2:15:27,  5.20s/it]2026-04-12 00:45:30,655 | INFO | train_grpo_train | metrics_logged mode=train step=1738

                                                       
{'loss': 0.0006, 'grad_norm': 0.2504282295703888, 'learning_rate': 4.736363636363637e-06, 'num_tokens': 3913550.0, 'completions/mean_length': 172.5, 'completions/min_length': 169.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 172.5, 'completions/min_terminated_length': 169.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.994848370552063, 'rewards/meter/std': 0.00046264444245025516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7124999761581421, 'rewards/repeat_penalty/std': 0.0353553481400013, 'rewards/total_composite/mean': 0.7088274955749512, 'rewards/total_composite/std': 0.03512411192059517, 'reward': 0.7088274955749512, 'reward_std': 0.03512410447001457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00417359871789813, 'sampling/sampling_logp_difference/max': 1.570167064666748, 'sampling/importance_sampling_ratio/min': 0.20801043510437012, 'sampling/importance_sampling_ratio/mean': 0.9992320537567139, 'sampling/importance_sampling_ratio/max': 1.0626802444458008, 'entropy': 0.014313822146505117, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0014450866729021072, 'clip_ratio/high_max': 0.0014450866729021072, 'clip_ratio/region_mean': 0.0014450866729021072, 'reward_total_mean': 0.7088274955749512, 'reward_meter_mean': 0.994848370552063, 'reward_meter_std': 0.00046264444245025516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7124999761581421, 'reward_repeat_penalty_std': 0.0353553481400013, 'reward_total_composite_mean': 0.7088274955749512, 'reward_total_composite_std': 0.03512411192059517, 'epoch': 0.07}

 53%|█████▎    | 1738/3300 [2:08:39<2:15:27,  5.20s/it]INFO 04-12 00:45:30 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1739/3300 [2:08:44<2:11:52,  5.07s/it]2026-04-12 00:45:35,411 | INFO | train_grpo_train | metrics_logged mode=train step=1739

                                                       
{'loss': -0.0102, 'grad_norm': 4.318221092224121, 'learning_rate': 4.7333333333333335e-06, 'num_tokens': 3915644.0, 'completions/mean_length': 81.75, 'completions/min_length': 79.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.75, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.9966109991073608, 'rewards/meter/std': 0.0017567714676260948, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9966109991073608, 'rewards/total_composite/std': 0.0017567714676260948, 'reward': 0.9966109991073608, 'reward_std': 0.0017567694885656238, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04741675779223442, 'sampling/sampling_logp_difference/max': 0.8447170257568359, 'sampling/importance_sampling_ratio/min': 0.42967894673347473, 'sampling/importance_sampling_ratio/mean': 1.0067728757858276, 'sampling/importance_sampling_ratio/max': 1.6328699588775635, 'entropy': 0.46590081229805946, 'clip_ratio/low_mean': 0.006289557088166475, 'clip_ratio/low_min': 0.006289557088166475, 'clip_ratio/high_mean': 0.028822204330936074, 'clip_ratio/high_max': 0.028822204330936074, 'clip_ratio/region_mean': 0.03511176141910255, 'reward_total_mean': 0.9966109991073608, 'reward_meter_mean': 0.9966109991073608, 'reward_meter_std': 0.0017567714676260948, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9966109991073608, 'reward_total_composite_std': 0.0017567714676260948, 'epoch': 0.07}

 53%|█████▎    | 1739/3300 [2:08:44<2:11:52,  5.07s/it]INFO 04-12 00:45:35 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1740/3300 [2:08:48<2:05:32,  4.83s/it]2026-04-12 00:45:39,681 | INFO | train_grpo_train | metrics_logged mode=train step=1740

                                                       
{'loss': -0.009, 'grad_norm': 7.030989646911621, 'learning_rate': 4.730303030303031e-06, 'num_tokens': 3917352.0, 'completions/mean_length': 57.5, 'completions/min_length': 55.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8923461437225342, 'rewards/meter/std': 0.2858184278011322, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8923461437225342, 'rewards/total_composite/std': 0.2858184278011322, 'reward': 0.8923461437225342, 'reward_std': 0.2858184278011322, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006146615371108055, 'sampling/sampling_logp_difference/max': 0.3172045946121216, 'sampling/importance_sampling_ratio/min': 0.8774904012680054, 'sampling/importance_sampling_ratio/mean': 1.0057528018951416, 'sampling/importance_sampling_ratio/max': 1.3732835054397583, 'entropy': 0.028293918527197093, 'clip_ratio/low_mean': 0.00909090880304575, 'clip_ratio/low_min': 0.00909090880304575, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.00909090880304575, 'reward_total_mean': 0.8923461437225342, 'reward_meter_mean': 0.8923461437225342, 'reward_meter_std': 0.2858184278011322, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8923461437225342, 'reward_total_composite_std': 0.2858184278011322, 'epoch': 0.07}

 53%|█████▎    | 1740/3300 [2:08:48<2:05:32,  4.83s/it]INFO 04-12 00:45:39 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1741/3300 [2:08:56<2:27:03,  5.66s/it]2026-04-12 00:45:47,281 | INFO | train_grpo_train | metrics_logged mode=train step=1741

                                                       
{'loss': 0.0527, 'grad_norm': 1.3457516431808472, 'learning_rate': 4.727272727272728e-06, 'num_tokens': 3920951.0, 'completions/mean_length': 273.875, 'completions/min_length': 252.0, 'completions/max_length': 290.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 273.875, 'completions/min_terminated_length': 252.0, 'completions/max_terminated_length': 290.0, 'rewards/meter/mean': 0.9979905486106873, 'rewards/meter/std': 0.0006236601620912552, 'rewards/count_adherence/mean': 0.910714328289032, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6429487466812134, 'rewards/repeat_penalty/std': 0.04667491465806961, 'rewards/total_composite/mean': 0.5869752168655396, 'rewards/total_composite/std': 0.08794166892766953, 'reward': 0.5869752168655396, 'reward_std': 0.08794166892766953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01812179759144783, 'sampling/sampling_logp_difference/max': 5.0853352546691895, 'sampling/importance_sampling_ratio/min': 0.006186812650412321, 'sampling/importance_sampling_ratio/mean': 1.0031613111495972, 'sampling/importance_sampling_ratio/max': 1.4954962730407715, 'entropy': 0.11419002618640661, 'clip_ratio/low_mean': 0.0034997850307263434, 'clip_ratio/low_min': 0.0034997850307263434, 'clip_ratio/high_mean': 0.004889143165200949, 'clip_ratio/high_max': 0.004889143165200949, 'clip_ratio/region_mean': 0.008388928195927292, 'reward_total_mean': 0.5869752168655396, 'reward_meter_mean': 0.9979905486106873, 'reward_meter_std': 0.0006236601620912552, 'reward_count_adherence_mean': 0.910714328289032, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6429487466812134, 'reward_repeat_penalty_std': 0.04667491465806961, 'reward_total_composite_mean': 0.5869752168655396, 'reward_total_composite_std': 0.08794166892766953, 'epoch': 0.07}

 53%|█████▎    | 1741/3300 [2:08:56<2:27:03,  5.66s/it]INFO 04-12 00:45:47 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1742/3300 [2:09:01<2:23:03,  5.51s/it]2026-04-12 00:45:52,438 | INFO | train_grpo_train | metrics_logged mode=train step=1742

                                                       
{'loss': 0.0179, 'grad_norm': 2.8153226375579834, 'learning_rate': 4.724242424242424e-06, 'num_tokens': 3923203.0, 'completions/mean_length': 119.5, 'completions/min_length': 114.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.5, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9969746470451355, 'rewards/meter/std': 0.0006134378490969539, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9471026659011841, 'rewards/total_composite/std': 0.09206004440784454, 'reward': 0.9471026659011841, 'reward_std': 0.09206003695726395, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04116864502429962, 'sampling/sampling_logp_difference/max': 1.3741235733032227, 'sampling/importance_sampling_ratio/min': 0.25306129455566406, 'sampling/importance_sampling_ratio/mean': 1.0121161937713623, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.397213451564312, 'clip_ratio/low_mean': 0.0020493179326877, 'clip_ratio/low_min': 0.0020493179326877, 'clip_ratio/high_mean': 0.02945910906419158, 'clip_ratio/high_max': 0.02945910906419158, 'clip_ratio/region_mean': 0.03150842699687928, 'reward_total_mean': 0.9471026659011841, 'reward_meter_mean': 0.9969746470451355, 'reward_meter_std': 0.0006134378490969539, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9471026659011841, 'reward_total_composite_std': 0.09206004440784454, 'epoch': 0.07}

 53%|█████▎    | 1742/3300 [2:09:01<2:23:03,  5.51s/it]INFO 04-12 00:45:52 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1743/3300 [2:09:06<2:19:56,  5.39s/it]2026-04-12 00:45:57,560 | INFO | train_grpo_train | metrics_logged mode=train step=1743

                                                       
{'loss': -0.0255, 'grad_norm': 3.2874183654785156, 'learning_rate': 4.721212121212122e-06, 'num_tokens': 3925534.0, 'completions/mean_length': 115.375, 'completions/min_length': 114.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.375, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9939801692962646, 'rewards/meter/std': 0.0004661441489588469, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.71875, 'rewards/repeat_penalty/std': 0.012626901268959045, 'rewards/total_composite/mean': 0.7144180536270142, 'rewards/total_composite/std': 0.012203346937894821, 'reward': 0.7144180536270142, 'reward_std': 0.012203331105411053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005304521415382624, 'sampling/sampling_logp_difference/max': 2.7986814975738525, 'sampling/importance_sampling_ratio/min': 0.06089029461145401, 'sampling/importance_sampling_ratio/mean': 0.9994580745697021, 'sampling/importance_sampling_ratio/max': 1.4329923391342163, 'entropy': 0.009788335533812642, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0010000000474974513, 'clip_ratio/high_max': 0.0010000000474974513, 'clip_ratio/region_mean': 0.0010000000474974513, 'reward_total_mean': 0.7144180536270142, 'reward_meter_mean': 0.9939801692962646, 'reward_meter_std': 0.0004661441489588469, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.71875, 'reward_repeat_penalty_std': 0.012626901268959045, 'reward_total_composite_mean': 0.7144180536270142, 'reward_total_composite_std': 0.012203346937894821, 'epoch': 0.07}

 53%|█████▎    | 1743/3300 [2:09:06<2:19:56,  5.39s/it]INFO 04-12 00:45:57 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1744/3300 [2:09:11<2:17:49,  5.31s/it]2026-04-12 00:46:02,689 | INFO | train_grpo_train | metrics_logged mode=train step=1744

                                                       
{'loss': 0.0389, 'grad_norm': 3.9866888523101807, 'learning_rate': 4.718181818181818e-06, 'num_tokens': 3927884.0, 'completions/mean_length': 116.75, 'completions/min_length': 111.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.75, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9970704913139343, 'rewards/meter/std': 0.0010195234790444374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970704913139343, 'rewards/total_composite/std': 0.0010195234790444374, 'reward': 0.9970704913139343, 'reward_std': 0.0010194990318268538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05034464970231056, 'sampling/sampling_logp_difference/max': 1.4906072616577148, 'sampling/importance_sampling_ratio/min': 0.22523583471775055, 'sampling/importance_sampling_ratio/mean': 1.0086452960968018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44544921815395355, 'clip_ratio/low_mean': 0.011764100869186223, 'clip_ratio/low_min': 0.011764100869186223, 'clip_ratio/high_mean': 0.024856957141309977, 'clip_ratio/high_max': 0.024856957141309977, 'clip_ratio/region_mean': 0.0366210580104962, 'reward_total_mean': 0.9970704913139343, 'reward_meter_mean': 0.9970704913139343, 'reward_meter_std': 0.0010195234790444374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970704913139343, 'reward_total_composite_std': 0.0010195234790444374, 'epoch': 0.07}

 53%|█████▎    | 1744/3300 [2:09:11<2:17:49,  5.31s/it]INFO 04-12 00:46:02 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1745/3300 [2:09:16<2:11:55,  5.09s/it]2026-04-12 00:46:07,258 | INFO | train_grpo_train | metrics_logged mode=train step=1745

                                                       
{'loss': 0.0183, 'grad_norm': 5.250048637390137, 'learning_rate': 4.715151515151515e-06, 'num_tokens': 3929691.0, 'completions/mean_length': 71.875, 'completions/min_length': 65.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8960655927658081, 'rewards/meter/std': 0.09317556023597717, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8960655927658081, 'rewards/total_composite/std': 0.09317556023597717, 'reward': 0.8960655927658081, 'reward_std': 0.09317556023597717, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04669762775301933, 'sampling/sampling_logp_difference/max': 1.2144498825073242, 'sampling/importance_sampling_ratio/min': 0.2968732714653015, 'sampling/importance_sampling_ratio/mean': 1.0071378946304321, 'sampling/importance_sampling_ratio/max': 1.7080049514770508, 'entropy': 0.4467601887881756, 'clip_ratio/low_mean': 0.01703576883301139, 'clip_ratio/low_min': 0.01703576883301139, 'clip_ratio/high_mean': 0.026507998118177056, 'clip_ratio/high_max': 0.026507998118177056, 'clip_ratio/region_mean': 0.043543766951188445, 'reward_total_mean': 0.8960655927658081, 'reward_meter_mean': 0.8960655927658081, 'reward_meter_std': 0.09317556023597717, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8960655927658081, 'reward_total_composite_std': 0.09317556023597717, 'epoch': 0.07}

 53%|█████▎    | 1745/3300 [2:09:16<2:11:55,  5.09s/it]INFO 04-12 00:46:07 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1746/3300 [2:09:21<2:13:37,  5.16s/it]2026-04-12 00:46:12,601 | INFO | train_grpo_train | metrics_logged mode=train step=1746

                                                       
{'loss': 0.0131, 'grad_norm': 2.7793028354644775, 'learning_rate': 4.7121212121212126e-06, 'num_tokens': 3932034.0, 'completions/mean_length': 118.875, 'completions/min_length': 114.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.875, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9958000183105469, 'rewards/meter/std': 0.0027918845880776644, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9958000183105469, 'rewards/total_composite/std': 0.0027918845880776644, 'reward': 0.9958000183105469, 'reward_std': 0.0027918978594243526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03678244352340698, 'sampling/sampling_logp_difference/max': 1.1251811981201172, 'sampling/importance_sampling_ratio/min': 0.3245936334133148, 'sampling/importance_sampling_ratio/mean': 1.0113009214401245, 'sampling/importance_sampling_ratio/max': 1.7213435173034668, 'entropy': 0.3711636923253536, 'clip_ratio/low_mean': 0.010288415476679802, 'clip_ratio/low_min': 0.010288415476679802, 'clip_ratio/high_mean': 0.01485425140708685, 'clip_ratio/high_max': 0.01485425140708685, 'clip_ratio/region_mean': 0.02514266688376665, 'reward_total_mean': 0.9958000183105469, 'reward_meter_mean': 0.9958000183105469, 'reward_meter_std': 0.0027918845880776644, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9958000183105469, 'reward_total_composite_std': 0.0027918845880776644, 'epoch': 0.07}

 53%|█████▎    | 1746/3300 [2:09:21<2:13:37,  5.16s/it]INFO 04-12 00:46:12 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1747/3300 [2:09:27<2:18:44,  5.36s/it]2026-04-12 00:46:18,420 | INFO | train_grpo_train | metrics_logged mode=train step=1747

                                                       
{'loss': 0.0052, 'grad_norm': 2.528543710708618, 'learning_rate': 4.709090909090909e-06, 'num_tokens': 3934342.0, 'completions/mean_length': 133.5, 'completions/min_length': 130.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.5, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9185582995414734, 'rewards/meter/std': 0.03270983695983887, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7857142686843872, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.7207491397857666, 'rewards/total_composite/std': 0.0627276748418808, 'reward': 0.7207491397857666, 'reward_std': 0.0627276822924614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018684184178709984, 'sampling/sampling_logp_difference/max': 0.9996829032897949, 'sampling/importance_sampling_ratio/min': 0.36799612641334534, 'sampling/importance_sampling_ratio/mean': 1.0059216022491455, 'sampling/importance_sampling_ratio/max': 1.7609641551971436, 'entropy': 0.14490803238004446, 'clip_ratio/low_mean': 0.005481442145537585, 'clip_ratio/low_min': 0.005481442145537585, 'clip_ratio/high_mean': 0.00946290313731879, 'clip_ratio/high_max': 0.00946290313731879, 'clip_ratio/region_mean': 0.014944345282856375, 'reward_total_mean': 0.7207491397857666, 'reward_meter_mean': 0.9185582995414734, 'reward_meter_std': 0.03270983695983887, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7857142686843872, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.7207491397857666, 'reward_total_composite_std': 0.0627276748418808, 'epoch': 0.07}

 53%|█████▎    | 1747/3300 [2:09:27<2:18:44,  5.36s/it]INFO 04-12 00:46:18 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1748/3300 [2:09:32<2:14:55,  5.22s/it]2026-04-12 00:46:23,394 | INFO | train_grpo_train | metrics_logged mode=train step=1748

                                                       
{'loss': 0.0045, 'grad_norm': 5.013108253479004, 'learning_rate': 4.706060606060606e-06, 'num_tokens': 3936075.0, 'completions/mean_length': 70.625, 'completions/min_length': 67.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9558204412460327, 'rewards/meter/std': 0.03730607032775879, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9558204412460327, 'rewards/total_composite/std': 0.03730607032775879, 'reward': 0.9558204412460327, 'reward_std': 0.03730607405304909, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.055451519787311554, 'sampling/sampling_logp_difference/max': 1.048119068145752, 'sampling/importance_sampling_ratio/min': 0.3505965769290924, 'sampling/importance_sampling_ratio/mean': 1.006996989250183, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45604707673192024, 'clip_ratio/low_mean': 0.01255695940926671, 'clip_ratio/low_min': 0.01255695940926671, 'clip_ratio/high_mean': 0.02681165118701756, 'clip_ratio/high_max': 0.02681165118701756, 'clip_ratio/region_mean': 0.03936861059628427, 'reward_total_mean': 0.9558204412460327, 'reward_meter_mean': 0.9558204412460327, 'reward_meter_std': 0.03730607032775879, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9558204412460327, 'reward_total_composite_std': 0.03730607032775879, 'epoch': 0.07}

 53%|█████▎    | 1748/3300 [2:09:32<2:14:55,  5.22s/it]INFO 04-12 00:46:23 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1749/3300 [2:09:37<2:13:24,  5.16s/it]2026-04-12 00:46:28,325 | INFO | train_grpo_train | metrics_logged mode=train step=1749

                                                       
{'loss': 0.0118, 'grad_norm': 6.071586608886719, 'learning_rate': 4.7030303030303035e-06, 'num_tokens': 3937998.0, 'completions/mean_length': 76.375, 'completions/min_length': 73.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9766583442687988, 'rewards/meter/std': 0.02674211747944355, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9766583442687988, 'rewards/total_composite/std': 0.02674211747944355, 'reward': 0.9766583442687988, 'reward_std': 0.02674211747944355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08423186093568802, 'sampling/sampling_logp_difference/max': 2.003994941711426, 'sampling/importance_sampling_ratio/min': 0.13479571044445038, 'sampling/importance_sampling_ratio/mean': 1.0096642971038818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6718878448009491, 'clip_ratio/low_mean': 0.023322351276874542, 'clip_ratio/low_min': 0.023322351276874542, 'clip_ratio/high_mean': 0.048262338852509856, 'clip_ratio/high_max': 0.048262338852509856, 'clip_ratio/region_mean': 0.0715846901293844, 'reward_total_mean': 0.9766583442687988, 'reward_meter_mean': 0.9766583442687988, 'reward_meter_std': 0.02674211747944355, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9766583442687988, 'reward_total_composite_std': 0.02674211747944355, 'epoch': 0.07}

 53%|█████▎    | 1749/3300 [2:09:37<2:13:24,  5.16s/it]INFO 04-12 00:46:28 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1750/3300 [2:09:42<2:10:44,  5.06s/it]2026-04-12 00:46:33,295 | INFO | train_grpo_train | metrics_logged mode=train step=1750

                                                       
{'loss': -0.0126, 'grad_norm': 2.597658157348633, 'learning_rate': 4.7e-06, 'num_tokens': 3939868.0, 'completions/mean_length': 74.75, 'completions/min_length': 72.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.998544454574585, 'rewards/meter/std': 0.00028903057682327926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998544454574585, 'rewards/total_composite/std': 0.00028903057682327926, 'reward': 0.998544454574585, 'reward_std': 0.00028902888880111277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04236089065670967, 'sampling/sampling_logp_difference/max': 1.5543270111083984, 'sampling/importance_sampling_ratio/min': 0.2113315463066101, 'sampling/importance_sampling_ratio/mean': 1.0116918087005615, 'sampling/importance_sampling_ratio/max': 1.9856641292572021, 'entropy': 0.3266411516815424, 'clip_ratio/low_mean': 0.018771880073472857, 'clip_ratio/low_min': 0.018771880073472857, 'clip_ratio/high_mean': 0.021297869854606688, 'clip_ratio/high_max': 0.021297869854606688, 'clip_ratio/region_mean': 0.040069749928079545, 'reward_total_mean': 0.998544454574585, 'reward_meter_mean': 0.998544454574585, 'reward_meter_std': 0.00028903057682327926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998544454574585, 'reward_total_composite_std': 0.00028903057682327926, 'epoch': 0.07}

 53%|█████▎    | 1750/3300 [2:09:42<2:10:44,  5.06s/it]INFO 04-12 00:46:33 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:30,  2.77s/it]

 23%|██▎       | 3/13 [00:08<00:29,  2.97s/it]

 31%|███       | 4/13 [00:13<00:32,  3.64s/it]

 38%|███▊      | 5/13 [00:18<00:33,  4.21s/it]

 46%|████▌     | 6/13 [00:24<00:32,  4.60s/it]

 54%|█████▍    | 7/13 [00:29<00:28,  4.68s/it]

 62%|██████▏   | 8/13 [00:34<00:25,  5.01s/it]

 69%|██████▉   | 9/13 [00:40<00:21,  5.34s/it]

 77%|███████▋  | 10/13 [00:45<00:15,  5.08s/it]

 85%|████████▍ | 11/13 [00:50<00:10,  5.01s/it]

 92%|█████████▏| 12/13 [00:55<00:05,  5.07s/it]

100%|██████████| 13/13 [01:00<00:00,  5.04s/it]2026-04-12 00:47:39,804 | INFO | train_grpo_train | metrics_logged mode=eval step=1750

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 66.4368, 'eval_samples_per_second': 1.565, 'eval_steps_per_second': 0.196, 'eval_num_tokens': 3939868.0, 'eval_completions/mean_length': 198.26923076923077, 'eval_completions/min_length': 63.15384615384615, 'eval_completions/max_length': 347.38461538461536, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 198.26923076923077, 'eval_completions/min_terminated_length': 63.15384615384615, 'eval_completions/max_terminated_length': 347.38461538461536, 'eval_rewards/meter/mean': 0.7541542970217191, 'eval_rewards/meter/std': 0.3719308634216969, 'eval_rewards/count_adherence/mean': 0.9028738003510696, 'eval_rewards/count_adherence/std': 0.12786896412189191, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8075039799396808, 'eval_rewards/repeat_penalty/std': 0.16454027822384468, 'eval_rewards/total_composite/mean': 0.5515764791231889, 'eval_rewards/total_composite/std': 0.33034826585879695, 'eval_reward': 0.5515764791231889, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.018031520339158866, 'eval_sampling/sampling_logp_difference/max': 1.1275318952707143, 'eval_sampling/importance_sampling_ratio/min': 0.33156666732751405, 'eval_sampling/importance_sampling_ratio/mean': 1.004565248122582, 'eval_sampling/importance_sampling_ratio/max': 1.4309009680381188, 'eval_entropy': 0.1877427167044236, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5515764791231889, 'eval_reward_meter_mean': 0.7541542970217191, 'eval_reward_meter_std': 0.3719308634216969, 'eval_reward_count_adherence_mean': 0.9028738003510696, 'eval_reward_count_adherence_std': 0.12786896412189191, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8075039799396808, 'eval_reward_repeat_penalty_std': 0.16454027822384468, 'eval_reward_total_composite_mean': 0.5515764791231889, 'eval_reward_total_composite_std': 0.33034826585879695, 'epoch': 0.07}

 53%|█████▎    | 1750/3300 [2:10:48<2:10:44,  5.06s/it]

100%|██████████| 13/13 [01:00<00:00,  5.04s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 53%|█████▎    | 1751/3300 [2:10:56<11:04:36, 25.74s/it]2026-04-12 00:47:47,173 | INFO | train_grpo_train | metrics_logged mode=train step=1751

                                                        
{'loss': -0.0145, 'grad_norm': 3.8397207260131836, 'learning_rate': 4.696969696969698e-06, 'num_tokens': 3941579.0, 'completions/mean_length': 59.875, 'completions/min_length': 58.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.875, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.994328498840332, 'rewards/meter/std': 0.0016656159423291683, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.994328498840332, 'rewards/total_composite/std': 0.0016656159423291683, 'reward': 0.994328498840332, 'reward_std': 0.001665610121563077, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02080354280769825, 'sampling/sampling_logp_difference/max': 0.939934492111206, 'sampling/importance_sampling_ratio/min': 0.39065343141555786, 'sampling/importance_sampling_ratio/mean': 1.0038416385650635, 'sampling/importance_sampling_ratio/max': 1.5397275686264038, 'entropy': 0.12970400508493185, 'clip_ratio/low_mean': 0.00628531095571816, 'clip_ratio/low_min': 0.00628531095571816, 'clip_ratio/high_mean': 0.006016385043039918, 'clip_ratio/high_max': 0.006016385043039918, 'clip_ratio/region_mean': 0.012301695998758078, 'reward_total_mean': 0.994328498840332, 'reward_meter_mean': 0.994328498840332, 'reward_meter_std': 0.0016656159423291683, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.994328498840332, 'reward_total_composite_std': 0.0016656159423291683, 'epoch': 0.07}

 53%|█████▎    | 1751/3300 [2:10:56<11:04:36, 25.74s/it]INFO 04-12 00:47:47 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1752/3300 [2:11:01<8:25:28, 19.59s/it] 2026-04-12 00:47:52,390 | INFO | train_grpo_train | metrics_logged mode=train step=1752

                                                       
{'loss': 0.0264, 'grad_norm': 9.45784854888916, 'learning_rate': 4.693939393939394e-06, 'num_tokens': 3943517.0, 'completions/mean_length': 78.25, 'completions/min_length': 70.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.25, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9296289086341858, 'rewards/meter/std': 0.09625125676393509, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9296289086341858, 'rewards/total_composite/std': 0.09625125676393509, 'reward': 0.9296289086341858, 'reward_std': 0.09625126421451569, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08540250360965729, 'sampling/sampling_logp_difference/max': 1.4676272869110107, 'sampling/importance_sampling_ratio/min': 0.230471670627594, 'sampling/importance_sampling_ratio/mean': 1.0267534255981445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7139590308070183, 'clip_ratio/low_mean': 0.02355617005378008, 'clip_ratio/low_min': 0.02355617005378008, 'clip_ratio/high_mean': 0.04827680857852101, 'clip_ratio/high_max': 0.04827680857852101, 'clip_ratio/region_mean': 0.07183297863230109, 'reward_total_mean': 0.9296289086341858, 'reward_meter_mean': 0.9296289086341858, 'reward_meter_std': 0.09625125676393509, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9296289086341858, 'reward_total_composite_std': 0.09625125676393509, 'epoch': 0.07}

 53%|█████▎    | 1752/3300 [2:11:01<8:25:28, 19.59s/it]INFO 04-12 00:47:52 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1753/3300 [2:11:05<6:28:43, 15.08s/it]2026-04-12 00:47:56,928 | INFO | train_grpo_train | metrics_logged mode=train step=1753

                                                       
{'loss': -0.0159, 'grad_norm': 3.8942646980285645, 'learning_rate': 4.690909090909092e-06, 'num_tokens': 3945216.0, 'completions/mean_length': 61.375, 'completions/min_length': 59.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9958833456039429, 'rewards/meter/std': 0.0009836297249421477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9958833456039429, 'rewards/total_composite/std': 0.0009836297249421477, 'reward': 0.9958833456039429, 'reward_std': 0.0009836361277848482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01771390438079834, 'sampling/sampling_logp_difference/max': 1.0666460990905762, 'sampling/importance_sampling_ratio/min': 0.34416088461875916, 'sampling/importance_sampling_ratio/mean': 1.0048010349273682, 'sampling/importance_sampling_ratio/max': 1.43393075466156, 'entropy': 0.11703337356448174, 'clip_ratio/low_mean': 0.008269546087831259, 'clip_ratio/low_min': 0.008269546087831259, 'clip_ratio/high_mean': 0.005952381296083331, 'clip_ratio/high_max': 0.005952381296083331, 'clip_ratio/region_mean': 0.01422192738391459, 'reward_total_mean': 0.9958833456039429, 'reward_meter_mean': 0.9958833456039429, 'reward_meter_std': 0.0009836297249421477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9958833456039429, 'reward_total_composite_std': 0.0009836297249421477, 'epoch': 0.07}

 53%|█████▎    | 1753/3300 [2:11:05<6:28:43, 15.08s/it]INFO 04-12 00:47:57 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1754/3300 [2:11:10<5:09:08, 12.00s/it]2026-04-12 00:48:01,741 | INFO | train_grpo_train | metrics_logged mode=train step=1754

                                                       
{'loss': 0.0065, 'grad_norm': 3.0640108585357666, 'learning_rate': 4.687878787878788e-06, 'num_tokens': 3947246.0, 'completions/mean_length': 73.75, 'completions/min_length': 72.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.998239278793335, 'rewards/meter/std': 0.0011535151861608028, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998239278793335, 'rewards/total_composite/std': 0.0011535151861608028, 'reward': 0.998239278793335, 'reward_std': 0.0011535151861608028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.050084829330444336, 'sampling/sampling_logp_difference/max': 1.3049192428588867, 'sampling/importance_sampling_ratio/min': 0.2711944580078125, 'sampling/importance_sampling_ratio/mean': 1.0054799318313599, 'sampling/importance_sampling_ratio/max': 1.8171170949935913, 'entropy': 0.4345889072865248, 'clip_ratio/low_mean': 0.015182648552581668, 'clip_ratio/low_min': 0.015182648552581668, 'clip_ratio/high_mean': 0.030779469525441527, 'clip_ratio/high_max': 0.030779469525441527, 'clip_ratio/region_mean': 0.045962118078023195, 'reward_total_mean': 0.998239278793335, 'reward_meter_mean': 0.998239278793335, 'reward_meter_std': 0.0011535151861608028, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998239278793335, 'reward_total_composite_std': 0.0011535151861608028, 'epoch': 0.07}

 53%|█████▎    | 1754/3300 [2:11:10<5:09:08, 12.00s/it]INFO 04-12 00:48:01 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1755/3300 [2:11:15<4:16:56,  9.98s/it]2026-04-12 00:48:07,013 | INFO | train_grpo_train | metrics_logged mode=train step=1755

                                                       
{'loss': 0.0099, 'grad_norm': 1.097916841506958, 'learning_rate': 4.684848484848485e-06, 'num_tokens': 3949239.0, 'completions/mean_length': 68.125, 'completions/min_length': 67.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9972657561302185, 'rewards/meter/std': 0.0007608237792737782, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972657561302185, 'rewards/total_composite/std': 0.0007608237792737782, 'reward': 0.9972657561302185, 'reward_std': 0.0007608263404108584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009788503870368004, 'sampling/sampling_logp_difference/max': 0.9967951774597168, 'sampling/importance_sampling_ratio/min': 0.7174111604690552, 'sampling/importance_sampling_ratio/mean': 1.0076888799667358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06952884793281555, 'clip_ratio/low_mean': 0.0053571430034935474, 'clip_ratio/low_min': 0.0053571430034935474, 'clip_ratio/high_mean': 0.0036764706019312143, 'clip_ratio/high_max': 0.0036764706019312143, 'clip_ratio/region_mean': 0.009033613605424762, 'reward_total_mean': 0.9972657561302185, 'reward_meter_mean': 0.9972657561302185, 'reward_meter_std': 0.0007608237792737782, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972657561302185, 'reward_total_composite_std': 0.0007608237792737782, 'epoch': 0.07}

 53%|█████▎    | 1755/3300 [2:11:16<4:16:56,  9.98s/it]INFO 04-12 00:48:07 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1756/3300 [2:11:20<3:32:12,  8.25s/it]2026-04-12 00:48:11,214 | INFO | train_grpo_train | metrics_logged mode=train step=1756

                                                       
{'loss': 0.013, 'grad_norm': 5.3303751945495605, 'learning_rate': 4.681818181818183e-06, 'num_tokens': 3950770.0, 'completions/mean_length': 35.375, 'completions/min_length': 35.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9960504770278931, 'rewards/meter/std': 0.0025489823892712593, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9960504770278931, 'rewards/total_composite/std': 0.0025489823892712593, 'reward': 0.9960504770278931, 'reward_std': 0.0025490045081824064, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026823023334145546, 'sampling/sampling_logp_difference/max': 1.4001922607421875, 'sampling/importance_sampling_ratio/min': 0.2465495616197586, 'sampling/importance_sampling_ratio/mean': 1.0033628940582275, 'sampling/importance_sampling_ratio/max': 1.6020153760910034, 'entropy': 0.14098990987986326, 'clip_ratio/low_mean': 0.01736111124046147, 'clip_ratio/low_min': 0.01736111124046147, 'clip_ratio/high_mean': 0.010714285774156451, 'clip_ratio/high_max': 0.010714285774156451, 'clip_ratio/region_mean': 0.02807539701461792, 'reward_total_mean': 0.9960504770278931, 'reward_meter_mean': 0.9960504770278931, 'reward_meter_std': 0.0025489823892712593, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9960504770278931, 'reward_total_composite_std': 0.0025489823892712593, 'epoch': 0.07}

 53%|█████▎    | 1756/3300 [2:11:20<3:32:12,  8.25s/it]INFO 04-12 00:48:11 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1757/3300 [2:11:25<3:11:08,  7.43s/it]2026-04-12 00:48:16,749 | INFO | train_grpo_train | metrics_logged mode=train step=1757

                                                       
{'loss': 0.0099, 'grad_norm': 2.692962884902954, 'learning_rate': 4.678787878787879e-06, 'num_tokens': 3953321.0, 'completions/mean_length': 148.875, 'completions/min_length': 145.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.875, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.9979263544082642, 'rewards/meter/std': 0.0016821667086333036, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.09155284613370895, 'rewards/total_composite/mean': 0.8732737898826599, 'rewards/total_composite/std': 0.09220181405544281, 'reward': 0.8732737898826599, 'reward_std': 0.09220179915428162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04017876833677292, 'sampling/sampling_logp_difference/max': 1.1506986618041992, 'sampling/importance_sampling_ratio/min': 0.3164156377315521, 'sampling/importance_sampling_ratio/mean': 1.0055673122406006, 'sampling/importance_sampling_ratio/max': 1.883940577507019, 'entropy': 0.3133071381598711, 'clip_ratio/low_mean': 0.020188873866572976, 'clip_ratio/low_min': 0.020188873866572976, 'clip_ratio/high_mean': 0.010034454986453056, 'clip_ratio/high_max': 0.010034454986453056, 'clip_ratio/region_mean': 0.030223328853026032, 'reward_total_mean': 0.8732737898826599, 'reward_meter_mean': 0.9979263544082642, 'reward_meter_std': 0.0016821667086333036, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.09155284613370895, 'reward_total_composite_mean': 0.8732737898826599, 'reward_total_composite_std': 0.09220181405544281, 'epoch': 0.07}

 53%|█████▎    | 1757/3300 [2:11:25<3:11:08,  7.43s/it]INFO 04-12 00:48:16 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1758/3300 [2:11:30<2:50:21,  6.63s/it]2026-04-12 00:48:21,501 | INFO | train_grpo_train | metrics_logged mode=train step=1758

                                                       
{'loss': 0.0195, 'grad_norm': 6.611409664154053, 'learning_rate': 4.675757575757576e-06, 'num_tokens': 3955117.0, 'completions/mean_length': 76.5, 'completions/min_length': 75.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9615824222564697, 'rewards/meter/std': 0.03308320790529251, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9615824222564697, 'rewards/total_composite/std': 0.03308320790529251, 'reward': 0.9615824222564697, 'reward_std': 0.03308321535587311, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06566065549850464, 'sampling/sampling_logp_difference/max': 1.2004508972167969, 'sampling/importance_sampling_ratio/min': 0.30105844140052795, 'sampling/importance_sampling_ratio/mean': 1.0291359424591064, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7143527865409851, 'clip_ratio/low_mean': 0.02743737120181322, 'clip_ratio/low_min': 0.02743737120181322, 'clip_ratio/high_mean': 0.03448549238964915, 'clip_ratio/high_max': 0.03448549238964915, 'clip_ratio/region_mean': 0.061922863591462374, 'reward_total_mean': 0.9615824222564697, 'reward_meter_mean': 0.9615824222564697, 'reward_meter_std': 0.03308320790529251, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9615824222564697, 'reward_total_composite_std': 0.03308320790529251, 'epoch': 0.07}

 53%|█████▎    | 1758/3300 [2:11:30<2:50:21,  6.63s/it]INFO 04-12 00:48:21 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1759/3300 [2:11:34<2:31:46,  5.91s/it]2026-04-12 00:48:25,733 | INFO | train_grpo_train | metrics_logged mode=train step=1759

                                                       
{'loss': 0.0438, 'grad_norm': 5.4466166496276855, 'learning_rate': 4.6727272727272735e-06, 'num_tokens': 3956703.0, 'completions/mean_length': 36.25, 'completions/min_length': 35.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9933856725692749, 'rewards/meter/std': 0.007373345550149679, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9933856725692749, 'rewards/total_composite/std': 0.007373345550149679, 'reward': 0.9933856725692749, 'reward_std': 0.007373336236923933, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021740470081567764, 'sampling/sampling_logp_difference/max': 0.7743911743164062, 'sampling/importance_sampling_ratio/min': 0.46098437905311584, 'sampling/importance_sampling_ratio/mean': 1.0038979053497314, 'sampling/importance_sampling_ratio/max': 1.7787117958068848, 'entropy': 0.12411041604354978, 'clip_ratio/low_mean': 0.016025641234591603, 'clip_ratio/low_min': 0.016025641234591603, 'clip_ratio/high_mean': 0.0034722222480922937, 'clip_ratio/high_max': 0.0034722222480922937, 'clip_ratio/region_mean': 0.019497863482683897, 'reward_total_mean': 0.9933856725692749, 'reward_meter_mean': 0.9933856725692749, 'reward_meter_std': 0.007373345550149679, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9933856725692749, 'reward_total_composite_std': 0.007373345550149679, 'epoch': 0.07}

 53%|█████▎    | 1759/3300 [2:11:34<2:31:46,  5.91s/it]INFO 04-12 00:48:25 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1760/3300 [2:11:40<2:28:39,  5.79s/it]2026-04-12 00:48:31,251 | INFO | train_grpo_train | metrics_logged mode=train step=1760

                                                       
{'loss': -0.0871, 'grad_norm': 4.831878662109375, 'learning_rate': 4.66969696969697e-06, 'num_tokens': 3959470.0, 'completions/mean_length': 141.875, 'completions/min_length': 125.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.875, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.992239236831665, 'rewards/meter/std': 0.0005570180364884436, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.609375, 'rewards/repeat_penalty/std': 0.012938717380166054, 'rewards/total_composite/mean': 0.558111310005188, 'rewards/total_composite/std': 0.05107751861214638, 'reward': 0.558111310005188, 'reward_std': 0.051077522337436676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008833160623908043, 'sampling/sampling_logp_difference/max': 2.1678309440612793, 'sampling/importance_sampling_ratio/min': 0.11442553997039795, 'sampling/importance_sampling_ratio/mean': 0.9995898008346558, 'sampling/importance_sampling_ratio/max': 1.4811135530471802, 'entropy': 0.02644984540529549, 'clip_ratio/low_mean': 0.0020000000949949026, 'clip_ratio/low_min': 0.0020000000949949026, 'clip_ratio/high_mean': 0.003289473708719015, 'clip_ratio/high_max': 0.003289473708719015, 'clip_ratio/region_mean': 0.005289473803713918, 'reward_total_mean': 0.558111310005188, 'reward_meter_mean': 0.992239236831665, 'reward_meter_std': 0.0005570180364884436, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.609375, 'reward_repeat_penalty_std': 0.012938717380166054, 'reward_total_composite_mean': 0.558111310005188, 'reward_total_composite_std': 0.05107751861214638, 'epoch': 0.07}

 53%|█████▎    | 1760/3300 [2:11:40<2:28:39,  5.79s/it]INFO 04-12 00:48:31 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1761/3300 [2:11:44<2:17:48,  5.37s/it]2026-04-12 00:48:35,642 | INFO | train_grpo_train | metrics_logged mode=train step=1761

                                                       
{'loss': -0.0004, 'grad_norm': 0.03243406489491463, 'learning_rate': 4.666666666666667e-06, 'num_tokens': 3961206.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9981468319892883, 'rewards/meter/std': 7.291422207345022e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981468319892883, 'rewards/total_composite/std': 7.291422207345022e-06, 'reward': 0.9981468319892883, 'reward_std': 7.28539862393518e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0019470350816845894, 'sampling/sampling_logp_difference/max': 0.4015388488769531, 'sampling/importance_sampling_ratio/min': 0.8816590905189514, 'sampling/importance_sampling_ratio/mean': 1.0011060237884521, 'sampling/importance_sampling_ratio/max': 1.4941221475601196, 'entropy': 0.01341856678482145, 'clip_ratio/low_mean': 0.001953125, 'clip_ratio/low_min': 0.001953125, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.9981468319892883, 'reward_meter_mean': 0.9981468319892883, 'reward_meter_std': 7.291422207345022e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981468319892883, 'reward_total_composite_std': 7.291422207345022e-06, 'epoch': 0.07}

 53%|█████▎    | 1761/3300 [2:11:44<2:17:48,  5.37s/it]INFO 04-12 00:48:35 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1762/3300 [2:11:49<2:13:23,  5.20s/it]2026-04-12 00:48:40,453 | INFO | train_grpo_train | metrics_logged mode=train step=1762

                                                       
{'loss': 0.0019, 'grad_norm': 0.9674805998802185, 'learning_rate': 4.663636363636364e-06, 'num_tokens': 3963328.0, 'completions/mean_length': 95.25, 'completions/min_length': 95.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.25, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9970584511756897, 'rewards/meter/std': 0.00035686453338712454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8475098013877869, 'rewards/total_composite/std': 0.09242907166481018, 'reward': 0.8475098013877869, 'reward_std': 0.09242908656597137, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00739458529278636, 'sampling/sampling_logp_difference/max': 0.7102556228637695, 'sampling/importance_sampling_ratio/min': 0.4915185272693634, 'sampling/importance_sampling_ratio/mean': 1.0031148195266724, 'sampling/importance_sampling_ratio/max': 1.3529959917068481, 'entropy': 0.04998829588294029, 'clip_ratio/low_mean': 0.0026178729021921754, 'clip_ratio/low_min': 0.0026178729021921754, 'clip_ratio/high_mean': 0.0013157895300537348, 'clip_ratio/high_max': 0.0013157895300537348, 'clip_ratio/region_mean': 0.00393366243224591, 'reward_total_mean': 0.8475098013877869, 'reward_meter_mean': 0.9970584511756897, 'reward_meter_std': 0.00035686453338712454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8475098013877869, 'reward_total_composite_std': 0.09242907166481018, 'epoch': 0.07}

 53%|█████▎    | 1762/3300 [2:11:49<2:13:23,  5.20s/it]INFO 04-12 00:48:40 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1763/3300 [2:11:57<2:34:18,  6.02s/it]2026-04-12 00:48:48,391 | INFO | train_grpo_train | metrics_logged mode=train step=1763

                                                       
{'loss': -0.0018, 'grad_norm': 1.530092477798462, 'learning_rate': 4.660606060606061e-06, 'num_tokens': 3967632.0, 'completions/mean_length': 331.0, 'completions/min_length': 316.0, 'completions/max_length': 347.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 331.0, 'completions/min_terminated_length': 316.0, 'completions/max_terminated_length': 347.0, 'rewards/meter/mean': 0.9972841739654541, 'rewards/meter/std': 0.0010668373433873057, 'rewards/count_adherence/mean': 0.692307710647583, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8308823108673096, 'rewards/repeat_penalty/std': 0.06623479723930359, 'rewards/total_composite/mean': 0.5736789703369141, 'rewards/total_composite/std': 0.04592465981841087, 'reward': 0.5736789703369141, 'reward_std': 0.04592467471957207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030713124200701714, 'sampling/sampling_logp_difference/max': 1.0928246974945068, 'sampling/importance_sampling_ratio/min': 0.33526813983917236, 'sampling/importance_sampling_ratio/mean': 1.0089269876480103, 'sampling/importance_sampling_ratio/max': 1.8286789655685425, 'entropy': 0.308476896956563, 'clip_ratio/low_mean': 0.013588423724286258, 'clip_ratio/low_min': 0.013588423724286258, 'clip_ratio/high_mean': 0.011261471780017018, 'clip_ratio/high_max': 0.011261471780017018, 'clip_ratio/region_mean': 0.024849895504303277, 'reward_total_mean': 0.5736789703369141, 'reward_meter_mean': 0.9972841739654541, 'reward_meter_std': 0.0010668373433873057, 'reward_count_adherence_mean': 0.692307710647583, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8308823108673096, 'reward_repeat_penalty_std': 0.06623479723930359, 'reward_total_composite_mean': 0.5736789703369141, 'reward_total_composite_std': 0.04592465981841087, 'epoch': 0.07}

 53%|█████▎    | 1763/3300 [2:11:57<2:34:18,  6.02s/it]INFO 04-12 00:48:48 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1764/3300 [2:12:01<2:20:05,  5.47s/it]2026-04-12 00:48:52,575 | INFO | train_grpo_train | metrics_logged mode=train step=1764

                                                       
{'loss': -0.0047, 'grad_norm': 5.83869743347168, 'learning_rate': 4.657575757575758e-06, 'num_tokens': 3969088.0, 'completions/mean_length': 35.0, 'completions/min_length': 34.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9969139099121094, 'rewards/meter/std': 0.002184208482503891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969139099121094, 'rewards/total_composite/std': 0.002184208482503891, 'reward': 0.9969139099121094, 'reward_std': 0.002184197073802352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01693633943796158, 'sampling/sampling_logp_difference/max': 0.8935699462890625, 'sampling/importance_sampling_ratio/min': 0.40919235348701477, 'sampling/importance_sampling_ratio/mean': 0.9999939203262329, 'sampling/importance_sampling_ratio/max': 1.648281455039978, 'entropy': 0.09531538654118776, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9969139099121094, 'reward_meter_mean': 0.9969139099121094, 'reward_meter_std': 0.002184208482503891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969139099121094, 'reward_total_composite_std': 0.002184208482503891, 'epoch': 0.07}

 53%|█████▎    | 1764/3300 [2:12:01<2:20:05,  5.47s/it]INFO 04-12 00:48:52 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1765/3300 [2:12:10<2:45:57,  6.49s/it]2026-04-12 00:49:01,433 | INFO | train_grpo_train | metrics_logged mode=train step=1765

                                                       
{'loss': -0.0026, 'grad_norm': 1.4083707332611084, 'learning_rate': 4.654545454545455e-06, 'num_tokens': 3973789.0, 'completions/mean_length': 400.625, 'completions/min_length': 381.0, 'completions/max_length': 426.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 400.625, 'completions/min_terminated_length': 381.0, 'completions/max_terminated_length': 426.0, 'rewards/meter/mean': 0.9959809184074402, 'rewards/meter/std': 0.002331522526219487, 'rewards/count_adherence/mean': 0.6176470518112183, 'rewards/count_adherence/std': 0.03144249692559242, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8089442253112793, 'rewards/repeat_penalty/std': 0.07961878180503845, 'rewards/total_composite/mean': 0.4966247081756592, 'rewards/total_composite/std': 0.04498037323355675, 'reward': 0.4966247081756592, 'reward_std': 0.04498036950826645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030164016410708427, 'sampling/sampling_logp_difference/max': 1.2716550827026367, 'sampling/importance_sampling_ratio/min': 0.2803671956062317, 'sampling/importance_sampling_ratio/mean': 1.0090073347091675, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30641947500407696, 'clip_ratio/low_mean': 0.013535281177610159, 'clip_ratio/low_min': 0.013535281177610159, 'clip_ratio/high_mean': 0.00853747595101595, 'clip_ratio/high_max': 0.00853747595101595, 'clip_ratio/region_mean': 0.022072757128626108, 'reward_total_mean': 0.4966247081756592, 'reward_meter_mean': 0.9959809184074402, 'reward_meter_std': 0.002331522526219487, 'reward_count_adherence_mean': 0.6176470518112183, 'reward_count_adherence_std': 0.03144249692559242, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8089442253112793, 'reward_repeat_penalty_std': 0.07961878180503845, 'reward_total_composite_mean': 0.4966247081756592, 'reward_total_composite_std': 0.04498037323355675, 'epoch': 0.07}

 53%|█████▎    | 1765/3300 [2:12:10<2:45:57,  6.49s/it]INFO 04-12 00:49:01 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1766/3300 [2:12:15<2:35:19,  6.08s/it]2026-04-12 00:49:06,553 | INFO | train_grpo_train | metrics_logged mode=train step=1766

                                                       
{'loss': 0.0029, 'grad_norm': 3.6053273677825928, 'learning_rate': 4.651515151515152e-06, 'num_tokens': 3976025.0, 'completions/mean_length': 110.5, 'completions/min_length': 107.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.5, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.9980085492134094, 'rewards/meter/std': 0.0012038704007863998, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9480756521224976, 'rewards/total_composite/std': 0.09206356108188629, 'reward': 0.9480756521224976, 'reward_std': 0.09206356853246689, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04403796046972275, 'sampling/sampling_logp_difference/max': 1.918081283569336, 'sampling/importance_sampling_ratio/min': 0.14688852429389954, 'sampling/importance_sampling_ratio/mean': 1.0053863525390625, 'sampling/importance_sampling_ratio/max': 1.5807472467422485, 'entropy': 0.38782429322600365, 'clip_ratio/low_mean': 0.00802111136727035, 'clip_ratio/low_min': 0.00802111136727035, 'clip_ratio/high_mean': 0.03843654436059296, 'clip_ratio/high_max': 0.03843654436059296, 'clip_ratio/region_mean': 0.04645765572786331, 'reward_total_mean': 0.9480756521224976, 'reward_meter_mean': 0.9980085492134094, 'reward_meter_std': 0.0012038704007863998, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9480756521224976, 'reward_total_composite_std': 0.09206356108188629, 'epoch': 0.07}

 54%|█████▎    | 1766/3300 [2:12:15<2:35:19,  6.08s/it]INFO 04-12 00:49:06 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1767/3300 [2:12:19<2:22:50,  5.59s/it]2026-04-12 00:49:11,014 | INFO | train_grpo_train | metrics_logged mode=train step=1767

                                                       
{'loss': 0.0007, 'grad_norm': 0.5792075991630554, 'learning_rate': 4.648484848484849e-06, 'num_tokens': 3977817.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9981694221496582, 'rewards/meter/std': 9.33830306166783e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981694221496582, 'rewards/total_composite/std': 9.33830306166783e-05, 'reward': 0.9981694221496582, 'reward_std': 9.336639777757227e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006093635223805904, 'sampling/sampling_logp_difference/max': 0.6290621757507324, 'sampling/importance_sampling_ratio/min': 0.5330914855003357, 'sampling/importance_sampling_ratio/mean': 0.9992143511772156, 'sampling/importance_sampling_ratio/max': 1.7626099586486816, 'entropy': 0.018734007026068866, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.9981694221496582, 'reward_meter_mean': 0.9981694221496582, 'reward_meter_std': 9.33830306166783e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981694221496582, 'reward_total_composite_std': 9.33830306166783e-05, 'epoch': 0.07}

 54%|█████▎    | 1767/3300 [2:12:20<2:22:50,  5.59s/it]INFO 04-12 00:49:11 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1768/3300 [2:12:24<2:13:54,  5.24s/it]2026-04-12 00:49:15,442 | INFO | train_grpo_train | metrics_logged mode=train step=1768

                                                       
{'loss': 0.0077, 'grad_norm': 4.614588737487793, 'learning_rate': 4.645454545454545e-06, 'num_tokens': 3979649.0, 'completions/mean_length': 68.0, 'completions/min_length': 68.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9972659349441528, 'rewards/meter/std': 0.0006626376416534185, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972659349441528, 'rewards/total_composite/std': 0.0006626376416534185, 'reward': 0.9972659349441528, 'reward_std': 0.0006626351969316602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020455962046980858, 'sampling/sampling_logp_difference/max': 0.9924750328063965, 'sampling/importance_sampling_ratio/min': 0.37065815925598145, 'sampling/importance_sampling_ratio/mean': 1.0017422437667847, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11503921588882804, 'clip_ratio/low_mean': 0.005514706019312143, 'clip_ratio/low_min': 0.005514706019312143, 'clip_ratio/high_mean': 0.009191176504828036, 'clip_ratio/high_max': 0.009191176504828036, 'clip_ratio/region_mean': 0.01470588252414018, 'reward_total_mean': 0.9972659349441528, 'reward_meter_mean': 0.9972659349441528, 'reward_meter_std': 0.0006626376416534185, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972659349441528, 'reward_total_composite_std': 0.0006626376416534185, 'epoch': 0.07}

 54%|█████▎    | 1768/3300 [2:12:24<2:13:54,  5.24s/it]INFO 04-12 00:49:15 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1769/3300 [2:12:28<2:07:03,  4.98s/it]2026-04-12 00:49:19,803 | INFO | train_grpo_train | metrics_logged mode=train step=1769

                                                       
{'loss': 0.0015, 'grad_norm': 1.5824861526489258, 'learning_rate': 4.642424242424243e-06, 'num_tokens': 3981601.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9967877864837646, 'rewards/meter/std': 0.00022891137632541358, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9967877864837646, 'rewards/total_composite/std': 0.00022891137632541358, 'reward': 0.9967877864837646, 'reward_std': 0.00022891460685059428, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00849367305636406, 'sampling/sampling_logp_difference/max': 0.3904110789299011, 'sampling/importance_sampling_ratio/min': 0.6767785549163818, 'sampling/importance_sampling_ratio/mean': 1.003195881843567, 'sampling/importance_sampling_ratio/max': 1.2106847763061523, 'entropy': 0.060156718362122774, 'clip_ratio/low_mean': 0.003968254197388887, 'clip_ratio/low_min': 0.003968254197388887, 'clip_ratio/high_mean': 0.005952381296083331, 'clip_ratio/high_max': 0.005952381296083331, 'clip_ratio/region_mean': 0.009920635493472219, 'reward_total_mean': 0.9967877864837646, 'reward_meter_mean': 0.9967877864837646, 'reward_meter_std': 0.00022891137632541358, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9967877864837646, 'reward_total_composite_std': 0.00022891137632541358, 'epoch': 0.07}

 54%|█████▎    | 1769/3300 [2:12:28<2:07:03,  4.98s/it]INFO 04-12 00:49:20 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1770/3300 [2:12:33<2:07:34,  5.00s/it]2026-04-12 00:49:24,857 | INFO | train_grpo_train | metrics_logged mode=train step=1770

                                                       
{'loss': 0.1371, 'grad_norm': 7.558566093444824, 'learning_rate': 4.63939393939394e-06, 'num_tokens': 3983404.0, 'completions/mean_length': 81.375, 'completions/min_length': 73.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.375, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9489094018936157, 'rewards/meter/std': 0.09742892533540726, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.886616051197052, 'rewards/total_composite/std': 0.18366359174251556, 'reward': 0.886616051197052, 'reward_std': 0.18366359174251556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06127572059631348, 'sampling/sampling_logp_difference/max': 1.5139522552490234, 'sampling/importance_sampling_ratio/min': 0.22003860771656036, 'sampling/importance_sampling_ratio/mean': 1.0119951963424683, 'sampling/importance_sampling_ratio/max': 1.8940720558166504, 'entropy': 0.5354950428009033, 'clip_ratio/low_mean': 0.012420599116012454, 'clip_ratio/low_min': 0.012420599116012454, 'clip_ratio/high_mean': 0.029366800910793245, 'clip_ratio/high_max': 0.029366800910793245, 'clip_ratio/region_mean': 0.0417874000268057, 'reward_total_mean': 0.886616051197052, 'reward_meter_mean': 0.9489094018936157, 'reward_meter_std': 0.09742892533540726, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.886616051197052, 'reward_total_composite_std': 0.18366359174251556, 'epoch': 0.07}

 54%|█████▎    | 1770/3300 [2:12:33<2:07:34,  5.00s/it]INFO 04-12 00:49:25 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1771/3300 [2:12:38<2:04:51,  4.90s/it]2026-04-12 00:49:29,520 | INFO | train_grpo_train | metrics_logged mode=train step=1771

                                                       
{'loss': 0.006, 'grad_norm': 2.8212172985076904, 'learning_rate': 4.636363636363636e-06, 'num_tokens': 3985659.0, 'completions/mean_length': 94.875, 'completions/min_length': 94.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.875, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9980397820472717, 'rewards/meter/std': 0.0001069462377927266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8732877969741821, 'rewards/total_composite/std': 0.10333773493766785, 'reward': 0.8732877969741821, 'reward_std': 0.10333773493766785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009597206488251686, 'sampling/sampling_logp_difference/max': 0.9711115956306458, 'sampling/importance_sampling_ratio/min': 0.3786618709564209, 'sampling/importance_sampling_ratio/mean': 0.9989795684814453, 'sampling/importance_sampling_ratio/max': 1.4593793153762817, 'entropy': 0.03504605917260051, 'clip_ratio/low_mean': 0.003947368590161204, 'clip_ratio/low_min': 0.003947368590161204, 'clip_ratio/high_mean': 0.0066069429740309715, 'clip_ratio/high_max': 0.0066069429740309715, 'clip_ratio/region_mean': 0.010554311564192176, 'reward_total_mean': 0.8732877969741821, 'reward_meter_mean': 0.9980397820472717, 'reward_meter_std': 0.0001069462377927266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8732877969741821, 'reward_total_composite_std': 0.10333773493766785, 'epoch': 0.07}

 54%|█████▎    | 1771/3300 [2:12:38<2:04:51,  4.90s/it]INFO 04-12 00:49:29 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1772/3300 [2:12:42<2:00:23,  4.73s/it]2026-04-12 00:49:33,851 | INFO | train_grpo_train | metrics_logged mode=train step=1772

                                                       
{'loss': -0.0013, 'grad_norm': 9.031041145324707, 'learning_rate': 4.633333333333334e-06, 'num_tokens': 3987488.0, 'completions/mean_length': 62.625, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9955183267593384, 'rewards/meter/std': 0.0038252200465649366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955183267593384, 'rewards/total_composite/std': 0.0038252200465649366, 'reward': 0.9955183267593384, 'reward_std': 0.0038252256345003843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02198421023786068, 'sampling/sampling_logp_difference/max': 1.470834732055664, 'sampling/importance_sampling_ratio/min': 0.2297336310148239, 'sampling/importance_sampling_ratio/mean': 0.9984506964683533, 'sampling/importance_sampling_ratio/max': 1.5951910018920898, 'entropy': 0.10218251217156649, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.007936508394777775, 'clip_ratio/high_max': 0.007936508394777775, 'clip_ratio/region_mean': 0.010019841836765409, 'reward_total_mean': 0.9955183267593384, 'reward_meter_mean': 0.9955183267593384, 'reward_meter_std': 0.0038252200465649366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955183267593384, 'reward_total_composite_std': 0.0038252200465649366, 'epoch': 0.07}

 54%|█████▎    | 1772/3300 [2:12:42<2:00:23,  4.73s/it]INFO 04-12 00:49:34 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1773/3300 [2:12:49<2:12:52,  5.22s/it]2026-04-12 00:49:40,220 | INFO | train_grpo_train | metrics_logged mode=train step=1773

                                                       
{'loss': 0.0489, 'grad_norm': 2.2303996086120605, 'learning_rate': 4.630303030303031e-06, 'num_tokens': 3990562.0, 'completions/mean_length': 200.25, 'completions/min_length': 191.0, 'completions/max_length': 223.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 200.25, 'completions/min_terminated_length': 191.0, 'completions/max_terminated_length': 223.0, 'rewards/meter/mean': 0.9357019662857056, 'rewards/meter/std': 0.17194922268390656, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7526223659515381, 'rewards/repeat_penalty/std': 0.07428380101919174, 'rewards/total_composite/mean': 0.6658661961555481, 'rewards/total_composite/std': 0.11478758603334427, 'reward': 0.6658661961555481, 'reward_std': 0.11478757858276367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01970170997083187, 'sampling/sampling_logp_difference/max': 2.9212241172790527, 'sampling/importance_sampling_ratio/min': 0.053867705166339874, 'sampling/importance_sampling_ratio/mean': 1.0023951530456543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11565544595941901, 'clip_ratio/low_mean': 0.006157553056254983, 'clip_ratio/low_min': 0.006157553056254983, 'clip_ratio/high_mean': 0.005890052416361868, 'clip_ratio/high_max': 0.005890052416361868, 'clip_ratio/region_mean': 0.012047605472616851, 'reward_total_mean': 0.6658661961555481, 'reward_meter_mean': 0.9357019662857056, 'reward_meter_std': 0.17194922268390656, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7526223659515381, 'reward_repeat_penalty_std': 0.07428380101919174, 'reward_total_composite_mean': 0.6658661961555481, 'reward_total_composite_std': 0.11478758603334427, 'epoch': 0.07}

 54%|█████▎    | 1773/3300 [2:12:49<2:12:52,  5.22s/it]INFO 04-12 00:49:40 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1774/3300 [2:12:54<2:12:48,  5.22s/it]2026-04-12 00:49:45,441 | INFO | train_grpo_train | metrics_logged mode=train step=1774

                                                       
{'loss': 0.0009, 'grad_norm': 2.5007715225219727, 'learning_rate': 4.627272727272727e-06, 'num_tokens': 3992891.0, 'completions/mean_length': 127.125, 'completions/min_length': 127.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.125, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.997309684753418, 'rewards/meter/std': 0.00012155534204794094, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8392857313156128, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8370258808135986, 'rewards/total_composite/std': 0.05034034699201584, 'reward': 0.8370258808135986, 'reward_std': 0.05034033954143524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00933446828275919, 'sampling/sampling_logp_difference/max': 1.754063606262207, 'sampling/importance_sampling_ratio/min': 0.17306923866271973, 'sampling/importance_sampling_ratio/mean': 1.0009863376617432, 'sampling/importance_sampling_ratio/max': 1.2224539518356323, 'entropy': 0.05814863136038184, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0019685039296746254, 'clip_ratio/high_max': 0.0019685039296746254, 'clip_ratio/region_mean': 0.0019685039296746254, 'reward_total_mean': 0.8370258808135986, 'reward_meter_mean': 0.997309684753418, 'reward_meter_std': 0.00012155534204794094, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8392857313156128, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8370258808135986, 'reward_total_composite_std': 0.05034034699201584, 'epoch': 0.07}

 54%|█████▍    | 1774/3300 [2:12:54<2:12:48,  5.22s/it]INFO 04-12 00:49:45 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1775/3300 [2:12:59<2:08:18,  5.05s/it]2026-04-12 00:49:50,087 | INFO | train_grpo_train | metrics_logged mode=train step=1775

                                                       
{'loss': -0.0231, 'grad_norm': 6.361879348754883, 'learning_rate': 4.6242424242424245e-06, 'num_tokens': 3994961.0, 'completions/mean_length': 75.75, 'completions/min_length': 71.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9477696418762207, 'rewards/meter/std': 0.10307468473911285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9477696418762207, 'rewards/total_composite/std': 0.10307468473911285, 'reward': 0.9477696418762207, 'reward_std': 0.10307466983795166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07908248156309128, 'sampling/sampling_logp_difference/max': 3.3076226711273193, 'sampling/importance_sampling_ratio/min': 0.036603085696697235, 'sampling/importance_sampling_ratio/mean': 1.0081671476364136, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6292642988264561, 'clip_ratio/low_mean': 0.007042253389954567, 'clip_ratio/low_min': 0.007042253389954567, 'clip_ratio/high_mean': 0.040617201710119843, 'clip_ratio/high_max': 0.040617201710119843, 'clip_ratio/region_mean': 0.04765945510007441, 'reward_total_mean': 0.9477696418762207, 'reward_meter_mean': 0.9477696418762207, 'reward_meter_std': 0.10307468473911285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9477696418762207, 'reward_total_composite_std': 0.10307468473911285, 'epoch': 0.07}

 54%|█████▍    | 1775/3300 [2:12:59<2:08:18,  5.05s/it]INFO 04-12 00:49:50 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1776/3300 [2:13:03<2:04:21,  4.90s/it]2026-04-12 00:49:54,628 | INFO | train_grpo_train | metrics_logged mode=train step=1776

                                                       
{'loss': -0.0062, 'grad_norm': 9.041850090026855, 'learning_rate': 4.621212121212122e-06, 'num_tokens': 3996754.0, 'completions/mean_length': 56.125, 'completions/min_length': 56.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8854995965957642, 'rewards/meter/std': 0.043195273727178574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8854995965957642, 'rewards/total_composite/std': 0.043195273727178574, 'reward': 0.8854995965957642, 'reward_std': 0.04319525882601738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007866953499615192, 'sampling/sampling_logp_difference/max': 1.9500985145568848, 'sampling/importance_sampling_ratio/min': 0.7719552516937256, 'sampling/importance_sampling_ratio/mean': 1.0033729076385498, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.028014210518449545, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.8854995965957642, 'reward_meter_mean': 0.8854995965957642, 'reward_meter_std': 0.043195273727178574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8854995965957642, 'reward_total_composite_std': 0.043195273727178574, 'epoch': 0.07}

 54%|█████▍    | 1776/3300 [2:13:03<2:04:21,  4.90s/it]INFO 04-12 00:49:54 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1777/3300 [2:13:08<2:02:36,  4.83s/it]2026-04-12 00:49:59,306 | INFO | train_grpo_train | metrics_logged mode=train step=1777

                                                       
{'loss': 0.013, 'grad_norm': 4.8217244148254395, 'learning_rate': 4.618181818181818e-06, 'num_tokens': 3998604.0, 'completions/mean_length': 71.25, 'completions/min_length': 69.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9572109580039978, 'rewards/meter/std': 0.052123501896858215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9572109580039978, 'rewards/total_composite/std': 0.052123501896858215, 'reward': 0.9572109580039978, 'reward_std': 0.05212350934743881, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.057501811534166336, 'sampling/sampling_logp_difference/max': 1.8455209732055664, 'sampling/importance_sampling_ratio/min': 0.1579430252313614, 'sampling/importance_sampling_ratio/mean': 1.0080149173736572, 'sampling/importance_sampling_ratio/max': 1.790594458580017, 'entropy': 0.4883102737367153, 'clip_ratio/low_mean': 0.012251984560862184, 'clip_ratio/low_min': 0.012251984560862184, 'clip_ratio/high_mean': 0.03510406916029751, 'clip_ratio/high_max': 0.03510406916029751, 'clip_ratio/region_mean': 0.0473560537211597, 'reward_total_mean': 0.9572109580039978, 'reward_meter_mean': 0.9572109580039978, 'reward_meter_std': 0.052123501896858215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9572109580039978, 'reward_total_composite_std': 0.052123501896858215, 'epoch': 0.07}

 54%|█████▍    | 1777/3300 [2:13:08<2:02:36,  4.83s/it]INFO 04-12 00:49:59 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1778/3300 [2:13:14<2:12:20,  5.22s/it]2026-04-12 00:50:05,432 | INFO | train_grpo_train | metrics_logged mode=train step=1778

                                                       
{'loss': 0.0316, 'grad_norm': 4.067593097686768, 'learning_rate': 4.615151515151515e-06, 'num_tokens': 4001210.0, 'completions/mean_length': 160.75, 'completions/min_length': 146.0, 'completions/max_length': 206.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.75, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 206.0, 'rewards/meter/mean': 0.7919793128967285, 'rewards/meter/std': 0.23337964713573456, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9665178656578064, 'rewards/repeat_penalty/std': 0.062180306762456894, 'rewards/total_composite/mean': 0.7290425300598145, 'rewards/total_composite/std': 0.19308261573314667, 'reward': 0.7290425300598145, 'reward_std': 0.19308261573314667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08566864579916, 'sampling/sampling_logp_difference/max': 8.59061336517334, 'sampling/importance_sampling_ratio/min': 0.00018584205827210099, 'sampling/importance_sampling_ratio/mean': 1.014849066734314, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7425523065030575, 'clip_ratio/low_mean': 0.017992550507187843, 'clip_ratio/low_min': 0.017992550507187843, 'clip_ratio/high_mean': 0.04102262854576111, 'clip_ratio/high_max': 0.04102262854576111, 'clip_ratio/region_mean': 0.05901517905294895, 'reward_total_mean': 0.7290425300598145, 'reward_meter_mean': 0.7919793128967285, 'reward_meter_std': 0.23337964713573456, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9665178656578064, 'reward_repeat_penalty_std': 0.062180306762456894, 'reward_total_composite_mean': 0.7290425300598145, 'reward_total_composite_std': 0.19308261573314667, 'epoch': 0.07}

 54%|█████▍    | 1778/3300 [2:13:14<2:12:20,  5.22s/it]INFO 04-12 00:50:05 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1779/3300 [2:13:19<2:12:36,  5.23s/it]2026-04-12 00:50:10,692 | INFO | train_grpo_train | metrics_logged mode=train step=1779

                                                       
{'loss': 0.0089, 'grad_norm': 2.5793981552124023, 'learning_rate': 4.612121212121212e-06, 'num_tokens': 4003558.0, 'completions/mean_length': 134.5, 'completions/min_length': 133.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.5, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.9970629215240479, 'rewards/meter/std': 0.0009571582195349038, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.854625403881073, 'rewards/total_composite/std': 0.0008204205660149455, 'reward': 0.854625403881073, 'reward_std': 0.0008204328478313982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010115544311702251, 'sampling/sampling_logp_difference/max': 0.8139721155166626, 'sampling/importance_sampling_ratio/min': 0.44309455156326294, 'sampling/importance_sampling_ratio/mean': 1.0006462335586548, 'sampling/importance_sampling_ratio/max': 1.5591769218444824, 'entropy': 0.052033907268196344, 'clip_ratio/low_mean': 0.0027173913549631834, 'clip_ratio/low_min': 0.0027173913549631834, 'clip_ratio/high_mean': 0.0074559845379553735, 'clip_ratio/high_max': 0.0074559845379553735, 'clip_ratio/region_mean': 0.010173375892918557, 'reward_total_mean': 0.854625403881073, 'reward_meter_mean': 0.9970629215240479, 'reward_meter_std': 0.0009571582195349038, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.854625403881073, 'reward_total_composite_std': 0.0008204205660149455, 'epoch': 0.07}

 54%|█████▍    | 1779/3300 [2:13:19<2:12:36,  5.23s/it]INFO 04-12 00:50:10 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1780/3300 [2:13:27<2:29:16,  5.89s/it]2026-04-12 00:50:18,122 | INFO | train_grpo_train | metrics_logged mode=train step=1780

                                                       
{'loss': 0.0069, 'grad_norm': 2.3572487831115723, 'learning_rate': 4.60909090909091e-06, 'num_tokens': 4007791.0, 'completions/mean_length': 304.125, 'completions/min_length': 290.0, 'completions/max_length': 313.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 304.125, 'completions/min_terminated_length': 290.0, 'completions/max_terminated_length': 313.0, 'rewards/meter/mean': 0.9952999353408813, 'rewards/meter/std': 0.0025832315441221, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9177083373069763, 'rewards/repeat_penalty/std': 0.08480106294155121, 'rewards/total_composite/mean': 0.7979733347892761, 'rewards/total_composite/std': 0.06831313669681549, 'reward': 0.7979733347892761, 'reward_std': 0.06831313669681549, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.045330848544836044, 'sampling/sampling_logp_difference/max': 2.610581398010254, 'sampling/importance_sampling_ratio/min': 0.07349180430173874, 'sampling/importance_sampling_ratio/mean': 1.010035753250122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4535849690437317, 'clip_ratio/low_mean': 0.01071572583168745, 'clip_ratio/low_min': 0.01071572583168745, 'clip_ratio/high_mean': 0.018438909435644746, 'clip_ratio/high_max': 0.018438909435644746, 'clip_ratio/region_mean': 0.029154635267332196, 'reward_total_mean': 0.7979733347892761, 'reward_meter_mean': 0.9952999353408813, 'reward_meter_std': 0.0025832315441221, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9177083373069763, 'reward_repeat_penalty_std': 0.08480106294155121, 'reward_total_composite_mean': 0.7979733347892761, 'reward_total_composite_std': 0.06831313669681549, 'epoch': 0.07}

 54%|█████▍    | 1780/3300 [2:13:27<2:29:16,  5.89s/it]INFO 04-12 00:50:18 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1781/3300 [2:13:31<2:16:24,  5.39s/it]2026-04-12 00:50:22,340 | INFO | train_grpo_train | metrics_logged mode=train step=1781

                                                       
{'loss': -0.0006, 'grad_norm': 2.1961991786956787, 'learning_rate': 4.606060606060606e-06, 'num_tokens': 4009447.0, 'completions/mean_length': 50.0, 'completions/min_length': 50.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9375852346420288, 'rewards/meter/std': 0.0008122828439809382, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9375852346420288, 'rewards/total_composite/std': 0.0008122828439809382, 'reward': 0.9375852346420288, 'reward_std': 0.0008122779545374215, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011180485598742962, 'sampling/sampling_logp_difference/max': 0.815919816493988, 'sampling/importance_sampling_ratio/min': 0.4422323405742645, 'sampling/importance_sampling_ratio/mean': 1.0008137226104736, 'sampling/importance_sampling_ratio/max': 1.5095360279083252, 'entropy': 0.08982300292700529, 'clip_ratio/low_mean': 0.0024999999441206455, 'clip_ratio/low_min': 0.0024999999441206455, 'clip_ratio/high_mean': 0.0024999999441206455, 'clip_ratio/high_max': 0.0024999999441206455, 'clip_ratio/region_mean': 0.004999999888241291, 'reward_total_mean': 0.9375852346420288, 'reward_meter_mean': 0.9375852346420288, 'reward_meter_std': 0.0008122828439809382, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9375852346420288, 'reward_total_composite_std': 0.0008122828439809382, 'epoch': 0.07}

 54%|█████▍    | 1781/3300 [2:13:31<2:16:24,  5.39s/it]INFO 04-12 00:50:22 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1782/3300 [2:13:35<2:11:09,  5.18s/it]2026-04-12 00:50:27,042 | INFO | train_grpo_train | metrics_logged mode=train step=1782

                                                       
{'loss': -0.0032, 'grad_norm': 2.5923538208007812, 'learning_rate': 4.603030303030304e-06, 'num_tokens': 4011350.0, 'completions/mean_length': 80.875, 'completions/min_length': 79.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.875, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.9967532753944397, 'rewards/meter/std': 0.0014706900110468268, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8722769021987915, 'rewards/total_composite/std': 0.3524559736251831, 'reward': 0.8722769021987915, 'reward_std': 0.3524559438228607, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05614684522151947, 'sampling/sampling_logp_difference/max': 1.5454130172729492, 'sampling/importance_sampling_ratio/min': 0.21322380006313324, 'sampling/importance_sampling_ratio/mean': 1.0133930444717407, 'sampling/importance_sampling_ratio/max': 1.5346218347549438, 'entropy': 0.5954937487840652, 'clip_ratio/low_mean': 0.004746835213154554, 'clip_ratio/low_min': 0.004746835213154554, 'clip_ratio/high_mean': 0.03219789918512106, 'clip_ratio/high_max': 0.03219789918512106, 'clip_ratio/region_mean': 0.036944734398275614, 'reward_total_mean': 0.8722769021987915, 'reward_meter_mean': 0.9967532753944397, 'reward_meter_std': 0.0014706900110468268, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8722769021987915, 'reward_total_composite_std': 0.3524559736251831, 'epoch': 0.07}

 54%|█████▍    | 1782/3300 [2:13:36<2:11:09,  5.18s/it]INFO 04-12 00:50:27 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1783/3300 [2:13:41<2:12:48,  5.25s/it]2026-04-12 00:50:32,456 | INFO | train_grpo_train | metrics_logged mode=train step=1783

                                                       
{'loss': 0.0012, 'grad_norm': 0.8668001294136047, 'learning_rate': 4.600000000000001e-06, 'num_tokens': 4014214.0, 'completions/mean_length': 152.0, 'completions/min_length': 152.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 152.0, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9920245409011841, 'rewards/meter/std': 0.00013670595944859087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.612500011920929, 'rewards/repeat_penalty/std': 0.035355325788259506, 'rewards/total_composite/mean': 0.6076171398162842, 'rewards/total_composite/std': 0.035118672996759415, 'reward': 0.6076171398162842, 'reward_std': 0.03511865437030792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007640105206519365, 'sampling/sampling_logp_difference/max': 2.3813014030456543, 'sampling/importance_sampling_ratio/min': 0.09243020415306091, 'sampling/importance_sampling_ratio/mean': 0.9984169006347656, 'sampling/importance_sampling_ratio/max': 1.160125970840454, 'entropy': 0.024257719283923507, 'clip_ratio/low_mean': 0.004111842135898769, 'clip_ratio/low_min': 0.004111842135898769, 'clip_ratio/high_mean': 0.0008223684271797538, 'clip_ratio/high_max': 0.0008223684271797538, 'clip_ratio/region_mean': 0.004934210563078523, 'reward_total_mean': 0.6076171398162842, 'reward_meter_mean': 0.9920245409011841, 'reward_meter_std': 0.00013670595944859087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.612500011920929, 'reward_repeat_penalty_std': 0.035355325788259506, 'reward_total_composite_mean': 0.6076171398162842, 'reward_total_composite_std': 0.035118672996759415, 'epoch': 0.07}

 54%|█████▍    | 1783/3300 [2:13:41<2:12:48,  5.25s/it]INFO 04-12 00:50:32 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1784/3300 [2:13:45<2:07:26,  5.04s/it]2026-04-12 00:50:37,011 | INFO | train_grpo_train | metrics_logged mode=train step=1784

                                                       
{'loss': 0.0087, 'grad_norm': 4.6875691413879395, 'learning_rate': 4.596969696969697e-06, 'num_tokens': 4015926.0, 'completions/mean_length': 65.0, 'completions/min_length': 62.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8023771643638611, 'rewards/meter/std': 0.2034316211938858, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8023771643638611, 'rewards/total_composite/std': 0.2034316211938858, 'reward': 0.8023771643638611, 'reward_std': 0.203431636095047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0649370476603508, 'sampling/sampling_logp_difference/max': 1.2312498092651367, 'sampling/importance_sampling_ratio/min': 0.2919274866580963, 'sampling/importance_sampling_ratio/mean': 1.0129239559173584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5869800895452499, 'clip_ratio/low_mean': 0.01740056835114956, 'clip_ratio/low_min': 0.01740056835114956, 'clip_ratio/high_mean': 0.02884695027023554, 'clip_ratio/high_max': 0.02884695027023554, 'clip_ratio/region_mean': 0.0462475186213851, 'reward_total_mean': 0.8023771643638611, 'reward_meter_mean': 0.8023771643638611, 'reward_meter_std': 0.2034316211938858, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8023771643638611, 'reward_total_composite_std': 0.2034316211938858, 'epoch': 0.07}

 54%|█████▍    | 1784/3300 [2:13:46<2:07:26,  5.04s/it]INFO 04-12 00:50:37 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1785/3300 [2:13:50<2:03:42,  4.90s/it]2026-04-12 00:50:41,578 | INFO | train_grpo_train | metrics_logged mode=train step=1785

                                                       
{'loss': -0.0151, 'grad_norm': 2.411994457244873, 'learning_rate': 4.5939393939393945e-06, 'num_tokens': 4017713.0, 'completions/mean_length': 62.375, 'completions/min_length': 59.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.375, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9964926242828369, 'rewards/meter/std': 0.0008941100095398724, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9964926242828369, 'rewards/total_composite/std': 0.0008941100095398724, 'reward': 0.9964926242828369, 'reward_std': 0.0008941145497374237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01294564176350832, 'sampling/sampling_logp_difference/max': 1.444681167602539, 'sampling/importance_sampling_ratio/min': 0.23582126200199127, 'sampling/importance_sampling_ratio/mean': 1.0034265518188477, 'sampling/importance_sampling_ratio/max': 1.3098853826522827, 'entropy': 0.08539297711104155, 'clip_ratio/low_mean': 0.004167824285104871, 'clip_ratio/low_min': 0.004167824285104871, 'clip_ratio/high_mean': 0.005921379197388887, 'clip_ratio/high_max': 0.005921379197388887, 'clip_ratio/region_mean': 0.010089203482493758, 'reward_total_mean': 0.9964926242828369, 'reward_meter_mean': 0.9964926242828369, 'reward_meter_std': 0.0008941100095398724, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9964926242828369, 'reward_total_composite_std': 0.0008941100095398724, 'epoch': 0.07}

 54%|█████▍    | 1785/3300 [2:13:50<2:03:42,  4.90s/it]INFO 04-12 00:50:41 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1786/3300 [2:13:56<2:10:06,  5.16s/it]2026-04-12 00:50:47,330 | INFO | train_grpo_train | metrics_logged mode=train step=1786

                                                       
{'loss': -0.0127, 'grad_norm': 2.4926347732543945, 'learning_rate': 4.590909090909092e-06, 'num_tokens': 4020442.0, 'completions/mean_length': 152.125, 'completions/min_length': 151.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 152.125, 'completions/min_terminated_length': 151.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9910025596618652, 'rewards/meter/std': 0.0011056133080273867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6749999523162842, 'rewards/repeat_penalty/std': 0.103509820997715, 'rewards/total_composite/mean': 0.6688380241394043, 'rewards/total_composite/std': 0.10174151510000229, 'reward': 0.6688380241394043, 'reward_std': 0.1017415001988411, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008964202366769314, 'sampling/sampling_logp_difference/max': 1.3915038108825684, 'sampling/importance_sampling_ratio/min': 0.24870102107524872, 'sampling/importance_sampling_ratio/mean': 1.0039024353027344, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04743872885592282, 'clip_ratio/low_mean': 0.0016556291375309229, 'clip_ratio/low_min': 0.0016556291375309229, 'clip_ratio/high_mean': 0.004780629184097052, 'clip_ratio/high_max': 0.004780629184097052, 'clip_ratio/region_mean': 0.0064362583216279745, 'reward_total_mean': 0.6688380241394043, 'reward_meter_mean': 0.9910025596618652, 'reward_meter_std': 0.0011056133080273867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6749999523162842, 'reward_repeat_penalty_std': 0.103509820997715, 'reward_total_composite_mean': 0.6688380241394043, 'reward_total_composite_std': 0.10174151510000229, 'epoch': 0.07}

 54%|█████▍    | 1786/3300 [2:13:56<2:10:06,  5.16s/it]INFO 04-12 00:50:47 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1787/3300 [2:14:01<2:08:13,  5.09s/it]2026-04-12 00:50:52,247 | INFO | train_grpo_train | metrics_logged mode=train step=1787

                                                       
{'loss': 0.0007, 'grad_norm': 1.5127758979797363, 'learning_rate': 4.587878787878788e-06, 'num_tokens': 4022570.0, 'completions/mean_length': 95.0, 'completions/min_length': 94.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.0, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9981739521026611, 'rewards/meter/std': 6.629144627368078e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9732200503349304, 'rewards/total_composite/std': 0.07058726996183395, 'reward': 0.9732200503349304, 'reward_std': 0.07058726251125336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010238762944936752, 'sampling/sampling_logp_difference/max': 1.506108283996582, 'sampling/importance_sampling_ratio/min': 0.22177138924598694, 'sampling/importance_sampling_ratio/mean': 1.000563383102417, 'sampling/importance_sampling_ratio/max': 1.5315839052200317, 'entropy': 0.03921856731176376, 'clip_ratio/low_mean': 0.003947368357330561, 'clip_ratio/low_min': 0.003947368357330561, 'clip_ratio/high_mean': 0.005263741128146648, 'clip_ratio/high_max': 0.005263741128146648, 'clip_ratio/region_mean': 0.009211109485477209, 'reward_total_mean': 0.9732200503349304, 'reward_meter_mean': 0.9981739521026611, 'reward_meter_std': 6.629144627368078e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9732200503349304, 'reward_total_composite_std': 0.07058726996183395, 'epoch': 0.07}

 54%|█████▍    | 1787/3300 [2:14:01<2:08:13,  5.09s/it]INFO 04-12 00:50:52 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1788/3300 [2:14:07<2:14:54,  5.35s/it]2026-04-12 00:50:58,229 | INFO | train_grpo_train | metrics_logged mode=train step=1788

                                                       
{'loss': 0.0073, 'grad_norm': 2.8437070846557617, 'learning_rate': 4.5848484848484854e-06, 'num_tokens': 4025609.0, 'completions/mean_length': 193.875, 'completions/min_length': 190.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 193.875, 'completions/min_terminated_length': 190.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.9956002235412598, 'rewards/meter/std': 0.0012398899998515844, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9679561853408813, 'rewards/total_composite/std': 0.05145645886659622, 'reward': 0.9679561853408813, 'reward_std': 0.05145645514130592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05295930430293083, 'sampling/sampling_logp_difference/max': 1.3258538246154785, 'sampling/importance_sampling_ratio/min': 0.2655761241912842, 'sampling/importance_sampling_ratio/mean': 1.008656620979309, 'sampling/importance_sampling_ratio/max': 1.9465110301971436, 'entropy': 0.5489388853311539, 'clip_ratio/low_mean': 0.0070457912515848875, 'clip_ratio/low_min': 0.0070457912515848875, 'clip_ratio/high_mean': 0.038041369058191776, 'clip_ratio/high_max': 0.038041369058191776, 'clip_ratio/region_mean': 0.045087160309776664, 'reward_total_mean': 0.9679561853408813, 'reward_meter_mean': 0.9956002235412598, 'reward_meter_std': 0.0012398899998515844, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9679561853408813, 'reward_total_composite_std': 0.05145645886659622, 'epoch': 0.07}

 54%|█████▍    | 1788/3300 [2:14:07<2:14:54,  5.35s/it]INFO 04-12 00:50:58 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1789/3300 [2:14:11<2:08:58,  5.12s/it]2026-04-12 00:51:02,811 | INFO | train_grpo_train | metrics_logged mode=train step=1789

                                                       
{'loss': 0.0114, 'grad_norm': 3.5395383834838867, 'learning_rate': 4.581818181818183e-06, 'num_tokens': 4027414.0, 'completions/mean_length': 67.625, 'completions/min_length': 64.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9397947192192078, 'rewards/meter/std': 0.026229292154312134, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9397947192192078, 'rewards/total_composite/std': 0.026229292154312134, 'reward': 0.9397947192192078, 'reward_std': 0.026229292154312134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04796122387051582, 'sampling/sampling_logp_difference/max': 1.1403248310089111, 'sampling/importance_sampling_ratio/min': 0.3197151720523834, 'sampling/importance_sampling_ratio/mean': 1.0007556676864624, 'sampling/importance_sampling_ratio/max': 1.5871034860610962, 'entropy': 0.38035739958286285, 'clip_ratio/low_mean': 0.011368778301402926, 'clip_ratio/low_min': 0.011368778301402926, 'clip_ratio/high_mean': 0.016516561503522098, 'clip_ratio/high_max': 0.016516561503522098, 'clip_ratio/region_mean': 0.027885339804925025, 'reward_total_mean': 0.9397947192192078, 'reward_meter_mean': 0.9397947192192078, 'reward_meter_std': 0.026229292154312134, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9397947192192078, 'reward_total_composite_std': 0.026229292154312134, 'epoch': 0.07}

 54%|█████▍    | 1789/3300 [2:14:11<2:08:58,  5.12s/it]INFO 04-12 00:51:03 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1790/3300 [2:14:17<2:10:06,  5.17s/it]2026-04-12 00:51:08,107 | INFO | train_grpo_train | metrics_logged mode=train step=1790

                                                       
{'loss': 0.0094, 'grad_norm': 2.5047359466552734, 'learning_rate': 4.578787878787879e-06, 'num_tokens': 4029603.0, 'completions/mean_length': 110.625, 'completions/min_length': 107.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.625, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9982960224151611, 'rewards/meter/std': 0.000550406810361892, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982960224151611, 'rewards/total_composite/std': 0.000550406810361892, 'reward': 0.9982960224151611, 'reward_std': 0.0005504076834768057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048308465629816055, 'sampling/sampling_logp_difference/max': 1.0485877990722656, 'sampling/importance_sampling_ratio/min': 0.35043227672576904, 'sampling/importance_sampling_ratio/mean': 1.005360722541809, 'sampling/importance_sampling_ratio/max': 1.5884928703308105, 'entropy': 0.4972764067351818, 'clip_ratio/low_mean': 0.019280804554000497, 'clip_ratio/low_min': 0.019280804554000497, 'clip_ratio/high_mean': 0.013679954456165433, 'clip_ratio/high_max': 0.013679954456165433, 'clip_ratio/region_mean': 0.03296075901016593, 'reward_total_mean': 0.9982960224151611, 'reward_meter_mean': 0.9982960224151611, 'reward_meter_std': 0.000550406810361892, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982960224151611, 'reward_total_composite_std': 0.000550406810361892, 'epoch': 0.07}

 54%|█████▍    | 1790/3300 [2:14:17<2:10:06,  5.17s/it]INFO 04-12 00:51:08 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1791/3300 [2:14:24<2:27:01,  5.85s/it]2026-04-12 00:51:15,525 | INFO | train_grpo_train | metrics_logged mode=train step=1791

                                                       
{'loss': -0.0086, 'grad_norm': 2.072072982788086, 'learning_rate': 4.575757575757576e-06, 'num_tokens': 4033409.0, 'completions/mean_length': 280.75, 'completions/min_length': 270.0, 'completions/max_length': 292.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 280.75, 'completions/min_terminated_length': 270.0, 'completions/max_terminated_length': 292.0, 'rewards/meter/mean': 0.9915696978569031, 'rewards/meter/std': 0.01817433536052704, 'rewards/count_adherence/mean': 0.921875, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8535714149475098, 'rewards/repeat_penalty/std': 0.07463330775499344, 'rewards/total_composite/mean': 0.7775777578353882, 'rewards/total_composite/std': 0.05556074157357216, 'reward': 0.7775777578353882, 'reward_std': 0.05556074529886246, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03709190711379051, 'sampling/sampling_logp_difference/max': 1.9197864532470703, 'sampling/importance_sampling_ratio/min': 0.14663827419281006, 'sampling/importance_sampling_ratio/mean': 1.0069139003753662, 'sampling/importance_sampling_ratio/max': 1.9100106954574585, 'entropy': 0.39633841440081596, 'clip_ratio/low_mean': 0.009515275072772056, 'clip_ratio/low_min': 0.009515275072772056, 'clip_ratio/high_mean': 0.015547977527603507, 'clip_ratio/high_max': 0.015547977527603507, 'clip_ratio/region_mean': 0.025063252600375563, 'reward_total_mean': 0.7775777578353882, 'reward_meter_mean': 0.9915696978569031, 'reward_meter_std': 0.01817433536052704, 'reward_count_adherence_mean': 0.921875, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8535714149475098, 'reward_repeat_penalty_std': 0.07463330775499344, 'reward_total_composite_mean': 0.7775777578353882, 'reward_total_composite_std': 0.05556074157357216, 'epoch': 0.07}

 54%|█████▍    | 1791/3300 [2:14:24<2:27:01,  5.85s/it]INFO 04-12 00:51:15 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1792/3300 [2:14:29<2:18:31,  5.51s/it]2026-04-12 00:51:20,255 | INFO | train_grpo_train | metrics_logged mode=train step=1792

                                                       
{'loss': 0.0012, 'grad_norm': 0.36933040618896484, 'learning_rate': 4.572727272727273e-06, 'num_tokens': 4035226.0, 'completions/mean_length': 68.125, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9975280165672302, 'rewards/meter/std': 4.185540819889866e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975280165672302, 'rewards/total_composite/std': 4.185540819889866e-05, 'reward': 0.9975280165672302, 'reward_std': 4.18471208831761e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0064887660555541515, 'sampling/sampling_logp_difference/max': 0.6608014106750488, 'sampling/importance_sampling_ratio/min': 0.516437292098999, 'sampling/importance_sampling_ratio/mean': 1.0018044710159302, 'sampling/importance_sampling_ratio/max': 1.4639054536819458, 'entropy': 0.04893740452826023, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9975280165672302, 'reward_meter_mean': 0.9975280165672302, 'reward_meter_std': 4.185540819889866e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975280165672302, 'reward_total_composite_std': 4.185540819889866e-05, 'epoch': 0.07}

 54%|█████▍    | 1792/3300 [2:14:29<2:18:31,  5.51s/it]INFO 04-12 00:51:20 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1793/3300 [2:14:34<2:14:30,  5.36s/it]2026-04-12 00:51:25,255 | INFO | train_grpo_train | metrics_logged mode=train step=1793

                                                       
{'loss': 0.0024, 'grad_norm': 2.3365345001220703, 'learning_rate': 4.56969696969697e-06, 'num_tokens': 4037239.0, 'completions/mean_length': 94.625, 'completions/min_length': 94.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.625, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9981467723846436, 'rewards/meter/std': 4.623903805622831e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8733782768249512, 'rewards/total_composite/std': 0.1033172756433487, 'reward': 0.8733782768249512, 'reward_std': 0.10331728309392929, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003089956007897854, 'sampling/sampling_logp_difference/max': 0.6266226768493652, 'sampling/importance_sampling_ratio/min': 0.5343936085700989, 'sampling/importance_sampling_ratio/mean': 1.0002309083938599, 'sampling/importance_sampling_ratio/max': 1.2057961225509644, 'entropy': 0.021367451641708612, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0013157895300537348, 'clip_ratio/high_max': 0.0013157895300537348, 'clip_ratio/region_mean': 0.0013157895300537348, 'reward_total_mean': 0.8733782768249512, 'reward_meter_mean': 0.9981467723846436, 'reward_meter_std': 4.623903805622831e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8733782768249512, 'reward_total_composite_std': 0.1033172756433487, 'epoch': 0.07}

 54%|█████▍    | 1793/3300 [2:14:34<2:14:30,  5.36s/it]INFO 04-12 00:51:25 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1794/3300 [2:14:40<2:23:27,  5.72s/it]2026-04-12 00:51:31,804 | INFO | train_grpo_train | metrics_logged mode=train step=1794

                                                       
{'loss': -0.0075, 'grad_norm': 1.2437056303024292, 'learning_rate': 4.566666666666667e-06, 'num_tokens': 4040774.0, 'completions/mean_length': 205.875, 'completions/min_length': 204.0, 'completions/max_length': 213.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 205.875, 'completions/min_terminated_length': 204.0, 'completions/max_terminated_length': 213.0, 'rewards/meter/mean': 0.9908320307731628, 'rewards/meter/std': 0.000867572205606848, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6607142686843872, 'rewards/repeat_penalty/std': 0.06331465393304825, 'rewards/total_composite/mean': 0.6546941995620728, 'rewards/total_composite/std': 0.06324289739131927, 'reward': 0.6546941995620728, 'reward_std': 0.06324288249015808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009390618652105331, 'sampling/sampling_logp_difference/max': 1.3305549621582031, 'sampling/importance_sampling_ratio/min': 0.26433053612709045, 'sampling/importance_sampling_ratio/mean': 1.0008214712142944, 'sampling/importance_sampling_ratio/max': 1.7237244844436646, 'entropy': 0.0551890037022531, 'clip_ratio/low_mean': 0.0030517695122398436, 'clip_ratio/low_min': 0.0030517695122398436, 'clip_ratio/high_mean': 0.002393221133388579, 'clip_ratio/high_max': 0.002393221133388579, 'clip_ratio/region_mean': 0.0054449906456284225, 'reward_total_mean': 0.6546941995620728, 'reward_meter_mean': 0.9908320307731628, 'reward_meter_std': 0.000867572205606848, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6607142686843872, 'reward_repeat_penalty_std': 0.06331465393304825, 'reward_total_composite_mean': 0.6546941995620728, 'reward_total_composite_std': 0.06324289739131927, 'epoch': 0.07}

 54%|█████▍    | 1794/3300 [2:14:40<2:23:27,  5.72s/it]INFO 04-12 00:51:32 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1795/3300 [2:14:46<2:21:51,  5.66s/it]2026-04-12 00:51:37,313 | INFO | train_grpo_train | metrics_logged mode=train step=1795

                                                       
{'loss': 0.0056, 'grad_norm': 2.7669808864593506, 'learning_rate': 4.563636363636364e-06, 'num_tokens': 4043010.0, 'completions/mean_length': 118.5, 'completions/min_length': 116.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.5, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9971722364425659, 'rewards/meter/std': 0.0011399141512811184, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9722900390625, 'rewards/total_composite/std': 0.0711437240242958, 'reward': 0.9722900390625, 'reward_std': 0.07114371657371521, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04540203511714935, 'sampling/sampling_logp_difference/max': 1.2590808868408203, 'sampling/importance_sampling_ratio/min': 0.28391486406326294, 'sampling/importance_sampling_ratio/mean': 1.014491319656372, 'sampling/importance_sampling_ratio/max': 1.7191333770751953, 'entropy': 0.46814702078700066, 'clip_ratio/low_mean': 0.004201680887490511, 'clip_ratio/low_min': 0.004201680887490511, 'clip_ratio/high_mean': 0.030286709661595523, 'clip_ratio/high_max': 0.030286709661595523, 'clip_ratio/region_mean': 0.034488390549086034, 'reward_total_mean': 0.9722900390625, 'reward_meter_mean': 0.9971722364425659, 'reward_meter_std': 0.0011399141512811184, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9722900390625, 'reward_total_composite_std': 0.0711437240242958, 'epoch': 0.07}

 54%|█████▍    | 1795/3300 [2:14:46<2:21:51,  5.66s/it]INFO 04-12 00:51:37 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1796/3300 [2:14:52<2:26:36,  5.85s/it]2026-04-12 00:51:43,623 | INFO | train_grpo_train | metrics_logged mode=train step=1796

                                                       
{'loss': 0.0069, 'grad_norm': 2.253188371658325, 'learning_rate': 4.560606060606061e-06, 'num_tokens': 4045960.0, 'completions/mean_length': 184.75, 'completions/min_length': 168.0, 'completions/max_length': 199.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 184.75, 'completions/min_terminated_length': 168.0, 'completions/max_terminated_length': 199.0, 'rewards/meter/mean': 0.9870539903640747, 'rewards/meter/std': 0.024071279913187027, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.8888888955116272, 'rewards/repeat_penalty/std': 0.10286889225244522, 'rewards/total_composite/mean': 0.7317129969596863, 'rewards/total_composite/std': 0.30691879987716675, 'reward': 0.7317129969596863, 'reward_std': 0.30691882967948914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05264398828148842, 'sampling/sampling_logp_difference/max': 1.2949190139770508, 'sampling/importance_sampling_ratio/min': 0.27392005920410156, 'sampling/importance_sampling_ratio/mean': 1.0125421285629272, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5114907324314117, 'clip_ratio/low_mean': 0.002659574383869767, 'clip_ratio/low_min': 0.002659574383869767, 'clip_ratio/high_mean': 0.024542337749153376, 'clip_ratio/high_max': 0.024542337749153376, 'clip_ratio/region_mean': 0.027201912133023143, 'reward_total_mean': 0.7317129969596863, 'reward_meter_mean': 0.9870539903640747, 'reward_meter_std': 0.024071279913187027, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.8888888955116272, 'reward_repeat_penalty_std': 0.10286889225244522, 'reward_total_composite_mean': 0.7317129969596863, 'reward_total_composite_std': 0.30691879987716675, 'epoch': 0.07}

 54%|█████▍    | 1796/3300 [2:14:52<2:26:36,  5.85s/it]INFO 04-12 00:51:43 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1797/3300 [2:15:01<2:53:20,  6.92s/it]2026-04-12 00:51:53,032 | INFO | train_grpo_train | metrics_logged mode=train step=1797

                                                       
{'loss': 0.0079, 'grad_norm': 1.9448902606964111, 'learning_rate': 4.557575757575758e-06, 'num_tokens': 4051243.0, 'completions/mean_length': 420.375, 'completions/min_length': 395.0, 'completions/max_length': 458.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 420.375, 'completions/min_terminated_length': 395.0, 'completions/max_terminated_length': 458.0, 'rewards/meter/mean': 0.448710560798645, 'rewards/meter/std': 0.4174151122570038, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.023145509883761406, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7848790884017944, 'rewards/repeat_penalty/std': 0.13375473022460938, 'rewards/total_composite/mean': 0.2035287767648697, 'rewards/total_composite/std': 0.17783083021640778, 'reward': 0.2035287767648697, 'reward_std': 0.17783081531524658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036439694464206696, 'sampling/sampling_logp_difference/max': 7.995334625244141, 'sampling/importance_sampling_ratio/min': 0.0003370313497725874, 'sampling/importance_sampling_ratio/mean': 1.0021564960479736, 'sampling/importance_sampling_ratio/max': 1.8132941722869873, 'entropy': 0.22779023833572865, 'clip_ratio/low_mean': 0.010334618214983493, 'clip_ratio/low_min': 0.010334618214983493, 'clip_ratio/high_mean': 0.01244131033308804, 'clip_ratio/high_max': 0.01244131033308804, 'clip_ratio/region_mean': 0.022775928548071533, 'reward_total_mean': 0.2035287767648697, 'reward_meter_mean': 0.448710560798645, 'reward_meter_std': 0.4174151122570038, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.023145509883761406, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7848790884017944, 'reward_repeat_penalty_std': 0.13375473022460938, 'reward_total_composite_mean': 0.2035287767648697, 'reward_total_composite_std': 0.17783083021640778, 'epoch': 0.07}

 54%|█████▍    | 1797/3300 [2:15:02<2:53:20,  6.92s/it]INFO 04-12 00:51:53 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1798/3300 [2:15:08<2:47:49,  6.70s/it]2026-04-12 00:51:59,230 | INFO | train_grpo_train | metrics_logged mode=train step=1798

                                                       
{'loss': 0.0113, 'grad_norm': 2.2561349868774414, 'learning_rate': 4.554545454545455e-06, 'num_tokens': 4054231.0, 'completions/mean_length': 182.5, 'completions/min_length': 172.0, 'completions/max_length': 188.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 182.5, 'completions/min_terminated_length': 172.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.9977061152458191, 'rewards/meter/std': 0.001272720517590642, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9305555820465088, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.9284070730209351, 'rewards/total_composite/std': 0.0822766125202179, 'reward': 0.9284070730209351, 'reward_std': 0.08227662742137909, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03487614914774895, 'sampling/sampling_logp_difference/max': 1.8383588790893555, 'sampling/importance_sampling_ratio/min': 0.15907828509807587, 'sampling/importance_sampling_ratio/mean': 1.008242130279541, 'sampling/importance_sampling_ratio/max': 1.7942215204238892, 'entropy': 0.3780089020729065, 'clip_ratio/low_mean': 0.013545665598940104, 'clip_ratio/low_min': 0.013545665598940104, 'clip_ratio/high_mean': 0.017403612146154046, 'clip_ratio/high_max': 0.017403612146154046, 'clip_ratio/region_mean': 0.03094927774509415, 'reward_total_mean': 0.9284070730209351, 'reward_meter_mean': 0.9977061152458191, 'reward_meter_std': 0.001272720517590642, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9305555820465088, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.9284070730209351, 'reward_total_composite_std': 0.0822766125202179, 'epoch': 0.07}

 54%|█████▍    | 1798/3300 [2:15:08<2:47:49,  6.70s/it]INFO 04-12 00:51:59 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1799/3300 [2:15:13<2:33:44,  6.15s/it]2026-04-12 00:52:04,075 | INFO | train_grpo_train | metrics_logged mode=train step=1799

                                                       
{'loss': -0.0358, 'grad_norm': 5.813068866729736, 'learning_rate': 4.551515151515152e-06, 'num_tokens': 4056037.0, 'completions/mean_length': 68.75, 'completions/min_length': 63.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.7557946443557739, 'rewards/meter/std': 0.27249956130981445, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7557946443557739, 'rewards/total_composite/std': 0.27249956130981445, 'reward': 0.7557946443557739, 'reward_std': 0.27249956130981445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054108865559101105, 'sampling/sampling_logp_difference/max': 0.9117650985717773, 'sampling/importance_sampling_ratio/min': 0.40181437134742737, 'sampling/importance_sampling_ratio/mean': 1.0174649953842163, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5148973725736141, 'clip_ratio/low_mean': 0.018777355086058378, 'clip_ratio/low_min': 0.018777355086058378, 'clip_ratio/high_mean': 0.019246643991209567, 'clip_ratio/high_max': 0.019246643991209567, 'clip_ratio/region_mean': 0.038023999077267945, 'reward_total_mean': 0.7557946443557739, 'reward_meter_mean': 0.7557946443557739, 'reward_meter_std': 0.27249956130981445, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7557946443557739, 'reward_total_composite_std': 0.27249956130981445, 'epoch': 0.07}

 55%|█████▍    | 1799/3300 [2:15:13<2:33:44,  6.15s/it]INFO 04-12 00:52:04 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1800/3300 [2:15:17<2:19:50,  5.59s/it]2026-04-12 00:52:08,382 | INFO | train_grpo_train | metrics_logged mode=train step=1800

                                                       
{'loss': 0.0696, 'grad_norm': 13.604365348815918, 'learning_rate': 4.548484848484849e-06, 'num_tokens': 4057576.0, 'completions/mean_length': 40.375, 'completions/min_length': 39.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9156763553619385, 'rewards/meter/std': 0.20896102488040924, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9156763553619385, 'rewards/total_composite/std': 0.20896102488040924, 'reward': 0.9156763553619385, 'reward_std': 0.20896100997924805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07238361239433289, 'sampling/sampling_logp_difference/max': 1.5261836051940918, 'sampling/importance_sampling_ratio/min': 0.217363640666008, 'sampling/importance_sampling_ratio/mean': 1.0134040117263794, 'sampling/importance_sampling_ratio/max': 1.8849577903747559, 'entropy': 0.5478580147027969, 'clip_ratio/low_mean': 0.01666666753590107, 'clip_ratio/low_min': 0.01666666753590107, 'clip_ratio/high_mean': 0.05649038520641625, 'clip_ratio/high_max': 0.05649038520641625, 'clip_ratio/region_mean': 0.07315705274231732, 'reward_total_mean': 0.9156763553619385, 'reward_meter_mean': 0.9156763553619385, 'reward_meter_std': 0.20896102488040924, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9156763553619385, 'reward_total_composite_std': 0.20896102488040924, 'epoch': 0.07}

 55%|█████▍    | 1800/3300 [2:15:17<2:19:50,  5.59s/it]INFO 04-12 00:52:08 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:29,  2.64s/it]

 23%|██▎       | 3/13 [00:08<00:30,  3.00s/it]

 31%|███       | 4/13 [00:13<00:32,  3.66s/it]

 38%|███▊      | 5/13 [00:18<00:32,  4.00s/it]

 46%|████▌     | 6/13 [00:23<00:30,  4.42s/it]

 54%|█████▍    | 7/13 [00:28<00:27,  4.62s/it]

 62%|██████▏   | 8/13 [00:34<00:25,  5.11s/it]

 69%|██████▉   | 9/13 [00:40<00:20,  5.23s/it]

 77%|███████▋  | 10/13 [00:44<00:14,  4.97s/it]

 85%|████████▍ | 11/13 [00:49<00:10,  5.06s/it]

 92%|█████████▏| 12/13 [00:54<00:05,  5.04s/it]

100%|██████████| 13/13 [00:59<00:00,  4.93s/it]2026-04-12 00:53:14,465 | INFO | train_grpo_train | metrics_logged mode=eval step=1800

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 66.0173, 'eval_samples_per_second': 1.575, 'eval_steps_per_second': 0.197, 'eval_num_tokens': 4057576.0, 'eval_completions/mean_length': 192.78846153846155, 'eval_completions/min_length': 63.84615384615385, 'eval_completions/max_length': 341.0, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 192.78846153846155, 'eval_completions/min_terminated_length': 63.84615384615385, 'eval_completions/max_terminated_length': 341.0, 'eval_rewards/meter/mean': 0.6883980471354264, 'eval_rewards/meter/std': 0.388204580746018, 'eval_rewards/count_adherence/mean': 0.8753787233279302, 'eval_rewards/count_adherence/std': 0.13936235010623932, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.8605746856102576, 'eval_rewards/repeat_penalty/std': 0.13569565231983477, 'eval_rewards/total_composite/mean': 0.5133193800082574, 'eval_rewards/total_composite/std': 0.346407216328841, 'eval_reward': 0.5133193800082574, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.02931356322593414, 'eval_sampling/sampling_logp_difference/max': 1.0275482031015248, 'eval_sampling/importance_sampling_ratio/min': 0.36259872638262236, 'eval_sampling/importance_sampling_ratio/mean': 1.0084307743952825, 'eval_sampling/importance_sampling_ratio/max': 1.5431275459436269, 'eval_entropy': 0.3313331672778496, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5133193800082574, 'eval_reward_meter_mean': 0.6883980471354264, 'eval_reward_meter_std': 0.388204580746018, 'eval_reward_count_adherence_mean': 0.8753787233279302, 'eval_reward_count_adherence_std': 0.13936235010623932, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.8605746856102576, 'eval_reward_repeat_penalty_std': 0.13569565231983477, 'eval_reward_total_composite_mean': 0.5133193800082574, 'eval_reward_total_composite_std': 0.346407216328841, 'epoch': 0.07}

 55%|█████▍    | 1800/3300 [2:16:23<2:19:50,  5.59s/it]

100%|██████████| 13/13 [00:59<00:00,  4.93s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 55%|█████▍    | 1801/3300 [2:16:32<11:03:09, 26.54s/it]2026-04-12 00:53:23,808 | INFO | train_grpo_train | metrics_logged mode=train step=1801

                                                        
{'loss': 0.0329, 'grad_norm': 4.170017719268799, 'learning_rate': 4.5454545454545455e-06, 'num_tokens': 4061109.0, 'completions/mean_length': 241.625, 'completions/min_length': 227.0, 'completions/max_length': 254.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 241.625, 'completions/min_terminated_length': 227.0, 'completions/max_terminated_length': 254.0, 'rewards/meter/mean': 0.5351053476333618, 'rewards/meter/std': 0.38701698184013367, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9261363744735718, 'rewards/repeat_penalty/std': 0.05368336662650108, 'rewards/total_composite/mean': 0.39037013053894043, 'rewards/total_composite/std': 0.3217414319515228, 'reward': 0.39037013053894043, 'reward_std': 0.3217414319515228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0780494287610054, 'sampling/sampling_logp_difference/max': 2.623555898666382, 'sampling/importance_sampling_ratio/min': 0.07254444062709808, 'sampling/importance_sampling_ratio/mean': 1.0214029550552368, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6957518979907036, 'clip_ratio/low_mean': 0.02372721955180168, 'clip_ratio/low_min': 0.02372721955180168, 'clip_ratio/high_mean': 0.027773853624239564, 'clip_ratio/high_max': 0.027773853624239564, 'clip_ratio/region_mean': 0.051501073176041245, 'reward_total_mean': 0.39037013053894043, 'reward_meter_mean': 0.5351053476333618, 'reward_meter_std': 0.38701698184013367, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9261363744735718, 'reward_repeat_penalty_std': 0.05368336662650108, 'reward_total_composite_mean': 0.39037013053894043, 'reward_total_composite_std': 0.3217414319515228, 'epoch': 0.07}

 55%|█████▍    | 1801/3300 [2:16:32<11:03:09, 26.54s/it]INFO 04-12 00:53:24 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1802/3300 [2:16:38<8:23:34, 20.17s/it] 2026-04-12 00:53:29,102 | INFO | train_grpo_train | metrics_logged mode=train step=1802

                                                       
{'loss': 0.038, 'grad_norm': 4.134006500244141, 'learning_rate': 4.542424242424243e-06, 'num_tokens': 4063157.0, 'completions/mean_length': 103.0, 'completions/min_length': 98.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9173542261123657, 'rewards/meter/std': 0.045704178512096405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.8030074238777161, 'rewards/total_composite/std': 0.10497334599494934, 'reward': 0.8030074238777161, 'reward_std': 0.10497336089611053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036168623715639114, 'sampling/sampling_logp_difference/max': 1.6687889099121094, 'sampling/importance_sampling_ratio/min': 0.18847519159317017, 'sampling/importance_sampling_ratio/mean': 1.0087733268737793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3216414675116539, 'clip_ratio/low_mean': 0.017737353453412652, 'clip_ratio/low_min': 0.017737353453412652, 'clip_ratio/high_mean': 0.01266491471324116, 'clip_ratio/high_max': 0.01266491471324116, 'clip_ratio/region_mean': 0.030402268166653812, 'reward_total_mean': 0.8030074238777161, 'reward_meter_mean': 0.9173542261123657, 'reward_meter_std': 0.045704178512096405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.8030074238777161, 'reward_total_composite_std': 0.10497334599494934, 'epoch': 0.07}

 55%|█████▍    | 1802/3300 [2:16:38<8:23:34, 20.17s/it]INFO 04-12 00:53:29 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1803/3300 [2:16:42<6:24:15, 15.40s/it]2026-04-12 00:53:33,374 | INFO | train_grpo_train | metrics_logged mode=train step=1803

                                                       
{'loss': -0.0189, 'grad_norm': 7.020390510559082, 'learning_rate': 4.539393939393939e-06, 'num_tokens': 4064817.0, 'completions/mean_length': 34.5, 'completions/min_length': 33.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9750309586524963, 'rewards/meter/std': 0.01676092855632305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9750309586524963, 'rewards/total_composite/std': 0.01676092855632305, 'reward': 0.9750309586524963, 'reward_std': 0.01676092855632305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06769955903291702, 'sampling/sampling_logp_difference/max': 1.4559955596923828, 'sampling/importance_sampling_ratio/min': 0.2331681251525879, 'sampling/importance_sampling_ratio/mean': 1.008634090423584, 'sampling/importance_sampling_ratio/max': 1.694462537765503, 'entropy': 0.6431399546563625, 'clip_ratio/low_mean': 0.022176598431542516, 'clip_ratio/low_min': 0.022176598431542516, 'clip_ratio/high_mean': 0.02186147216707468, 'clip_ratio/high_max': 0.02186147216707468, 'clip_ratio/region_mean': 0.044038070598617196, 'reward_total_mean': 0.9750309586524963, 'reward_meter_mean': 0.9750309586524963, 'reward_meter_std': 0.01676092855632305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9750309586524963, 'reward_total_composite_std': 0.01676092855632305, 'epoch': 0.07}

 55%|█████▍    | 1803/3300 [2:16:42<6:24:15, 15.40s/it]INFO 04-12 00:53:33 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1804/3300 [2:16:47<5:05:21, 12.25s/it]2026-04-12 00:53:38,265 | INFO | train_grpo_train | metrics_logged mode=train step=1804

                                                       
{'loss': 0.0014, 'grad_norm': 2.0322115421295166, 'learning_rate': 4.5363636363636364e-06, 'num_tokens': 4066459.0, 'completions/mean_length': 57.25, 'completions/min_length': 57.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.25, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9932329654693604, 'rewards/meter/std': 0.0008318190230056643, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9932329654693604, 'rewards/total_composite/std': 0.0008318190230056643, 'reward': 0.9932329654693604, 'reward_std': 0.0008318254840560257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026665696874260902, 'sampling/sampling_logp_difference/max': 1.751237392425537, 'sampling/importance_sampling_ratio/min': 0.4126149117946625, 'sampling/importance_sampling_ratio/mean': 1.0055011510849, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13894780911505222, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010927102295681834, 'clip_ratio/high_max': 0.010927102295681834, 'clip_ratio/region_mean': 0.010927102295681834, 'reward_total_mean': 0.9932329654693604, 'reward_meter_mean': 0.9932329654693604, 'reward_meter_std': 0.0008318190230056643, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9932329654693604, 'reward_total_composite_std': 0.0008318190230056643, 'epoch': 0.07}

 55%|█████▍    | 1804/3300 [2:16:47<5:05:21, 12.25s/it]INFO 04-12 00:53:38 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1805/3300 [2:16:51<4:06:21,  9.89s/it]2026-04-12 00:53:42,657 | INFO | train_grpo_train | metrics_logged mode=train step=1805

                                                       
{'loss': -0.0002, 'grad_norm': 2.5561840534210205, 'learning_rate': 4.533333333333334e-06, 'num_tokens': 4068131.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9917310476303101, 'rewards/meter/std': 0.014378287829458714, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9917310476303101, 'rewards/total_composite/std': 0.014378287829458714, 'reward': 0.9917310476303101, 'reward_std': 0.014378308318555355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010517938062548637, 'sampling/sampling_logp_difference/max': 0.489483118057251, 'sampling/importance_sampling_ratio/min': 0.612943172454834, 'sampling/importance_sampling_ratio/mean': 1.0059313774108887, 'sampling/importance_sampling_ratio/max': 1.4607477188110352, 'entropy': 0.09062223881483078, 'clip_ratio/low_mean': 0.0019841270986944437, 'clip_ratio/low_min': 0.0019841270986944437, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0019841270986944437, 'reward_total_mean': 0.9917310476303101, 'reward_meter_mean': 0.9917310476303101, 'reward_meter_std': 0.014378287829458714, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9917310476303101, 'reward_total_composite_std': 0.014378287829458714, 'epoch': 0.07}

 55%|█████▍    | 1805/3300 [2:16:51<4:06:21,  9.89s/it]INFO 04-12 00:53:42 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1806/3300 [2:16:56<3:28:31,  8.37s/it]2026-04-12 00:53:47,490 | INFO | train_grpo_train | metrics_logged mode=train step=1806

                                                       
{'loss': 0.0586, 'grad_norm': 5.9204182624816895, 'learning_rate': 4.53030303030303e-06, 'num_tokens': 4069918.0, 'completions/mean_length': 80.375, 'completions/min_length': 74.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.774915874004364, 'rewards/meter/std': 0.3431837856769562, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6509301066398621, 'rewards/total_composite/std': 0.42339858412742615, 'reward': 0.6509301066398621, 'reward_std': 0.42339858412742615, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09225737303495407, 'sampling/sampling_logp_difference/max': 3.772925615310669, 'sampling/importance_sampling_ratio/min': 0.02298472262918949, 'sampling/importance_sampling_ratio/mean': 1.022961139678955, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9013412185013294, 'clip_ratio/low_mean': 0.013943745056167245, 'clip_ratio/low_min': 0.013943745056167245, 'clip_ratio/high_mean': 0.03447378391865641, 'clip_ratio/high_max': 0.03447378391865641, 'clip_ratio/region_mean': 0.048417528974823654, 'reward_total_mean': 0.6509301066398621, 'reward_meter_mean': 0.774915874004364, 'reward_meter_std': 0.3431837856769562, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6509301066398621, 'reward_total_composite_std': 0.42339858412742615, 'epoch': 0.07}

 55%|█████▍    | 1806/3300 [2:16:56<3:28:31,  8.37s/it]INFO 04-12 00:53:47 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1807/3300 [2:17:00<2:59:06,  7.20s/it]2026-04-12 00:53:51,947 | INFO | train_grpo_train | metrics_logged mode=train step=1807

                                                       
{'loss': 0.0059, 'grad_norm': 3.1104016304016113, 'learning_rate': 4.527272727272727e-06, 'num_tokens': 4071641.0, 'completions/mean_length': 57.375, 'completions/min_length': 57.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.375, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9939553737640381, 'rewards/meter/std': 0.0008149455534294248, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9939553737640381, 'rewards/total_composite/std': 0.0008149455534294248, 'reward': 0.9939553737640381, 'reward_std': 0.0008149410132318735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021058987826108932, 'sampling/sampling_logp_difference/max': 0.8907175064086914, 'sampling/importance_sampling_ratio/min': 0.41036123037338257, 'sampling/importance_sampling_ratio/mean': 1.0032254457473755, 'sampling/importance_sampling_ratio/max': 1.5543711185455322, 'entropy': 0.13474871311336756, 'clip_ratio/low_mean': 0.008621971355751157, 'clip_ratio/low_min': 0.008621971355751157, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.008621971355751157, 'reward_total_mean': 0.9939553737640381, 'reward_meter_mean': 0.9939553737640381, 'reward_meter_std': 0.0008149455534294248, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9939553737640381, 'reward_total_composite_std': 0.0008149455534294248, 'epoch': 0.07}

 55%|█████▍    | 1807/3300 [2:17:00<2:59:06,  7.20s/it]INFO 04-12 00:53:52 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1808/3300 [2:17:06<2:48:21,  6.77s/it]2026-04-12 00:53:57,722 | INFO | train_grpo_train | metrics_logged mode=train step=1808

                                                       
{'loss': -0.0419, 'grad_norm': 4.073712348937988, 'learning_rate': 4.524242424242425e-06, 'num_tokens': 4074374.0, 'completions/mean_length': 157.625, 'completions/min_length': 132.0, 'completions/max_length': 169.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.625, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 169.0, 'rewards/meter/mean': 0.8494125604629517, 'rewards/meter/std': 0.2693995535373688, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7510942220687866, 'rewards/total_composite/std': 0.41580459475517273, 'reward': 0.7510942220687866, 'reward_std': 0.41580459475517273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0892598107457161, 'sampling/sampling_logp_difference/max': 1.8835735321044922, 'sampling/importance_sampling_ratio/min': 0.15204578638076782, 'sampling/importance_sampling_ratio/mean': 1.0168873071670532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9014198370277882, 'clip_ratio/low_mean': 0.01045380299910903, 'clip_ratio/low_min': 0.01045380299910903, 'clip_ratio/high_mean': 0.039301327895373106, 'clip_ratio/high_max': 0.039301327895373106, 'clip_ratio/region_mean': 0.049755130894482136, 'reward_total_mean': 0.7510942220687866, 'reward_meter_mean': 0.8494125604629517, 'reward_meter_std': 0.2693995535373688, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7510942220687866, 'reward_total_composite_std': 0.41580459475517273, 'epoch': 0.07}

 55%|█████▍    | 1808/3300 [2:17:06<2:48:21,  6.77s/it]INFO 04-12 00:53:57 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1809/3300 [2:17:12<2:41:04,  6.48s/it]2026-04-12 00:54:03,520 | INFO | train_grpo_train | metrics_logged mode=train step=1809

                                                       
{'loss': 0.0062, 'grad_norm': 4.39199161529541, 'learning_rate': 4.521212121212122e-06, 'num_tokens': 4077130.0, 'completions/mean_length': 160.5, 'completions/min_length': 152.0, 'completions/max_length': 172.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.5, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 172.0, 'rewards/meter/mean': 0.9320476055145264, 'rewards/meter/std': 0.11087671667337418, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9022413492202759, 'rewards/total_composite/std': 0.13386203348636627, 'reward': 0.9022413492202759, 'reward_std': 0.13386201858520508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10231172293424606, 'sampling/sampling_logp_difference/max': 1.5400352478027344, 'sampling/importance_sampling_ratio/min': 0.21437355875968933, 'sampling/importance_sampling_ratio/mean': 1.0246384143829346, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1915706768631935, 'clip_ratio/low_mean': 0.017330652568489313, 'clip_ratio/low_min': 0.017330652568489313, 'clip_ratio/high_mean': 0.044713106006383896, 'clip_ratio/high_max': 0.044713106006383896, 'clip_ratio/region_mean': 0.06204375857487321, 'reward_total_mean': 0.9022413492202759, 'reward_meter_mean': 0.9320476055145264, 'reward_meter_std': 0.11087671667337418, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9022413492202759, 'reward_total_composite_std': 0.13386203348636627, 'epoch': 0.07}

 55%|█████▍    | 1809/3300 [2:17:12<2:41:04,  6.48s/it]INFO 04-12 00:54:03 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1810/3300 [2:17:17<2:33:05,  6.16s/it]2026-04-12 00:54:08,952 | INFO | train_grpo_train | metrics_logged mode=train step=1810

                                                       
{'loss': -0.0001, 'grad_norm': 2.1702988147735596, 'learning_rate': 4.518181818181819e-06, 'num_tokens': 4079618.0, 'completions/mean_length': 136.0, 'completions/min_length': 132.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.0, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.8782405853271484, 'rewards/meter/std': 0.031040111556649208, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8035714030265808, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.7060449123382568, 'rewards/total_composite/std': 0.07351230084896088, 'reward': 0.7060449123382568, 'reward_std': 0.07351230829954147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022378606721758842, 'sampling/sampling_logp_difference/max': 1.1933460235595703, 'sampling/importance_sampling_ratio/min': 0.30320504307746887, 'sampling/importance_sampling_ratio/mean': 1.0071299076080322, 'sampling/importance_sampling_ratio/max': 1.6531912088394165, 'entropy': 0.24091620557010174, 'clip_ratio/low_mean': 0.006453010253608227, 'clip_ratio/low_min': 0.006453010253608227, 'clip_ratio/high_mean': 0.012699689657893032, 'clip_ratio/high_max': 0.012699689657893032, 'clip_ratio/region_mean': 0.01915269991150126, 'reward_total_mean': 0.7060449123382568, 'reward_meter_mean': 0.8782405853271484, 'reward_meter_std': 0.031040111556649208, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8035714030265808, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.7060449123382568, 'reward_total_composite_std': 0.07351230084896088, 'epoch': 0.07}

 55%|█████▍    | 1810/3300 [2:17:17<2:33:05,  6.16s/it]INFO 04-12 00:54:09 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1811/3300 [2:17:22<2:25:01,  5.84s/it]2026-04-12 00:54:14,044 | INFO | train_grpo_train | metrics_logged mode=train step=1811

                                                       
{'loss': 0.0067, 'grad_norm': 4.185992240905762, 'learning_rate': 4.5151515151515155e-06, 'num_tokens': 4081773.0, 'completions/mean_length': 104.375, 'completions/min_length': 99.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.375, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.8305845260620117, 'rewards/meter/std': 0.25914400815963745, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8305845260620117, 'rewards/total_composite/std': 0.25914400815963745, 'reward': 0.8305845260620117, 'reward_std': 0.25914397835731506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058025941252708435, 'sampling/sampling_logp_difference/max': 1.5509662628173828, 'sampling/importance_sampling_ratio/min': 0.21204298734664917, 'sampling/importance_sampling_ratio/mean': 1.0125762224197388, 'sampling/importance_sampling_ratio/max': 1.7020655870437622, 'entropy': 0.5101095288991928, 'clip_ratio/low_mean': 0.006021729204803705, 'clip_ratio/low_min': 0.006021729204803705, 'clip_ratio/high_mean': 0.037365528754889965, 'clip_ratio/high_max': 0.037365528754889965, 'clip_ratio/region_mean': 0.04338725795969367, 'reward_total_mean': 0.8305845260620117, 'reward_meter_mean': 0.8305845260620117, 'reward_meter_std': 0.25914400815963745, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8305845260620117, 'reward_total_composite_std': 0.25914400815963745, 'epoch': 0.07}

 55%|█████▍    | 1811/3300 [2:17:23<2:25:01,  5.84s/it]INFO 04-12 00:54:14 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1812/3300 [2:17:27<2:16:05,  5.49s/it]2026-04-12 00:54:18,703 | INFO | train_grpo_train | metrics_logged mode=train step=1812

                                                       
{'loss': 0.0156, 'grad_norm': 7.030357360839844, 'learning_rate': 4.512121212121213e-06, 'num_tokens': 4083458.0, 'completions/mean_length': 62.625, 'completions/min_length': 59.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.7619152069091797, 'rewards/meter/std': 0.4342895746231079, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7619152069091797, 'rewards/total_composite/std': 0.4342895746231079, 'reward': 0.7619152069091797, 'reward_std': 0.4342895746231079, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024434061720967293, 'sampling/sampling_logp_difference/max': 1.121795654296875, 'sampling/importance_sampling_ratio/min': 0.3256944417953491, 'sampling/importance_sampling_ratio/mean': 1.0049937963485718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13391203712671995, 'clip_ratio/low_mean': 0.003937252098694444, 'clip_ratio/low_min': 0.003937252098694444, 'clip_ratio/high_mean': 0.009920635493472219, 'clip_ratio/high_max': 0.009920635493472219, 'clip_ratio/region_mean': 0.013857887592166662, 'reward_total_mean': 0.7619152069091797, 'reward_meter_mean': 0.7619152069091797, 'reward_meter_std': 0.4342895746231079, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7619152069091797, 'reward_total_composite_std': 0.4342895746231079, 'epoch': 0.07}

 55%|█████▍    | 1812/3300 [2:17:27<2:16:05,  5.49s/it]INFO 04-12 00:54:18 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1813/3300 [2:17:33<2:20:16,  5.66s/it]2026-04-12 00:54:24,766 | INFO | train_grpo_train | metrics_logged mode=train step=1813

                                                       
{'loss': -0.0181, 'grad_norm': 3.172938346862793, 'learning_rate': 4.50909090909091e-06, 'num_tokens': 4086534.0, 'completions/mean_length': 176.5, 'completions/min_length': 160.0, 'completions/max_length': 195.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 176.5, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 195.0, 'rewards/meter/mean': 0.9555693864822388, 'rewards/meter/std': 0.09552352130413055, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8140916228294373, 'rewards/total_composite/std': 0.132888063788414, 'reward': 0.8140916228294373, 'reward_std': 0.132888063788414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07848730683326721, 'sampling/sampling_logp_difference/max': 1.3485078811645508, 'sampling/importance_sampling_ratio/min': 0.2596273720264435, 'sampling/importance_sampling_ratio/mean': 1.0195351839065552, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7889967784285545, 'clip_ratio/low_mean': 0.030510480049997568, 'clip_ratio/low_min': 0.030510480049997568, 'clip_ratio/high_mean': 0.015669515822082758, 'clip_ratio/high_max': 0.015669515822082758, 'clip_ratio/region_mean': 0.046179995872080326, 'reward_total_mean': 0.8140916228294373, 'reward_meter_mean': 0.9555693864822388, 'reward_meter_std': 0.09552352130413055, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8140916228294373, 'reward_total_composite_std': 0.132888063788414, 'epoch': 0.07}

 55%|█████▍    | 1813/3300 [2:17:33<2:20:16,  5.66s/it]INFO 04-12 00:54:25 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1814/3300 [2:17:38<2:14:53,  5.45s/it]2026-04-12 00:54:29,708 | INFO | train_grpo_train | metrics_logged mode=train step=1814

                                                       
{'loss': -0.0068, 'grad_norm': 2.741095781326294, 'learning_rate': 4.5060606060606065e-06, 'num_tokens': 4088604.0, 'completions/mean_length': 95.75, 'completions/min_length': 93.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.75, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9976317882537842, 'rewards/meter/std': 0.0016386967618018389, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976317882537842, 'rewards/total_composite/std': 0.0016386967618018389, 'reward': 0.9976317882537842, 'reward_std': 0.0016386950155720115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006327650509774685, 'sampling/sampling_logp_difference/max': 0.28154802322387695, 'sampling/importance_sampling_ratio/min': 0.8114128112792969, 'sampling/importance_sampling_ratio/mean': 1.0037214756011963, 'sampling/importance_sampling_ratio/max': 1.3251795768737793, 'entropy': 0.04789540730416775, 'clip_ratio/low_mean': 0.0013440860202535987, 'clip_ratio/low_min': 0.0013440860202535987, 'clip_ratio/high_mean': 0.0025775935500860214, 'clip_ratio/high_max': 0.0025775935500860214, 'clip_ratio/region_mean': 0.00392167957033962, 'reward_total_mean': 0.9976317882537842, 'reward_meter_mean': 0.9976317882537842, 'reward_meter_std': 0.0016386967618018389, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976317882537842, 'reward_total_composite_std': 0.0016386967618018389, 'epoch': 0.07}

 55%|█████▍    | 1814/3300 [2:17:38<2:14:53,  5.45s/it]INFO 04-12 00:54:29 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1815/3300 [2:17:44<2:18:02,  5.58s/it]2026-04-12 00:54:35,595 | INFO | train_grpo_train | metrics_logged mode=train step=1815

                                                       
{'loss': 0.0014, 'grad_norm': 2.079590320587158, 'learning_rate': 4.503030303030304e-06, 'num_tokens': 4091580.0, 'completions/mean_length': 175.0, 'completions/min_length': 163.0, 'completions/max_length': 185.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 175.0, 'completions/min_terminated_length': 163.0, 'completions/max_terminated_length': 185.0, 'rewards/meter/mean': 0.9942752122879028, 'rewards/meter/std': 0.0037329280748963356, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8152778148651123, 'rewards/repeat_penalty/std': 0.05002203211188316, 'rewards/total_composite/mean': 0.6754500865936279, 'rewards/total_composite/std': 0.04033276066184044, 'reward': 0.6754500865936279, 'reward_std': 0.04033275321125984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019572407007217407, 'sampling/sampling_logp_difference/max': 2.9577598571777344, 'sampling/importance_sampling_ratio/min': 0.05193512886762619, 'sampling/importance_sampling_ratio/mean': 1.0022673606872559, 'sampling/importance_sampling_ratio/max': 1.8784008026123047, 'entropy': 0.12415077816694975, 'clip_ratio/low_mean': 0.005773193319328129, 'clip_ratio/low_min': 0.005773193319328129, 'clip_ratio/high_mean': 0.004235348082147539, 'clip_ratio/high_max': 0.004235348082147539, 'clip_ratio/region_mean': 0.010008541401475668, 'reward_total_mean': 0.6754500865936279, 'reward_meter_mean': 0.9942752122879028, 'reward_meter_std': 0.0037329280748963356, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8152778148651123, 'reward_repeat_penalty_std': 0.05002203211188316, 'reward_total_composite_mean': 0.6754500865936279, 'reward_total_composite_std': 0.04033276066184044, 'epoch': 0.07}

 55%|█████▌    | 1815/3300 [2:17:44<2:18:02,  5.58s/it]INFO 04-12 00:54:35 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1816/3300 [2:17:49<2:10:51,  5.29s/it]2026-04-12 00:54:40,216 | INFO | train_grpo_train | metrics_logged mode=train step=1816

                                                       
{'loss': 0.3243, 'grad_norm': 7.065499782562256, 'learning_rate': 4.5e-06, 'num_tokens': 4093095.0, 'completions/mean_length': 51.375, 'completions/min_length': 36.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9962137937545776, 'rewards/meter/std': 0.0032140789553523064, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6230374574661255, 'rewards/total_composite/std': 0.5159306526184082, 'reward': 0.6230374574661255, 'reward_std': 0.5159306526184082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08005251735448837, 'sampling/sampling_logp_difference/max': 1.0488758087158203, 'sampling/importance_sampling_ratio/min': 0.3503313660621643, 'sampling/importance_sampling_ratio/mean': 1.0141249895095825, 'sampling/importance_sampling_ratio/max': 1.9115855693817139, 'entropy': 0.709318995475769, 'clip_ratio/low_mean': 0.022342659067362547, 'clip_ratio/low_min': 0.022342659067362547, 'clip_ratio/high_mean': 0.03601037664338946, 'clip_ratio/high_max': 0.03601037664338946, 'clip_ratio/region_mean': 0.05835303571075201, 'reward_total_mean': 0.6230374574661255, 'reward_meter_mean': 0.9962137937545776, 'reward_meter_std': 0.0032140789553523064, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6230374574661255, 'reward_total_composite_std': 0.5159306526184082, 'epoch': 0.07}

 55%|█████▌    | 1816/3300 [2:17:49<2:10:51,  5.29s/it]INFO 04-12 00:54:40 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1817/3300 [2:17:53<2:06:11,  5.11s/it]2026-04-12 00:54:44,888 | INFO | train_grpo_train | metrics_logged mode=train step=1817

                                                       
{'loss': 0.0293, 'grad_norm': 9.501838684082031, 'learning_rate': 4.496969696969697e-06, 'num_tokens': 4094892.0, 'completions/mean_length': 72.625, 'completions/min_length': 68.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.8915443420410156, 'rewards/meter/std': 0.21483440697193146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8915443420410156, 'rewards/total_composite/std': 0.21483440697193146, 'reward': 0.8915443420410156, 'reward_std': 0.21483437716960907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08924838155508041, 'sampling/sampling_logp_difference/max': 2.1535000801086426, 'sampling/importance_sampling_ratio/min': 0.11607716232538223, 'sampling/importance_sampling_ratio/mean': 1.027990460395813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8766130320727825, 'clip_ratio/low_mean': 0.011928289197385311, 'clip_ratio/low_min': 0.011928289197385311, 'clip_ratio/high_mean': 0.050623598508536816, 'clip_ratio/high_max': 0.050623598508536816, 'clip_ratio/region_mean': 0.06255188770592213, 'reward_total_mean': 0.8915443420410156, 'reward_meter_mean': 0.8915443420410156, 'reward_meter_std': 0.21483440697193146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8915443420410156, 'reward_total_composite_std': 0.21483440697193146, 'epoch': 0.07}

 55%|█████▌    | 1817/3300 [2:17:53<2:06:11,  5.11s/it]INFO 04-12 00:54:45 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1818/3300 [2:17:58<2:04:15,  5.03s/it]2026-04-12 00:54:49,746 | INFO | train_grpo_train | metrics_logged mode=train step=1818

                                                       
{'loss': -0.0031, 'grad_norm': 0.42083650827407837, 'learning_rate': 4.493939393939395e-06, 'num_tokens': 4097204.0, 'completions/mean_length': 102.0, 'completions/min_length': 100.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.0, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9974691271781921, 'rewards/meter/std': 6.317263614619151e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974691271781921, 'rewards/total_composite/std': 6.317263614619151e-05, 'reward': 0.9974691271781921, 'reward_std': 6.316715735010803e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009578624740242958, 'sampling/sampling_logp_difference/max': 0.9488290548324585, 'sampling/importance_sampling_ratio/min': 0.3871941566467285, 'sampling/importance_sampling_ratio/mean': 1.001711130142212, 'sampling/importance_sampling_ratio/max': 1.6696717739105225, 'entropy': 0.04315004916861653, 'clip_ratio/low_mean': 0.006177184404805303, 'clip_ratio/low_min': 0.006177184404805303, 'clip_ratio/high_mean': 0.0036526747280731797, 'clip_ratio/high_max': 0.0036526747280731797, 'clip_ratio/region_mean': 0.009829859132878482, 'reward_total_mean': 0.9974691271781921, 'reward_meter_mean': 0.9974691271781921, 'reward_meter_std': 6.317263614619151e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974691271781921, 'reward_total_composite_std': 6.317263614619151e-05, 'epoch': 0.07}

 55%|█████▌    | 1818/3300 [2:17:58<2:04:15,  5.03s/it]INFO 04-12 00:54:49 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1819/3300 [2:18:03<2:03:25,  5.00s/it]2026-04-12 00:54:54,675 | INFO | train_grpo_train | metrics_logged mode=train step=1819

                                                       
{'loss': -0.0058, 'grad_norm': 2.0940115451812744, 'learning_rate': 4.490909090909091e-06, 'num_tokens': 4099258.0, 'completions/mean_length': 93.75, 'completions/min_length': 92.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.75, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.8540617823600769, 'rewards/meter/std': 0.2617916762828827, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8056372404098511, 'rewards/total_composite/std': 0.25278210639953613, 'reward': 0.8056372404098511, 'reward_std': 0.25278210639953613, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029167180880904198, 'sampling/sampling_logp_difference/max': 0.9276943206787109, 'sampling/importance_sampling_ratio/min': 0.3954644799232483, 'sampling/importance_sampling_ratio/mean': 1.0129790306091309, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23089196532964706, 'clip_ratio/low_mean': 0.006778868613764644, 'clip_ratio/low_min': 0.006778868613764644, 'clip_ratio/high_mean': 0.020909853279590607, 'clip_ratio/high_max': 0.020909853279590607, 'clip_ratio/region_mean': 0.02768872189335525, 'reward_total_mean': 0.8056372404098511, 'reward_meter_mean': 0.8540617823600769, 'reward_meter_std': 0.2617916762828827, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8056372404098511, 'reward_total_composite_std': 0.25278210639953613, 'epoch': 0.07}

 55%|█████▌    | 1819/3300 [2:18:03<2:03:25,  5.00s/it]INFO 04-12 00:54:54 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1820/3300 [2:18:08<1:59:26,  4.84s/it]2026-04-12 00:54:59,149 | INFO | train_grpo_train | metrics_logged mode=train step=1820

                                                       
{'loss': 0.0047, 'grad_norm': 2.0801198482513428, 'learning_rate': 4.487878787878788e-06, 'num_tokens': 4101001.0, 'completions/mean_length': 57.875, 'completions/min_length': 57.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9942807555198669, 'rewards/meter/std': 0.0004856908926740289, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9942807555198669, 'rewards/total_composite/std': 0.0004856908926740289, 'reward': 0.9942807555198669, 'reward_std': 0.00048569359933026135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017049286514520645, 'sampling/sampling_logp_difference/max': 1.2708497047424316, 'sampling/importance_sampling_ratio/min': 0.2805930972099304, 'sampling/importance_sampling_ratio/mean': 1.0053722858428955, 'sampling/importance_sampling_ratio/max': 1.5113404989242554, 'entropy': 0.10346860811114311, 'clip_ratio/low_mean': 0.006392460549250245, 'clip_ratio/low_min': 0.006392460549250245, 'clip_ratio/high_mean': 0.008658499689772725, 'clip_ratio/high_max': 0.008658499689772725, 'clip_ratio/region_mean': 0.01505096023902297, 'reward_total_mean': 0.9942807555198669, 'reward_meter_mean': 0.9942807555198669, 'reward_meter_std': 0.0004856908926740289, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9942807555198669, 'reward_total_composite_std': 0.0004856908926740289, 'epoch': 0.07}

 55%|█████▌    | 1820/3300 [2:18:08<1:59:26,  4.84s/it]INFO 04-12 00:54:59 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1821/3300 [2:18:12<1:57:52,  4.78s/it]2026-04-12 00:55:03,793 | INFO | train_grpo_train | metrics_logged mode=train step=1821

                                                       
{'loss': 0.0157, 'grad_norm': 5.139247417449951, 'learning_rate': 4.4848484848484855e-06, 'num_tokens': 4102797.0, 'completions/mean_length': 74.5, 'completions/min_length': 73.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9978154897689819, 'rewards/meter/std': 0.0017045722343027592, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978154897689819, 'rewards/total_composite/std': 0.0017045722343027592, 'reward': 0.9978154897689819, 'reward_std': 0.0017045725835487247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.055102791637182236, 'sampling/sampling_logp_difference/max': 1.0592074394226074, 'sampling/importance_sampling_ratio/min': 0.346730500459671, 'sampling/importance_sampling_ratio/mean': 1.015410304069519, 'sampling/importance_sampling_ratio/max': 1.806069016456604, 'entropy': 0.6388323791325092, 'clip_ratio/low_mean': 0.016557018272578716, 'clip_ratio/low_min': 0.016557018272578716, 'clip_ratio/high_mean': 0.040383500047028065, 'clip_ratio/high_max': 0.040383500047028065, 'clip_ratio/region_mean': 0.05694051831960678, 'reward_total_mean': 0.9978154897689819, 'reward_meter_mean': 0.9978154897689819, 'reward_meter_std': 0.0017045722343027592, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978154897689819, 'reward_total_composite_std': 0.0017045722343027592, 'epoch': 0.07}

 55%|█████▌    | 1821/3300 [2:18:12<1:57:52,  4.78s/it]INFO 04-12 00:55:04 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1822/3300 [2:18:17<1:56:31,  4.73s/it]2026-04-12 00:55:08,402 | INFO | train_grpo_train | metrics_logged mode=train step=1822

                                                       
{'loss': -0.0115, 'grad_norm': 5.289615154266357, 'learning_rate': 4.481818181818182e-06, 'num_tokens': 4104667.0, 'completions/mean_length': 73.75, 'completions/min_length': 70.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.75, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9970654249191284, 'rewards/meter/std': 0.0017285288777202368, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970654249191284, 'rewards/total_composite/std': 0.0017285288777202368, 'reward': 0.9970654249191284, 'reward_std': 0.0017285366775467992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05023278295993805, 'sampling/sampling_logp_difference/max': 1.2027688026428223, 'sampling/importance_sampling_ratio/min': 0.30036142468452454, 'sampling/importance_sampling_ratio/mean': 1.0059843063354492, 'sampling/importance_sampling_ratio/max': 1.728150725364685, 'entropy': 0.5206275209784508, 'clip_ratio/low_mean': 0.020897203590720892, 'clip_ratio/low_min': 0.020897203590720892, 'clip_ratio/high_mean': 0.018180417479015887, 'clip_ratio/high_max': 0.018180417479015887, 'clip_ratio/region_mean': 0.03907762106973678, 'reward_total_mean': 0.9970654249191284, 'reward_meter_mean': 0.9970654249191284, 'reward_meter_std': 0.0017285288777202368, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970654249191284, 'reward_total_composite_std': 0.0017285288777202368, 'epoch': 0.07}

 55%|█████▌    | 1822/3300 [2:18:17<1:56:31,  4.73s/it]INFO 04-12 00:55:08 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1823/3300 [2:18:21<1:54:45,  4.66s/it]2026-04-12 00:55:12,903 | INFO | train_grpo_train | metrics_logged mode=train step=1823

                                                       
{'loss': -0.0029, 'grad_norm': 2.6084089279174805, 'learning_rate': 4.478787878787879e-06, 'num_tokens': 4106511.0, 'completions/mean_length': 64.5, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9984657764434814, 'rewards/meter/std': 0.0002891587500926107, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984657764434814, 'rewards/total_composite/std': 0.0002891587500926107, 'reward': 0.9984657764434814, 'reward_std': 0.0002891614567488432, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007291331887245178, 'sampling/sampling_logp_difference/max': 0.45546579360961914, 'sampling/importance_sampling_ratio/min': 0.6341525316238403, 'sampling/importance_sampling_ratio/mean': 0.9986276626586914, 'sampling/importance_sampling_ratio/max': 1.241689920425415, 'entropy': 0.033252993831411004, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007692307699471712, 'clip_ratio/high_max': 0.007692307699471712, 'clip_ratio/region_mean': 0.007692307699471712, 'reward_total_mean': 0.9984657764434814, 'reward_meter_mean': 0.9984657764434814, 'reward_meter_std': 0.0002891587500926107, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984657764434814, 'reward_total_composite_std': 0.0002891587500926107, 'epoch': 0.07}

 55%|█████▌    | 1823/3300 [2:18:21<1:54:45,  4.66s/it]INFO 04-12 00:55:13 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1824/3300 [2:18:26<1:52:20,  4.57s/it]2026-04-12 00:55:17,248 | INFO | train_grpo_train | metrics_logged mode=train step=1824

                                                       
{'loss': -0.0206, 'grad_norm': 11.332904815673828, 'learning_rate': 4.4757575757575765e-06, 'num_tokens': 4108093.0, 'completions/mean_length': 37.75, 'completions/min_length': 35.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9941079616546631, 'rewards/meter/std': 0.012935124337673187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9941079616546631, 'rewards/total_composite/std': 0.012935124337673187, 'reward': 0.9941079616546631, 'reward_std': 0.012935123406350613, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0540194995701313, 'sampling/sampling_logp_difference/max': 1.3527865409851074, 'sampling/importance_sampling_ratio/min': 0.25851887464523315, 'sampling/importance_sampling_ratio/mean': 1.0034513473510742, 'sampling/importance_sampling_ratio/max': 1.575133204460144, 'entropy': 0.39845724031329155, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.044933649245649576, 'clip_ratio/high_max': 0.044933649245649576, 'clip_ratio/region_mean': 0.04850507783703506, 'reward_total_mean': 0.9941079616546631, 'reward_meter_mean': 0.9941079616546631, 'reward_meter_std': 0.012935124337673187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9941079616546631, 'reward_total_composite_std': 0.012935124337673187, 'epoch': 0.07}

 55%|█████▌    | 1824/3300 [2:18:26<1:52:20,  4.57s/it]INFO 04-12 00:55:17 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1825/3300 [2:18:31<1:55:03,  4.68s/it]2026-04-12 00:55:22,192 | INFO | train_grpo_train | metrics_logged mode=train step=1825

                                                       
{'loss': 0.0106, 'grad_norm': 4.047023296356201, 'learning_rate': 4.472727272727273e-06, 'num_tokens': 4110202.0, 'completions/mean_length': 94.625, 'completions/min_length': 90.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.625, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.587390661239624, 'rewards/meter/std': 0.44364675879478455, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.554187536239624, 'rewards/total_composite/std': 0.42821574211120605, 'reward': 0.554187536239624, 'reward_std': 0.42821574211120605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03696272522211075, 'sampling/sampling_logp_difference/max': 0.9904708862304688, 'sampling/importance_sampling_ratio/min': 0.37140175700187683, 'sampling/importance_sampling_ratio/mean': 1.0154355764389038, 'sampling/importance_sampling_ratio/max': 1.9479297399520874, 'entropy': 0.3725603432394564, 'clip_ratio/low_mean': 0.02371319057419896, 'clip_ratio/low_min': 0.02371319057419896, 'clip_ratio/high_mean': 0.0026315790601074696, 'clip_ratio/high_max': 0.0026315790601074696, 'clip_ratio/region_mean': 0.02634476963430643, 'reward_total_mean': 0.554187536239624, 'reward_meter_mean': 0.587390661239624, 'reward_meter_std': 0.44364675879478455, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.554187536239624, 'reward_total_composite_std': 0.42821574211120605, 'epoch': 0.07}

 55%|█████▌    | 1825/3300 [2:18:31<1:55:03,  4.68s/it]INFO 04-12 00:55:22 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1826/3300 [2:18:36<1:56:39,  4.75s/it]2026-04-12 00:55:27,100 | INFO | train_grpo_train | metrics_logged mode=train step=1826

                                                       
{'loss': 0.0078, 'grad_norm': 3.477172613143921, 'learning_rate': 4.46969696969697e-06, 'num_tokens': 4112208.0, 'completions/mean_length': 92.75, 'completions/min_length': 91.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.75, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9020949602127075, 'rewards/meter/std': 0.12145007401704788, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8535131216049194, 'rewards/total_composite/std': 0.12295859307050705, 'reward': 0.8535131216049194, 'reward_std': 0.12295857816934586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02582640014588833, 'sampling/sampling_logp_difference/max': 1.0430831909179688, 'sampling/importance_sampling_ratio/min': 0.3523665964603424, 'sampling/importance_sampling_ratio/mean': 1.0086493492126465, 'sampling/importance_sampling_ratio/max': 1.6056605577468872, 'entropy': 0.20322295650839806, 'clip_ratio/low_mean': 0.0066495382925495505, 'clip_ratio/low_min': 0.0066495382925495505, 'clip_ratio/high_mean': 0.004032257944345474, 'clip_ratio/high_max': 0.004032257944345474, 'clip_ratio/region_mean': 0.010681796236895025, 'reward_total_mean': 0.8535131216049194, 'reward_meter_mean': 0.9020949602127075, 'reward_meter_std': 0.12145007401704788, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8535131216049194, 'reward_total_composite_std': 0.12295859307050705, 'epoch': 0.07}

 55%|█████▌    | 1826/3300 [2:18:36<1:56:39,  4.75s/it]INFO 04-12 00:55:27 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1827/3300 [2:18:41<1:58:16,  4.82s/it]2026-04-12 00:55:32,079 | INFO | train_grpo_train | metrics_logged mode=train step=1827

                                                       
{'loss': -0.0104, 'grad_norm': 3.58325457572937, 'learning_rate': 4.4666666666666665e-06, 'num_tokens': 4114241.0, 'completions/mean_length': 100.125, 'completions/min_length': 96.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.7311835289001465, 'rewards/meter/std': 0.286437064409256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.7074779868125916, 'rewards/total_composite/std': 0.27346280217170715, 'reward': 0.7074779868125916, 'reward_std': 0.27346280217170715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04709630832076073, 'sampling/sampling_logp_difference/max': 0.8693370819091797, 'sampling/importance_sampling_ratio/min': 0.4192293882369995, 'sampling/importance_sampling_ratio/mean': 1.018310546875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37415438517928123, 'clip_ratio/low_mean': 0.015318386955186725, 'clip_ratio/low_min': 0.015318386955186725, 'clip_ratio/high_mean': 0.03203791263513267, 'clip_ratio/high_max': 0.03203791263513267, 'clip_ratio/region_mean': 0.047356299590319395, 'reward_total_mean': 0.7074779868125916, 'reward_meter_mean': 0.7311835289001465, 'reward_meter_std': 0.286437064409256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.7074779868125916, 'reward_total_composite_std': 0.27346280217170715, 'epoch': 0.07}

 55%|█████▌    | 1827/3300 [2:18:41<1:58:16,  4.82s/it]INFO 04-12 00:55:32 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1828/3300 [2:18:45<1:58:50,  4.84s/it]2026-04-12 00:55:36,984 | INFO | train_grpo_train | metrics_logged mode=train step=1828

                                                       
{'loss': 0.0007, 'grad_norm': 1.8939130306243896, 'learning_rate': 4.463636363636364e-06, 'num_tokens': 4116559.0, 'completions/mean_length': 101.75, 'completions/min_length': 99.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.75, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9971822500228882, 'rewards/meter/std': 0.0007533096941187978, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971822500228882, 'rewards/total_composite/std': 0.0007533096941187978, 'reward': 0.9971822500228882, 'reward_std': 0.0007533166790381074, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011065986938774586, 'sampling/sampling_logp_difference/max': 0.6848697662353516, 'sampling/importance_sampling_ratio/min': 0.5041558742523193, 'sampling/importance_sampling_ratio/mean': 1.002759337425232, 'sampling/importance_sampling_ratio/max': 1.46595299243927, 'entropy': 0.06690388498827815, 'clip_ratio/low_mean': 0.0012254902394488454, 'clip_ratio/low_min': 0.0012254902394488454, 'clip_ratio/high_mean': 0.007329145446419716, 'clip_ratio/high_max': 0.007329145446419716, 'clip_ratio/region_mean': 0.008554635685868561, 'reward_total_mean': 0.9971822500228882, 'reward_meter_mean': 0.9971822500228882, 'reward_meter_std': 0.0007533096941187978, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971822500228882, 'reward_total_composite_std': 0.0007533096941187978, 'epoch': 0.07}

 55%|█████▌    | 1828/3300 [2:18:45<1:58:50,  4.84s/it]INFO 04-12 00:55:37 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1829/3300 [2:18:51<2:07:31,  5.20s/it]2026-04-12 00:55:43,022 | INFO | train_grpo_train | metrics_logged mode=train step=1829

                                                       
{'loss': -0.0037, 'grad_norm': 6.232636451721191, 'learning_rate': 4.460606060606061e-06, 'num_tokens': 4119597.0, 'completions/mean_length': 185.75, 'completions/min_length': 183.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 185.75, 'completions/min_terminated_length': 183.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9841963052749634, 'rewards/meter/std': 0.033278584480285645, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7750000357627869, 'rewards/repeat_penalty/std': 0.0707106813788414, 'rewards/total_composite/mean': 0.6353453397750854, 'rewards/total_composite/std': 0.05924127250909805, 'reward': 0.6353453397750854, 'reward_std': 0.05924128741025925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0172612052410841, 'sampling/sampling_logp_difference/max': 1.5019760131835938, 'sampling/importance_sampling_ratio/min': 0.222689688205719, 'sampling/importance_sampling_ratio/mean': 1.0010361671447754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10350964730605483, 'clip_ratio/low_mean': 0.002019958512391895, 'clip_ratio/low_min': 0.002019958512391895, 'clip_ratio/high_mean': 0.0033174321288242936, 'clip_ratio/high_max': 0.0033174321288242936, 'clip_ratio/region_mean': 0.005337390641216189, 'reward_total_mean': 0.6353453397750854, 'reward_meter_mean': 0.9841963052749634, 'reward_meter_std': 0.033278584480285645, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7750000357627869, 'reward_repeat_penalty_std': 0.0707106813788414, 'reward_total_composite_mean': 0.6353453397750854, 'reward_total_composite_std': 0.05924127250909805, 'epoch': 0.07}

 55%|█████▌    | 1829/3300 [2:18:52<2:07:31,  5.20s/it]INFO 04-12 00:55:43 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1830/3300 [2:18:56<2:05:15,  5.11s/it]2026-04-12 00:55:47,960 | INFO | train_grpo_train | metrics_logged mode=train step=1830

                                                       
{'loss': 0.0001, 'grad_norm': 0.13791632652282715, 'learning_rate': 4.4575757575757575e-06, 'num_tokens': 4121701.0, 'completions/mean_length': 99.0, 'completions/min_length': 99.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.0, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9990502595901489, 'rewards/meter/std': 1.3007632333028596e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990502595901489, 'rewards/total_composite/std': 1.3007632333028596e-05, 'reward': 0.9990502595901489, 'reward_std': 1.300759322475642e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00393805792555213, 'sampling/sampling_logp_difference/max': 0.512671709060669, 'sampling/importance_sampling_ratio/min': 0.6270684599876404, 'sampling/importance_sampling_ratio/mean': 1.0018277168273926, 'sampling/importance_sampling_ratio/max': 1.6697462797164917, 'entropy': 0.020989181706681848, 'clip_ratio/low_mean': 0.0012626262614503503, 'clip_ratio/low_min': 0.0012626262614503503, 'clip_ratio/high_mean': 0.0012626262614503503, 'clip_ratio/high_max': 0.0012626262614503503, 'clip_ratio/region_mean': 0.0025252525229007006, 'reward_total_mean': 0.9990502595901489, 'reward_meter_mean': 0.9990502595901489, 'reward_meter_std': 1.3007632333028596e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990502595901489, 'reward_total_composite_std': 1.3007632333028596e-05, 'epoch': 0.07}

 55%|█████▌    | 1830/3300 [2:18:56<2:05:15,  5.11s/it]INFO 04-12 00:55:48 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1831/3300 [2:19:02<2:07:29,  5.21s/it]2026-04-12 00:55:53,354 | INFO | train_grpo_train | metrics_logged mode=train step=1831

                                                       
{'loss': 0.0079, 'grad_norm': 3.588621139526367, 'learning_rate': 4.454545454545455e-06, 'num_tokens': 4123721.0, 'completions/mean_length': 89.5, 'completions/min_length': 85.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.5, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.9949823021888733, 'rewards/meter/std': 0.0008774827001616359, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949823021888733, 'rewards/total_composite/std': 0.0008774827001616359, 'reward': 0.9949823021888733, 'reward_std': 0.0008774733869358897, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02186514623463154, 'sampling/sampling_logp_difference/max': 1.0048408508300781, 'sampling/importance_sampling_ratio/min': 0.3661029040813446, 'sampling/importance_sampling_ratio/mean': 1.007697343826294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14992150850594044, 'clip_ratio/low_mean': 0.010989011265337467, 'clip_ratio/low_min': 0.010989011265337467, 'clip_ratio/high_mean': 0.0055555556900799274, 'clip_ratio/high_max': 0.0055555556900799274, 'clip_ratio/region_mean': 0.016544566955417395, 'reward_total_mean': 0.9949823021888733, 'reward_meter_mean': 0.9949823021888733, 'reward_meter_std': 0.0008774827001616359, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949823021888733, 'reward_total_composite_std': 0.0008774827001616359, 'epoch': 0.07}

 55%|█████▌    | 1831/3300 [2:19:02<2:07:29,  5.21s/it]INFO 04-12 00:55:53 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1832/3300 [2:19:07<2:03:52,  5.06s/it]2026-04-12 00:55:58,082 | INFO | train_grpo_train | metrics_logged mode=train step=1832

                                                       
{'loss': 0.0014, 'grad_norm': 1.7238988876342773, 'learning_rate': 4.451515151515152e-06, 'num_tokens': 4125807.0, 'completions/mean_length': 89.75, 'completions/min_length': 88.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.75, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9952454566955566, 'rewards/meter/std': 0.00010376512364018708, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952454566955566, 'rewards/total_composite/std': 0.00010376512364018708, 'reward': 0.9952454566955566, 'reward_std': 0.00010377082071499899, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014209388755261898, 'sampling/sampling_logp_difference/max': 1.3864455223083496, 'sampling/importance_sampling_ratio/min': 0.2499622255563736, 'sampling/importance_sampling_ratio/mean': 1.001192331314087, 'sampling/importance_sampling_ratio/max': 1.3277589082717896, 'entropy': 0.09856001799926162, 'clip_ratio/low_mean': 0.008428030530922115, 'clip_ratio/low_min': 0.008428030530922115, 'clip_ratio/high_mean': 0.0055555556900799274, 'clip_ratio/high_max': 0.0055555556900799274, 'clip_ratio/region_mean': 0.013983586221002042, 'reward_total_mean': 0.9952454566955566, 'reward_meter_mean': 0.9952454566955566, 'reward_meter_std': 0.00010376512364018708, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952454566955566, 'reward_total_composite_std': 0.00010376512364018708, 'epoch': 0.07}

 56%|█████▌    | 1832/3300 [2:19:07<2:03:52,  5.06s/it]INFO 04-12 00:55:58 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1833/3300 [2:19:11<2:00:59,  4.95s/it]2026-04-12 00:56:02,761 | INFO | train_grpo_train | metrics_logged mode=train step=1833

                                                       
{'loss': 0.0092, 'grad_norm': 3.797224283218384, 'learning_rate': 4.448484848484848e-06, 'num_tokens': 4127747.0, 'completions/mean_length': 75.5, 'completions/min_length': 72.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9976780414581299, 'rewards/meter/std': 0.0009906893828883767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976780414581299, 'rewards/total_composite/std': 0.0009906893828883767, 'reward': 0.9976780414581299, 'reward_std': 0.0009906899649649858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03965409845113754, 'sampling/sampling_logp_difference/max': 0.8967013359069824, 'sampling/importance_sampling_ratio/min': 0.4079129993915558, 'sampling/importance_sampling_ratio/mean': 1.0070687532424927, 'sampling/importance_sampling_ratio/max': 1.881462812423706, 'entropy': 0.36707402765750885, 'clip_ratio/low_mean': 0.018417607876472175, 'clip_ratio/low_min': 0.018417607876472175, 'clip_ratio/high_mean': 0.016408111667260528, 'clip_ratio/high_max': 0.016408111667260528, 'clip_ratio/region_mean': 0.0348257195437327, 'reward_total_mean': 0.9976780414581299, 'reward_meter_mean': 0.9976780414581299, 'reward_meter_std': 0.0009906893828883767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976780414581299, 'reward_total_composite_std': 0.0009906893828883767, 'epoch': 0.07}

 56%|█████▌    | 1833/3300 [2:19:11<2:00:59,  4.95s/it]INFO 04-12 00:56:02 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1834/3300 [2:19:17<2:03:41,  5.06s/it]2026-04-12 00:56:08,092 | INFO | train_grpo_train | metrics_logged mode=train step=1834

                                                       
{'loss': -0.0051, 'grad_norm': 3.744053840637207, 'learning_rate': 4.445454545454546e-06, 'num_tokens': 4130264.0, 'completions/mean_length': 133.625, 'completions/min_length': 130.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.625, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9959684014320374, 'rewards/meter/std': 0.0031451100949198008, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.9247440099716187, 'rewards/total_composite/std': 0.07493867725133896, 'reward': 0.9247440099716187, 'reward_std': 0.07493869215250015, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06190461292862892, 'sampling/sampling_logp_difference/max': 2.0104918479919434, 'sampling/importance_sampling_ratio/min': 0.1339227855205536, 'sampling/importance_sampling_ratio/mean': 1.0151046514511108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.501116044819355, 'clip_ratio/low_mean': 0.015092328016180545, 'clip_ratio/low_min': 0.015092328016180545, 'clip_ratio/high_mean': 0.03248523222282529, 'clip_ratio/high_max': 0.03248523222282529, 'clip_ratio/region_mean': 0.047577560239005834, 'reward_total_mean': 0.9247440099716187, 'reward_meter_mean': 0.9959684014320374, 'reward_meter_std': 0.0031451100949198008, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.9247440099716187, 'reward_total_composite_std': 0.07493867725133896, 'epoch': 0.07}

 56%|█████▌    | 1834/3300 [2:19:17<2:03:41,  5.06s/it]INFO 04-12 00:56:08 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1835/3300 [2:19:21<1:59:44,  4.90s/it]2026-04-12 00:56:12,622 | INFO | train_grpo_train | metrics_logged mode=train step=1835

                                                       
{'loss': -0.0192, 'grad_norm': 5.053841590881348, 'learning_rate': 4.442424242424243e-06, 'num_tokens': 4132027.0, 'completions/mean_length': 62.375, 'completions/min_length': 58.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.375, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8166337013244629, 'rewards/meter/std': 0.3090912401676178, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8166337013244629, 'rewards/total_composite/std': 0.3090912401676178, 'reward': 0.8166337013244629, 'reward_std': 0.3090912103652954, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038962990045547485, 'sampling/sampling_logp_difference/max': 1.3373403549194336, 'sampling/importance_sampling_ratio/min': 0.2625430226325989, 'sampling/importance_sampling_ratio/mean': 0.9984228610992432, 'sampling/importance_sampling_ratio/max': 1.7029978036880493, 'entropy': 0.26288264244794846, 'clip_ratio/low_mean': 0.008216594811528921, 'clip_ratio/low_min': 0.008216594811528921, 'clip_ratio/high_mean': 0.023937532445415854, 'clip_ratio/high_max': 0.023937532445415854, 'clip_ratio/region_mean': 0.032154127256944776, 'reward_total_mean': 0.8166337013244629, 'reward_meter_mean': 0.8166337013244629, 'reward_meter_std': 0.3090912401676178, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8166337013244629, 'reward_total_composite_std': 0.3090912401676178, 'epoch': 0.07}

 56%|█████▌    | 1835/3300 [2:19:21<1:59:44,  4.90s/it]INFO 04-12 00:56:12 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1836/3300 [2:19:26<2:01:43,  4.99s/it]2026-04-12 00:56:17,816 | INFO | train_grpo_train | metrics_logged mode=train step=1836

                                                       
{'loss': 0.0132, 'grad_norm': 5.317627906799316, 'learning_rate': 4.43939393939394e-06, 'num_tokens': 4134279.0, 'completions/mean_length': 116.5, 'completions/min_length': 111.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.5, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9806186556816101, 'rewards/meter/std': 0.017093582078814507, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9559872150421143, 'rewards/total_composite/std': 0.06989036500453949, 'reward': 0.9559872150421143, 'reward_std': 0.06989037245512009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07094138115644455, 'sampling/sampling_logp_difference/max': 1.1895244121551514, 'sampling/importance_sampling_ratio/min': 0.30436596274375916, 'sampling/importance_sampling_ratio/mean': 1.0162169933319092, 'sampling/importance_sampling_ratio/max': 1.9605425596237183, 'entropy': 0.6946443617343903, 'clip_ratio/low_mean': 0.00966143561527133, 'clip_ratio/low_min': 0.00966143561527133, 'clip_ratio/high_mean': 0.04100045142695308, 'clip_ratio/high_max': 0.04100045142695308, 'clip_ratio/region_mean': 0.05066188704222441, 'reward_total_mean': 0.9559872150421143, 'reward_meter_mean': 0.9806186556816101, 'reward_meter_std': 0.017093582078814507, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9559872150421143, 'reward_total_composite_std': 0.06989036500453949, 'epoch': 0.07}

 56%|█████▌    | 1836/3300 [2:19:26<2:01:43,  4.99s/it]INFO 04-12 00:56:18 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1837/3300 [2:19:32<2:06:23,  5.18s/it]2026-04-12 00:56:23,448 | INFO | train_grpo_train | metrics_logged mode=train step=1837

                                                       
{'loss': 0.0003, 'grad_norm': 1.2245804071426392, 'learning_rate': 4.436363636363637e-06, 'num_tokens': 4137119.0, 'completions/mean_length': 170.0, 'completions/min_length': 170.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.0, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9958264231681824, 'rewards/meter/std': 6.799297989346087e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.762499988079071, 'rewards/repeat_penalty/std': 0.05175492912530899, 'rewards/total_composite/mean': 0.7593173980712891, 'rewards/total_composite/std': 0.05153491348028183, 'reward': 0.7593173980712891, 'reward_std': 0.051534902304410934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005329389125108719, 'sampling/sampling_logp_difference/max': 0.6398299336433411, 'sampling/importance_sampling_ratio/min': 0.5273821353912354, 'sampling/importance_sampling_ratio/mean': 1.0017375946044922, 'sampling/importance_sampling_ratio/max': 1.4358134269714355, 'entropy': 0.04040496004745364, 'clip_ratio/low_mean': 0.000735294132027775, 'clip_ratio/low_min': 0.000735294132027775, 'clip_ratio/high_mean': 0.0029411765281111, 'clip_ratio/high_max': 0.0029411765281111, 'clip_ratio/region_mean': 0.0036764706601388752, 'reward_total_mean': 0.7593173980712891, 'reward_meter_mean': 0.9958264231681824, 'reward_meter_std': 6.799297989346087e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.762499988079071, 'reward_repeat_penalty_std': 0.05175492912530899, 'reward_total_composite_mean': 0.7593173980712891, 'reward_total_composite_std': 0.05153491348028183, 'epoch': 0.07}

 56%|█████▌    | 1837/3300 [2:19:32<2:06:23,  5.18s/it]INFO 04-12 00:56:23 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1838/3300 [2:19:38<2:16:17,  5.59s/it]2026-04-12 00:56:30,001 | INFO | train_grpo_train | metrics_logged mode=train step=1838

                                                       
{'loss': -0.019, 'grad_norm': 3.743595838546753, 'learning_rate': 4.433333333333334e-06, 'num_tokens': 4140579.0, 'completions/mean_length': 228.5, 'completions/min_length': 195.0, 'completions/max_length': 240.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 228.5, 'completions/min_terminated_length': 195.0, 'completions/max_terminated_length': 240.0, 'rewards/meter/mean': 0.7707350254058838, 'rewards/meter/std': 0.3024085462093353, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9791666865348816, 'rewards/repeat_penalty/std': 0.038575831800699234, 'rewards/total_composite/mean': 0.7390586137771606, 'rewards/total_composite/std': 0.29844897985458374, 'reward': 0.7390586137771606, 'reward_std': 0.29844897985458374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06499035656452179, 'sampling/sampling_logp_difference/max': 1.2022161483764648, 'sampling/importance_sampling_ratio/min': 0.3005274534225464, 'sampling/importance_sampling_ratio/mean': 1.015356183052063, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.668959241360426, 'clip_ratio/low_mean': 0.022611827589571476, 'clip_ratio/low_min': 0.022611827589571476, 'clip_ratio/high_mean': 0.022940170136280358, 'clip_ratio/high_max': 0.022940170136280358, 'clip_ratio/region_mean': 0.045551997725851834, 'reward_total_mean': 0.7390586137771606, 'reward_meter_mean': 0.7707350254058838, 'reward_meter_std': 0.3024085462093353, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9791666865348816, 'reward_repeat_penalty_std': 0.038575831800699234, 'reward_total_composite_mean': 0.7390586137771606, 'reward_total_composite_std': 0.29844897985458374, 'epoch': 0.07}

 56%|█████▌    | 1838/3300 [2:19:38<2:16:17,  5.59s/it]INFO 04-12 00:56:30 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1839/3300 [2:19:43<2:09:59,  5.34s/it]2026-04-12 00:56:34,743 | INFO | train_grpo_train | metrics_logged mode=train step=1839

                                                       
{'loss': -0.03, 'grad_norm': 5.02839994430542, 'learning_rate': 4.430303030303031e-06, 'num_tokens': 4142461.0, 'completions/mean_length': 78.25, 'completions/min_length': 73.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.25, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9820351600646973, 'rewards/meter/std': 0.018020369112491608, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9820351600646973, 'rewards/total_composite/std': 0.018020369112491608, 'reward': 0.9820351600646973, 'reward_std': 0.018020374700427055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07023806869983673, 'sampling/sampling_logp_difference/max': 1.4098834991455078, 'sampling/importance_sampling_ratio/min': 0.24417172372341156, 'sampling/importance_sampling_ratio/mean': 1.0230660438537598, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7259627729654312, 'clip_ratio/low_mean': 0.01651052851229906, 'clip_ratio/low_min': 0.01651052851229906, 'clip_ratio/high_mean': 0.02181874285452068, 'clip_ratio/high_max': 0.02181874285452068, 'clip_ratio/region_mean': 0.03832927136681974, 'reward_total_mean': 0.9820351600646973, 'reward_meter_mean': 0.9820351600646973, 'reward_meter_std': 0.018020369112491608, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9820351600646973, 'reward_total_composite_std': 0.018020369112491608, 'epoch': 0.07}

 56%|█████▌    | 1839/3300 [2:19:43<2:09:59,  5.34s/it]INFO 04-12 00:56:34 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1840/3300 [2:19:48<2:02:35,  5.04s/it]2026-04-12 00:56:39,079 | INFO | train_grpo_train | metrics_logged mode=train step=1840

                                                       
{'loss': 0.0184, 'grad_norm': 6.407735347747803, 'learning_rate': 4.4272727272727275e-06, 'num_tokens': 4144162.0, 'completions/mean_length': 49.625, 'completions/min_length': 46.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.625, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7431851029396057, 'rewards/meter/std': 0.294132262468338, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7431851029396057, 'rewards/total_composite/std': 0.294132262468338, 'reward': 0.7431851029396057, 'reward_std': 0.2941322326660156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05073566362261772, 'sampling/sampling_logp_difference/max': 1.3865375518798828, 'sampling/importance_sampling_ratio/min': 0.2499392181634903, 'sampling/importance_sampling_ratio/mean': 1.0074049234390259, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24762486666440964, 'clip_ratio/low_mean': 0.022959183901548386, 'clip_ratio/low_min': 0.022959183901548386, 'clip_ratio/high_mean': 0.027999775717034936, 'clip_ratio/high_max': 0.027999775717034936, 'clip_ratio/region_mean': 0.05095895961858332, 'reward_total_mean': 0.7431851029396057, 'reward_meter_mean': 0.7431851029396057, 'reward_meter_std': 0.294132262468338, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7431851029396057, 'reward_total_composite_std': 0.294132262468338, 'epoch': 0.07}

 56%|█████▌    | 1840/3300 [2:19:48<2:02:35,  5.04s/it]INFO 04-12 00:56:39 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1841/3300 [2:19:53<2:03:22,  5.07s/it]2026-04-12 00:56:44,241 | INFO | train_grpo_train | metrics_logged mode=train step=1841

                                                       
{'loss': 0.012, 'grad_norm': 3.6605427265167236, 'learning_rate': 4.424242424242425e-06, 'num_tokens': 4146697.0, 'completions/mean_length': 119.875, 'completions/min_length': 116.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.875, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9946545362472534, 'rewards/meter/std': 0.0017242503818124533, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9058225154876709, 'rewards/total_composite/std': 0.07321521639823914, 'reward': 0.9058225154876709, 'reward_std': 0.07321520894765854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016406936571002007, 'sampling/sampling_logp_difference/max': 1.0592851638793945, 'sampling/importance_sampling_ratio/min': 0.34670352935791016, 'sampling/importance_sampling_ratio/mean': 0.9984482526779175, 'sampling/importance_sampling_ratio/max': 1.6830854415893555, 'entropy': 0.08843067381531, 'clip_ratio/low_mean': 0.008309659664519131, 'clip_ratio/low_min': 0.008309659664519131, 'clip_ratio/high_mean': 0.00737299001775682, 'clip_ratio/high_max': 0.00737299001775682, 'clip_ratio/region_mean': 0.01568264968227595, 'reward_total_mean': 0.9058225154876709, 'reward_meter_mean': 0.9946545362472534, 'reward_meter_std': 0.0017242503818124533, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9058225154876709, 'reward_total_composite_std': 0.07321521639823914, 'epoch': 0.07}

 56%|█████▌    | 1841/3300 [2:19:53<2:03:22,  5.07s/it]INFO 04-12 00:56:44 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1842/3300 [2:19:57<1:59:32,  4.92s/it]2026-04-12 00:56:48,796 | INFO | train_grpo_train | metrics_logged mode=train step=1842

                                                       
{'loss': 0.0062, 'grad_norm': 4.1317853927612305, 'learning_rate': 4.421212121212122e-06, 'num_tokens': 4148545.0, 'completions/mean_length': 65.0, 'completions/min_length': 63.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.8449044227600098, 'rewards/meter/std': 0.19496634602546692, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8449044227600098, 'rewards/total_composite/std': 0.19496634602546692, 'reward': 0.8449044227600098, 'reward_std': 0.19496633112430573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03727368265390396, 'sampling/sampling_logp_difference/max': 1.3517389297485352, 'sampling/importance_sampling_ratio/min': 0.2587898373603821, 'sampling/importance_sampling_ratio/mean': 1.0063512325286865, 'sampling/importance_sampling_ratio/max': 1.8676364421844482, 'entropy': 0.2709883898496628, 'clip_ratio/low_mean': 0.012987380847334862, 'clip_ratio/low_min': 0.012987380847334862, 'clip_ratio/high_mean': 0.00787545822095126, 'clip_ratio/high_max': 0.00787545822095126, 'clip_ratio/region_mean': 0.02086283906828612, 'reward_total_mean': 0.8449044227600098, 'reward_meter_mean': 0.8449044227600098, 'reward_meter_std': 0.19496634602546692, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8449044227600098, 'reward_total_composite_std': 0.19496634602546692, 'epoch': 0.07}

 56%|█████▌    | 1842/3300 [2:19:57<1:59:32,  4.92s/it]INFO 04-12 00:56:49 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1843/3300 [2:20:02<1:56:45,  4.81s/it]2026-04-12 00:56:53,349 | INFO | train_grpo_train | metrics_logged mode=train step=1843

                                                       
{'loss': 0.0033, 'grad_norm': 2.7024903297424316, 'learning_rate': 4.418181818181818e-06, 'num_tokens': 4150316.0, 'completions/mean_length': 69.375, 'completions/min_length': 69.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9967325925827026, 'rewards/meter/std': 0.0019061544444411993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9967325925827026, 'rewards/total_composite/std': 0.0019061544444411993, 'reward': 0.9967325925827026, 'reward_std': 0.0019061638740822673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015712035819888115, 'sampling/sampling_logp_difference/max': 1.4424948692321777, 'sampling/importance_sampling_ratio/min': 0.23633739352226257, 'sampling/importance_sampling_ratio/mean': 1.0016083717346191, 'sampling/importance_sampling_ratio/max': 1.3403892517089844, 'entropy': 0.09468106366693974, 'clip_ratio/low_mean': 0.005357142887078226, 'clip_ratio/low_min': 0.005357142887078226, 'clip_ratio/high_mean': 0.005434782709926367, 'clip_ratio/high_max': 0.005434782709926367, 'clip_ratio/region_mean': 0.010791925597004592, 'reward_total_mean': 0.9967325925827026, 'reward_meter_mean': 0.9967325925827026, 'reward_meter_std': 0.0019061544444411993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9967325925827026, 'reward_total_composite_std': 0.0019061544444411993, 'epoch': 0.07}

 56%|█████▌    | 1843/3300 [2:20:02<1:56:45,  4.81s/it]INFO 04-12 00:56:53 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1844/3300 [2:20:07<1:57:01,  4.82s/it]2026-04-12 00:56:58,201 | INFO | train_grpo_train | metrics_logged mode=train step=1844

                                                       
{'loss': 0.0055, 'grad_norm': 4.094827175140381, 'learning_rate': 4.415151515151516e-06, 'num_tokens': 4152604.0, 'completions/mean_length': 94.0, 'completions/min_length': 91.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.0, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9527285695075989, 'rewards/meter/std': 0.020587077364325523, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9527285695075989, 'rewards/total_composite/std': 0.020587077364325523, 'reward': 0.9527285695075989, 'reward_std': 0.02058705873787403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0326327309012413, 'sampling/sampling_logp_difference/max': 2.7240052223205566, 'sampling/importance_sampling_ratio/min': 0.06561143696308136, 'sampling/importance_sampling_ratio/mean': 1.006330132484436, 'sampling/importance_sampling_ratio/max': 1.6599549055099487, 'entropy': 0.2582856100052595, 'clip_ratio/low_mean': 0.00394825276453048, 'clip_ratio/low_min': 0.00394825276453048, 'clip_ratio/high_mean': 0.01599274785257876, 'clip_ratio/high_max': 0.01599274785257876, 'clip_ratio/region_mean': 0.01994100061710924, 'reward_total_mean': 0.9527285695075989, 'reward_meter_mean': 0.9527285695075989, 'reward_meter_std': 0.020587077364325523, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9527285695075989, 'reward_total_composite_std': 0.020587077364325523, 'epoch': 0.07}

 56%|█████▌    | 1844/3300 [2:20:07<1:57:01,  4.82s/it]INFO 04-12 00:56:58 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1845/3300 [2:20:11<1:55:16,  4.75s/it]2026-04-12 00:57:02,794 | INFO | train_grpo_train | metrics_logged mode=train step=1845

                                                       
{'loss': -0.0044, 'grad_norm': 6.061025619506836, 'learning_rate': 4.412121212121213e-06, 'num_tokens': 4154452.0, 'completions/mean_length': 69.0, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9951037168502808, 'rewards/meter/std': 0.0059710158966481686, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9951037168502808, 'rewards/total_composite/std': 0.0059710158966481686, 'reward': 0.9951037168502808, 'reward_std': 0.005971014499664307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009792874567210674, 'sampling/sampling_logp_difference/max': 0.5869016647338867, 'sampling/importance_sampling_ratio/min': 0.5560474395751953, 'sampling/importance_sampling_ratio/mean': 1.0022023916244507, 'sampling/importance_sampling_ratio/max': 1.1276072263717651, 'entropy': 0.07721387594938278, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005434782709926367, 'clip_ratio/high_max': 0.005434782709926367, 'clip_ratio/region_mean': 0.005434782709926367, 'reward_total_mean': 0.9951037168502808, 'reward_meter_mean': 0.9951037168502808, 'reward_meter_std': 0.0059710158966481686, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9951037168502808, 'reward_total_composite_std': 0.0059710158966481686, 'epoch': 0.07}

 56%|█████▌    | 1845/3300 [2:20:11<1:55:16,  4.75s/it]INFO 04-12 00:57:03 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1846/3300 [2:20:17<2:01:18,  5.01s/it]2026-04-12 00:57:08,389 | INFO | train_grpo_train | metrics_logged mode=train step=1846

                                                       
{'loss': 0.0137, 'grad_norm': 2.195159912109375, 'learning_rate': 4.409090909090909e-06, 'num_tokens': 4157062.0, 'completions/mean_length': 152.25, 'completions/min_length': 150.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 152.25, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.8805509805679321, 'rewards/meter/std': 0.11502306908369064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.05143444612622261, 'rewards/total_composite/mean': 0.7611056566238403, 'rewards/total_composite/std': 0.12567628920078278, 'reward': 0.7611056566238403, 'reward_std': 0.12567628920078278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02993069775402546, 'sampling/sampling_logp_difference/max': 1.2740850448608398, 'sampling/importance_sampling_ratio/min': 0.27968674898147583, 'sampling/importance_sampling_ratio/mean': 1.0067634582519531, 'sampling/importance_sampling_ratio/max': 1.9269274473190308, 'entropy': 0.28932092152535915, 'clip_ratio/low_mean': 0.00973088713362813, 'clip_ratio/low_min': 0.00973088713362813, 'clip_ratio/high_mean': 0.015674303169362247, 'clip_ratio/high_max': 0.015674303169362247, 'clip_ratio/region_mean': 0.025405190302990377, 'reward_total_mean': 0.7611056566238403, 'reward_meter_mean': 0.8805509805679321, 'reward_meter_std': 0.11502306908369064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.05143444612622261, 'reward_total_composite_mean': 0.7611056566238403, 'reward_total_composite_std': 0.12567628920078278, 'epoch': 0.07}

 56%|█████▌    | 1846/3300 [2:20:17<2:01:18,  5.01s/it]INFO 04-12 00:57:08 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1847/3300 [2:20:22<2:03:40,  5.11s/it]2026-04-12 00:57:13,734 | INFO | train_grpo_train | metrics_logged mode=train step=1847

                                                       
{'loss': 0.002, 'grad_norm': 1.8122365474700928, 'learning_rate': 4.4060606060606066e-06, 'num_tokens': 4159504.0, 'completions/mean_length': 137.25, 'completions/min_length': 137.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.25, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.9956631660461426, 'rewards/meter/std': 0.004949868656694889, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8709865212440491, 'rewards/total_composite/std': 0.04542805626988411, 'reward': 0.8709865212440491, 'reward_std': 0.04542805254459381, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00922767911106348, 'sampling/sampling_logp_difference/max': 1.0127711296081543, 'sampling/importance_sampling_ratio/min': 0.36321109533309937, 'sampling/importance_sampling_ratio/mean': 1.0003859996795654, 'sampling/importance_sampling_ratio/max': 1.290838360786438, 'entropy': 0.058579874224960804, 'clip_ratio/low_mean': 0.0027372262557037175, 'clip_ratio/low_min': 0.0027372262557037175, 'clip_ratio/high_mean': 0.0018115942366421223, 'clip_ratio/high_max': 0.0018115942366421223, 'clip_ratio/region_mean': 0.00454882049234584, 'reward_total_mean': 0.8709865212440491, 'reward_meter_mean': 0.9956631660461426, 'reward_meter_std': 0.004949868656694889, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8709865212440491, 'reward_total_composite_std': 0.04542805626988411, 'epoch': 0.07}

 56%|█████▌    | 1847/3300 [2:20:22<2:03:40,  5.11s/it]INFO 04-12 00:57:13 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1848/3300 [2:20:29<2:13:40,  5.52s/it]2026-04-12 00:57:20,229 | INFO | train_grpo_train | metrics_logged mode=train step=1848

                                                       
{'loss': -0.0158, 'grad_norm': 3.0020229816436768, 'learning_rate': 4.403030303030304e-06, 'num_tokens': 4162692.0, 'completions/mean_length': 193.5, 'completions/min_length': 181.0, 'completions/max_length': 226.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 193.5, 'completions/min_terminated_length': 181.0, 'completions/max_terminated_length': 226.0, 'rewards/meter/mean': 0.9245690107345581, 'rewards/meter/std': 0.12092939764261246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9245690107345581, 'rewards/total_composite/std': 0.12092939764261246, 'reward': 0.9245690107345581, 'reward_std': 0.12092941999435425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06919539719820023, 'sampling/sampling_logp_difference/max': 1.4243488311767578, 'sampling/importance_sampling_ratio/min': 0.24066513776779175, 'sampling/importance_sampling_ratio/mean': 1.01200270652771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6156773902475834, 'clip_ratio/low_mean': 0.01064768130891025, 'clip_ratio/low_min': 0.01064768130891025, 'clip_ratio/high_mean': 0.03866560058668256, 'clip_ratio/high_max': 0.03866560058668256, 'clip_ratio/region_mean': 0.04931328189559281, 'reward_total_mean': 0.9245690107345581, 'reward_meter_mean': 0.9245690107345581, 'reward_meter_std': 0.12092939764261246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9245690107345581, 'reward_total_composite_std': 0.12092939764261246, 'epoch': 0.07}

 56%|█████▌    | 1848/3300 [2:20:29<2:13:40,  5.52s/it]INFO 04-12 00:57:20 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1849/3300 [2:20:34<2:08:45,  5.32s/it]2026-04-12 00:57:25,086 | INFO | train_grpo_train | metrics_logged mode=train step=1849

                                                       
{'loss': 0.0019, 'grad_norm': 0.7777861952781677, 'learning_rate': 4.4e-06, 'num_tokens': 4165004.0, 'completions/mean_length': 103.0, 'completions/min_length': 103.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.0, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9962564706802368, 'rewards/meter/std': 0.003385263029485941, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9962564706802368, 'rewards/total_composite/std': 0.003385263029485941, 'reward': 0.9962564706802368, 'reward_std': 0.0033852593041956425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006807548459619284, 'sampling/sampling_logp_difference/max': 0.6204808950424194, 'sampling/importance_sampling_ratio/min': 0.5530079007148743, 'sampling/importance_sampling_ratio/mean': 1.0040092468261719, 'sampling/importance_sampling_ratio/max': 1.859822154045105, 'entropy': 0.04185265023261309, 'clip_ratio/low_mean': 0.0012135922443121672, 'clip_ratio/low_min': 0.0012135922443121672, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0012135922443121672, 'reward_total_mean': 0.9962564706802368, 'reward_meter_mean': 0.9962564706802368, 'reward_meter_std': 0.003385263029485941, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9962564706802368, 'reward_total_composite_std': 0.003385263029485941, 'epoch': 0.07}

 56%|█████▌    | 1849/3300 [2:20:34<2:08:45,  5.32s/it]INFO 04-12 00:57:25 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1850/3300 [2:20:38<2:02:23,  5.06s/it]2026-04-12 00:57:29,546 | INFO | train_grpo_train | metrics_logged mode=train step=1850

                                                       
{'loss': -0.0062, 'grad_norm': 5.336696624755859, 'learning_rate': 4.3969696969696975e-06, 'num_tokens': 4166714.0, 'completions/mean_length': 65.75, 'completions/min_length': 64.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9984281063079834, 'rewards/meter/std': 0.0004244510782882571, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984281063079834, 'rewards/total_composite/std': 0.0004244510782882571, 'reward': 0.9984281063079834, 'reward_std': 0.0004244422307237983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02056110091507435, 'sampling/sampling_logp_difference/max': 1.2651333808898926, 'sampling/importance_sampling_ratio/min': 0.2822016477584839, 'sampling/importance_sampling_ratio/mean': 0.9998464584350586, 'sampling/importance_sampling_ratio/max': 1.7678344249725342, 'entropy': 0.06614176463335752, 'clip_ratio/low_mean': 0.005741003900766373, 'clip_ratio/low_min': 0.005741003900766373, 'clip_ratio/high_mean': 0.011363636702299118, 'clip_ratio/high_max': 0.011363636702299118, 'clip_ratio/region_mean': 0.01710464060306549, 'reward_total_mean': 0.9984281063079834, 'reward_meter_mean': 0.9984281063079834, 'reward_meter_std': 0.0004244510782882571, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984281063079834, 'reward_total_composite_std': 0.0004244510782882571, 'epoch': 0.07}

 56%|█████▌    | 1850/3300 [2:20:38<2:02:23,  5.06s/it]INFO 04-12 00:57:29 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:26,  2.41s/it]

 23%|██▎       | 3/13 [00:08<00:27,  2.75s/it]

 31%|███       | 4/13 [00:12<00:31,  3.49s/it]

 38%|███▊      | 5/13 [00:17<00:31,  3.91s/it]

 46%|████▌     | 6/13 [00:22<00:30,  4.39s/it]

 54%|█████▍    | 7/13 [00:27<00:27,  4.56s/it]

 62%|██████▏   | 8/13 [00:33<00:24,  4.93s/it]

 69%|██████▉   | 9/13 [00:39<00:20,  5.13s/it]

 77%|███████▋  | 10/13 [00:43<00:14,  4.85s/it]

 85%|████████▍ | 11/13 [00:47<00:09,  4.79s/it]

 92%|█████████▏| 12/13 [00:52<00:04,  4.79s/it]

100%|██████████| 13/13 [00:57<00:00,  4.79s/it]2026-04-12 00:58:32,306 | INFO | train_grpo_train | metrics_logged mode=eval step=1850

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 62.689, 'eval_samples_per_second': 1.659, 'eval_steps_per_second': 0.207, 'eval_num_tokens': 4166714.0, 'eval_completions/mean_length': 193.1153846153846, 'eval_completions/min_length': 63.76923076923077, 'eval_completions/max_length': 327.2307692307692, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 193.1153846153846, 'eval_completions/min_terminated_length': 63.76923076923077, 'eval_completions/max_terminated_length': 327.2307692307692, 'eval_rewards/meter/mean': 0.6479656077348269, 'eval_rewards/meter/std': 0.4532018624819242, 'eval_rewards/count_adherence/mean': 0.8968340616959792, 'eval_rewards/count_adherence/std': 0.13571923226118088, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8583002182153555, 'eval_rewards/repeat_penalty/std': 0.13356439138834292, 'eval_rewards/total_composite/mean': 0.510624709037634, 'eval_rewards/total_composite/std': 0.38263802230358124, 'eval_reward': 0.510624709037634, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.023962400352152493, 'eval_sampling/sampling_logp_difference/max': 1.0324598183998694, 'eval_sampling/importance_sampling_ratio/min': 0.3671792642428325, 'eval_sampling/importance_sampling_ratio/mean': 1.0072369300402129, 'eval_sampling/importance_sampling_ratio/max': 1.4861979392858653, 'eval_entropy': 0.2572394087910652, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.510624709037634, 'eval_reward_meter_mean': 0.6479656077348269, 'eval_reward_meter_std': 0.4532018624819242, 'eval_reward_count_adherence_mean': 0.8968340616959792, 'eval_reward_count_adherence_std': 0.13571923226118088, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8583002182153555, 'eval_reward_repeat_penalty_std': 0.13356439138834292, 'eval_reward_total_composite_mean': 0.510624709037634, 'eval_reward_total_composite_std': 0.38263802230358124, 'epoch': 0.07}

 56%|█████▌    | 1850/3300 [2:21:41<2:02:23,  5.06s/it]

100%|██████████| 13/13 [00:57<00:00,  4.79s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 56%|█████▌    | 1851/3300 [2:21:49<9:58:02, 24.76s/it]2026-04-12 00:58:40,290 | INFO | train_grpo_train | metrics_logged mode=train step=1851

                                                       
{'loss': 0.0113, 'grad_norm': 4.0140767097473145, 'learning_rate': 4.393939393939394e-06, 'num_tokens': 4168913.0, 'completions/mean_length': 110.875, 'completions/min_length': 103.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.875, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.37161165475845337, 'rewards/meter/std': 0.27739158272743225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.37161165475845337, 'rewards/total_composite/std': 0.27739158272743225, 'reward': 0.37161165475845337, 'reward_std': 0.27739158272743225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06999441236257553, 'sampling/sampling_logp_difference/max': 1.3712263107299805, 'sampling/importance_sampling_ratio/min': 0.2537955343723297, 'sampling/importance_sampling_ratio/mean': 1.0157358646392822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5153284594416618, 'clip_ratio/low_mean': 0.044483832316473126, 'clip_ratio/low_min': 0.044483832316473126, 'clip_ratio/high_mean': 0.00905440840870142, 'clip_ratio/high_max': 0.00905440840870142, 'clip_ratio/region_mean': 0.053538240725174546, 'reward_total_mean': 0.37161165475845337, 'reward_meter_mean': 0.37161165475845337, 'reward_meter_std': 0.27739158272743225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.37161165475845337, 'reward_total_composite_std': 0.27739158272743225, 'epoch': 0.07}

 56%|█████▌    | 1851/3300 [2:21:49<9:58:02, 24.76s/it]INFO 04-12 00:58:40 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1852/3300 [2:21:56<7:50:57, 19.52s/it]2026-04-12 00:58:47,559 | INFO | train_grpo_train | metrics_logged mode=train step=1852

                                                       
{'loss': 0.0023, 'grad_norm': 3.773585557937622, 'learning_rate': 4.390909090909091e-06, 'num_tokens': 4172359.0, 'completions/mean_length': 247.75, 'completions/min_length': 222.0, 'completions/max_length': 273.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 247.75, 'completions/min_terminated_length': 222.0, 'completions/max_terminated_length': 273.0, 'rewards/meter/mean': 0.6540890336036682, 'rewards/meter/std': 0.43809670209884644, 'rewards/count_adherence/mean': 0.9464285969734192, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9910714626312256, 'rewards/repeat_penalty/std': 0.025253823027014732, 'rewards/total_composite/mean': 0.6160822510719299, 'rewards/total_composite/std': 0.42173710465431213, 'reward': 0.6160822510719299, 'reward_std': 0.42173710465431213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09465256333351135, 'sampling/sampling_logp_difference/max': 5.8384318351745605, 'sampling/importance_sampling_ratio/min': 0.002913407515734434, 'sampling/importance_sampling_ratio/mean': 1.012731671333313, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7441882863640785, 'clip_ratio/low_mean': 0.01895719300955534, 'clip_ratio/low_min': 0.01895719300955534, 'clip_ratio/high_mean': 0.04731742758303881, 'clip_ratio/high_max': 0.04731742758303881, 'clip_ratio/region_mean': 0.06627462059259415, 'reward_total_mean': 0.6160822510719299, 'reward_meter_mean': 0.6540890336036682, 'reward_meter_std': 0.43809670209884644, 'reward_count_adherence_mean': 0.9464285969734192, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9910714626312256, 'reward_repeat_penalty_std': 0.025253823027014732, 'reward_total_composite_mean': 0.6160822510719299, 'reward_total_composite_std': 0.42173710465431213, 'epoch': 0.07}

 56%|█████▌    | 1852/3300 [2:21:56<7:50:57, 19.52s/it]INFO 04-12 00:58:47 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1853/3300 [2:22:00<6:01:59, 15.01s/it]2026-04-12 00:58:52,045 | INFO | train_grpo_train | metrics_logged mode=train step=1853

                                                       
{'loss': 0.006, 'grad_norm': 11.522212982177734, 'learning_rate': 4.387878787878788e-06, 'num_tokens': 4174010.0, 'completions/mean_length': 35.375, 'completions/min_length': 35.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9776397943496704, 'rewards/meter/std': 0.01006159745156765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9776397943496704, 'rewards/total_composite/std': 0.01006159745156765, 'reward': 0.9776397943496704, 'reward_std': 0.010061599314212799, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01866281032562256, 'sampling/sampling_logp_difference/max': 0.5338115692138672, 'sampling/importance_sampling_ratio/min': 0.6540494561195374, 'sampling/importance_sampling_ratio/mean': 1.0117498636245728, 'sampling/importance_sampling_ratio/max': 1.7054202556610107, 'entropy': 0.18969010561704636, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0035714285913854837, 'reward_total_mean': 0.9776397943496704, 'reward_meter_mean': 0.9776397943496704, 'reward_meter_std': 0.01006159745156765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9776397943496704, 'reward_total_composite_std': 0.01006159745156765, 'epoch': 0.07}

 56%|█████▌    | 1853/3300 [2:22:01<6:01:59, 15.01s/it]INFO 04-12 00:58:52 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1854/3300 [2:22:06<4:50:48, 12.07s/it]2026-04-12 00:58:57,235 | INFO | train_grpo_train | metrics_logged mode=train step=1854

                                                       
{'loss': 0.0082, 'grad_norm': 4.39829158782959, 'learning_rate': 4.384848484848485e-06, 'num_tokens': 4176279.0, 'completions/mean_length': 114.625, 'completions/min_length': 110.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.625, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9875234365463257, 'rewards/meter/std': 0.01168668083846569, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9875234365463257, 'rewards/total_composite/std': 0.01168668083846569, 'reward': 0.9875234365463257, 'reward_std': 0.01168669294565916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054695580154657364, 'sampling/sampling_logp_difference/max': 1.2924197912216187, 'sampling/importance_sampling_ratio/min': 0.27460548281669617, 'sampling/importance_sampling_ratio/mean': 1.0194075107574463, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5414317175745964, 'clip_ratio/low_mean': 0.009980237111449242, 'clip_ratio/low_min': 0.009980237111449242, 'clip_ratio/high_mean': 0.03236477100290358, 'clip_ratio/high_max': 0.03236477100290358, 'clip_ratio/region_mean': 0.04234500811435282, 'reward_total_mean': 0.9875234365463257, 'reward_meter_mean': 0.9875234365463257, 'reward_meter_std': 0.01168668083846569, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9875234365463257, 'reward_total_composite_std': 0.01168668083846569, 'epoch': 0.07}

 56%|█████▌    | 1854/3300 [2:22:06<4:50:48, 12.07s/it]INFO 04-12 00:58:57 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1855/3300 [2:22:10<3:55:41,  9.79s/it]2026-04-12 00:59:01,706 | INFO | train_grpo_train | metrics_logged mode=train step=1855

                                                       
{'loss': 0.0017, 'grad_norm': 2.625075101852417, 'learning_rate': 4.381818181818182e-06, 'num_tokens': 4178031.0, 'completions/mean_length': 56.0, 'completions/min_length': 56.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9880209565162659, 'rewards/meter/std': 0.007883170619606972, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9880209565162659, 'rewards/total_composite/std': 0.007883170619606972, 'reward': 0.9880209565162659, 'reward_std': 0.0078831622377038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010663346387445927, 'sampling/sampling_logp_difference/max': 0.7668309211730957, 'sampling/importance_sampling_ratio/min': 0.46448272466659546, 'sampling/importance_sampling_ratio/mean': 1.0048121213912964, 'sampling/importance_sampling_ratio/max': 1.396715521812439, 'entropy': 0.0864438284188509, 'clip_ratio/low_mean': 0.0022321429569274187, 'clip_ratio/low_min': 0.0022321429569274187, 'clip_ratio/high_mean': 0.008928571827709675, 'clip_ratio/high_max': 0.008928571827709675, 'clip_ratio/region_mean': 0.011160714784637094, 'reward_total_mean': 0.9880209565162659, 'reward_meter_mean': 0.9880209565162659, 'reward_meter_std': 0.007883170619606972, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9880209565162659, 'reward_total_composite_std': 0.007883170619606972, 'epoch': 0.07}

 56%|█████▌    | 1855/3300 [2:22:10<3:55:41,  9.79s/it]INFO 04-12 00:59:01 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1856/3300 [2:22:15<3:21:53,  8.39s/it]2026-04-12 00:59:06,833 | INFO | train_grpo_train | metrics_logged mode=train step=1856

                                                       
{'loss': 0.0108, 'grad_norm': 3.003889799118042, 'learning_rate': 4.378787878787879e-06, 'num_tokens': 4180182.0, 'completions/mean_length': 111.875, 'completions/min_length': 109.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.875, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9974768757820129, 'rewards/meter/std': 0.0010170461609959602, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8977147340774536, 'rewards/total_composite/std': 0.10649988800287247, 'reward': 0.8977147340774536, 'reward_std': 0.10649988800287247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04684070870280266, 'sampling/sampling_logp_difference/max': 1.5592315196990967, 'sampling/importance_sampling_ratio/min': 0.2102976143360138, 'sampling/importance_sampling_ratio/mean': 1.004133701324463, 'sampling/importance_sampling_ratio/max': 1.5614416599273682, 'entropy': 0.35822908766567707, 'clip_ratio/low_mean': 0.007754115387797356, 'clip_ratio/low_min': 0.007754115387797356, 'clip_ratio/high_mean': 0.023573311744257808, 'clip_ratio/high_max': 0.023573311744257808, 'clip_ratio/region_mean': 0.03132742713205516, 'reward_total_mean': 0.8977147340774536, 'reward_meter_mean': 0.9974768757820129, 'reward_meter_std': 0.0010170461609959602, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8977147340774536, 'reward_total_composite_std': 0.10649988800287247, 'epoch': 0.07}

 56%|█████▌    | 1856/3300 [2:22:15<3:21:53,  8.39s/it]INFO 04-12 00:59:07 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1857/3300 [2:22:20<2:54:12,  7.24s/it]2026-04-12 00:59:11,405 | INFO | train_grpo_train | metrics_logged mode=train step=1857

                                                       
{'loss': -0.0105, 'grad_norm': 1.938165307044983, 'learning_rate': 4.375757575757576e-06, 'num_tokens': 4181844.0, 'completions/mean_length': 56.75, 'completions/min_length': 56.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9916501641273499, 'rewards/meter/std': 0.001931402483023703, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9916501641273499, 'rewards/total_composite/std': 0.001931402483023703, 'reward': 0.9916501641273499, 'reward_std': 0.0019314087694510818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013949423097074032, 'sampling/sampling_logp_difference/max': 1.0583875179290771, 'sampling/importance_sampling_ratio/min': 0.3470149040222168, 'sampling/importance_sampling_ratio/mean': 1.0090504884719849, 'sampling/importance_sampling_ratio/max': 1.4867905378341675, 'entropy': 0.09877756051719189, 'clip_ratio/low_mean': 0.017778822453692555, 'clip_ratio/low_min': 0.017778822453692555, 'clip_ratio/high_mean': 0.0021551724057644606, 'clip_ratio/high_max': 0.0021551724057644606, 'clip_ratio/region_mean': 0.019933994859457016, 'reward_total_mean': 0.9916501641273499, 'reward_meter_mean': 0.9916501641273499, 'reward_meter_std': 0.001931402483023703, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9916501641273499, 'reward_total_composite_std': 0.001931402483023703, 'epoch': 0.07}

 56%|█████▋    | 1857/3300 [2:22:20<2:54:12,  7.24s/it]INFO 04-12 00:59:11 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1858/3300 [2:22:25<2:36:47,  6.52s/it]2026-04-12 00:59:16,250 | INFO | train_grpo_train | metrics_logged mode=train step=1858

                                                       
{'loss': -0.0057, 'grad_norm': 2.222642421722412, 'learning_rate': 4.372727272727273e-06, 'num_tokens': 4183958.0, 'completions/mean_length': 88.25, 'completions/min_length': 82.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.25, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9945516586303711, 'rewards/meter/std': 0.0012705601984634995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9945516586303711, 'rewards/total_composite/std': 0.0012705601984634995, 'reward': 0.9945516586303711, 'reward_std': 0.0012705654371529818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012971967458724976, 'sampling/sampling_logp_difference/max': 0.851701021194458, 'sampling/importance_sampling_ratio/min': 0.4266884922981262, 'sampling/importance_sampling_ratio/mean': 1.0038167238235474, 'sampling/importance_sampling_ratio/max': 1.6056901216506958, 'entropy': 0.09635706804692745, 'clip_ratio/low_mean': 0.0028572361916303635, 'clip_ratio/low_min': 0.0028572361916303635, 'clip_ratio/high_mean': 0.007095551351085305, 'clip_ratio/high_max': 0.007095551351085305, 'clip_ratio/region_mean': 0.009952787542715669, 'reward_total_mean': 0.9945516586303711, 'reward_meter_mean': 0.9945516586303711, 'reward_meter_std': 0.0012705601984634995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9945516586303711, 'reward_total_composite_std': 0.0012705601984634995, 'epoch': 0.07}

 56%|█████▋    | 1858/3300 [2:22:25<2:36:47,  6.52s/it]INFO 04-12 00:59:16 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1859/3300 [2:22:29<2:22:37,  5.94s/it]2026-04-12 00:59:20,822 | INFO | train_grpo_train | metrics_logged mode=train step=1859

                                                       
{'loss': -0.0001, 'grad_norm': 4.79885721206665, 'learning_rate': 4.36969696969697e-06, 'num_tokens': 4185702.0, 'completions/mean_length': 51.0, 'completions/min_length': 51.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9259762763977051, 'rewards/meter/std': 0.0015891619259491563, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9259762763977051, 'rewards/total_composite/std': 0.0015891619259491563, 'reward': 0.9259762763977051, 'reward_std': 0.001589166815392673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012182224541902542, 'sampling/sampling_logp_difference/max': 0.820074200630188, 'sampling/importance_sampling_ratio/min': 0.4403989613056183, 'sampling/importance_sampling_ratio/mean': 0.9998874068260193, 'sampling/importance_sampling_ratio/max': 1.6099872589111328, 'entropy': 0.049468324054032564, 'clip_ratio/low_mean': 0.007352941436693072, 'clip_ratio/low_min': 0.007352941436693072, 'clip_ratio/high_mean': 0.0049019609577953815, 'clip_ratio/high_max': 0.0049019609577953815, 'clip_ratio/region_mean': 0.012254902394488454, 'reward_total_mean': 0.9259762763977051, 'reward_meter_mean': 0.9259762763977051, 'reward_meter_std': 0.0015891619259491563, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9259762763977051, 'reward_total_composite_std': 0.0015891619259491563, 'epoch': 0.07}

 56%|█████▋    | 1859/3300 [2:22:29<2:22:37,  5.94s/it]INFO 04-12 00:59:21 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1860/3300 [2:22:35<2:19:06,  5.80s/it]2026-04-12 00:59:26,289 | INFO | train_grpo_train | metrics_logged mode=train step=1860

                                                       
{'loss': -0.0032, 'grad_norm': 2.1606459617614746, 'learning_rate': 4.366666666666667e-06, 'num_tokens': 4188182.0, 'completions/mean_length': 135.0, 'completions/min_length': 132.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.0, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.989547610282898, 'rewards/meter/std': 0.007958509027957916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8834315538406372, 'rewards/total_composite/std': 0.06423989683389664, 'reward': 0.8834315538406372, 'reward_std': 0.06423989683389664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01665782928466797, 'sampling/sampling_logp_difference/max': 1.199911117553711, 'sampling/importance_sampling_ratio/min': 0.30122095346450806, 'sampling/importance_sampling_ratio/mean': 1.0032159090042114, 'sampling/importance_sampling_ratio/max': 1.8788100481033325, 'entropy': 0.100110930390656, 'clip_ratio/low_mean': 0.008354876074008644, 'clip_ratio/low_min': 0.008354876074008644, 'clip_ratio/high_mean': 0.006407182663679123, 'clip_ratio/high_max': 0.006407182663679123, 'clip_ratio/region_mean': 0.014762058737687767, 'reward_total_mean': 0.8834315538406372, 'reward_meter_mean': 0.989547610282898, 'reward_meter_std': 0.007958509027957916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8834315538406372, 'reward_total_composite_std': 0.06423989683389664, 'epoch': 0.07}

 56%|█████▋    | 1860/3300 [2:22:35<2:19:06,  5.80s/it]INFO 04-12 00:59:26 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1861/3300 [2:22:40<2:13:38,  5.57s/it]2026-04-12 00:59:31,341 | INFO | train_grpo_train | metrics_logged mode=train step=1861

                                                       
{'loss': 0.0208, 'grad_norm': 2.7288973331451416, 'learning_rate': 4.363636363636364e-06, 'num_tokens': 4190278.0, 'completions/mean_length': 101.0, 'completions/min_length': 97.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.7360309958457947, 'rewards/meter/std': 0.27524590492248535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.7065380811691284, 'rewards/total_composite/std': 0.27462705969810486, 'reward': 0.7065380811691284, 'reward_std': 0.27462708950042725, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0317775197327137, 'sampling/sampling_logp_difference/max': 2.5332324504852295, 'sampling/importance_sampling_ratio/min': 0.07940194010734558, 'sampling/importance_sampling_ratio/mean': 1.0014276504516602, 'sampling/importance_sampling_ratio/max': 1.5228219032287598, 'entropy': 0.16979045886546373, 'clip_ratio/low_mean': 0.009549640817567706, 'clip_ratio/low_min': 0.009549640817567706, 'clip_ratio/high_mean': 0.0124762476189062, 'clip_ratio/high_max': 0.0124762476189062, 'clip_ratio/region_mean': 0.022025888436473906, 'reward_total_mean': 0.7065380811691284, 'reward_meter_mean': 0.7360309958457947, 'reward_meter_std': 0.27524590492248535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.7065380811691284, 'reward_total_composite_std': 0.27462705969810486, 'epoch': 0.07}

 56%|█████▋    | 1861/3300 [2:22:40<2:13:38,  5.57s/it]INFO 04-12 00:59:31 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1862/3300 [2:22:46<2:20:13,  5.85s/it]2026-04-12 00:59:37,840 | INFO | train_grpo_train | metrics_logged mode=train step=1862

                                                       
{'loss': 0.0047, 'grad_norm': 1.0823556184768677, 'learning_rate': 4.36060606060606e-06, 'num_tokens': 4193661.0, 'completions/mean_length': 221.875, 'completions/min_length': 221.0, 'completions/max_length': 225.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 221.875, 'completions/min_terminated_length': 221.0, 'completions/max_terminated_length': 225.0, 'rewards/meter/mean': 0.9964575171470642, 'rewards/meter/std': 0.0009934601839631796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6643049716949463, 'rewards/total_composite/std': 0.0006622962537221611, 'reward': 0.6643049716949463, 'reward_std': 0.000662296952214092, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01341662835329771, 'sampling/sampling_logp_difference/max': 9.43514347076416, 'sampling/importance_sampling_ratio/min': 7.986734999576584e-05, 'sampling/importance_sampling_ratio/mean': 1.0002436637878418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03992933081462979, 'clip_ratio/low_mean': 0.0011111111380159855, 'clip_ratio/low_min': 0.0011111111380159855, 'clip_ratio/high_mean': 0.005077758920378983, 'clip_ratio/high_max': 0.005077758920378983, 'clip_ratio/region_mean': 0.0061888700583949685, 'reward_total_mean': 0.6643049716949463, 'reward_meter_mean': 0.9964575171470642, 'reward_meter_std': 0.0009934601839631796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6643049716949463, 'reward_total_composite_std': 0.0006622962537221611, 'epoch': 0.07}

 56%|█████▋    | 1862/3300 [2:22:46<2:20:13,  5.85s/it]INFO 04-12 00:59:38 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1863/3300 [2:22:51<2:11:44,  5.50s/it]2026-04-12 00:59:42,521 | INFO | train_grpo_train | metrics_logged mode=train step=1863

                                                       
{'loss': 0.0087, 'grad_norm': 5.034853458404541, 'learning_rate': 4.3575757575757576e-06, 'num_tokens': 4195615.0, 'completions/mean_length': 73.25, 'completions/min_length': 67.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9600696563720703, 'rewards/meter/std': 0.09926661849021912, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9600696563720703, 'rewards/total_composite/std': 0.09926661849021912, 'reward': 0.9600696563720703, 'reward_std': 0.09926661849021912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07004112005233765, 'sampling/sampling_logp_difference/max': 2.483992576599121, 'sampling/importance_sampling_ratio/min': 0.08340954035520554, 'sampling/importance_sampling_ratio/mean': 1.0097596645355225, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6060812920331955, 'clip_ratio/low_mean': 0.006666666828095913, 'clip_ratio/low_min': 0.006666666828095913, 'clip_ratio/high_mean': 0.05100303632207215, 'clip_ratio/high_max': 0.05100303632207215, 'clip_ratio/region_mean': 0.05766970315016806, 'reward_total_mean': 0.9600696563720703, 'reward_meter_mean': 0.9600696563720703, 'reward_meter_std': 0.09926661849021912, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9600696563720703, 'reward_total_composite_std': 0.09926661849021912, 'epoch': 0.07}

 56%|█████▋    | 1863/3300 [2:22:51<2:11:44,  5.50s/it]INFO 04-12 00:59:42 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1864/3300 [2:22:56<2:06:18,  5.28s/it]2026-04-12 00:59:47,279 | INFO | train_grpo_train | metrics_logged mode=train step=1864

                                                       
{'loss': -0.0061, 'grad_norm': 1.9747538566589355, 'learning_rate': 4.354545454545455e-06, 'num_tokens': 4197677.0, 'completions/mean_length': 89.75, 'completions/min_length': 88.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.75, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9951763153076172, 'rewards/meter/std': 0.0005524771986529231, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9951763153076172, 'rewards/total_composite/std': 0.0005524771986529231, 'reward': 0.9951763153076172, 'reward_std': 0.0005524645675905049, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008235161192715168, 'sampling/sampling_logp_difference/max': 1.078155517578125, 'sampling/importance_sampling_ratio/min': 0.34022247791290283, 'sampling/importance_sampling_ratio/mean': 1.0047131776809692, 'sampling/importance_sampling_ratio/max': 1.4608083963394165, 'entropy': 0.05834774626418948, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0013888889225199819, 'clip_ratio/high_max': 0.0013888889225199819, 'clip_ratio/region_mean': 0.0013888889225199819, 'reward_total_mean': 0.9951763153076172, 'reward_meter_mean': 0.9951763153076172, 'reward_meter_std': 0.0005524771986529231, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9951763153076172, 'reward_total_composite_std': 0.0005524771986529231, 'epoch': 0.07}

 56%|█████▋    | 1864/3300 [2:22:56<2:06:18,  5.28s/it]INFO 04-12 00:59:47 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1865/3300 [2:23:00<2:01:44,  5.09s/it]2026-04-12 00:59:51,931 | INFO | train_grpo_train | metrics_logged mode=train step=1865

                                                       
{'loss': 0.0111, 'grad_norm': 3.9621639251708984, 'learning_rate': 4.351515151515152e-06, 'num_tokens': 4199512.0, 'completions/mean_length': 69.375, 'completions/min_length': 69.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9972849488258362, 'rewards/meter/std': 0.0004286824550945312, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972849488258362, 'rewards/total_composite/std': 0.0004286824550945312, 'reward': 0.9972849488258362, 'reward_std': 0.000428676517913118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015396162867546082, 'sampling/sampling_logp_difference/max': 0.8150291442871094, 'sampling/importance_sampling_ratio/min': 0.4426264464855194, 'sampling/importance_sampling_ratio/mean': 1.003419280052185, 'sampling/importance_sampling_ratio/max': 1.5455111265182495, 'entropy': 0.12075466383248568, 'clip_ratio/low_mean': 0.010716472752392292, 'clip_ratio/low_min': 0.010716472752392292, 'clip_ratio/high_mean': 0.0036231884732842445, 'clip_ratio/high_max': 0.0036231884732842445, 'clip_ratio/region_mean': 0.014339661225676537, 'reward_total_mean': 0.9972849488258362, 'reward_meter_mean': 0.9972849488258362, 'reward_meter_std': 0.0004286824550945312, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972849488258362, 'reward_total_composite_std': 0.0004286824550945312, 'epoch': 0.07}

 57%|█████▋    | 1865/3300 [2:23:00<2:01:44,  5.09s/it]INFO 04-12 00:59:52 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1866/3300 [2:23:05<2:01:01,  5.06s/it]2026-04-12 00:59:56,931 | INFO | train_grpo_train | metrics_logged mode=train step=1866

                                                       
{'loss': 0.0048, 'grad_norm': 1.6598762273788452, 'learning_rate': 4.348484848484849e-06, 'num_tokens': 4201587.0, 'completions/mean_length': 103.375, 'completions/min_length': 103.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.375, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9974133968353271, 'rewards/meter/std': 0.00014544985606335104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974133968353271, 'rewards/total_composite/std': 0.00014544985606335104, 'reward': 0.9974133968353271, 'reward_std': 0.00014544471923727542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009199890308082104, 'sampling/sampling_logp_difference/max': 0.9815373420715332, 'sampling/importance_sampling_ratio/min': 0.3747345805168152, 'sampling/importance_sampling_ratio/mean': 0.9997120499610901, 'sampling/importance_sampling_ratio/max': 1.4277647733688354, 'entropy': 0.03981838142499328, 'clip_ratio/low_mean': 0.0036407767329365015, 'clip_ratio/low_min': 0.0036407767329365015, 'clip_ratio/high_mean': 0.0036407767329365015, 'clip_ratio/high_max': 0.0036407767329365015, 'clip_ratio/region_mean': 0.007281553465873003, 'reward_total_mean': 0.9974133968353271, 'reward_meter_mean': 0.9974133968353271, 'reward_meter_std': 0.00014544985606335104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974133968353271, 'reward_total_composite_std': 0.00014544985606335104, 'epoch': 0.07}

 57%|█████▋    | 1866/3300 [2:23:05<2:01:01,  5.06s/it]INFO 04-12 00:59:57 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1867/3300 [2:23:14<2:25:15,  6.08s/it]2026-04-12 01:00:05,391 | INFO | train_grpo_train | metrics_logged mode=train step=1867

                                                       
{'loss': 0.0447, 'grad_norm': 2.981818675994873, 'learning_rate': 4.345454545454546e-06, 'num_tokens': 4206088.0, 'completions/mean_length': 356.625, 'completions/min_length': 331.0, 'completions/max_length': 391.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 356.625, 'completions/min_terminated_length': 331.0, 'completions/max_terminated_length': 391.0, 'rewards/meter/mean': 0.904925525188446, 'rewards/meter/std': 0.24405907094478607, 'rewards/count_adherence/mean': 0.6333333253860474, 'rewards/count_adherence/std': 0.035634830594062805, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8377193212509155, 'rewards/repeat_penalty/std': 0.11382605135440826, 'rewards/total_composite/mean': 0.46806132793426514, 'rewards/total_composite/std': 0.1208736002445221, 'reward': 0.46806132793426514, 'reward_std': 0.1208736002445221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05986325070261955, 'sampling/sampling_logp_difference/max': 14.691458702087402, 'sampling/importance_sampling_ratio/min': 4.1646697468422644e-07, 'sampling/importance_sampling_ratio/mean': 1.0076172351837158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6004703845828772, 'clip_ratio/low_mean': 0.01117079914547503, 'clip_ratio/low_min': 0.01117079914547503, 'clip_ratio/high_mean': 0.01905520213767886, 'clip_ratio/high_max': 0.01905520213767886, 'clip_ratio/region_mean': 0.03022600128315389, 'reward_total_mean': 0.46806132793426514, 'reward_meter_mean': 0.904925525188446, 'reward_meter_std': 0.24405907094478607, 'reward_count_adherence_mean': 0.6333333253860474, 'reward_count_adherence_std': 0.035634830594062805, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8377193212509155, 'reward_repeat_penalty_std': 0.11382605135440826, 'reward_total_composite_mean': 0.46806132793426514, 'reward_total_composite_std': 0.1208736002445221, 'epoch': 0.07}

 57%|█████▋    | 1867/3300 [2:23:14<2:25:15,  6.08s/it]INFO 04-12 01:00:05 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1868/3300 [2:23:19<2:16:19,  5.71s/it]2026-04-12 01:00:10,233 | INFO | train_grpo_train | metrics_logged mode=train step=1868

                                                       
{'loss': -0.0004, 'grad_norm': 0.8906328678131104, 'learning_rate': 4.342424242424243e-06, 'num_tokens': 4208151.0, 'completions/mean_length': 89.875, 'completions/min_length': 88.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.875, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.9952363967895508, 'rewards/meter/std': 0.00016089908604044467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952363967895508, 'rewards/total_composite/std': 0.00016089908604044467, 'reward': 0.9952363967895508, 'reward_std': 0.0001608873571967706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009152467362582684, 'sampling/sampling_logp_difference/max': 0.7741870880126953, 'sampling/importance_sampling_ratio/min': 0.4610784351825714, 'sampling/importance_sampling_ratio/mean': 1.002822756767273, 'sampling/importance_sampling_ratio/max': 1.6893128156661987, 'entropy': 0.06600452493876219, 'clip_ratio/low_mean': 0.0028093435103073716, 'clip_ratio/low_min': 0.0028093435103073716, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0028093435103073716, 'reward_total_mean': 0.9952363967895508, 'reward_meter_mean': 0.9952363967895508, 'reward_meter_std': 0.00016089908604044467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952363967895508, 'reward_total_composite_std': 0.00016089908604044467, 'epoch': 0.08}

 57%|█████▋    | 1868/3300 [2:23:19<2:16:19,  5.71s/it]INFO 04-12 01:00:10 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1869/3300 [2:23:24<2:11:07,  5.50s/it]2026-04-12 01:00:15,239 | INFO | train_grpo_train | metrics_logged mode=train step=1869

                                                       
{'loss': 0.0019, 'grad_norm': 2.3382408618927, 'learning_rate': 4.33939393939394e-06, 'num_tokens': 4210238.0, 'completions/mean_length': 102.875, 'completions/min_length': 102.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.875, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9974008798599243, 'rewards/meter/std': 0.00030872138449922204, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974008798599243, 'rewards/total_composite/std': 0.00030872138449922204, 'reward': 0.9974008798599243, 'reward_std': 0.00030873037758283317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009639319032430649, 'sampling/sampling_logp_difference/max': 1.461493968963623, 'sampling/importance_sampling_ratio/min': 0.23188959062099457, 'sampling/importance_sampling_ratio/mean': 0.9990761280059814, 'sampling/importance_sampling_ratio/max': 1.2280147075653076, 'entropy': 0.03081246791407466, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0012135922443121672, 'clip_ratio/high_max': 0.0012135922443121672, 'clip_ratio/region_mean': 0.0012135922443121672, 'reward_total_mean': 0.9974008798599243, 'reward_meter_mean': 0.9974008798599243, 'reward_meter_std': 0.00030872138449922204, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974008798599243, 'reward_total_composite_std': 0.00030872138449922204, 'epoch': 0.08}

 57%|█████▋    | 1869/3300 [2:23:24<2:11:07,  5.50s/it]INFO 04-12 01:00:15 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1870/3300 [2:23:28<2:06:14,  5.30s/it]2026-04-12 01:00:20,060 | INFO | train_grpo_train | metrics_logged mode=train step=1870

                                                       
{'loss': -0.0106, 'grad_norm': 4.046680927276611, 'learning_rate': 4.336363636363637e-06, 'num_tokens': 4212193.0, 'completions/mean_length': 76.375, 'completions/min_length': 70.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9944717288017273, 'rewards/meter/std': 0.003807853441685438, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9944717288017273, 'rewards/total_composite/std': 0.003807853441685438, 'reward': 0.9944717288017273, 'reward_std': 0.003807859495282173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04645724594593048, 'sampling/sampling_logp_difference/max': 1.8717412948608398, 'sampling/importance_sampling_ratio/min': 0.15385551750659943, 'sampling/importance_sampling_ratio/mean': 1.011504054069519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4157518967986107, 'clip_ratio/low_mean': 0.02176366886124015, 'clip_ratio/low_min': 0.02176366886124015, 'clip_ratio/high_mean': 0.011222697328776121, 'clip_ratio/high_max': 0.011222697328776121, 'clip_ratio/region_mean': 0.03298636619001627, 'reward_total_mean': 0.9944717288017273, 'reward_meter_mean': 0.9944717288017273, 'reward_meter_std': 0.003807853441685438, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9944717288017273, 'reward_total_composite_std': 0.003807853441685438, 'epoch': 0.08}

 57%|█████▋    | 1870/3300 [2:23:29<2:06:14,  5.30s/it]INFO 04-12 01:00:20 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1871/3300 [2:23:33<2:00:55,  5.08s/it]2026-04-12 01:00:24,636 | INFO | train_grpo_train | metrics_logged mode=train step=1871

                                                       
{'loss': -0.0038, 'grad_norm': 2.565702438354492, 'learning_rate': 4.333333333333334e-06, 'num_tokens': 4213938.0, 'completions/mean_length': 59.125, 'completions/min_length': 58.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9952658414840698, 'rewards/meter/std': 0.0007007194799371064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952658414840698, 'rewards/total_composite/std': 0.0007007194799371064, 'reward': 0.9952658414840698, 'reward_std': 0.000700704287737608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013889300636947155, 'sampling/sampling_logp_difference/max': 0.7738056182861328, 'sampling/importance_sampling_ratio/min': 0.461254358291626, 'sampling/importance_sampling_ratio/mean': 0.996799886226654, 'sampling/importance_sampling_ratio/max': 1.3448500633239746, 'entropy': 0.057940175756812096, 'clip_ratio/low_mean': 0.0021186440717428923, 'clip_ratio/low_min': 0.0021186440717428923, 'clip_ratio/high_mean': 0.004167824285104871, 'clip_ratio/high_max': 0.004167824285104871, 'clip_ratio/region_mean': 0.006286468356847763, 'reward_total_mean': 0.9952658414840698, 'reward_meter_mean': 0.9952658414840698, 'reward_meter_std': 0.0007007194799371064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952658414840698, 'reward_total_composite_std': 0.0007007194799371064, 'epoch': 0.08}

 57%|█████▋    | 1871/3300 [2:23:33<2:00:55,  5.08s/it]INFO 04-12 01:00:24 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1872/3300 [2:23:43<2:33:03,  6.43s/it]2026-04-12 01:00:34,221 | INFO | train_grpo_train | metrics_logged mode=train step=1872

                                                       
{'loss': -0.3047, 'grad_norm': 1.4247676134109497, 'learning_rate': 4.330303030303031e-06, 'num_tokens': 4216745.0, 'completions/mean_length': 294.875, 'completions/min_length': 212.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 222.5, 'completions/min_terminated_length': 212.0, 'completions/max_terminated_length': 233.0, 'rewards/meter/mean': 0.8686028718948364, 'rewards/meter/std': 0.3345178961753845, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 0.9460227489471436, 'rewards/repeat_penalty/std': 0.06318090111017227, 'rewards/total_composite/mean': 0.6971731185913086, 'rewards/total_composite/std': 0.433406800031662, 'reward': 0.6971731185913086, 'reward_std': 0.433406800031662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0717155784368515, 'sampling/sampling_logp_difference/max': 1.2843422889709473, 'sampling/importance_sampling_ratio/min': 0.27683258056640625, 'sampling/importance_sampling_ratio/mean': 1.018416404724121, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5619211308658123, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03716663923114538, 'clip_ratio/high_max': 0.03716663923114538, 'clip_ratio/region_mean': 0.03716663923114538, 'reward_total_mean': 0.6971731185913086, 'reward_meter_mean': 0.8686028718948364, 'reward_meter_std': 0.3345178961753845, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 0.9460227489471436, 'reward_repeat_penalty_std': 0.06318090111017227, 'reward_total_composite_mean': 0.6971731185913086, 'reward_total_composite_std': 0.433406800031662, 'epoch': 0.08}

 57%|█████▋    | 1872/3300 [2:23:43<2:33:03,  6.43s/it]INFO 04-12 01:00:34 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1873/3300 [2:23:52<2:55:49,  7.39s/it]2026-04-12 01:00:43,858 | INFO | train_grpo_train | metrics_logged mode=train step=1873

                                                       
{'loss': -0.2461, 'grad_norm': 1.2419461011886597, 'learning_rate': 4.327272727272728e-06, 'num_tokens': 4219511.0, 'completions/mean_length': 235.75, 'completions/min_length': 190.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 196.2857208251953, 'completions/min_terminated_length': 190.0, 'completions/max_terminated_length': 203.0, 'rewards/meter/mean': 0.8880919814109802, 'rewards/meter/std': 0.2659795582294464, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9611110687255859, 'rewards/repeat_penalty/std': 0.07582584023475647, 'rewards/total_composite/mean': 0.851571798324585, 'rewards/total_composite/std': 0.26690804958343506, 'reward': 0.851571798324585, 'reward_std': 0.26690801978111267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058800045400857925, 'sampling/sampling_logp_difference/max': 1.3028392791748047, 'sampling/importance_sampling_ratio/min': 0.2717590928077698, 'sampling/importance_sampling_ratio/mean': 1.0172072649002075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5523100271821022, 'clip_ratio/low_mean': 0.0032552082557231188, 'clip_ratio/low_min': 0.0032552082557231188, 'clip_ratio/high_mean': 0.02799722831696272, 'clip_ratio/high_max': 0.02799722831696272, 'clip_ratio/region_mean': 0.03125243657268584, 'reward_total_mean': 0.851571798324585, 'reward_meter_mean': 0.8880919814109802, 'reward_meter_std': 0.2659795582294464, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9611110687255859, 'reward_repeat_penalty_std': 0.07582584023475647, 'reward_total_composite_mean': 0.851571798324585, 'reward_total_composite_std': 0.26690804958343506, 'epoch': 0.08}

 57%|█████▋    | 1873/3300 [2:23:52<2:55:49,  7.39s/it]INFO 04-12 01:00:44 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1874/3300 [2:23:58<2:40:12,  6.74s/it]2026-04-12 01:00:49,077 | INFO | train_grpo_train | metrics_logged mode=train step=1874

                                                       
{'loss': -0.0071, 'grad_norm': 6.65327787399292, 'learning_rate': 4.324242424242425e-06, 'num_tokens': 4221629.0, 'completions/mean_length': 105.75, 'completions/min_length': 99.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.75, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9625563621520996, 'rewards/meter/std': 0.03560182824730873, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9133010506629944, 'rewards/total_composite/std': 0.08394214510917664, 'reward': 0.9133010506629944, 'reward_std': 0.08394213020801544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06265782564878464, 'sampling/sampling_logp_difference/max': 1.7746086120605469, 'sampling/importance_sampling_ratio/min': 0.16954979300498962, 'sampling/importance_sampling_ratio/mean': 0.9978815317153931, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40878177247941494, 'clip_ratio/low_mean': 0.012383450288325548, 'clip_ratio/low_min': 0.012383450288325548, 'clip_ratio/high_mean': 0.0333196020219475, 'clip_ratio/high_max': 0.0333196020219475, 'clip_ratio/region_mean': 0.04570305231027305, 'reward_total_mean': 0.9133010506629944, 'reward_meter_mean': 0.9625563621520996, 'reward_meter_std': 0.03560182824730873, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9133010506629944, 'reward_total_composite_std': 0.08394214510917664, 'epoch': 0.08}

 57%|█████▋    | 1874/3300 [2:23:58<2:40:12,  6.74s/it]INFO 04-12 01:00:49 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1875/3300 [2:24:04<2:38:11,  6.66s/it]2026-04-12 01:00:55,548 | INFO | train_grpo_train | metrics_logged mode=train step=1875

                                                       
{'loss': 0.0297, 'grad_norm': 3.1858253479003906, 'learning_rate': 4.321212121212121e-06, 'num_tokens': 4224968.0, 'completions/mean_length': 202.375, 'completions/min_length': 173.0, 'completions/max_length': 221.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 202.375, 'completions/min_terminated_length': 173.0, 'completions/max_terminated_length': 221.0, 'rewards/meter/mean': 0.888024091720581, 'rewards/meter/std': 0.2775907516479492, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9608585834503174, 'rewards/repeat_penalty/std': 0.08052574098110199, 'rewards/total_composite/mean': 0.8535768985748291, 'rewards/total_composite/std': 0.29564857482910156, 'reward': 0.8535768985748291, 'reward_std': 0.29564857482910156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05819166079163551, 'sampling/sampling_logp_difference/max': 1.7184324264526367, 'sampling/importance_sampling_ratio/min': 0.17934706807136536, 'sampling/importance_sampling_ratio/mean': 1.0147325992584229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6360356472432613, 'clip_ratio/low_mean': 0.008985401596873999, 'clip_ratio/low_min': 0.008985401596873999, 'clip_ratio/high_mean': 0.02583647519350052, 'clip_ratio/high_max': 0.02583647519350052, 'clip_ratio/region_mean': 0.03482187679037452, 'reward_total_mean': 0.8535768985748291, 'reward_meter_mean': 0.888024091720581, 'reward_meter_std': 0.2775907516479492, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9608585834503174, 'reward_repeat_penalty_std': 0.08052574098110199, 'reward_total_composite_mean': 0.8535768985748291, 'reward_total_composite_std': 0.29564857482910156, 'epoch': 0.08}

 57%|█████▋    | 1875/3300 [2:24:04<2:38:11,  6.66s/it]INFO 04-12 01:00:55 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1876/3300 [2:24:09<2:24:13,  6.08s/it]2026-04-12 01:01:00,263 | INFO | train_grpo_train | metrics_logged mode=train step=1876

                                                       
{'loss': 0.0171, 'grad_norm': 5.3259077072143555, 'learning_rate': 4.3181818181818185e-06, 'num_tokens': 4227160.0, 'completions/mean_length': 72.0, 'completions/min_length': 69.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.919043779373169, 'rewards/meter/std': 0.19243919849395752, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.919043779373169, 'rewards/total_composite/std': 0.19243919849395752, 'reward': 0.919043779373169, 'reward_std': 0.19243919849395752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04635407403111458, 'sampling/sampling_logp_difference/max': 1.1690726280212402, 'sampling/importance_sampling_ratio/min': 0.310654878616333, 'sampling/importance_sampling_ratio/mean': 1.0002164840698242, 'sampling/importance_sampling_ratio/max': 1.5501129627227783, 'entropy': 0.309705201536417, 'clip_ratio/low_mean': 0.010135134682059288, 'clip_ratio/low_min': 0.010135134682059288, 'clip_ratio/high_mean': 0.034461831324733794, 'clip_ratio/high_max': 0.034461831324733794, 'clip_ratio/region_mean': 0.04459696600679308, 'reward_total_mean': 0.919043779373169, 'reward_meter_mean': 0.919043779373169, 'reward_meter_std': 0.19243919849395752, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.919043779373169, 'reward_total_composite_std': 0.19243919849395752, 'epoch': 0.08}

 57%|█████▋    | 1876/3300 [2:24:09<2:24:13,  6.08s/it]INFO 04-12 01:01:00 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1877/3300 [2:24:15<2:27:36,  6.22s/it]2026-04-12 01:01:06,842 | INFO | train_grpo_train | metrics_logged mode=train step=1877

                                                       
{'loss': -0.0123, 'grad_norm': 2.582993745803833, 'learning_rate': 4.315151515151516e-06, 'num_tokens': 4230422.0, 'completions/mean_length': 204.75, 'completions/min_length': 192.0, 'completions/max_length': 227.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 204.75, 'completions/min_terminated_length': 192.0, 'completions/max_terminated_length': 227.0, 'rewards/meter/mean': 0.9647862911224365, 'rewards/meter/std': 0.05925486981868744, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9597222208976746, 'rewards/repeat_penalty/std': 0.055694278329610825, 'rewards/total_composite/mean': 0.9254087209701538, 'rewards/total_composite/std': 0.07272467762231827, 'reward': 0.9254087209701538, 'reward_std': 0.07272466272115707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05267081782221794, 'sampling/sampling_logp_difference/max': 1.6134209632873535, 'sampling/importance_sampling_ratio/min': 0.1992049664258957, 'sampling/importance_sampling_ratio/mean': 1.0190201997756958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5658724009990692, 'clip_ratio/low_mean': 0.015004753833636642, 'clip_ratio/low_min': 0.015004753833636642, 'clip_ratio/high_mean': 0.015678069554269314, 'clip_ratio/high_max': 0.015678069554269314, 'clip_ratio/region_mean': 0.030682823387905955, 'reward_total_mean': 0.9254087209701538, 'reward_meter_mean': 0.9647862911224365, 'reward_meter_std': 0.05925486981868744, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9597222208976746, 'reward_repeat_penalty_std': 0.055694278329610825, 'reward_total_composite_mean': 0.9254087209701538, 'reward_total_composite_std': 0.07272467762231827, 'epoch': 0.08}

 57%|█████▋    | 1877/3300 [2:24:15<2:27:36,  6.22s/it]INFO 04-12 01:01:07 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1878/3300 [2:24:25<2:54:01,  7.34s/it]2026-04-12 01:01:16,816 | INFO | train_grpo_train | metrics_logged mode=train step=1878

                                                       
{'loss': -0.2576, 'grad_norm': 1.26083242893219, 'learning_rate': 4.312121212121212e-06, 'num_tokens': 4234422.0, 'completions/mean_length': 332.0, 'completions/min_length': 285.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 306.2857360839844, 'completions/min_terminated_length': 285.0, 'completions/max_terminated_length': 319.0, 'rewards/meter/mean': 0.8925167322158813, 'rewards/meter/std': 0.15122567117214203, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.05345224589109421, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.919905424118042, 'rewards/repeat_penalty/std': 0.1122746393084526, 'rewards/total_composite/mean': 0.5566210746765137, 'rewards/total_composite/std': 0.25293970108032227, 'reward': 0.5566210746765137, 'reward_std': 0.25293970108032227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07019904255867004, 'sampling/sampling_logp_difference/max': 14.930951118469238, 'sampling/importance_sampling_ratio/min': 3.277708344739949e-07, 'sampling/importance_sampling_ratio/mean': 1.012736439704895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48640119284391403, 'clip_ratio/low_mean': 0.011000885395333171, 'clip_ratio/low_min': 0.011000885395333171, 'clip_ratio/high_mean': 0.032999717397615314, 'clip_ratio/high_max': 0.032999717397615314, 'clip_ratio/region_mean': 0.044000602792948484, 'reward_total_mean': 0.5566210746765137, 'reward_meter_mean': 0.8925167322158813, 'reward_meter_std': 0.15122567117214203, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.05345224589109421, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.919905424118042, 'reward_repeat_penalty_std': 0.1122746393084526, 'reward_total_composite_mean': 0.5566210746765137, 'reward_total_composite_std': 0.25293970108032227, 'epoch': 0.08}

 57%|█████▋    | 1878/3300 [2:24:25<2:54:01,  7.34s/it]INFO 04-12 01:01:17 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1879/3300 [2:24:35<3:11:13,  8.07s/it]2026-04-12 01:01:26,568 | INFO | train_grpo_train | metrics_logged mode=train step=1879

                                                       
{'loss': -0.2401, 'grad_norm': 0.8991881012916565, 'learning_rate': 4.309090909090909e-06, 'num_tokens': 4236988.0, 'completions/mean_length': 199.75, 'completions/min_length': 146.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 155.1428680419922, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.8689680099487305, 'rewards/meter/std': 0.3431207239627838, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.866482138633728, 'rewards/total_composite/std': 0.3501509428024292, 'reward': 0.866482138633728, 'reward_std': 0.3501509428024292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.052908286452293396, 'sampling/sampling_logp_difference/max': 1.6577634811401367, 'sampling/importance_sampling_ratio/min': 0.19056470692157745, 'sampling/importance_sampling_ratio/mean': 1.0059314966201782, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3837529495358467, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0337692714529112, 'clip_ratio/high_max': 0.0337692714529112, 'clip_ratio/region_mean': 0.0337692714529112, 'reward_total_mean': 0.866482138633728, 'reward_meter_mean': 0.8689680099487305, 'reward_meter_std': 0.3431207239627838, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.866482138633728, 'reward_total_composite_std': 0.3501509428024292, 'epoch': 0.08}

 57%|█████▋    | 1879/3300 [2:24:35<3:11:13,  8.07s/it]INFO 04-12 01:01:26 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1880/3300 [2:24:40<2:47:11,  7.06s/it]2026-04-12 01:01:31,275 | INFO | train_grpo_train | metrics_logged mode=train step=1880

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.306060606060607e-06, 'num_tokens': 4238773.0, 'completions/mean_length': 69.125, 'completions/min_length': 69.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.997582197189331, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997582197189331, 'rewards/total_composite/std': 0.0, 'reward': 0.997582197189331, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0021389967296272516, 'sampling/sampling_logp_difference/max': 0.24272990226745605, 'sampling/importance_sampling_ratio/min': 0.7844833731651306, 'sampling/importance_sampling_ratio/mean': 1.000409722328186, 'sampling/importance_sampling_ratio/max': 1.1089041233062744, 'entropy': 0.013600267469882965, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.997582197189331, 'reward_meter_mean': 0.997582197189331, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997582197189331, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 57%|█████▋    | 1880/3300 [2:24:40<2:47:11,  7.06s/it]INFO 04-12 01:01:31 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1881/3300 [2:24:44<2:29:32,  6.32s/it]2026-04-12 01:01:35,876 | INFO | train_grpo_train | metrics_logged mode=train step=1881

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.303030303030303e-06, 'num_tokens': 4240565.0, 'completions/mean_length': 69.0, 'completions/min_length': 69.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.997582197189331, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997582197189331, 'rewards/total_composite/std': 0.0, 'reward': 0.997582197189331, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.001675679231993854, 'sampling/sampling_logp_difference/max': 0.10361464321613312, 'sampling/importance_sampling_ratio/min': 0.9542538523674011, 'sampling/importance_sampling_ratio/mean': 1.0012025833129883, 'sampling/importance_sampling_ratio/max': 1.1091729402542114, 'entropy': 0.014390837168321013, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.997582197189331, 'reward_meter_mean': 0.997582197189331, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997582197189331, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 57%|█████▋    | 1881/3300 [2:24:44<2:29:32,  6.32s/it]INFO 04-12 01:01:36 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1882/3300 [2:24:49<2:15:37,  5.74s/it]2026-04-12 01:01:40,245 | INFO | train_grpo_train | metrics_logged mode=train step=1882

                                                       
{'loss': 0.0199, 'grad_norm': 2.6807198524475098, 'learning_rate': 4.3e-06, 'num_tokens': 4242138.0, 'completions/mean_length': 42.625, 'completions/min_length': 40.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9950450658798218, 'rewards/meter/std': 0.00384511542506516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9950450658798218, 'rewards/total_composite/std': 0.00384511542506516, 'reward': 0.9950450658798218, 'reward_std': 0.0038451338186860085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03229609876871109, 'sampling/sampling_logp_difference/max': 0.9218476414680481, 'sampling/importance_sampling_ratio/min': 0.39778339862823486, 'sampling/importance_sampling_ratio/mean': 1.0034741163253784, 'sampling/importance_sampling_ratio/max': 1.5244181156158447, 'entropy': 0.2804165482521057, 'clip_ratio/low_mean': 0.005681818351149559, 'clip_ratio/low_min': 0.005681818351149559, 'clip_ratio/high_mean': 0.0058139534667134285, 'clip_ratio/high_max': 0.0058139534667134285, 'clip_ratio/region_mean': 0.011495771817862988, 'reward_total_mean': 0.9950450658798218, 'reward_meter_mean': 0.9950450658798218, 'reward_meter_std': 0.00384511542506516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9950450658798218, 'reward_total_composite_std': 0.00384511542506516, 'epoch': 0.08}

 57%|█████▋    | 1882/3300 [2:24:49<2:15:37,  5.74s/it]INFO 04-12 01:01:40 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1883/3300 [2:24:54<2:15:42,  5.75s/it]2026-04-12 01:01:46,008 | INFO | train_grpo_train | metrics_logged mode=train step=1883

                                                       
{'loss': -0.0012, 'grad_norm': 2.456183433532715, 'learning_rate': 4.296969696969698e-06, 'num_tokens': 4244703.0, 'completions/mean_length': 160.625, 'completions/min_length': 154.0, 'completions/max_length': 166.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.625, 'completions/min_terminated_length': 154.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.9891646504402161, 'rewards/meter/std': 0.004805354867130518, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9540842771530151, 'rewards/total_composite/std': 0.06925938278436661, 'reward': 0.9540842771530151, 'reward_std': 0.06925938278436661, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.037685707211494446, 'sampling/sampling_logp_difference/max': 0.9908924102783203, 'sampling/importance_sampling_ratio/min': 0.37124526500701904, 'sampling/importance_sampling_ratio/mean': 1.0097475051879883, 'sampling/importance_sampling_ratio/max': 1.8647836446762085, 'entropy': 0.38958779722452164, 'clip_ratio/low_mean': 0.007823501946404576, 'clip_ratio/low_min': 0.007823501946404576, 'clip_ratio/high_mean': 0.02487414190545678, 'clip_ratio/high_max': 0.02487414190545678, 'clip_ratio/region_mean': 0.03269764385186136, 'reward_total_mean': 0.9540842771530151, 'reward_meter_mean': 0.9891646504402161, 'reward_meter_std': 0.004805354867130518, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9540842771530151, 'reward_total_composite_std': 0.06925938278436661, 'epoch': 0.08}

 57%|█████▋    | 1883/3300 [2:24:55<2:15:42,  5.75s/it]INFO 04-12 01:01:46 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1884/3300 [2:24:59<2:07:58,  5.42s/it]2026-04-12 01:01:50,674 | INFO | train_grpo_train | metrics_logged mode=train step=1884

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.293939393939394e-06, 'num_tokens': 4246477.0, 'completions/mean_length': 74.75, 'completions/min_length': 73.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9963148832321167, 'rewards/meter/std': 0.0056813484989106655, 'rewards/count_adherence/mean': 0.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.0, 'rewards/total_composite/std': 0.0, 'reward': 0.0, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.04152185842394829, 'sampling/sampling_logp_difference/max': 1.5260066986083984, 'sampling/importance_sampling_ratio/min': 0.21740210056304932, 'sampling/importance_sampling_ratio/mean': 1.003008246421814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2754708882421255, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.0, 'reward_meter_mean': 0.9963148832321167, 'reward_meter_std': 0.0056813484989106655, 'reward_count_adherence_mean': 0.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.0, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 57%|█████▋    | 1884/3300 [2:24:59<2:07:58,  5.42s/it]INFO 04-12 01:01:50 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1885/3300 [2:25:06<2:18:17,  5.86s/it]2026-04-12 01:01:57,572 | INFO | train_grpo_train | metrics_logged mode=train step=1885

                                                       
{'loss': -0.0235, 'grad_norm': 2.4556612968444824, 'learning_rate': 4.290909090909091e-06, 'num_tokens': 4250056.0, 'completions/mean_length': 247.375, 'completions/min_length': 234.0, 'completions/max_length': 260.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 247.375, 'completions/min_terminated_length': 234.0, 'completions/max_terminated_length': 260.0, 'rewards/meter/mean': 0.9978925585746765, 'rewards/meter/std': 0.00047845288645476103, 'rewards/count_adherence/mean': 0.9464285969734192, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8693909645080566, 'rewards/repeat_penalty/std': 0.11218275874853134, 'rewards/total_composite/mean': 0.8244522213935852, 'rewards/total_composite/std': 0.14857427775859833, 'reward': 0.8244522213935852, 'reward_std': 0.14857426285743713, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031580884009599686, 'sampling/sampling_logp_difference/max': 1.2896442413330078, 'sampling/importance_sampling_ratio/min': 0.27536872029304504, 'sampling/importance_sampling_ratio/mean': 1.0049325227737427, 'sampling/importance_sampling_ratio/max': 1.8980882167816162, 'entropy': 0.29367757216095924, 'clip_ratio/low_mean': 0.010505896178074181, 'clip_ratio/low_min': 0.010505896178074181, 'clip_ratio/high_mean': 0.01472888421267271, 'clip_ratio/high_max': 0.01472888421267271, 'clip_ratio/region_mean': 0.02523478039074689, 'reward_total_mean': 0.8244522213935852, 'reward_meter_mean': 0.9978925585746765, 'reward_meter_std': 0.00047845288645476103, 'reward_count_adherence_mean': 0.9464285969734192, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8693909645080566, 'reward_repeat_penalty_std': 0.11218275874853134, 'reward_total_composite_mean': 0.8244522213935852, 'reward_total_composite_std': 0.14857427775859833, 'epoch': 0.08}

 57%|█████▋    | 1885/3300 [2:25:06<2:18:17,  5.86s/it]INFO 04-12 01:01:57 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1886/3300 [2:25:12<2:22:12,  6.03s/it]2026-04-12 01:02:03,998 | INFO | train_grpo_train | metrics_logged mode=train step=1886

                                                       
{'loss': -0.0178, 'grad_norm': 2.436864137649536, 'learning_rate': 4.287878787878788e-06, 'num_tokens': 4253377.0, 'completions/mean_length': 214.125, 'completions/min_length': 204.0, 'completions/max_length': 222.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 214.125, 'completions/min_terminated_length': 204.0, 'completions/max_terminated_length': 222.0, 'rewards/meter/mean': 0.9979674220085144, 'rewards/meter/std': 0.00041360704926773906, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8818181753158569, 'rewards/repeat_penalty/std': 0.07008175551891327, 'rewards/total_composite/mean': 0.8467762470245361, 'rewards/total_composite/std': 0.12215393781661987, 'reward': 0.8467762470245361, 'reward_std': 0.12215393036603928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03404805809259415, 'sampling/sampling_logp_difference/max': 2.1384077072143555, 'sampling/importance_sampling_ratio/min': 0.11784233152866364, 'sampling/importance_sampling_ratio/mean': 1.0003207921981812, 'sampling/importance_sampling_ratio/max': 1.72610342502594, 'entropy': 0.27855938114225864, 'clip_ratio/low_mean': 0.00671045109629631, 'clip_ratio/low_min': 0.00671045109629631, 'clip_ratio/high_mean': 0.020219704834744334, 'clip_ratio/high_max': 0.020219704834744334, 'clip_ratio/region_mean': 0.026930155931040645, 'reward_total_mean': 0.8467762470245361, 'reward_meter_mean': 0.9979674220085144, 'reward_meter_std': 0.00041360704926773906, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8818181753158569, 'reward_repeat_penalty_std': 0.07008175551891327, 'reward_total_composite_mean': 0.8467762470245361, 'reward_total_composite_std': 0.12215393781661987, 'epoch': 0.08}

 57%|█████▋    | 1886/3300 [2:25:12<2:22:12,  6.03s/it]INFO 04-12 01:02:04 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1887/3300 [2:25:18<2:18:12,  5.87s/it]2026-04-12 01:02:09,491 | INFO | train_grpo_train | metrics_logged mode=train step=1887

                                                       
{'loss': -0.0062, 'grad_norm': 2.4621639251708984, 'learning_rate': 4.284848484848485e-06, 'num_tokens': 4255950.0, 'completions/mean_length': 144.625, 'completions/min_length': 141.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 144.625, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.997658908367157, 'rewards/meter/std': 0.0006960767204873264, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.10101525485515594, 'rewards/total_composite/mean': 0.8907864093780518, 'rewards/total_composite/std': 0.10098132491111755, 'reward': 0.8907864093780518, 'reward_std': 0.10098133236169815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02850562147796154, 'sampling/sampling_logp_difference/max': 1.2042560577392578, 'sampling/importance_sampling_ratio/min': 0.29991504549980164, 'sampling/importance_sampling_ratio/mean': 1.0084306001663208, 'sampling/importance_sampling_ratio/max': 1.7592705488204956, 'entropy': 0.25286005809903145, 'clip_ratio/low_mean': 0.01478908269200474, 'clip_ratio/low_min': 0.01478908269200474, 'clip_ratio/high_mean': 0.006850266130641103, 'clip_ratio/high_max': 0.006850266130641103, 'clip_ratio/region_mean': 0.021639348822645843, 'reward_total_mean': 0.8907864093780518, 'reward_meter_mean': 0.997658908367157, 'reward_meter_std': 0.0006960767204873264, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.10101525485515594, 'reward_total_composite_mean': 0.8907864093780518, 'reward_total_composite_std': 0.10098132491111755, 'epoch': 0.08}

 57%|█████▋    | 1887/3300 [2:25:18<2:18:12,  5.87s/it]INFO 04-12 01:02:09 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1888/3300 [2:25:25<2:30:06,  6.38s/it]2026-04-12 01:02:17,051 | INFO | train_grpo_train | metrics_logged mode=train step=1888

                                                       
{'loss': -0.0096, 'grad_norm': 2.372718334197998, 'learning_rate': 4.281818181818182e-06, 'num_tokens': 4259851.0, 'completions/mean_length': 296.625, 'completions/min_length': 279.0, 'completions/max_length': 317.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 296.625, 'completions/min_terminated_length': 279.0, 'completions/max_terminated_length': 317.0, 'rewards/meter/mean': 0.9935998320579529, 'rewards/meter/std': 0.005851938389241695, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8760073184967041, 'rewards/repeat_penalty/std': 0.07234964519739151, 'rewards/total_composite/mean': 0.7886214256286621, 'rewards/total_composite/std': 0.07851462066173553, 'reward': 0.7886214256286621, 'reward_std': 0.07851462066173553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034045975655317307, 'sampling/sampling_logp_difference/max': 2.5086705684661865, 'sampling/importance_sampling_ratio/min': 0.08137635141611099, 'sampling/importance_sampling_ratio/mean': 1.0055367946624756, 'sampling/importance_sampling_ratio/max': 1.7818008661270142, 'entropy': 0.28176482766866684, 'clip_ratio/low_mean': 0.005528143374249339, 'clip_ratio/low_min': 0.005528143374249339, 'clip_ratio/high_mean': 0.019573589437641203, 'clip_ratio/high_max': 0.019573589437641203, 'clip_ratio/region_mean': 0.025101732811890543, 'reward_total_mean': 0.7886214256286621, 'reward_meter_mean': 0.9935998320579529, 'reward_meter_std': 0.005851938389241695, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8760073184967041, 'reward_repeat_penalty_std': 0.07234964519739151, 'reward_total_composite_mean': 0.7886214256286621, 'reward_total_composite_std': 0.07851462066173553, 'epoch': 0.08}

 57%|█████▋    | 1888/3300 [2:25:26<2:30:06,  6.38s/it]INFO 04-12 01:02:17 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1889/3300 [2:25:33<2:40:10,  6.81s/it]2026-04-12 01:02:24,873 | INFO | train_grpo_train | metrics_logged mode=train step=1889

                                                       
{'loss': -0.0161, 'grad_norm': 1.960667610168457, 'learning_rate': 4.278787878787879e-06, 'num_tokens': 4264323.0, 'completions/mean_length': 328.0, 'completions/min_length': 305.0, 'completions/max_length': 341.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 328.0, 'completions/min_terminated_length': 305.0, 'completions/max_terminated_length': 341.0, 'rewards/meter/mean': 0.9920902252197266, 'rewards/meter/std': 0.005891723092645407, 'rewards/count_adherence/mean': 0.6057692766189575, 'rewards/count_adherence/std': 0.027196412906050682, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9280506372451782, 'rewards/repeat_penalty/std': 0.05198076739907265, 'rewards/total_composite/mean': 0.5576076507568359, 'rewards/total_composite/std': 0.03838847950100899, 'reward': 0.5576076507568359, 'reward_std': 0.03838847950100899, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03868624567985535, 'sampling/sampling_logp_difference/max': 2.517657995223999, 'sampling/importance_sampling_ratio/min': 0.0806482657790184, 'sampling/importance_sampling_ratio/mean': 1.0100011825561523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35988450422883034, 'clip_ratio/low_mean': 0.010531234613154083, 'clip_ratio/low_min': 0.010531234613154083, 'clip_ratio/high_mean': 0.02012083982117474, 'clip_ratio/high_max': 0.02012083982117474, 'clip_ratio/region_mean': 0.030652074434328824, 'reward_total_mean': 0.5576076507568359, 'reward_meter_mean': 0.9920902252197266, 'reward_meter_std': 0.005891723092645407, 'reward_count_adherence_mean': 0.6057692766189575, 'reward_count_adherence_std': 0.027196412906050682, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9280506372451782, 'reward_repeat_penalty_std': 0.05198076739907265, 'reward_total_composite_mean': 0.5576076507568359, 'reward_total_composite_std': 0.03838847950100899, 'epoch': 0.08}

 57%|█████▋    | 1889/3300 [2:25:33<2:40:10,  6.81s/it]INFO 04-12 01:02:25 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1890/3300 [2:25:38<2:23:25,  6.10s/it]2026-04-12 01:02:29,321 | INFO | train_grpo_train | metrics_logged mode=train step=1890

                                                       
{'loss': -0.0005, 'grad_norm': 1.555624008178711, 'learning_rate': 4.275757575757576e-06, 'num_tokens': 4266105.0, 'completions/mean_length': 62.75, 'completions/min_length': 61.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.75, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9969385862350464, 'rewards/meter/std': 0.00011288504174444824, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969385862350464, 'rewards/total_composite/std': 0.00011288504174444824, 'reward': 0.9969385862350464, 'reward_std': 0.0001128727599279955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005147209390997887, 'sampling/sampling_logp_difference/max': 0.49872732162475586, 'sampling/importance_sampling_ratio/min': 0.7584096193313599, 'sampling/importance_sampling_ratio/mean': 1.002489447593689, 'sampling/importance_sampling_ratio/max': 1.6466243267059326, 'entropy': 0.02277997531928122, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.003968254197388887, 'clip_ratio/high_max': 0.003968254197388887, 'clip_ratio/region_mean': 0.006017434410750866, 'reward_total_mean': 0.9969385862350464, 'reward_meter_mean': 0.9969385862350464, 'reward_meter_std': 0.00011288504174444824, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969385862350464, 'reward_total_composite_std': 0.00011288504174444824, 'epoch': 0.08}

 57%|█████▋    | 1890/3300 [2:25:38<2:23:25,  6.10s/it]INFO 04-12 01:02:29 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1891/3300 [2:25:42<2:10:41,  5.57s/it]2026-04-12 01:02:33,638 | INFO | train_grpo_train | metrics_logged mode=train step=1891

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.272727272727273e-06, 'num_tokens': 4267721.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9946010708808899, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946010708808899, 'rewards/total_composite/std': 0.0, 'reward': 0.9946010708808899, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0011367781553417444, 'sampling/sampling_logp_difference/max': 0.0915985256433487, 'sampling/importance_sampling_ratio/min': 0.9592980146408081, 'sampling/importance_sampling_ratio/mean': 1.0008796453475952, 'sampling/importance_sampling_ratio/max': 1.095924735069275, 'entropy': 0.011020259640645236, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9946010708808899, 'reward_meter_mean': 0.9946010708808899, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946010708808899, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 57%|█████▋    | 1891/3300 [2:25:42<2:10:41,  5.57s/it]INFO 04-12 01:02:33 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1892/3300 [2:25:46<2:00:51,  5.15s/it]2026-04-12 01:02:37,814 | INFO | train_grpo_train | metrics_logged mode=train step=1892

                                                       
{'loss': 0.0082, 'grad_norm': 4.663846969604492, 'learning_rate': 4.2696969696969695e-06, 'num_tokens': 4269442.0, 'completions/mean_length': 33.125, 'completions/min_length': 33.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9981597661972046, 'rewards/meter/std': 0.0018737530335783958, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981597661972046, 'rewards/total_composite/std': 0.0018737530335783958, 'reward': 0.9981597661972046, 'reward_std': 0.0018737498903647065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011638238094747066, 'sampling/sampling_logp_difference/max': 0.7902803421020508, 'sampling/importance_sampling_ratio/min': 0.4537176191806793, 'sampling/importance_sampling_ratio/mean': 0.9993104338645935, 'sampling/importance_sampling_ratio/max': 1.9366061687469482, 'entropy': 0.02548716589808464, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.007575757801532745, 'clip_ratio/high_max': 0.007575757801532745, 'clip_ratio/region_mean': 0.01125222840346396, 'reward_total_mean': 0.9981597661972046, 'reward_meter_mean': 0.9981597661972046, 'reward_meter_std': 0.0018737530335783958, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981597661972046, 'reward_total_composite_std': 0.0018737530335783958, 'epoch': 0.08}

 57%|█████▋    | 1892/3300 [2:25:46<2:00:51,  5.15s/it]INFO 04-12 01:02:38 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1893/3300 [2:25:52<2:03:25,  5.26s/it]2026-04-12 01:02:43,342 | INFO | train_grpo_train | metrics_logged mode=train step=1893

                                                       
{'loss': 0.012, 'grad_norm': 4.968896865844727, 'learning_rate': 4.266666666666668e-06, 'num_tokens': 4272121.0, 'completions/mean_length': 147.875, 'completions/min_length': 142.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.875, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.997498095035553, 'rewards/meter/std': 0.0016895781736820936, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9796580076217651, 'rewards/total_composite/std': 0.0498599074780941, 'reward': 0.9796580076217651, 'reward_std': 0.049859896302223206, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05557761713862419, 'sampling/sampling_logp_difference/max': 1.7668085098266602, 'sampling/importance_sampling_ratio/min': 0.17087747156620026, 'sampling/importance_sampling_ratio/mean': 1.002869725227356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3808891177177429, 'clip_ratio/low_mean': 0.004999999888241291, 'clip_ratio/low_min': 0.004999999888241291, 'clip_ratio/high_mean': 0.043705128598958254, 'clip_ratio/high_max': 0.043705128598958254, 'clip_ratio/region_mean': 0.048705128487199545, 'reward_total_mean': 0.9796580076217651, 'reward_meter_mean': 0.997498095035553, 'reward_meter_std': 0.0016895781736820936, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9796580076217651, 'reward_total_composite_std': 0.0498599074780941, 'epoch': 0.08}

 57%|█████▋    | 1893/3300 [2:25:52<2:03:25,  5.26s/it]INFO 04-12 01:02:43 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1894/3300 [2:25:58<2:07:58,  5.46s/it]2026-04-12 01:02:49,268 | INFO | train_grpo_train | metrics_logged mode=train step=1894

                                                       
{'loss': 0.01, 'grad_norm': 3.8838560581207275, 'learning_rate': 4.263636363636364e-06, 'num_tokens': 4274755.0, 'completions/mean_length': 184.25, 'completions/min_length': 177.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 184.25, 'completions/min_terminated_length': 177.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9978748559951782, 'rewards/meter/std': 0.0017333579016849399, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.08399210125207901, 'rewards/total_composite/mean': 0.9423543214797974, 'rewards/total_composite/std': 0.08282524347305298, 'reward': 0.9423543214797974, 'reward_std': 0.08282525092363358, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049167633056640625, 'sampling/sampling_logp_difference/max': 2.1750478744506836, 'sampling/importance_sampling_ratio/min': 0.11360271275043488, 'sampling/importance_sampling_ratio/mean': 1.002802848815918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33961667120456696, 'clip_ratio/low_mean': 0.012207416351884604, 'clip_ratio/low_min': 0.012207416351884604, 'clip_ratio/high_mean': 0.03263740334659815, 'clip_ratio/high_max': 0.03263740334659815, 'clip_ratio/region_mean': 0.04484481969848275, 'reward_total_mean': 0.9423543214797974, 'reward_meter_mean': 0.9978748559951782, 'reward_meter_std': 0.0017333579016849399, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.08399210125207901, 'reward_total_composite_mean': 0.9423543214797974, 'reward_total_composite_std': 0.08282524347305298, 'epoch': 0.08}

 57%|█████▋    | 1894/3300 [2:25:58<2:07:58,  5.46s/it]INFO 04-12 01:02:49 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1895/3300 [2:26:04<2:13:05,  5.68s/it]2026-04-12 01:02:55,477 | INFO | train_grpo_train | metrics_logged mode=train step=1895

                                                       
{'loss': -0.0041, 'grad_norm': 2.22912859916687, 'learning_rate': 4.260606060606061e-06, 'num_tokens': 4277761.0, 'completions/mean_length': 200.75, 'completions/min_length': 195.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 200.75, 'completions/min_terminated_length': 195.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.9934834837913513, 'rewards/meter/std': 0.003708732081577182, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.8416043519973755, 'rewards/total_composite/std': 0.34394556283950806, 'reward': 0.8416043519973755, 'reward_std': 0.34394556283950806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.037548985332250595, 'sampling/sampling_logp_difference/max': 1.3242321014404297, 'sampling/importance_sampling_ratio/min': 0.2660071551799774, 'sampling/importance_sampling_ratio/mean': 1.0138887166976929, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3611125461757183, 'clip_ratio/low_mean': 0.0031887756194919348, 'clip_ratio/low_min': 0.0031887756194919348, 'clip_ratio/high_mean': 0.022966360847931355, 'clip_ratio/high_max': 0.022966360847931355, 'clip_ratio/region_mean': 0.02615513646742329, 'reward_total_mean': 0.8416043519973755, 'reward_meter_mean': 0.9934834837913513, 'reward_meter_std': 0.003708732081577182, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.8416043519973755, 'reward_total_composite_std': 0.34394556283950806, 'epoch': 0.08}

 57%|█████▋    | 1895/3300 [2:26:04<2:13:05,  5.68s/it]INFO 04-12 01:02:55 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1896/3300 [2:26:09<2:12:18,  5.65s/it]2026-04-12 01:03:01,053 | INFO | train_grpo_train | metrics_logged mode=train step=1896

                                                       
{'loss': 0.0005, 'grad_norm': 0.0076941619627177715, 'learning_rate': 4.2575757575757585e-06, 'num_tokens': 4280665.0, 'completions/mean_length': 165.0, 'completions/min_length': 165.0, 'completions/max_length': 165.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.0, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.9991617202758789, 'rewards/meter/std': 2.346204610148561e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7771258354187012, 'rewards/total_composite/std': 1.8245253841087106e-06, 'reward': 0.7771258354187012, 'reward_std': 1.8390023797110189e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0018403534777462482, 'sampling/sampling_logp_difference/max': 1.3649120330810547, 'sampling/importance_sampling_ratio/min': 0.25540316104888916, 'sampling/importance_sampling_ratio/mean': 1.0000821352005005, 'sampling/importance_sampling_ratio/max': 1.0792498588562012, 'entropy': 0.007689857040531933, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0007575757335871458, 'clip_ratio/high_max': 0.0007575757335871458, 'clip_ratio/region_mean': 0.0007575757335871458, 'reward_total_mean': 0.7771258354187012, 'reward_meter_mean': 0.9991617202758789, 'reward_meter_std': 2.346204610148561e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7771258354187012, 'reward_total_composite_std': 1.8245253841087106e-06, 'epoch': 0.08}

 57%|█████▋    | 1896/3300 [2:26:10<2:12:18,  5.65s/it]INFO 04-12 01:03:01 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1897/3300 [2:26:15<2:11:24,  5.62s/it]2026-04-12 01:03:06,593 | INFO | train_grpo_train | metrics_logged mode=train step=1897

                                                       
{'loss': 0.0006, 'grad_norm': 4.121653079986572, 'learning_rate': 4.254545454545455e-06, 'num_tokens': 4283357.0, 'completions/mean_length': 152.5, 'completions/min_length': 144.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 152.5, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.9909997582435608, 'rewards/meter/std': 0.008123436011373997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9732140302658081, 'rewards/total_composite/std': 0.04892229288816452, 'reward': 0.9732140302658081, 'reward_std': 0.04892229288816452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04408065602183342, 'sampling/sampling_logp_difference/max': 2.633777618408203, 'sampling/importance_sampling_ratio/min': 0.07180669158697128, 'sampling/importance_sampling_ratio/mean': 1.006994605064392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34401146695017815, 'clip_ratio/low_mean': 0.0025167784187942743, 'clip_ratio/low_min': 0.0025167784187942743, 'clip_ratio/high_mean': 0.04165800241753459, 'clip_ratio/high_max': 0.04165800241753459, 'clip_ratio/region_mean': 0.044174780836328864, 'reward_total_mean': 0.9732140302658081, 'reward_meter_mean': 0.9909997582435608, 'reward_meter_std': 0.008123436011373997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9732140302658081, 'reward_total_composite_std': 0.04892229288816452, 'epoch': 0.08}

 57%|█████▋    | 1897/3300 [2:26:15<2:11:24,  5.62s/it]INFO 04-12 01:03:06 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1898/3300 [2:26:20<2:08:14,  5.49s/it]2026-04-12 01:03:11,773 | INFO | train_grpo_train | metrics_logged mode=train step=1898

                                                       
{'loss': -0.001, 'grad_norm': 2.702057361602783, 'learning_rate': 4.251515151515152e-06, 'num_tokens': 4285658.0, 'completions/mean_length': 119.625, 'completions/min_length': 116.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.625, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9938209652900696, 'rewards/meter/std': 0.002714785048738122, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9938209652900696, 'rewards/total_composite/std': 0.002714785048738122, 'reward': 0.9938209652900696, 'reward_std': 0.0027147859800606966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039888497442007065, 'sampling/sampling_logp_difference/max': 1.169196605682373, 'sampling/importance_sampling_ratio/min': 0.3106164038181305, 'sampling/importance_sampling_ratio/mean': 1.0123997926712036, 'sampling/importance_sampling_ratio/max': 1.5793664455413818, 'entropy': 0.3594598062336445, 'clip_ratio/low_mean': 0.017732840729877353, 'clip_ratio/low_min': 0.017732840729877353, 'clip_ratio/high_mean': 0.01466671982780099, 'clip_ratio/high_max': 0.01466671982780099, 'clip_ratio/region_mean': 0.03239956055767834, 'reward_total_mean': 0.9938209652900696, 'reward_meter_mean': 0.9938209652900696, 'reward_meter_std': 0.002714785048738122, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9938209652900696, 'reward_total_composite_std': 0.002714785048738122, 'epoch': 0.08}

 58%|█████▊    | 1898/3300 [2:26:20<2:08:14,  5.49s/it]INFO 04-12 01:03:12 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1899/3300 [2:26:25<2:01:35,  5.21s/it]2026-04-12 01:03:16,326 | INFO | train_grpo_train | metrics_logged mode=train step=1899

                                                       
{'loss': 0.0146, 'grad_norm': 5.739034175872803, 'learning_rate': 4.248484848484849e-06, 'num_tokens': 4287482.0, 'completions/mean_length': 63.0, 'completions/min_length': 62.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9764682650566101, 'rewards/meter/std': 0.05796706676483154, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9764682650566101, 'rewards/total_composite/std': 0.05796706676483154, 'reward': 0.9764682650566101, 'reward_std': 0.057967059314250946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010971495881676674, 'sampling/sampling_logp_difference/max': 1.089432716369629, 'sampling/importance_sampling_ratio/min': 0.33640727400779724, 'sampling/importance_sampling_ratio/mean': 0.9993007183074951, 'sampling/importance_sampling_ratio/max': 1.515744924545288, 'entropy': 0.05075907311402261, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.002016128972172737, 'clip_ratio/high_max': 0.002016128972172737, 'clip_ratio/region_mean': 0.002016128972172737, 'reward_total_mean': 0.9764682650566101, 'reward_meter_mean': 0.9764682650566101, 'reward_meter_std': 0.05796706676483154, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9764682650566101, 'reward_total_composite_std': 0.05796706676483154, 'epoch': 0.08}

 58%|█████▊    | 1899/3300 [2:26:25<2:01:35,  5.21s/it]INFO 04-12 01:03:16 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1900/3300 [2:26:30<1:59:09,  5.11s/it]2026-04-12 01:03:21,203 | INFO | train_grpo_train | metrics_logged mode=train step=1900

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.245454545454546e-06, 'num_tokens': 4289610.0, 'completions/mean_length': 99.0, 'completions/min_length': 99.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.0, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9990620613098145, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990620613098145, 'rewards/total_composite/std': 0.0, 'reward': 0.9990620613098145, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0004978616489097476, 'sampling/sampling_logp_difference/max': 0.014289772137999535, 'sampling/importance_sampling_ratio/min': 0.9927101135253906, 'sampling/importance_sampling_ratio/mean': 1.00044846534729, 'sampling/importance_sampling_ratio/max': 1.0143922567367554, 'entropy': 0.004198343900498003, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990620613098145, 'reward_meter_mean': 0.9990620613098145, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990620613098145, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 58%|█████▊    | 1900/3300 [2:26:30<1:59:09,  5.11s/it]INFO 04-12 01:03:21 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:28,  2.59s/it]

 23%|██▎       | 3/13 [00:08<00:28,  2.87s/it]

 31%|███       | 4/13 [00:13<00:32,  3.57s/it]

 38%|███▊      | 5/13 [00:18<00:32,  4.02s/it]

 46%|████▌     | 6/13 [00:23<00:31,  4.46s/it]

 54%|█████▍    | 7/13 [00:28<00:27,  4.60s/it]

 62%|██████▏   | 8/13 [00:34<00:24,  4.97s/it]

 69%|██████▉   | 9/13 [00:39<00:20,  5.13s/it]

 77%|███████▋  | 10/13 [00:44<00:14,  4.95s/it]

 85%|████████▍ | 11/13 [00:49<00:09,  5.00s/it]

 92%|█████████▏| 12/13 [00:54<00:04,  4.94s/it]

100%|██████████| 13/13 [00:58<00:00,  4.90s/it]2026-04-12 01:04:25,446 | INFO | train_grpo_train | metrics_logged mode=eval step=1900

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 64.1724, 'eval_samples_per_second': 1.621, 'eval_steps_per_second': 0.203, 'eval_num_tokens': 4289610.0, 'eval_completions/mean_length': 199.51923076923077, 'eval_completions/min_length': 71.53846153846153, 'eval_completions/max_length': 335.6923076923077, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 199.51923076923077, 'eval_completions/min_terminated_length': 71.53846153846153, 'eval_completions/max_terminated_length': 335.6923076923077, 'eval_rewards/meter/mean': 0.7173460401021517, 'eval_rewards/meter/std': 0.4227825838785905, 'eval_rewards/count_adherence/mean': 0.8646397361388574, 'eval_rewards/count_adherence/std': 0.19298870517657354, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.8317501590802119, 'eval_rewards/repeat_penalty/std': 0.14178751237117326, 'eval_rewards/total_composite/mean': 0.5190799373846787, 'eval_rewards/total_composite/std': 0.36580440631279576, 'eval_reward': 0.5190799373846787, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.014502167307700101, 'eval_sampling/sampling_logp_difference/max': 0.989627324617826, 'eval_sampling/importance_sampling_ratio/min': 0.3849313259124756, 'eval_sampling/importance_sampling_ratio/mean': 1.0033589509817271, 'eval_sampling/importance_sampling_ratio/max': 1.388241483614995, 'eval_entropy': 0.15330308188612646, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5190799373846787, 'eval_reward_meter_mean': 0.7173460401021517, 'eval_reward_meter_std': 0.4227825838785905, 'eval_reward_count_adherence_mean': 0.8646397361388574, 'eval_reward_count_adherence_std': 0.19298870517657354, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.8317501590802119, 'eval_reward_repeat_penalty_std': 0.14178751237117326, 'eval_reward_total_composite_mean': 0.5190799373846787, 'eval_reward_total_composite_std': 0.36580440631279576, 'epoch': 0.08}

 58%|█████▊    | 1900/3300 [2:27:34<1:59:09,  5.11s/it]

100%|██████████| 13/13 [00:58<00:00,  4.90s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 58%|█████▊    | 1901/3300 [2:27:41<9:40:43, 24.91s/it]2026-04-12 01:04:32,305 | INFO | train_grpo_train | metrics_logged mode=train step=1901

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.242424242424243e-06, 'num_tokens': 4291354.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9946010708808899, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946010708808899, 'rewards/total_composite/std': 0.0, 'reward': 0.9946010708808899, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0012767898151651025, 'sampling/sampling_logp_difference/max': 0.17115801572799683, 'sampling/importance_sampling_ratio/min': 0.8426884412765503, 'sampling/importance_sampling_ratio/mean': 1.00002121925354, 'sampling/importance_sampling_ratio/max': 1.0319656133651733, 'entropy': 0.013817269122228026, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9946010708808899, 'reward_meter_mean': 0.9946010708808899, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946010708808899, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 58%|█████▊    | 1901/3300 [2:27:41<9:40:43, 24.91s/it]INFO 04-12 01:04:32 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1902/3300 [2:27:45<7:18:25, 18.82s/it]2026-04-12 01:04:36,917 | INFO | train_grpo_train | metrics_logged mode=train step=1902

                                                       
{'loss': 0.0286, 'grad_norm': 5.961615085601807, 'learning_rate': 4.2393939393939395e-06, 'num_tokens': 4293032.0, 'completions/mean_length': 50.75, 'completions/min_length': 49.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9352545142173767, 'rewards/meter/std': 0.01202213205397129, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9352545142173767, 'rewards/total_composite/std': 0.01202213205397129, 'reward': 0.9352545142173767, 'reward_std': 0.012022126466035843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016524022445082664, 'sampling/sampling_logp_difference/max': 0.9960556030273438, 'sampling/importance_sampling_ratio/min': 0.3693333864212036, 'sampling/importance_sampling_ratio/mean': 0.9972783327102661, 'sampling/importance_sampling_ratio/max': 1.7655236721038818, 'entropy': 0.0905577321536839, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005102040711790323, 'clip_ratio/high_max': 0.005102040711790323, 'clip_ratio/region_mean': 0.005102040711790323, 'reward_total_mean': 0.9352545142173767, 'reward_meter_mean': 0.9352545142173767, 'reward_meter_std': 0.01202213205397129, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9352545142173767, 'reward_total_composite_std': 0.01202213205397129, 'epoch': 0.08}

 58%|█████▊    | 1902/3300 [2:27:45<7:18:25, 18.82s/it]INFO 04-12 01:04:37 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1903/3300 [2:27:51<5:43:22, 14.75s/it]2026-04-12 01:04:42,160 | INFO | train_grpo_train | metrics_logged mode=train step=1903

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.236363636363637e-06, 'num_tokens': 4295672.0, 'completions/mean_length': 127.0, 'completions/min_length': 127.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.0, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.9974547624588013, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8549612164497375, 'rewards/total_composite/std': 0.0, 'reward': 0.8549612164497375, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.000930364360101521, 'sampling/sampling_logp_difference/max': 0.02307342365384102, 'sampling/importance_sampling_ratio/min': 0.9878833293914795, 'sampling/importance_sampling_ratio/mean': 1.0008466243743896, 'sampling/importance_sampling_ratio/max': 1.0233416557312012, 'entropy': 0.008461171586532146, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.8549612164497375, 'reward_meter_mean': 0.9974547624588013, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8549612164497375, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 58%|█████▊    | 1903/3300 [2:27:51<5:43:22, 14.75s/it]INFO 04-12 01:04:42 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1904/3300 [2:27:55<4:32:50, 11.73s/it]2026-04-12 01:04:46,845 | INFO | train_grpo_train | metrics_logged mode=train step=1904

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.233333333333334e-06, 'num_tokens': 4297480.0, 'completions/mean_length': 69.0, 'completions/min_length': 69.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.997582197189331, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997582197189331, 'rewards/total_composite/std': 0.0, 'reward': 0.997582197189331, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.000656367396004498, 'sampling/sampling_logp_difference/max': 0.00906725786626339, 'sampling/importance_sampling_ratio/min': 0.990973711013794, 'sampling/importance_sampling_ratio/mean': 1.0005828142166138, 'sampling/importance_sampling_ratio/max': 1.0086549520492554, 'entropy': 0.006679805053863674, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.997582197189331, 'reward_meter_mean': 0.997582197189331, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997582197189331, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 58%|█████▊    | 1904/3300 [2:27:55<4:32:50, 11.73s/it]INFO 04-12 01:04:47 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1905/3300 [2:28:00<3:46:21,  9.74s/it]2026-04-12 01:04:51,938 | INFO | train_grpo_train | metrics_logged mode=train step=1905

                                                       
{'loss': 0.0031, 'grad_norm': 2.374253034591675, 'learning_rate': 4.2303030303030304e-06, 'num_tokens': 4299561.0, 'completions/mean_length': 95.125, 'completions/min_length': 95.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.125, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9972648620605469, 'rewards/meter/std': 0.00012022549344692379, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972648620605469, 'rewards/total_composite/std': 0.00012022549344692379, 'reward': 0.9972648620605469, 'reward_std': 0.00012023184535792097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008180440403521061, 'sampling/sampling_logp_difference/max': 1.0717182159423828, 'sampling/importance_sampling_ratio/min': 0.34241965413093567, 'sampling/importance_sampling_ratio/mean': 1.0006614923477173, 'sampling/importance_sampling_ratio/max': 1.5949854850769043, 'entropy': 0.03163000999484211, 'clip_ratio/low_mean': 0.0026041667442768812, 'clip_ratio/low_min': 0.0026041667442768812, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0026041667442768812, 'reward_total_mean': 0.9972648620605469, 'reward_meter_mean': 0.9972648620605469, 'reward_meter_std': 0.00012022549344692379, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972648620605469, 'reward_total_composite_std': 0.00012022549344692379, 'epoch': 0.08}

 58%|█████▊    | 1905/3300 [2:28:00<3:46:21,  9.74s/it]INFO 04-12 01:04:52 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1906/3300 [2:28:05<3:09:36,  8.16s/it]2026-04-12 01:04:56,420 | INFO | train_grpo_train | metrics_logged mode=train step=1906

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.227272727272728e-06, 'num_tokens': 4301393.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.998939037322998, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 0.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.0, 'rewards/total_composite/std': 0.0, 'reward': 0.0, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003236948396079242, 'sampling/sampling_logp_difference/max': 0.011038463562726974, 'sampling/importance_sampling_ratio/min': 0.9955568313598633, 'sampling/importance_sampling_ratio/mean': 1.0002886056900024, 'sampling/importance_sampling_ratio/max': 1.0110995769500732, 'entropy': 0.0030257631733547896, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.0, 'reward_meter_mean': 0.998939037322998, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 0.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.0, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 58%|█████▊    | 1906/3300 [2:28:05<3:09:36,  8.16s/it]INFO 04-12 01:04:56 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1907/3300 [2:28:12<3:00:36,  7.78s/it]2026-04-12 01:05:03,307 | INFO | train_grpo_train | metrics_logged mode=train step=1907

                                                       
{'loss': 0.0114, 'grad_norm': 2.0788662433624268, 'learning_rate': 4.224242424242425e-06, 'num_tokens': 4305238.0, 'completions/mean_length': 251.625, 'completions/min_length': 243.0, 'completions/max_length': 258.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 251.625, 'completions/min_terminated_length': 243.0, 'completions/max_terminated_length': 258.0, 'rewards/meter/mean': 0.997933030128479, 'rewards/meter/std': 0.00043991892016492784, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7692307829856873, 'rewards/repeat_penalty/std': 0.07121692597866058, 'rewards/total_composite/mean': 0.7676311731338501, 'rewards/total_composite/std': 0.07094842940568924, 'reward': 0.7676311731338501, 'reward_std': 0.07094840705394745, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02055802382528782, 'sampling/sampling_logp_difference/max': 0.9864959716796875, 'sampling/importance_sampling_ratio/min': 0.3728809952735901, 'sampling/importance_sampling_ratio/mean': 1.0044364929199219, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17152301780879498, 'clip_ratio/low_mean': 0.0019704491132870317, 'clip_ratio/low_min': 0.0019704491132870317, 'clip_ratio/high_mean': 0.014517532312311232, 'clip_ratio/high_max': 0.014517532312311232, 'clip_ratio/region_mean': 0.016487981425598264, 'reward_total_mean': 0.7676311731338501, 'reward_meter_mean': 0.997933030128479, 'reward_meter_std': 0.00043991892016492784, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7692307829856873, 'reward_repeat_penalty_std': 0.07121692597866058, 'reward_total_composite_mean': 0.7676311731338501, 'reward_total_composite_std': 0.07094842940568924, 'epoch': 0.08}

 58%|█████▊    | 1907/3300 [2:28:12<3:00:36,  7.78s/it]INFO 04-12 01:05:03 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1908/3300 [2:28:16<2:38:40,  6.84s/it]2026-04-12 01:05:07,958 | INFO | train_grpo_train | metrics_logged mode=train step=1908

                                                       
{'loss': 0.0026, 'grad_norm': 4.272599697113037, 'learning_rate': 4.221212121212121e-06, 'num_tokens': 4307205.0, 'completions/mean_length': 73.875, 'completions/min_length': 72.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9967119693756104, 'rewards/meter/std': 0.001478171325288713, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9967119693756104, 'rewards/total_composite/std': 0.001478171325288713, 'reward': 0.9967119693756104, 'reward_std': 0.0014781749341636896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034489989280700684, 'sampling/sampling_logp_difference/max': 1.2823877334594727, 'sampling/importance_sampling_ratio/min': 0.2773742079734802, 'sampling/importance_sampling_ratio/mean': 1.0084983110427856, 'sampling/importance_sampling_ratio/max': 1.7452210187911987, 'entropy': 0.2469378486275673, 'clip_ratio/low_mean': 0.01705764839425683, 'clip_ratio/low_min': 0.01705764839425683, 'clip_ratio/high_mean': 0.013561643892899156, 'clip_ratio/high_max': 0.013561643892899156, 'clip_ratio/region_mean': 0.030619292287155986, 'reward_total_mean': 0.9967119693756104, 'reward_meter_mean': 0.9967119693756104, 'reward_meter_std': 0.001478171325288713, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9967119693756104, 'reward_total_composite_std': 0.001478171325288713, 'epoch': 0.08}

 58%|█████▊    | 1908/3300 [2:28:16<2:38:40,  6.84s/it]INFO 04-12 01:05:08 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1909/3300 [2:28:21<2:25:27,  6.27s/it]2026-04-12 01:05:12,912 | INFO | train_grpo_train | metrics_logged mode=train step=1909

                                                       
{'loss': -0.0005, 'grad_norm': 0.519289493560791, 'learning_rate': 4.218181818181819e-06, 'num_tokens': 4309357.0, 'completions/mean_length': 103.0, 'completions/min_length': 103.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.0, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9975078105926514, 'rewards/meter/std': 3.284694321337156e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975078105926514, 'rewards/total_composite/std': 3.284694321337156e-05, 'reward': 0.9975078105926514, 'reward_std': 3.283110709162429e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0023622734006494284, 'sampling/sampling_logp_difference/max': 0.6333708763122559, 'sampling/importance_sampling_ratio/min': 0.5307995080947876, 'sampling/importance_sampling_ratio/mean': 1.0008713006973267, 'sampling/importance_sampling_ratio/max': 1.714148759841919, 'entropy': 0.009892041911371052, 'clip_ratio/low_mean': 0.0012135922443121672, 'clip_ratio/low_min': 0.0012135922443121672, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0012135922443121672, 'reward_total_mean': 0.9975078105926514, 'reward_meter_mean': 0.9975078105926514, 'reward_meter_std': 3.284694321337156e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975078105926514, 'reward_total_composite_std': 3.284694321337156e-05, 'epoch': 0.08}

 58%|█████▊    | 1909/3300 [2:28:21<2:25:27,  6.27s/it]INFO 04-12 01:05:13 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1910/3300 [2:28:29<2:33:52,  6.64s/it]2026-04-12 01:05:20,410 | INFO | train_grpo_train | metrics_logged mode=train step=1910

                                                       
{'loss': -0.0326, 'grad_norm': 1.9287523031234741, 'learning_rate': 4.215151515151515e-06, 'num_tokens': 4313350.0, 'completions/mean_length': 291.125, 'completions/min_length': 258.0, 'completions/max_length': 307.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 291.125, 'completions/min_terminated_length': 258.0, 'completions/max_terminated_length': 307.0, 'rewards/meter/mean': 0.9932549595832825, 'rewards/meter/std': 0.0015650950372219086, 'rewards/count_adherence/mean': 0.8888888955116272, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6931459903717041, 'rewards/repeat_penalty/std': 0.010692903771996498, 'rewards/total_composite/mean': 0.6118792295455933, 'rewards/total_composite/std': 0.041350506246089935, 'reward': 0.6118792295455933, 'reward_std': 0.04135050252079964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010057836771011353, 'sampling/sampling_logp_difference/max': 1.9771476984024048, 'sampling/importance_sampling_ratio/min': 0.1384636014699936, 'sampling/importance_sampling_ratio/mean': 0.9992997646331787, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0457235153298825, 'clip_ratio/low_mean': 0.006014678278006613, 'clip_ratio/low_min': 0.006014678278006613, 'clip_ratio/high_mean': 0.0004071661096531898, 'clip_ratio/high_max': 0.0004071661096531898, 'clip_ratio/region_mean': 0.006421844387659803, 'reward_total_mean': 0.6118792295455933, 'reward_meter_mean': 0.9932549595832825, 'reward_meter_std': 0.0015650950372219086, 'reward_count_adherence_mean': 0.8888888955116272, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6931459903717041, 'reward_repeat_penalty_std': 0.010692903771996498, 'reward_total_composite_mean': 0.6118792295455933, 'reward_total_composite_std': 0.041350506246089935, 'epoch': 0.08}

 58%|█████▊    | 1910/3300 [2:28:29<2:33:52,  6.64s/it]INFO 04-12 01:05:20 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1911/3300 [2:28:33<2:19:03,  6.01s/it]2026-04-12 01:05:24,934 | INFO | train_grpo_train | metrics_logged mode=train step=1911

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.212121212121212e-06, 'num_tokens': 4315262.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.998939037322998, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998939037322998, 'rewards/total_composite/std': 0.0, 'reward': 0.998939037322998, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003693441394716501, 'sampling/sampling_logp_difference/max': 0.006262333132326603, 'sampling/importance_sampling_ratio/min': 0.9994460940361023, 'sampling/importance_sampling_ratio/mean': 1.0003656148910522, 'sampling/importance_sampling_ratio/max': 1.006282091140747, 'entropy': 0.004040350759169087, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.998939037322998, 'reward_meter_mean': 0.998939037322998, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998939037322998, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 58%|█████▊    | 1911/3300 [2:28:33<2:19:03,  6.01s/it]INFO 04-12 01:05:25 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1912/3300 [2:28:38<2:08:09,  5.54s/it]2026-04-12 01:05:29,386 | INFO | train_grpo_train | metrics_logged mode=train step=1912

                                                       
{'loss': 0.0001, 'grad_norm': 0.2683267891407013, 'learning_rate': 4.2090909090909095e-06, 'num_tokens': 4317134.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9989374279975891, 'rewards/meter/std': 4.551860001811292e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989374279975891, 'rewards/total_composite/std': 4.551860001811292e-06, 'reward': 0.9989374279975891, 'reward_std': 4.551859547063941e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002090727211907506, 'sampling/sampling_logp_difference/max': 0.19945931434631348, 'sampling/importance_sampling_ratio/min': 0.8604245185852051, 'sampling/importance_sampling_ratio/mean': 1.0016252994537354, 'sampling/importance_sampling_ratio/max': 1.2207425832748413, 'entropy': 0.011784833506681025, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.9989374279975891, 'reward_meter_mean': 0.9989374279975891, 'reward_meter_std': 4.551860001811292e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989374279975891, 'reward_total_composite_std': 4.551860001811292e-06, 'epoch': 0.08}

 58%|█████▊    | 1912/3300 [2:28:38<2:08:09,  5.54s/it]INFO 04-12 01:05:29 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1913/3300 [2:28:42<1:59:40,  5.18s/it]2026-04-12 01:05:33,710 | INFO | train_grpo_train | metrics_logged mode=train step=1913

                                                       
{'loss': -0.0002, 'grad_norm': 0.7913539409637451, 'learning_rate': 4.206060606060606e-06, 'num_tokens': 4318894.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9945913553237915, 'rewards/meter/std': 2.7374408091418445e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9945913553237915, 'rewards/total_composite/std': 2.7374408091418445e-05, 'reward': 0.9945913553237915, 'reward_std': 2.7371370379114524e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003088480094447732, 'sampling/sampling_logp_difference/max': 0.22362565994262695, 'sampling/importance_sampling_ratio/min': 0.8597717881202698, 'sampling/importance_sampling_ratio/mean': 1.0011041164398193, 'sampling/importance_sampling_ratio/max': 1.2506028413772583, 'entropy': 0.018943659961223602, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9945913553237915, 'reward_meter_mean': 0.9945913553237915, 'reward_meter_std': 2.7374408091418445e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9945913553237915, 'reward_total_composite_std': 2.7374408091418445e-05, 'epoch': 0.08}

 58%|█████▊    | 1913/3300 [2:28:42<1:59:40,  5.18s/it]INFO 04-12 01:05:33 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1914/3300 [2:28:47<1:54:36,  4.96s/it]2026-04-12 01:05:38,169 | INFO | train_grpo_train | metrics_logged mode=train step=1914

                                                       
{'loss': -0.0005, 'grad_norm': 0.36171841621398926, 'learning_rate': 4.203030303030303e-06, 'num_tokens': 4320646.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9989273548126221, 'rewards/meter/std': 3.3127438655355945e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989273548126221, 'rewards/total_composite/std': 3.3127438655355945e-05, 'reward': 0.9989273548126221, 'reward_std': 3.312742410344072e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004210924729704857, 'sampling/sampling_logp_difference/max': 1.9241247177124023, 'sampling/importance_sampling_ratio/min': 0.14600348472595215, 'sampling/importance_sampling_ratio/mean': 0.9987910985946655, 'sampling/importance_sampling_ratio/max': 1.0103963613510132, 'entropy': 0.005322357552358881, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9989273548126221, 'reward_meter_mean': 0.9989273548126221, 'reward_meter_std': 3.3127438655355945e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989273548126221, 'reward_total_composite_std': 3.3127438655355945e-05, 'epoch': 0.08}

 58%|█████▊    | 1914/3300 [2:28:47<1:54:36,  4.96s/it]INFO 04-12 01:05:38 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1915/3300 [2:28:52<1:54:50,  4.98s/it]2026-04-12 01:05:43,187 | INFO | train_grpo_train | metrics_logged mode=train step=1915

                                                       
{'loss': -0.0005, 'grad_norm': 0.24196073412895203, 'learning_rate': 4.2000000000000004e-06, 'num_tokens': 4323030.0, 'completions/mean_length': 122.0, 'completions/min_length': 122.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.0, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.995856523513794, 'rewards/meter/std': 2.842021240212489e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8535913228988647, 'rewards/total_composite/std': 2.4363118427572772e-05, 'reward': 0.8535913228988647, 'reward_std': 2.4356924768653698e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00429805601015687, 'sampling/sampling_logp_difference/max': 0.30840301513671875, 'sampling/importance_sampling_ratio/min': 0.7478480339050293, 'sampling/importance_sampling_ratio/mean': 1.0021631717681885, 'sampling/importance_sampling_ratio/max': 1.361249566078186, 'entropy': 0.02201851806603372, 'clip_ratio/low_mean': 0.0010245901066809893, 'clip_ratio/low_min': 0.0010245901066809893, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.003073770320042968, 'reward_total_mean': 0.8535913228988647, 'reward_meter_mean': 0.995856523513794, 'reward_meter_std': 2.842021240212489e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8535913228988647, 'reward_total_composite_std': 2.4363118427572772e-05, 'epoch': 0.08}

 58%|█████▊    | 1915/3300 [2:28:52<1:54:50,  4.98s/it]INFO 04-12 01:05:43 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1916/3300 [2:28:57<1:55:44,  5.02s/it]2026-04-12 01:05:48,298 | INFO | train_grpo_train | metrics_logged mode=train step=1916

                                                       
{'loss': 0.0035, 'grad_norm': 2.755591630935669, 'learning_rate': 4.196969696969697e-06, 'num_tokens': 4324864.0, 'completions/mean_length': 74.25, 'completions/min_length': 72.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.25, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9977705478668213, 'rewards/meter/std': 0.0007801271858625114, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977705478668213, 'rewards/total_composite/std': 0.0007801271858625114, 'reward': 0.9977705478668213, 'reward_std': 0.0007801253814250231, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02434798702597618, 'sampling/sampling_logp_difference/max': 1.0248136520385742, 'sampling/importance_sampling_ratio/min': 0.35886335372924805, 'sampling/importance_sampling_ratio/mean': 1.0088934898376465, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18484536185860634, 'clip_ratio/low_mean': 0.00856418942566961, 'clip_ratio/low_min': 0.00856418942566961, 'clip_ratio/high_mean': 0.005000000121071935, 'clip_ratio/high_max': 0.005000000121071935, 'clip_ratio/region_mean': 0.013564189546741545, 'reward_total_mean': 0.9977705478668213, 'reward_meter_mean': 0.9977705478668213, 'reward_meter_std': 0.0007801271858625114, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977705478668213, 'reward_total_composite_std': 0.0007801271858625114, 'epoch': 0.08}

 58%|█████▊    | 1916/3300 [2:28:57<1:55:44,  5.02s/it]INFO 04-12 01:05:48 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1917/3300 [2:29:02<1:55:42,  5.02s/it]2026-04-12 01:05:53,323 | INFO | train_grpo_train | metrics_logged mode=train step=1917

                                                       
{'loss': 0.1368, 'grad_norm': 7.223612308502197, 'learning_rate': 4.193939393939394e-06, 'num_tokens': 4326780.0, 'completions/mean_length': 81.5, 'completions/min_length': 75.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9922255277633667, 'rewards/meter/std': 0.006029736250638962, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9299638271331787, 'rewards/total_composite/std': 0.17459870874881744, 'reward': 0.9299638271331787, 'reward_std': 0.17459867894649506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043613094836473465, 'sampling/sampling_logp_difference/max': 3.7613840103149414, 'sampling/importance_sampling_ratio/min': 0.02325153723359108, 'sampling/importance_sampling_ratio/mean': 1.0046718120574951, 'sampling/importance_sampling_ratio/max': 1.5985634326934814, 'entropy': 0.30126412957906723, 'clip_ratio/low_mean': 0.006756756920367479, 'clip_ratio/low_min': 0.006756756920367479, 'clip_ratio/high_mean': 0.038703347789123654, 'clip_ratio/high_max': 0.038703347789123654, 'clip_ratio/region_mean': 0.045460104709491134, 'reward_total_mean': 0.9299638271331787, 'reward_meter_mean': 0.9922255277633667, 'reward_meter_std': 0.006029736250638962, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9299638271331787, 'reward_total_composite_std': 0.17459870874881744, 'epoch': 0.08}

 58%|█████▊    | 1917/3300 [2:29:02<1:55:42,  5.02s/it]INFO 04-12 01:05:53 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1918/3300 [2:29:06<1:53:00,  4.91s/it]2026-04-12 01:05:57,964 | INFO | train_grpo_train | metrics_logged mode=train step=1918

                                                       
{'loss': 0.0163, 'grad_norm': 3.7035961151123047, 'learning_rate': 4.190909090909091e-06, 'num_tokens': 4328592.0, 'completions/mean_length': 72.5, 'completions/min_length': 70.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.5, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.986924946308136, 'rewards/meter/std': 0.002686952007934451, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.986924946308136, 'rewards/total_composite/std': 0.002686952007934451, 'reward': 0.986924946308136, 'reward_std': 0.0026869599241763353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016851795837283134, 'sampling/sampling_logp_difference/max': 0.770622730255127, 'sampling/importance_sampling_ratio/min': 0.46272483468055725, 'sampling/importance_sampling_ratio/mean': 1.0001866817474365, 'sampling/importance_sampling_ratio/max': 1.7578001022338867, 'entropy': 0.15394389163702726, 'clip_ratio/low_mean': 0.005113846738822758, 'clip_ratio/low_min': 0.005113846738822758, 'clip_ratio/high_mean': 0.017437846050597727, 'clip_ratio/high_max': 0.017437846050597727, 'clip_ratio/region_mean': 0.022551692789420485, 'reward_total_mean': 0.986924946308136, 'reward_meter_mean': 0.986924946308136, 'reward_meter_std': 0.002686952007934451, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.986924946308136, 'reward_total_composite_std': 0.002686952007934451, 'epoch': 0.08}

 58%|█████▊    | 1918/3300 [2:29:06<1:53:00,  4.91s/it]INFO 04-12 01:05:58 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1919/3300 [2:29:11<1:50:00,  4.78s/it]2026-04-12 01:06:02,450 | INFO | train_grpo_train | metrics_logged mode=train step=1919

                                                       
{'loss': -0.0004, 'grad_norm': 1.28860342502594, 'learning_rate': 4.187878787878788e-06, 'num_tokens': 4330512.0, 'completions/mean_length': 69.0, 'completions/min_length': 69.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9976222515106201, 'rewards/meter/std': 4.285174509277567e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976222515106201, 'rewards/total_composite/std': 4.285174509277567e-05, 'reward': 0.9976222515106201, 'reward_std': 4.2851730540860444e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006954767741262913, 'sampling/sampling_logp_difference/max': 0.8643484711647034, 'sampling/importance_sampling_ratio/min': 0.4213259816169739, 'sampling/importance_sampling_ratio/mean': 0.9985252022743225, 'sampling/importance_sampling_ratio/max': 1.3529547452926636, 'entropy': 0.027878023218363523, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005434782709926367, 'clip_ratio/high_max': 0.005434782709926367, 'clip_ratio/region_mean': 0.005434782709926367, 'reward_total_mean': 0.9976222515106201, 'reward_meter_mean': 0.9976222515106201, 'reward_meter_std': 4.285174509277567e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976222515106201, 'reward_total_composite_std': 4.285174509277567e-05, 'epoch': 0.08}

 58%|█████▊    | 1919/3300 [2:29:11<1:50:00,  4.78s/it]INFO 04-12 01:06:02 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1920/3300 [2:29:16<1:53:02,  4.91s/it]2026-04-12 01:06:07,683 | INFO | train_grpo_train | metrics_logged mode=train step=1920

                                                       
{'loss': -0.0001, 'grad_norm': 0.23397041857242584, 'learning_rate': 4.184848484848485e-06, 'num_tokens': 4332959.0, 'completions/mean_length': 131.875, 'completions/min_length': 131.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.875, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.9991158246994019, 'rewards/meter/std': 1.4611580809287261e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8563849925994873, 'rewards/total_composite/std': 1.2507220162660815e-05, 'reward': 0.8563849925994873, 'reward_std': 1.2510253327491228e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004140977747738361, 'sampling/sampling_logp_difference/max': 1.6029167175292969, 'sampling/importance_sampling_ratio/min': 0.2013085037469864, 'sampling/importance_sampling_ratio/mean': 0.9987531304359436, 'sampling/importance_sampling_ratio/max': 1.0698660612106323, 'entropy': 0.01177009951788932, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.8563849925994873, 'reward_meter_mean': 0.9991158246994019, 'reward_meter_std': 1.4611580809287261e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8563849925994873, 'reward_total_composite_std': 1.2507220162660815e-05, 'epoch': 0.08}

 58%|█████▊    | 1920/3300 [2:29:16<1:53:02,  4.91s/it]INFO 04-12 01:06:07 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1921/3300 [2:29:21<1:54:19,  4.97s/it]2026-04-12 01:06:12,793 | INFO | train_grpo_train | metrics_logged mode=train step=1921

                                                       
{'loss': 0.0041, 'grad_norm': 3.0183725357055664, 'learning_rate': 4.181818181818182e-06, 'num_tokens': 4335289.0, 'completions/mean_length': 114.25, 'completions/min_length': 113.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.25, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9916590452194214, 'rewards/meter/std': 0.004383288789540529, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9916590452194214, 'rewards/total_composite/std': 0.004383288789540529, 'reward': 0.9916590452194214, 'reward_std': 0.004383287392556667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03317651152610779, 'sampling/sampling_logp_difference/max': 0.9611685276031494, 'sampling/importance_sampling_ratio/min': 0.4941423237323761, 'sampling/importance_sampling_ratio/mean': 1.0076040029525757, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3176443073898554, 'clip_ratio/low_mean': 0.011023133061826229, 'clip_ratio/low_min': 0.011023133061826229, 'clip_ratio/high_mean': 0.02495911391451955, 'clip_ratio/high_max': 0.02495911391451955, 'clip_ratio/region_mean': 0.03598224697634578, 'reward_total_mean': 0.9916590452194214, 'reward_meter_mean': 0.9916590452194214, 'reward_meter_std': 0.004383288789540529, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9916590452194214, 'reward_total_composite_std': 0.004383288789540529, 'epoch': 0.08}

 58%|█████▊    | 1921/3300 [2:29:21<1:54:19,  4.97s/it]INFO 04-12 01:06:13 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1922/3300 [2:29:26<1:51:09,  4.84s/it]2026-04-12 01:06:17,318 | INFO | train_grpo_train | metrics_logged mode=train step=1922

                                                       
{'loss': 0.0006, 'grad_norm': 0.0710740014910698, 'learning_rate': 4.1787878787878795e-06, 'num_tokens': 4337264.0, 'completions/mean_length': 65.875, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9989427924156189, 'rewards/meter/std': 1.0663165085134096e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989427924156189, 'rewards/total_composite/std': 1.0663165085134096e-05, 'reward': 0.9989427924156189, 'reward_std': 1.0657141501724254e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006586765870451927, 'sampling/sampling_logp_difference/max': 2.554171562194824, 'sampling/importance_sampling_ratio/min': 0.0777566209435463, 'sampling/importance_sampling_ratio/mean': 0.9995160102844238, 'sampling/importance_sampling_ratio/max': 1.1512835025787354, 'entropy': 0.015987101825885475, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.001923076924867928, 'clip_ratio/high_max': 0.001923076924867928, 'clip_ratio/region_mean': 0.001923076924867928, 'reward_total_mean': 0.9989427924156189, 'reward_meter_mean': 0.9989427924156189, 'reward_meter_std': 1.0663165085134096e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989427924156189, 'reward_total_composite_std': 1.0663165085134096e-05, 'epoch': 0.08}

 58%|█████▊    | 1922/3300 [2:29:26<1:51:09,  4.84s/it]INFO 04-12 01:06:17 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1923/3300 [2:29:32<1:59:36,  5.21s/it]2026-04-12 01:06:23,396 | INFO | train_grpo_train | metrics_logged mode=train step=1923

                                                       
{'loss': 0.0104, 'grad_norm': 0.6413597464561462, 'learning_rate': 4.175757575757576e-06, 'num_tokens': 4340257.0, 'completions/mean_length': 201.125, 'completions/min_length': 195.0, 'completions/max_length': 202.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 201.125, 'completions/min_terminated_length': 195.0, 'completions/max_terminated_length': 202.0, 'rewards/meter/mean': 0.9961963891983032, 'rewards/meter/std': 7.053603621898219e-05, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7291666865348816, 'rewards/repeat_penalty/std': 0.058925554156303406, 'rewards/total_composite/mean': 0.5811127424240112, 'rewards/total_composite/std': 0.0469353049993515, 'reward': 0.5811127424240112, 'reward_std': 0.0469353049993515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00489839119836688, 'sampling/sampling_logp_difference/max': 0.9933547973632812, 'sampling/importance_sampling_ratio/min': 0.3703322112560272, 'sampling/importance_sampling_ratio/mean': 0.9994967579841614, 'sampling/importance_sampling_ratio/max': 1.3452818393707275, 'entropy': 0.01974188955500722, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0031607006676495075, 'clip_ratio/high_max': 0.0031607006676495075, 'clip_ratio/region_mean': 0.0031607006676495075, 'reward_total_mean': 0.5811127424240112, 'reward_meter_mean': 0.9961963891983032, 'reward_meter_std': 7.053603621898219e-05, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7291666865348816, 'reward_repeat_penalty_std': 0.058925554156303406, 'reward_total_composite_mean': 0.5811127424240112, 'reward_total_composite_std': 0.0469353049993515, 'epoch': 0.08}

 58%|█████▊    | 1923/3300 [2:29:32<1:59:36,  5.21s/it]INFO 04-12 01:06:23 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1924/3300 [2:29:37<1:58:07,  5.15s/it]2026-04-12 01:06:28,411 | INFO | train_grpo_train | metrics_logged mode=train step=1924

                                                       
{'loss': 0.1384, 'grad_norm': 6.834925174713135, 'learning_rate': 4.172727272727273e-06, 'num_tokens': 4342199.0, 'completions/mean_length': 80.75, 'completions/min_length': 72.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9895592927932739, 'rewards/meter/std': 0.01372271403670311, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.928816556930542, 'rewards/total_composite/std': 0.17933166027069092, 'reward': 0.928816556930542, 'reward_std': 0.17933164536952972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048581015318632126, 'sampling/sampling_logp_difference/max': 4.14417028427124, 'sampling/importance_sampling_ratio/min': 0.015856586396694183, 'sampling/importance_sampling_ratio/mean': 1.009357213973999, 'sampling/importance_sampling_ratio/max': 1.9494916200637817, 'entropy': 0.3375435955822468, 'clip_ratio/low_mean': 0.004587155766785145, 'clip_ratio/low_min': 0.004587155766785145, 'clip_ratio/high_mean': 0.040673923096619546, 'clip_ratio/high_max': 0.040673923096619546, 'clip_ratio/region_mean': 0.04526107886340469, 'reward_total_mean': 0.928816556930542, 'reward_meter_mean': 0.9895592927932739, 'reward_meter_std': 0.01372271403670311, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.928816556930542, 'reward_total_composite_std': 0.17933166027069092, 'epoch': 0.08}

 58%|█████▊    | 1924/3300 [2:29:37<1:58:07,  5.15s/it]INFO 04-12 01:06:28 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1925/3300 [2:29:41<1:52:50,  4.92s/it]2026-04-12 01:06:32,807 | INFO | train_grpo_train | metrics_logged mode=train step=1925

                                                       
{'loss': 0.0103, 'grad_norm': 16.744361877441406, 'learning_rate': 4.1696969696969705e-06, 'num_tokens': 4343965.0, 'completions/mean_length': 57.75, 'completions/min_length': 56.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9946485757827759, 'rewards/meter/std': 0.00013940427743364125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946485757827759, 'rewards/total_composite/std': 0.00013940427743364125, 'reward': 0.9946485757827759, 'reward_std': 0.0001394047576468438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004949766676872969, 'sampling/sampling_logp_difference/max': 0.48597773909568787, 'sampling/importance_sampling_ratio/min': 0.6150954961776733, 'sampling/importance_sampling_ratio/mean': 0.9992939233779907, 'sampling/importance_sampling_ratio/max': 1.1374722719192505, 'entropy': 0.025414136587642133, 'clip_ratio/low_mean': 0.004387315362691879, 'clip_ratio/low_min': 0.004387315362691879, 'clip_ratio/high_mean': 0.004310344811528921, 'clip_ratio/high_max': 0.004310344811528921, 'clip_ratio/region_mean': 0.0086976601742208, 'reward_total_mean': 0.9946485757827759, 'reward_meter_mean': 0.9946485757827759, 'reward_meter_std': 0.00013940427743364125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946485757827759, 'reward_total_composite_std': 0.00013940427743364125, 'epoch': 0.08}

 58%|█████▊    | 1925/3300 [2:29:41<1:52:50,  4.92s/it]INFO 04-12 01:06:33 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1926/3300 [2:29:49<2:10:21,  5.69s/it]2026-04-12 01:06:40,286 | INFO | train_grpo_train | metrics_logged mode=train step=1926

                                                       
{'loss': 0.0014, 'grad_norm': 2.593313694000244, 'learning_rate': 4.166666666666667e-06, 'num_tokens': 4347997.0, 'completions/mean_length': 297.0, 'completions/min_length': 282.0, 'completions/max_length': 309.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 297.0, 'completions/min_terminated_length': 282.0, 'completions/max_terminated_length': 309.0, 'rewards/meter/mean': 0.9778913855552673, 'rewards/meter/std': 0.025907516479492188, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7741815447807312, 'rewards/repeat_penalty/std': 0.09523647278547287, 'rewards/total_composite/mean': 0.661956250667572, 'rewards/total_composite/std': 0.08313874155282974, 'reward': 0.661956250667572, 'reward_std': 0.08313874155282974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027429981157183647, 'sampling/sampling_logp_difference/max': 7.558941841125488, 'sampling/importance_sampling_ratio/min': 0.0005214267293922603, 'sampling/importance_sampling_ratio/mean': 1.0018694400787354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12555130943655968, 'clip_ratio/low_mean': 0.003884511475916952, 'clip_ratio/low_min': 0.003884511475916952, 'clip_ratio/high_mean': 0.012181702419184148, 'clip_ratio/high_max': 0.012181702419184148, 'clip_ratio/region_mean': 0.0160662138951011, 'reward_total_mean': 0.661956250667572, 'reward_meter_mean': 0.9778913855552673, 'reward_meter_std': 0.025907516479492188, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7741815447807312, 'reward_repeat_penalty_std': 0.09523647278547287, 'reward_total_composite_mean': 0.661956250667572, 'reward_total_composite_std': 0.08313874155282974, 'epoch': 0.08}

 58%|█████▊    | 1926/3300 [2:29:49<2:10:21,  5.69s/it]INFO 04-12 01:06:40 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1927/3300 [2:29:54<2:04:36,  5.45s/it]2026-04-12 01:06:45,160 | INFO | train_grpo_train | metrics_logged mode=train step=1927

                                                       
{'loss': -0.0006, 'grad_norm': 0.516058087348938, 'learning_rate': 4.163636363636364e-06, 'num_tokens': 4350132.0, 'completions/mean_length': 98.875, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.875, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9990072250366211, 'rewards/meter/std': 0.00012045173207297921, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990072250366211, 'rewards/total_composite/std': 0.00012045173207297921, 'reward': 0.9990072250366211, 'reward_std': 0.00012044955656165257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004538480192422867, 'sampling/sampling_logp_difference/max': 0.5121603012084961, 'sampling/importance_sampling_ratio/min': 0.5991997718811035, 'sampling/importance_sampling_ratio/mean': 0.9992603659629822, 'sampling/importance_sampling_ratio/max': 1.2373576164245605, 'entropy': 0.02114183083176613, 'clip_ratio/low_mean': 0.0025252525229007006, 'clip_ratio/low_min': 0.0025252525229007006, 'clip_ratio/high_mean': 0.0025252525229007006, 'clip_ratio/high_max': 0.0025252525229007006, 'clip_ratio/region_mean': 0.005050505045801401, 'reward_total_mean': 0.9990072250366211, 'reward_meter_mean': 0.9990072250366211, 'reward_meter_std': 0.00012045173207297921, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990072250366211, 'reward_total_composite_std': 0.00012045173207297921, 'epoch': 0.08}

 58%|█████▊    | 1927/3300 [2:29:54<2:04:36,  5.45s/it]INFO 04-12 01:06:45 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1928/3300 [2:29:58<1:58:16,  5.17s/it]2026-04-12 01:06:49,696 | INFO | train_grpo_train | metrics_logged mode=train step=1928

                                                       
{'loss': -0.0147, 'grad_norm': 3.781970977783203, 'learning_rate': 4.160606060606061e-06, 'num_tokens': 4351735.0, 'completions/mean_length': 61.375, 'completions/min_length': 60.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9964951276779175, 'rewards/meter/std': 0.00043850651127286255, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9964951276779175, 'rewards/total_composite/std': 0.00043850651127286255, 'reward': 0.9964951276779175, 'reward_std': 0.0004385166976135224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007140871603041887, 'sampling/sampling_logp_difference/max': 0.4663825035095215, 'sampling/importance_sampling_ratio/min': 0.6272673606872559, 'sampling/importance_sampling_ratio/mean': 1.0027474164962769, 'sampling/importance_sampling_ratio/max': 1.4405137300491333, 'entropy': 0.04211301123723388, 'clip_ratio/low_mean': 0.006048386916518211, 'clip_ratio/low_min': 0.006048386916518211, 'clip_ratio/high_mean': 0.0019841270986944437, 'clip_ratio/high_max': 0.0019841270986944437, 'clip_ratio/region_mean': 0.008032514015212655, 'reward_total_mean': 0.9964951276779175, 'reward_meter_mean': 0.9964951276779175, 'reward_meter_std': 0.00043850651127286255, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9964951276779175, 'reward_total_composite_std': 0.00043850651127286255, 'epoch': 0.08}

 58%|█████▊    | 1928/3300 [2:29:58<1:58:16,  5.17s/it]INFO 04-12 01:06:49 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1929/3300 [2:30:03<1:56:21,  5.09s/it]2026-04-12 01:06:54,593 | INFO | train_grpo_train | metrics_logged mode=train step=1929

                                                       
{'loss': -0.0459, 'grad_norm': 3.9287514686584473, 'learning_rate': 4.157575757575758e-06, 'num_tokens': 4353662.0, 'completions/mean_length': 81.875, 'completions/min_length': 78.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.875, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9926681518554688, 'rewards/meter/std': 0.00178977579344064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9926681518554688, 'rewards/total_composite/std': 0.00178977579344064, 'reward': 0.9926681518554688, 'reward_std': 0.0017897515790537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016444914042949677, 'sampling/sampling_logp_difference/max': 0.9670681953430176, 'sampling/importance_sampling_ratio/min': 0.38019606471061707, 'sampling/importance_sampling_ratio/mean': 1.0027321577072144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07624188321642578, 'clip_ratio/low_mean': 0.0126003761542961, 'clip_ratio/low_min': 0.0126003761542961, 'clip_ratio/high_mean': 0.0013888889225199819, 'clip_ratio/high_max': 0.0013888889225199819, 'clip_ratio/region_mean': 0.013989265076816082, 'reward_total_mean': 0.9926681518554688, 'reward_meter_mean': 0.9926681518554688, 'reward_meter_std': 0.00178977579344064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9926681518554688, 'reward_total_composite_std': 0.00178977579344064, 'epoch': 0.08}

 58%|█████▊    | 1929/3300 [2:30:03<1:56:21,  5.09s/it]INFO 04-12 01:06:54 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1930/3300 [2:30:08<1:52:41,  4.94s/it]2026-04-12 01:06:59,163 | INFO | train_grpo_train | metrics_logged mode=train step=1930

                                                       
{'loss': 0.0178, 'grad_norm': 3.3098363876342773, 'learning_rate': 4.154545454545455e-06, 'num_tokens': 4355551.0, 'completions/mean_length': 67.125, 'completions/min_length': 66.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.84641432762146, 'rewards/meter/std': 0.031771186739206314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.84641432762146, 'rewards/total_composite/std': 0.031771186739206314, 'reward': 0.84641432762146, 'reward_std': 0.03177119046449661, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027442539110779762, 'sampling/sampling_logp_difference/max': 1.9472160339355469, 'sampling/importance_sampling_ratio/min': 0.14267070591449738, 'sampling/importance_sampling_ratio/mean': 1.0048232078552246, 'sampling/importance_sampling_ratio/max': 1.4023113250732422, 'entropy': 0.19507121294736862, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.01119648793246597, 'clip_ratio/high_max': 0.01119648793246597, 'clip_ratio/region_mean': 0.014767916523851454, 'reward_total_mean': 0.84641432762146, 'reward_meter_mean': 0.84641432762146, 'reward_meter_std': 0.031771186739206314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.84641432762146, 'reward_total_composite_std': 0.031771186739206314, 'epoch': 0.08}

 58%|█████▊    | 1930/3300 [2:30:08<1:52:41,  4.94s/it]INFO 04-12 01:06:59 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1931/3300 [2:30:12<1:49:31,  4.80s/it]2026-04-12 01:07:03,652 | INFO | train_grpo_train | metrics_logged mode=train step=1931

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.151515151515152e-06, 'num_tokens': 4357311.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.998939037322998, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998939037322998, 'rewards/total_composite/std': 0.0, 'reward': 0.998939037322998, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0009288773289881647, 'sampling/sampling_logp_difference/max': 0.029468350112438202, 'sampling/importance_sampling_ratio/min': 0.9709616303443909, 'sampling/importance_sampling_ratio/mean': 1.0005451440811157, 'sampling/importance_sampling_ratio/max': 1.0244560241699219, 'entropy': 0.008458438969682902, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.998939037322998, 'reward_meter_mean': 0.998939037322998, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998939037322998, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 59%|█████▊    | 1931/3300 [2:30:12<1:49:31,  4.80s/it]INFO 04-12 01:07:03 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1932/3300 [2:30:17<1:49:17,  4.79s/it]2026-04-12 01:07:08,440 | INFO | train_grpo_train | metrics_logged mode=train step=1932

                                                       
{'loss': 0.0098, 'grad_norm': 3.314406156539917, 'learning_rate': 4.148484848484849e-06, 'num_tokens': 4359128.0, 'completions/mean_length': 73.125, 'completions/min_length': 70.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9976757764816284, 'rewards/meter/std': 0.0007166875875554979, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976757764816284, 'rewards/total_composite/std': 0.0007166875875554979, 'reward': 0.9976757764816284, 'reward_std': 0.0007166775758378208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029741626232862473, 'sampling/sampling_logp_difference/max': 1.3467721939086914, 'sampling/importance_sampling_ratio/min': 0.26007840037345886, 'sampling/importance_sampling_ratio/mean': 1.001599907875061, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17044041864573956, 'clip_ratio/low_mean': 0.0050675676902756095, 'clip_ratio/low_min': 0.0050675676902756095, 'clip_ratio/high_mean': 0.024061617092229426, 'clip_ratio/high_max': 0.024061617092229426, 'clip_ratio/region_mean': 0.029129184782505035, 'reward_total_mean': 0.9976757764816284, 'reward_meter_mean': 0.9976757764816284, 'reward_meter_std': 0.0007166875875554979, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976757764816284, 'reward_total_composite_std': 0.0007166875875554979, 'epoch': 0.08}

 59%|█████▊    | 1932/3300 [2:30:17<1:49:17,  4.79s/it]INFO 04-12 01:07:08 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1933/3300 [2:30:22<1:48:50,  4.78s/it]2026-04-12 01:07:13,173 | INFO | train_grpo_train | metrics_logged mode=train step=1933

                                                       
{'loss': 0.0007, 'grad_norm': 4.072077751159668, 'learning_rate': 4.145454545454546e-06, 'num_tokens': 4361033.0, 'completions/mean_length': 73.125, 'completions/min_length': 71.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9970039129257202, 'rewards/meter/std': 0.0011512924684211612, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970039129257202, 'rewards/total_composite/std': 0.0011512924684211612, 'reward': 0.9970039129257202, 'reward_std': 0.0011513038771227002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029263032600283623, 'sampling/sampling_logp_difference/max': 1.3621405363082886, 'sampling/importance_sampling_ratio/min': 0.2561119794845581, 'sampling/importance_sampling_ratio/mean': 1.0008363723754883, 'sampling/importance_sampling_ratio/max': 1.6174455881118774, 'entropy': 0.20059540774673223, 'clip_ratio/low_mean': 0.008469702675938606, 'clip_ratio/low_min': 0.008469702675938606, 'clip_ratio/high_mean': 0.0155549660557881, 'clip_ratio/high_max': 0.0155549660557881, 'clip_ratio/region_mean': 0.024024668731726706, 'reward_total_mean': 0.9970039129257202, 'reward_meter_mean': 0.9970039129257202, 'reward_meter_std': 0.0011512924684211612, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970039129257202, 'reward_total_composite_std': 0.0011512924684211612, 'epoch': 0.08}

 59%|█████▊    | 1933/3300 [2:30:22<1:48:50,  4.78s/it]INFO 04-12 01:07:13 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1934/3300 [2:30:28<2:01:25,  5.33s/it]2026-04-12 01:07:19,802 | INFO | train_grpo_train | metrics_logged mode=train step=1934

                                                       
{'loss': 0.0093, 'grad_norm': 3.0962040424346924, 'learning_rate': 4.142424242424243e-06, 'num_tokens': 4364465.0, 'completions/mean_length': 226.0, 'completions/min_length': 224.0, 'completions/max_length': 230.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 226.0, 'completions/min_terminated_length': 224.0, 'completions/max_terminated_length': 230.0, 'rewards/meter/mean': 0.8571816086769104, 'rewards/meter/std': 0.171591117978096, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8020833134651184, 'rewards/repeat_penalty/std': 0.0431290864944458, 'rewards/total_composite/mean': 0.6912178993225098, 'rewards/total_composite/std': 0.158027783036232, 'reward': 0.6912178993225098, 'reward_std': 0.1580277979373932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02388615906238556, 'sampling/sampling_logp_difference/max': 3.0924410820007324, 'sampling/importance_sampling_ratio/min': 0.0453910194337368, 'sampling/importance_sampling_ratio/mean': 0.9996902942657471, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10432031005620956, 'clip_ratio/low_mean': 0.006601066328585148, 'clip_ratio/low_min': 0.006601066328585148, 'clip_ratio/high_mean': 0.014982880558818579, 'clip_ratio/high_max': 0.014982880558818579, 'clip_ratio/region_mean': 0.021583946887403727, 'reward_total_mean': 0.6912178993225098, 'reward_meter_mean': 0.8571816086769104, 'reward_meter_std': 0.171591117978096, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8020833134651184, 'reward_repeat_penalty_std': 0.0431290864944458, 'reward_total_composite_mean': 0.6912178993225098, 'reward_total_composite_std': 0.158027783036232, 'epoch': 0.08}

 59%|█████▊    | 1934/3300 [2:30:28<2:01:25,  5.33s/it]INFO 04-12 01:07:20 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1935/3300 [2:30:33<1:57:26,  5.16s/it]2026-04-12 01:07:24,563 | INFO | train_grpo_train | metrics_logged mode=train step=1935

                                                       
{'loss': -0.0005, 'grad_norm': 3.562734365463257, 'learning_rate': 4.13939393939394e-06, 'num_tokens': 4366373.0, 'completions/mean_length': 67.5, 'completions/min_length': 66.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.72654128074646, 'rewards/meter/std': 0.2970520853996277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.72654128074646, 'rewards/total_composite/std': 0.2970520853996277, 'reward': 0.72654128074646, 'reward_std': 0.2970520555973053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03413822129368782, 'sampling/sampling_logp_difference/max': 1.4129490852355957, 'sampling/importance_sampling_ratio/min': 0.24342434108257294, 'sampling/importance_sampling_ratio/mean': 0.99456787109375, 'sampling/importance_sampling_ratio/max': 1.5096417665481567, 'entropy': 0.14524980820715427, 'clip_ratio/low_mean': 0.007359307492151856, 'clip_ratio/low_min': 0.007359307492151856, 'clip_ratio/high_mean': 0.025740933255292475, 'clip_ratio/high_max': 0.025740933255292475, 'clip_ratio/region_mean': 0.03310024074744433, 'reward_total_mean': 0.72654128074646, 'reward_meter_mean': 0.72654128074646, 'reward_meter_std': 0.2970520853996277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.72654128074646, 'reward_total_composite_std': 0.2970520853996277, 'epoch': 0.08}

 59%|█████▊    | 1935/3300 [2:30:33<1:57:26,  5.16s/it]INFO 04-12 01:07:24 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1936/3300 [2:30:38<1:56:21,  5.12s/it]2026-04-12 01:07:29,580 | INFO | train_grpo_train | metrics_logged mode=train step=1936

                                                       
{'loss': -0.0046, 'grad_norm': 3.389194965362549, 'learning_rate': 4.136363636363637e-06, 'num_tokens': 4368431.0, 'completions/mean_length': 99.25, 'completions/min_length': 98.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.25, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.8136383891105652, 'rewards/meter/std': 0.040068093687295914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.7909245491027832, 'rewards/total_composite/std': 0.028373466804623604, 'reward': 0.7909245491027832, 'reward_std': 0.028373470529913902, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01888117380440235, 'sampling/sampling_logp_difference/max': 1.0089292526245117, 'sampling/importance_sampling_ratio/min': 0.4062268137931824, 'sampling/importance_sampling_ratio/mean': 1.0059889554977417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15317998873069882, 'clip_ratio/low_mean': 0.006351783173158765, 'clip_ratio/low_min': 0.006351783173158765, 'clip_ratio/high_mean': 0.012465349864214659, 'clip_ratio/high_max': 0.012465349864214659, 'clip_ratio/region_mean': 0.018817133037373424, 'reward_total_mean': 0.7909245491027832, 'reward_meter_mean': 0.8136383891105652, 'reward_meter_std': 0.040068093687295914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.7909245491027832, 'reward_total_composite_std': 0.028373466804623604, 'epoch': 0.08}

 59%|█████▊    | 1936/3300 [2:30:38<1:56:21,  5.12s/it]INFO 04-12 01:07:29 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1937/3300 [2:30:42<1:51:24,  4.90s/it]2026-04-12 01:07:33,989 | INFO | train_grpo_train | metrics_logged mode=train step=1937

                                                       
{'loss': 0.0008, 'grad_norm': 1.4040637016296387, 'learning_rate': 4.133333333333333e-06, 'num_tokens': 4370119.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.994684100151062, 'rewards/meter/std': 5.123758455738425e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.994684100151062, 'rewards/total_composite/std': 5.123758455738425e-05, 'reward': 0.994684100151062, 'reward_std': 5.122838410898112e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004311581142246723, 'sampling/sampling_logp_difference/max': 0.4578728675842285, 'sampling/importance_sampling_ratio/min': 0.6326279044151306, 'sampling/importance_sampling_ratio/mean': 0.9982123374938965, 'sampling/importance_sampling_ratio/max': 1.146704077720642, 'entropy': 0.018475122982636094, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0021551724057644606, 'clip_ratio/high_max': 0.0021551724057644606, 'clip_ratio/region_mean': 0.0021551724057644606, 'reward_total_mean': 0.994684100151062, 'reward_meter_mean': 0.994684100151062, 'reward_meter_std': 5.123758455738425e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.994684100151062, 'reward_total_composite_std': 5.123758455738425e-05, 'epoch': 0.08}

 59%|█████▊    | 1937/3300 [2:30:42<1:51:24,  4.90s/it]INFO 04-12 01:07:34 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1938/3300 [2:30:47<1:50:36,  4.87s/it]2026-04-12 01:07:38,783 | INFO | train_grpo_train | metrics_logged mode=train step=1938

                                                       
{'loss': -0.0008, 'grad_norm': 3.30529522895813, 'learning_rate': 4.1303030303030305e-06, 'num_tokens': 4372001.0, 'completions/mean_length': 67.25, 'completions/min_length': 66.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8598266839981079, 'rewards/meter/std': 0.02622251957654953, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8598266839981079, 'rewards/total_composite/std': 0.02622251957654953, 'reward': 0.8598266839981079, 'reward_std': 0.02622251957654953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011166814714670181, 'sampling/sampling_logp_difference/max': 1.2125158309936523, 'sampling/importance_sampling_ratio/min': 0.2974480092525482, 'sampling/importance_sampling_ratio/mean': 1.0010411739349365, 'sampling/importance_sampling_ratio/max': 1.5219361782073975, 'entropy': 0.048515188973397017, 'clip_ratio/low_mean': 0.007407813798636198, 'clip_ratio/low_min': 0.007407813798636198, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.007407813798636198, 'reward_total_mean': 0.8598266839981079, 'reward_meter_mean': 0.8598266839981079, 'reward_meter_std': 0.02622251957654953, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8598266839981079, 'reward_total_composite_std': 0.02622251957654953, 'epoch': 0.08}

 59%|█████▊    | 1938/3300 [2:30:47<1:50:36,  4.87s/it]INFO 04-12 01:07:39 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1939/3300 [2:30:52<1:49:42,  4.84s/it]2026-04-12 01:07:43,538 | INFO | train_grpo_train | metrics_logged mode=train step=1939

                                                       
{'loss': 0.0039, 'grad_norm': 3.0253310203552246, 'learning_rate': 4.127272727272728e-06, 'num_tokens': 4373689.0, 'completions/mean_length': 72.0, 'completions/min_length': 69.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8803695440292358, 'rewards/meter/std': 0.31208541989326477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8803695440292358, 'rewards/total_composite/std': 0.31208541989326477, 'reward': 0.8803695440292358, 'reward_std': 0.3120853900909424, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035320769995450974, 'sampling/sampling_logp_difference/max': 0.9647955894470215, 'sampling/importance_sampling_ratio/min': 0.3810611069202423, 'sampling/importance_sampling_ratio/mean': 1.0067641735076904, 'sampling/importance_sampling_ratio/max': 1.6593668460845947, 'entropy': 0.28373553417623043, 'clip_ratio/low_mean': 0.0051369862630963326, 'clip_ratio/low_min': 0.0051369862630963326, 'clip_ratio/high_mean': 0.02452767826616764, 'clip_ratio/high_max': 0.02452767826616764, 'clip_ratio/region_mean': 0.029664664529263973, 'reward_total_mean': 0.8803695440292358, 'reward_meter_mean': 0.8803695440292358, 'reward_meter_std': 0.31208541989326477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8803695440292358, 'reward_total_composite_std': 0.31208541989326477, 'epoch': 0.08}

 59%|█████▉    | 1939/3300 [2:30:52<1:49:42,  4.84s/it]INFO 04-12 01:07:43 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1940/3300 [2:30:57<1:47:55,  4.76s/it]2026-04-12 01:07:48,131 | INFO | train_grpo_train | metrics_logged mode=train step=1940

                                                       
{'loss': -0.0001, 'grad_norm': 0.002891062991693616, 'learning_rate': 4.124242424242424e-06, 'num_tokens': 4375497.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9989398717880249, 'rewards/meter/std': 2.402423206149251e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989398717880249, 'rewards/total_composite/std': 2.402423206149251e-06, 'reward': 0.9989398717880249, 'reward_std': 2.396394847892225e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0017837814521044493, 'sampling/sampling_logp_difference/max': 0.4983246326446533, 'sampling/importance_sampling_ratio/min': 0.6075477004051208, 'sampling/importance_sampling_ratio/mean': 0.9999057054519653, 'sampling/importance_sampling_ratio/max': 1.0235236883163452, 'entropy': 0.010286982113029808, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.9989398717880249, 'reward_meter_mean': 0.9989398717880249, 'reward_meter_std': 2.402423206149251e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989398717880249, 'reward_total_composite_std': 2.402423206149251e-06, 'epoch': 0.08}

 59%|█████▉    | 1940/3300 [2:30:57<1:47:55,  4.76s/it]INFO 04-12 01:07:48 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1941/3300 [2:31:07<2:23:49,  6.35s/it]2026-04-12 01:07:58,181 | INFO | train_grpo_train | metrics_logged mode=train step=1941

                                                       
{'loss': -0.3041, 'grad_norm': 0.7915775775909424, 'learning_rate': 4.1212121212121215e-06, 'num_tokens': 4380383.0, 'completions/mean_length': 472.75, 'completions/min_length': 449.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 467.14288330078125, 'completions/min_terminated_length': 449.0, 'completions/max_terminated_length': 489.0, 'rewards/meter/mean': 0.9804998636245728, 'rewards/meter/std': 0.030505068600177765, 'rewards/count_adherence/mean': 0.6691176891326904, 'rewards/count_adherence/std': 0.062391772866249084, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9617094993591309, 'rewards/repeat_penalty/std': 0.04899247735738754, 'rewards/total_composite/mean': 0.5711581707000732, 'rewards/total_composite/std': 0.2336382418870926, 'reward': 0.5711581707000732, 'reward_std': 0.2336382418870926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.056584302335977554, 'sampling/sampling_logp_difference/max': 1.680664300918579, 'sampling/importance_sampling_ratio/min': 0.1862502098083496, 'sampling/importance_sampling_ratio/mean': 1.0137178897857666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5811280272901058, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03567814873531461, 'clip_ratio/high_max': 0.03567814873531461, 'clip_ratio/region_mean': 0.03567814873531461, 'reward_total_mean': 0.5711581707000732, 'reward_meter_mean': 0.9804998636245728, 'reward_meter_std': 0.030505068600177765, 'reward_count_adherence_mean': 0.6691176891326904, 'reward_count_adherence_std': 0.062391772866249084, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9617094993591309, 'reward_repeat_penalty_std': 0.04899247735738754, 'reward_total_composite_mean': 0.5711581707000732, 'reward_total_composite_std': 0.2336382418870926, 'epoch': 0.08}

 59%|█████▉    | 1941/3300 [2:31:07<2:23:49,  6.35s/it]INFO 04-12 01:07:58 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1942/3300 [2:31:11<2:12:51,  5.87s/it]2026-04-12 01:08:02,924 | INFO | train_grpo_train | metrics_logged mode=train step=1942

                                                       
{'loss': -0.008, 'grad_norm': 4.029205799102783, 'learning_rate': 4.118181818181819e-06, 'num_tokens': 4382253.0, 'completions/mean_length': 77.75, 'completions/min_length': 76.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.75, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9882785081863403, 'rewards/meter/std': 0.018005145713686943, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8649736642837524, 'rewards/total_composite/std': 0.34996482729911804, 'reward': 0.8649736642837524, 'reward_std': 0.34996482729911804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06439422816038132, 'sampling/sampling_logp_difference/max': 1.7122077941894531, 'sampling/importance_sampling_ratio/min': 0.1804669201374054, 'sampling/importance_sampling_ratio/mean': 1.0108014345169067, 'sampling/importance_sampling_ratio/max': 1.879008412361145, 'entropy': 0.5831749886274338, 'clip_ratio/low_mean': 0.004934210330247879, 'clip_ratio/low_min': 0.004934210330247879, 'clip_ratio/high_mean': 0.03670453419908881, 'clip_ratio/high_max': 0.03670453419908881, 'clip_ratio/region_mean': 0.04163874452933669, 'reward_total_mean': 0.8649736642837524, 'reward_meter_mean': 0.9882785081863403, 'reward_meter_std': 0.018005145713686943, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8649736642837524, 'reward_total_composite_std': 0.34996482729911804, 'epoch': 0.08}

 59%|█████▉    | 1942/3300 [2:31:11<2:12:51,  5.87s/it]INFO 04-12 01:08:03 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1943/3300 [2:31:16<2:06:52,  5.61s/it]2026-04-12 01:08:07,931 | INFO | train_grpo_train | metrics_logged mode=train step=1943

                                                       
{'loss': 0.0281, 'grad_norm': 2.609006643295288, 'learning_rate': 4.115151515151515e-06, 'num_tokens': 4384317.0, 'completions/mean_length': 96.0, 'completions/min_length': 95.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.0, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9970648288726807, 'rewards/meter/std': 0.0006929152878001332, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9721810817718506, 'rewards/total_composite/std': 0.07107478380203247, 'reward': 0.9721810817718506, 'reward_std': 0.07107476890087128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00461819302290678, 'sampling/sampling_logp_difference/max': 0.3813199996948242, 'sampling/importance_sampling_ratio/min': 0.6829593181610107, 'sampling/importance_sampling_ratio/mean': 1.0022590160369873, 'sampling/importance_sampling_ratio/max': 1.3738577365875244, 'entropy': 0.030685836798511446, 'clip_ratio/low_mean': 0.0012135922443121672, 'clip_ratio/low_min': 0.0012135922443121672, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0012135922443121672, 'reward_total_mean': 0.9721810817718506, 'reward_meter_mean': 0.9970648288726807, 'reward_meter_std': 0.0006929152878001332, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9721810817718506, 'reward_total_composite_std': 0.07107478380203247, 'epoch': 0.08}

 59%|█████▉    | 1943/3300 [2:31:16<2:06:52,  5.61s/it]INFO 04-12 01:08:08 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1944/3300 [2:31:21<1:59:23,  5.28s/it]2026-04-12 01:08:12,453 | INFO | train_grpo_train | metrics_logged mode=train step=1944

                                                       
{'loss': -0.0001, 'grad_norm': 0.053728338330984116, 'learning_rate': 4.112121212121212e-06, 'num_tokens': 4385917.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9970229864120483, 'rewards/meter/std': 8.450447239738423e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970229864120483, 'rewards/total_composite/std': 8.450447239738423e-06, 'reward': 0.9970229864120483, 'reward_std': 8.441417776339222e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0021621037740260363, 'sampling/sampling_logp_difference/max': 0.1747570037841797, 'sampling/importance_sampling_ratio/min': 0.8396610021591187, 'sampling/importance_sampling_ratio/mean': 1.0014358758926392, 'sampling/importance_sampling_ratio/max': 1.081364393234253, 'entropy': 0.015167628531344235, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0019841270986944437, 'clip_ratio/high_max': 0.0019841270986944437, 'clip_ratio/region_mean': 0.0019841270986944437, 'reward_total_mean': 0.9970229864120483, 'reward_meter_mean': 0.9970229864120483, 'reward_meter_std': 8.450447239738423e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970229864120483, 'reward_total_composite_std': 8.450447239738423e-06, 'epoch': 0.08}

 59%|█████▉    | 1944/3300 [2:31:21<1:59:23,  5.28s/it]INFO 04-12 01:08:12 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1945/3300 [2:31:25<1:54:22,  5.06s/it]2026-04-12 01:08:17,006 | INFO | train_grpo_train | metrics_logged mode=train step=1945

                                                       
{'loss': -0.0001, 'grad_norm': 0.013028639368712902, 'learning_rate': 4.10909090909091e-06, 'num_tokens': 4387733.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9989398717880249, 'rewards/meter/std': 2.402423206149251e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989398717880249, 'rewards/total_composite/std': 2.402423206149251e-06, 'reward': 0.9989398717880249, 'reward_std': 2.396394847892225e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002988132182508707, 'sampling/sampling_logp_difference/max': 0.1946016550064087, 'sampling/importance_sampling_ratio/min': 0.8231624960899353, 'sampling/importance_sampling_ratio/mean': 1.0004053115844727, 'sampling/importance_sampling_ratio/max': 1.1326637268066406, 'entropy': 0.019811521749943495, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.9989398717880249, 'reward_meter_mean': 0.9989398717880249, 'reward_meter_std': 2.402423206149251e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989398717880249, 'reward_total_composite_std': 2.402423206149251e-06, 'epoch': 0.08}

 59%|█████▉    | 1945/3300 [2:31:26<1:54:22,  5.06s/it]INFO 04-12 01:08:17 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1946/3300 [2:31:30<1:52:41,  4.99s/it]2026-04-12 01:08:21,842 | INFO | train_grpo_train | metrics_logged mode=train step=1946

                                                       
{'loss': 0.0117, 'grad_norm': 2.780980110168457, 'learning_rate': 4.106060606060606e-06, 'num_tokens': 4389585.0, 'completions/mean_length': 67.5, 'completions/min_length': 66.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8567509055137634, 'rewards/meter/std': 0.07118333131074905, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8567509055137634, 'rewards/total_composite/std': 0.07118333131074905, 'reward': 0.8567509055137634, 'reward_std': 0.07118332386016846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019841060042381287, 'sampling/sampling_logp_difference/max': 1.137892723083496, 'sampling/importance_sampling_ratio/min': 0.3204936981201172, 'sampling/importance_sampling_ratio/mean': 1.0084105730056763, 'sampling/importance_sampling_ratio/max': 1.5270895957946777, 'entropy': 0.18513774313032627, 'clip_ratio/low_mean': 0.003623949596658349, 'clip_ratio/low_min': 0.003623949596658349, 'clip_ratio/high_mean': 0.011114974273368716, 'clip_ratio/high_max': 0.011114974273368716, 'clip_ratio/region_mean': 0.014738923870027065, 'reward_total_mean': 0.8567509055137634, 'reward_meter_mean': 0.8567509055137634, 'reward_meter_std': 0.07118333131074905, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8567509055137634, 'reward_total_composite_std': 0.07118333131074905, 'epoch': 0.08}

 59%|█████▉    | 1946/3300 [2:31:30<1:52:41,  4.99s/it]INFO 04-12 01:08:22 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1947/3300 [2:31:35<1:50:52,  4.92s/it]2026-04-12 01:08:26,567 | INFO | train_grpo_train | metrics_logged mode=train step=1947

                                                       
{'loss': 0.003, 'grad_norm': 2.5665481090545654, 'learning_rate': 4.103030303030303e-06, 'num_tokens': 4391466.0, 'completions/mean_length': 73.125, 'completions/min_length': 73.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9979450702667236, 'rewards/meter/std': 0.00038199007394723594, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979450702667236, 'rewards/total_composite/std': 0.00038199007394723594, 'reward': 0.9979450702667236, 'reward_std': 0.00038197485264390707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02320614643394947, 'sampling/sampling_logp_difference/max': 1.1087846755981445, 'sampling/importance_sampling_ratio/min': 0.3299597203731537, 'sampling/importance_sampling_ratio/mean': 0.9995145201683044, 'sampling/importance_sampling_ratio/max': 1.5535597801208496, 'entropy': 0.1381580764427781, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.011963161756284535, 'clip_ratio/high_max': 0.011963161756284535, 'clip_ratio/region_mean': 0.011963161756284535, 'reward_total_mean': 0.9979450702667236, 'reward_meter_mean': 0.9979450702667236, 'reward_meter_std': 0.00038199007394723594, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979450702667236, 'reward_total_composite_std': 0.00038199007394723594, 'epoch': 0.08}

 59%|█████▉    | 1947/3300 [2:31:35<1:50:52,  4.92s/it]INFO 04-12 01:08:26 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1948/3300 [2:31:40<1:49:59,  4.88s/it]2026-04-12 01:08:31,376 | INFO | train_grpo_train | metrics_logged mode=train step=1948

                                                       
{'loss': 0.0029, 'grad_norm': 2.1724514961242676, 'learning_rate': 4.1e-06, 'num_tokens': 4393381.0, 'completions/mean_length': 73.375, 'completions/min_length': 72.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9980074167251587, 'rewards/meter/std': 0.0006349986069835722, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980074167251587, 'rewards/total_composite/std': 0.0006349986069835722, 'reward': 0.9980074167251587, 'reward_std': 0.0006350158364512026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025800568982958794, 'sampling/sampling_logp_difference/max': 0.789684534072876, 'sampling/importance_sampling_ratio/min': 0.4539879858493805, 'sampling/importance_sampling_ratio/mean': 1.0009372234344482, 'sampling/importance_sampling_ratio/max': 1.8275814056396484, 'entropy': 0.16229457408189774, 'clip_ratio/low_mean': 0.00856354646384716, 'clip_ratio/low_min': 0.00856354646384716, 'clip_ratio/high_mean': 0.012034508981741965, 'clip_ratio/high_max': 0.012034508981741965, 'clip_ratio/region_mean': 0.020598055445589125, 'reward_total_mean': 0.9980074167251587, 'reward_meter_mean': 0.9980074167251587, 'reward_meter_std': 0.0006349986069835722, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980074167251587, 'reward_total_composite_std': 0.0006349986069835722, 'epoch': 0.08}

 59%|█████▉    | 1948/3300 [2:31:40<1:49:59,  4.88s/it]INFO 04-12 01:08:31 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1949/3300 [2:31:45<1:55:07,  5.11s/it]2026-04-12 01:08:37,024 | INFO | train_grpo_train | metrics_logged mode=train step=1949

                                                       
{'loss': -0.0289, 'grad_norm': 5.015791893005371, 'learning_rate': 4.096969696969697e-06, 'num_tokens': 4395881.0, 'completions/mean_length': 122.5, 'completions/min_length': 110.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.5, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.10345004498958588, 'rewards/meter/std': 0.17283838987350464, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.851190447807312, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.0730920359492302, 'rewards/total_composite/std': 0.11985667049884796, 'reward': 0.0730920359492302, 'reward_std': 0.11985667049884796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03781670331954956, 'sampling/sampling_logp_difference/max': 1.3046746253967285, 'sampling/importance_sampling_ratio/min': 0.2712607979774475, 'sampling/importance_sampling_ratio/mean': 0.9997954368591309, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1550168227404356, 'clip_ratio/low_mean': 0.023918771417811513, 'clip_ratio/low_min': 0.023918771417811513, 'clip_ratio/high_mean': 0.012344544753432274, 'clip_ratio/high_max': 0.012344544753432274, 'clip_ratio/region_mean': 0.03626331617124379, 'reward_total_mean': 0.0730920359492302, 'reward_meter_mean': 0.10345004498958588, 'reward_meter_std': 0.17283838987350464, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.851190447807312, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.0730920359492302, 'reward_total_composite_std': 0.11985667049884796, 'epoch': 0.08}

 59%|█████▉    | 1949/3300 [2:31:46<1:55:07,  5.11s/it]INFO 04-12 01:08:37 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1950/3300 [2:31:51<1:55:49,  5.15s/it]2026-04-12 01:08:42,259 | INFO | train_grpo_train | metrics_logged mode=train step=1950

                                                       
{'loss': 0.3165, 'grad_norm': 6.135134220123291, 'learning_rate': 4.093939393939394e-06, 'num_tokens': 4397640.0, 'completions/mean_length': 54.875, 'completions/min_length': 37.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9981762170791626, 'rewards/meter/std': 0.0004922283114865422, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.49901658296585083, 'rewards/total_composite/std': 0.5334713459014893, 'reward': 0.49901658296585083, 'reward_std': 0.5334712862968445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02422521449625492, 'sampling/sampling_logp_difference/max': 1.499140977859497, 'sampling/importance_sampling_ratio/min': 0.22332191467285156, 'sampling/importance_sampling_ratio/mean': 1.0079329013824463, 'sampling/importance_sampling_ratio/max': 1.6159636974334717, 'entropy': 0.1553973164409399, 'clip_ratio/low_mean': 0.00868320802692324, 'clip_ratio/low_min': 0.00868320802692324, 'clip_ratio/high_mean': 0.0033783784601837397, 'clip_ratio/high_max': 0.0033783784601837397, 'clip_ratio/region_mean': 0.012061586487106979, 'reward_total_mean': 0.49901658296585083, 'reward_meter_mean': 0.9981762170791626, 'reward_meter_std': 0.0004922283114865422, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.49901658296585083, 'reward_total_composite_std': 0.5334713459014893, 'epoch': 0.08}

 59%|█████▉    | 1950/3300 [2:31:51<1:55:49,  5.15s/it]INFO 04-12 01:08:42 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:34,  3.10s/it]

 23%|██▎       | 3/13 [00:09<00:33,  3.38s/it]

 31%|███       | 4/13 [00:15<00:37,  4.16s/it]

 38%|███▊      | 5/13 [00:20<00:36,  4.54s/it]

 46%|████▌     | 6/13 [00:26<00:35,  5.02s/it]

 54%|█████▍    | 7/13 [00:31<00:30,  5.03s/it]

 62%|██████▏   | 8/13 [00:38<00:27,  5.49s/it]

 69%|██████▉   | 9/13 [00:45<00:24,  6.07s/it]

 77%|███████▋  | 10/13 [00:50<00:17,  5.79s/it]

 85%|████████▍ | 11/13 [00:56<00:11,  5.71s/it]

 92%|█████████▏| 12/13 [01:03<00:06,  6.10s/it]

100%|██████████| 13/13 [01:08<00:00,  5.79s/it]2026-04-12 01:09:56,474 | INFO | train_grpo_train | metrics_logged mode=eval step=1950

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 74.1396, 'eval_samples_per_second': 1.403, 'eval_steps_per_second': 0.175, 'eval_num_tokens': 4397640.0, 'eval_completions/mean_length': 219.20192307692307, 'eval_completions/min_length': 67.84615384615384, 'eval_completions/max_length': 383.15384615384613, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 216.13324209359976, 'eval_completions/min_terminated_length': 67.84615384615384, 'eval_completions/max_terminated_length': 372.84615384615387, 'eval_rewards/meter/mean': 0.7193602369381831, 'eval_rewards/meter/std': 0.41060519218444824, 'eval_rewards/count_adherence/mean': 0.8801831007003784, 'eval_rewards/count_adherence/std': 0.15413257680260217, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.8391693326143118, 'eval_rewards/repeat_penalty/std': 0.14583237583820635, 'eval_rewards/total_composite/mean': 0.5268221841408656, 'eval_rewards/total_composite/std': 0.36114954260679394, 'eval_reward': 0.5268221841408656, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.02039779959103236, 'eval_sampling/sampling_logp_difference/max': 1.0237883787888746, 'eval_sampling/importance_sampling_ratio/min': 0.3655441105365753, 'eval_sampling/importance_sampling_ratio/mean': 1.0057256405170147, 'eval_sampling/importance_sampling_ratio/max': 1.5754675223277166, 'eval_entropy': 0.2521360000738731, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5268221841408656, 'eval_reward_meter_mean': 0.7193602369381831, 'eval_reward_meter_std': 0.41060519218444824, 'eval_reward_count_adherence_mean': 0.8801831007003784, 'eval_reward_count_adherence_std': 0.15413257680260217, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.8391693326143118, 'eval_reward_repeat_penalty_std': 0.14583237583820635, 'eval_reward_total_composite_mean': 0.5268221841408656, 'eval_reward_total_composite_std': 0.36114954260679394, 'epoch': 0.08}

 59%|█████▉    | 1950/3300 [2:33:05<1:55:49,  5.15s/it]

100%|██████████| 13/13 [01:08<00:00,  5.79s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 59%|█████▉    | 1951/3300 [2:33:13<10:38:22, 28.39s/it]2026-04-12 01:10:04,883 | INFO | train_grpo_train | metrics_logged mode=train step=1951

                                                        
{'loss': 0.0003, 'grad_norm': 0.36703649163246155, 'learning_rate': 4.0909090909090915e-06, 'num_tokens': 4399744.0, 'completions/mean_length': 90.0, 'completions/min_length': 90.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.0, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9955518841743469, 'rewards/meter/std': 1.7999940610025078e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955518841743469, 'rewards/total_composite/std': 1.7999940610025078e-05, 'reward': 0.9955518841743469, 'reward_std': 1.8004628145718016e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004722072742879391, 'sampling/sampling_logp_difference/max': 0.8973627090454102, 'sampling/importance_sampling_ratio/min': 0.40764331817626953, 'sampling/importance_sampling_ratio/mean': 0.9992460012435913, 'sampling/importance_sampling_ratio/max': 1.191457748413086, 'entropy': 0.015661518438719213, 'clip_ratio/low_mean': 0.0027777778450399637, 'clip_ratio/low_min': 0.0027777778450399637, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0027777778450399637, 'reward_total_mean': 0.9955518841743469, 'reward_meter_mean': 0.9955518841743469, 'reward_meter_std': 1.7999940610025078e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955518841743469, 'reward_total_composite_std': 1.7999940610025078e-05, 'epoch': 0.08}

 59%|█████▉    | 1951/3300 [2:33:13<10:38:22, 28.39s/it]INFO 04-12 01:10:05 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1952/3300 [2:33:19<8:03:43, 21.53s/it] 2026-04-12 01:10:10,405 | INFO | train_grpo_train | metrics_logged mode=train step=1952

                                                       
{'loss': -0.0014, 'grad_norm': 4.403555870056152, 'learning_rate': 4.087878787878789e-06, 'num_tokens': 4401615.0, 'completions/mean_length': 71.875, 'completions/min_length': 70.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.987866997718811, 'rewards/meter/std': 0.015270087867975235, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.987866997718811, 'rewards/total_composite/std': 0.015270087867975235, 'reward': 0.987866997718811, 'reward_std': 0.015270096249878407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028592025861144066, 'sampling/sampling_logp_difference/max': 0.8273954391479492, 'sampling/importance_sampling_ratio/min': 0.43718650937080383, 'sampling/importance_sampling_ratio/mean': 1.002774715423584, 'sampling/importance_sampling_ratio/max': 1.4481229782104492, 'entropy': 0.23499170877039433, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.031067086150869727, 'clip_ratio/high_max': 0.031067086150869727, 'clip_ratio/region_mean': 0.031067086150869727, 'reward_total_mean': 0.987866997718811, 'reward_meter_mean': 0.987866997718811, 'reward_meter_std': 0.015270087867975235, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.987866997718811, 'reward_total_composite_std': 0.015270087867975235, 'epoch': 0.08}

 59%|█████▉    | 1952/3300 [2:33:19<8:03:43, 21.53s/it]INFO 04-12 01:10:10 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1953/3300 [2:33:24<6:12:55, 16.61s/it]2026-04-12 01:10:15,533 | INFO | train_grpo_train | metrics_logged mode=train step=1953

                                                       
{'loss': 0.0586, 'grad_norm': 6.583865165710449, 'learning_rate': 4.084848484848485e-06, 'num_tokens': 4403483.0, 'completions/mean_length': 74.5, 'completions/min_length': 59.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.43240049481391907, 'rewards/meter/std': 0.30809712409973145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.41915929317474365, 'rewards/total_composite/std': 0.30558672547340393, 'reward': 0.41915929317474365, 'reward_std': 0.30558669567108154, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043784063309431076, 'sampling/sampling_logp_difference/max': 1.1320176124572754, 'sampling/importance_sampling_ratio/min': 0.3223821520805359, 'sampling/importance_sampling_ratio/mean': 1.0041871070861816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2592748161405325, 'clip_ratio/low_mean': 0.012907008291222155, 'clip_ratio/low_min': 0.012907008291222155, 'clip_ratio/high_mean': 0.029118790524080396, 'clip_ratio/high_max': 0.029118790524080396, 'clip_ratio/region_mean': 0.04202579881530255, 'reward_total_mean': 0.41915929317474365, 'reward_meter_mean': 0.43240049481391907, 'reward_meter_std': 0.30809712409973145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.41915929317474365, 'reward_total_composite_std': 0.30558672547340393, 'epoch': 0.08}

 59%|█████▉    | 1953/3300 [2:33:24<6:12:55, 16.61s/it]INFO 04-12 01:10:15 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1954/3300 [2:33:29<4:54:02, 13.11s/it]2026-04-12 01:10:20,472 | INFO | train_grpo_train | metrics_logged mode=train step=1954

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.081818181818182e-06, 'num_tokens': 4405203.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9989815950393677, 'rewards/meter/std': 5.247284934739582e-05, 'rewards/count_adherence/mean': 0.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.0, 'rewards/total_composite/std': 0.0, 'reward': 0.0, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.007122131530195475, 'sampling/sampling_logp_difference/max': 0.5082035064697266, 'sampling/importance_sampling_ratio/min': 0.6015753746032715, 'sampling/importance_sampling_ratio/mean': 1.001094937324524, 'sampling/importance_sampling_ratio/max': 1.501068353652954, 'entropy': 0.02731157885864377, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.0, 'reward_meter_mean': 0.9989815950393677, 'reward_meter_std': 5.247284934739582e-05, 'reward_count_adherence_mean': 0.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.0, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 59%|█████▉    | 1954/3300 [2:33:29<4:54:02, 13.11s/it]INFO 04-12 01:10:20 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1955/3300 [2:33:37<4:18:40, 11.54s/it]2026-04-12 01:10:28,354 | INFO | train_grpo_train | metrics_logged mode=train step=1955

                                                       
{'loss': -0.0135, 'grad_norm': 1.9077023267745972, 'learning_rate': 4.07878787878788e-06, 'num_tokens': 4409222.0, 'completions/mean_length': 300.375, 'completions/min_length': 293.0, 'completions/max_length': 315.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 300.375, 'completions/min_terminated_length': 293.0, 'completions/max_terminated_length': 315.0, 'rewards/meter/mean': 0.9946515560150146, 'rewards/meter/std': 0.002225227653980255, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7741013169288635, 'rewards/repeat_penalty/std': 0.08203362673521042, 'rewards/total_composite/mean': 0.6738342642784119, 'rewards/total_composite/std': 0.07273107022047043, 'reward': 0.6738342642784119, 'reward_std': 0.07273106276988983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018673671409487724, 'sampling/sampling_logp_difference/max': 2.512004852294922, 'sampling/importance_sampling_ratio/min': 0.08110546320676804, 'sampling/importance_sampling_ratio/mean': 1.0040055513381958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12408511992543936, 'clip_ratio/low_mean': 0.007172186364186928, 'clip_ratio/low_min': 0.007172186364186928, 'clip_ratio/high_mean': 0.005274800234474242, 'clip_ratio/high_max': 0.005274800234474242, 'clip_ratio/region_mean': 0.01244698659866117, 'reward_total_mean': 0.6738342642784119, 'reward_meter_mean': 0.9946515560150146, 'reward_meter_std': 0.002225227653980255, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7741013169288635, 'reward_repeat_penalty_std': 0.08203362673521042, 'reward_total_composite_mean': 0.6738342642784119, 'reward_total_composite_std': 0.07273107022047043, 'epoch': 0.08}

 59%|█████▉    | 1955/3300 [2:33:37<4:18:40, 11.54s/it]INFO 04-12 01:10:28 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1956/3300 [2:33:41<3:31:26,  9.44s/it]2026-04-12 01:10:32,890 | INFO | train_grpo_train | metrics_logged mode=train step=1956

                                                       
{'loss': 0.0496, 'grad_norm': 8.814888000488281, 'learning_rate': 4.075757575757576e-06, 'num_tokens': 4410857.0, 'completions/mean_length': 47.375, 'completions/min_length': 43.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6137043237686157, 'rewards/meter/std': 0.33018791675567627, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6137043237686157, 'rewards/total_composite/std': 0.33018791675567627, 'reward': 0.6137043237686157, 'reward_std': 0.33018791675567627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06914983689785004, 'sampling/sampling_logp_difference/max': 1.6852831840515137, 'sampling/importance_sampling_ratio/min': 0.18539191782474518, 'sampling/importance_sampling_ratio/mean': 0.9924876689910889, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3848374430090189, 'clip_ratio/low_mean': 0.028407905949279666, 'clip_ratio/low_min': 0.028407905949279666, 'clip_ratio/high_mean': 0.041018035262823105, 'clip_ratio/high_max': 0.041018035262823105, 'clip_ratio/region_mean': 0.06942594121210277, 'reward_total_mean': 0.6137043237686157, 'reward_meter_mean': 0.6137043237686157, 'reward_meter_std': 0.33018791675567627, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6137043237686157, 'reward_total_composite_std': 0.33018791675567627, 'epoch': 0.08}

 59%|█████▉    | 1956/3300 [2:33:41<3:31:26,  9.44s/it]INFO 04-12 01:10:33 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1957/3300 [2:33:47<3:04:26,  8.24s/it]2026-04-12 01:10:38,334 | INFO | train_grpo_train | metrics_logged mode=train step=1957

                                                       
{'loss': 0.0042, 'grad_norm': 2.345047950744629, 'learning_rate': 4.072727272727273e-06, 'num_tokens': 4413295.0, 'completions/mean_length': 136.75, 'completions/min_length': 136.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.75, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.9962515830993652, 'rewards/meter/std': 0.0024396260268986225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8717420101165771, 'rewards/total_composite/std': 0.05084509775042534, 'reward': 0.8717420101165771, 'reward_std': 0.05084508657455444, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01802092418074608, 'sampling/sampling_logp_difference/max': 3.096953868865967, 'sampling/importance_sampling_ratio/min': 0.045186638832092285, 'sampling/importance_sampling_ratio/mean': 1.0027059316635132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0967075265944004, 'clip_ratio/low_mean': 0.009184467780869454, 'clip_ratio/low_min': 0.009184467780869454, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.011022703081835061, 'reward_total_mean': 0.8717420101165771, 'reward_meter_mean': 0.9962515830993652, 'reward_meter_std': 0.0024396260268986225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8717420101165771, 'reward_total_composite_std': 0.05084509775042534, 'epoch': 0.08}

 59%|█████▉    | 1957/3300 [2:33:47<3:04:26,  8.24s/it]INFO 04-12 01:10:38 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1958/3300 [2:33:51<2:36:33,  7.00s/it]2026-04-12 01:10:42,440 | INFO | train_grpo_train | metrics_logged mode=train step=1958

                                                       
{'loss': 0.0012, 'grad_norm': 10.1416654586792, 'learning_rate': 4.0696969696969706e-06, 'num_tokens': 4414671.0, 'completions/mean_length': 33.0, 'completions/min_length': 33.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9584056735038757, 'rewards/meter/std': 0.020386451855301857, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9584056735038757, 'rewards/total_composite/std': 0.020386451855301857, 'reward': 0.9584056735038757, 'reward_std': 0.020386451855301857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025948278605937958, 'sampling/sampling_logp_difference/max': 0.9829981327056885, 'sampling/importance_sampling_ratio/min': 0.37418755888938904, 'sampling/importance_sampling_ratio/mean': 1.0064359903335571, 'sampling/importance_sampling_ratio/max': 1.6090956926345825, 'entropy': 0.20033110305666924, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007575757801532745, 'clip_ratio/high_max': 0.007575757801532745, 'clip_ratio/region_mean': 0.007575757801532745, 'reward_total_mean': 0.9584056735038757, 'reward_meter_mean': 0.9584056735038757, 'reward_meter_std': 0.020386451855301857, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9584056735038757, 'reward_total_composite_std': 0.020386451855301857, 'epoch': 0.08}

 59%|█████▉    | 1958/3300 [2:33:51<2:36:33,  7.00s/it]INFO 04-12 01:10:42 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1959/3300 [2:33:56<2:22:03,  6.36s/it]2026-04-12 01:10:47,293 | INFO | train_grpo_train | metrics_logged mode=train step=1959

                                                       
{'loss': -0.0107, 'grad_norm': 8.695691108703613, 'learning_rate': 4.066666666666667e-06, 'num_tokens': 4416753.0, 'completions/mean_length': 76.25, 'completions/min_length': 70.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9713230133056641, 'rewards/meter/std': 0.06465231627225876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9713230133056641, 'rewards/total_composite/std': 0.06465231627225876, 'reward': 0.9713230133056641, 'reward_std': 0.06465231627225876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04479827731847763, 'sampling/sampling_logp_difference/max': 1.1926956176757812, 'sampling/importance_sampling_ratio/min': 0.303402304649353, 'sampling/importance_sampling_ratio/mean': 1.0147606134414673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44979994744062424, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.028867413057014346, 'clip_ratio/high_max': 0.028867413057014346, 'clip_ratio/region_mean': 0.03243884164839983, 'reward_total_mean': 0.9713230133056641, 'reward_meter_mean': 0.9713230133056641, 'reward_meter_std': 0.06465231627225876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9713230133056641, 'reward_total_composite_std': 0.06465231627225876, 'epoch': 0.08}

 59%|█████▉    | 1959/3300 [2:33:56<2:22:03,  6.36s/it]INFO 04-12 01:10:47 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1960/3300 [2:34:00<2:10:14,  5.83s/it]2026-04-12 01:10:51,902 | INFO | train_grpo_train | metrics_logged mode=train step=1960

                                                       
{'loss': -0.0106, 'grad_norm': 5.968605041503906, 'learning_rate': 4.063636363636364e-06, 'num_tokens': 4418613.0, 'completions/mean_length': 66.5, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7400355339050293, 'rewards/meter/std': 0.26013627648353577, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7400355339050293, 'rewards/total_composite/std': 0.26013627648353577, 'reward': 0.7400355339050293, 'reward_std': 0.26013627648353577, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03610033914446831, 'sampling/sampling_logp_difference/max': 1.647068977355957, 'sampling/importance_sampling_ratio/min': 0.1926136314868927, 'sampling/importance_sampling_ratio/mean': 1.0059431791305542, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.253327926620841, 'clip_ratio/low_mean': 0.007722355891019106, 'clip_ratio/low_min': 0.007722355891019106, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.01145369908772409, 'reward_total_mean': 0.7400355339050293, 'reward_meter_mean': 0.7400355339050293, 'reward_meter_std': 0.26013627648353577, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7400355339050293, 'reward_total_composite_std': 0.26013627648353577, 'epoch': 0.08}

 59%|█████▉    | 1960/3300 [2:34:00<2:10:14,  5.83s/it]INFO 04-12 01:10:52 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1961/3300 [2:34:05<2:01:39,  5.45s/it]2026-04-12 01:10:56,483 | INFO | train_grpo_train | metrics_logged mode=train step=1961

                                                       
{'loss': -0.0019, 'grad_norm': 0.6822566390037537, 'learning_rate': 4.060606060606061e-06, 'num_tokens': 4420309.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.996955156326294, 'rewards/meter/std': 0.0001835073926486075, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.996955156326294, 'rewards/total_composite/std': 0.0001835073926486075, 'reward': 0.996955156326294, 'reward_std': 0.00018350737809669226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003504997817799449, 'sampling/sampling_logp_difference/max': 0.9989252090454102, 'sampling/importance_sampling_ratio/min': 0.36827507615089417, 'sampling/importance_sampling_ratio/mean': 1.0001448392868042, 'sampling/importance_sampling_ratio/max': 1.0358319282531738, 'entropy': 0.014339913497678936, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.996955156326294, 'reward_meter_mean': 0.996955156326294, 'reward_meter_std': 0.0001835073926486075, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.996955156326294, 'reward_total_composite_std': 0.0001835073926486075, 'epoch': 0.08}

 59%|█████▉    | 1961/3300 [2:34:05<2:01:39,  5.45s/it]INFO 04-12 01:10:56 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1962/3300 [2:34:10<1:57:10,  5.25s/it]2026-04-12 01:11:01,262 | INFO | train_grpo_train | metrics_logged mode=train step=1962

                                                       
{'loss': 0.0256, 'grad_norm': 3.0988998413085938, 'learning_rate': 4.057575757575758e-06, 'num_tokens': 4422201.0, 'completions/mean_length': 77.5, 'completions/min_length': 73.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9931849837303162, 'rewards/meter/std': 0.005035399924963713, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9931849837303162, 'rewards/total_composite/std': 0.005035399924963713, 'reward': 0.9931849837303162, 'reward_std': 0.00503541948273778, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04507505148649216, 'sampling/sampling_logp_difference/max': 1.2003469467163086, 'sampling/importance_sampling_ratio/min': 0.30108973383903503, 'sampling/importance_sampling_ratio/mean': 1.0084387063980103, 'sampling/importance_sampling_ratio/max': 1.963693380355835, 'entropy': 0.5117241851985455, 'clip_ratio/low_mean': 0.006172839552164078, 'clip_ratio/low_min': 0.006172839552164078, 'clip_ratio/high_mean': 0.01814535935409367, 'clip_ratio/high_max': 0.01814535935409367, 'clip_ratio/region_mean': 0.02431819890625775, 'reward_total_mean': 0.9931849837303162, 'reward_meter_mean': 0.9931849837303162, 'reward_meter_std': 0.005035399924963713, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9931849837303162, 'reward_total_composite_std': 0.005035399924963713, 'epoch': 0.08}

 59%|█████▉    | 1962/3300 [2:34:10<1:57:10,  5.25s/it]INFO 04-12 01:11:01 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1963/3300 [2:34:15<1:54:11,  5.12s/it]2026-04-12 01:11:06,085 | INFO | train_grpo_train | metrics_logged mode=train step=1963

                                                       
{'loss': -0.01, 'grad_norm': 5.142487049102783, 'learning_rate': 4.054545454545455e-06, 'num_tokens': 4423998.0, 'completions/mean_length': 72.625, 'completions/min_length': 71.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.997757077217102, 'rewards/meter/std': 0.0013475378509610891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997757077217102, 'rewards/total_composite/std': 0.0013475378509610891, 'reward': 0.997757077217102, 'reward_std': 0.0013475335435941815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0390477180480957, 'sampling/sampling_logp_difference/max': 1.0736570358276367, 'sampling/importance_sampling_ratio/min': 0.3417564034461975, 'sampling/importance_sampling_ratio/mean': 0.9945013523101807, 'sampling/importance_sampling_ratio/max': 1.6228265762329102, 'entropy': 0.245115103200078, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.018698629923164845, 'clip_ratio/high_max': 0.018698629923164845, 'clip_ratio/region_mean': 0.022170852171257138, 'reward_total_mean': 0.997757077217102, 'reward_meter_mean': 0.997757077217102, 'reward_meter_std': 0.0013475378509610891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997757077217102, 'reward_total_composite_std': 0.0013475378509610891, 'epoch': 0.08}

 59%|█████▉    | 1963/3300 [2:34:15<1:54:11,  5.12s/it]INFO 04-12 01:11:06 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1964/3300 [2:34:21<2:00:28,  5.41s/it]2026-04-12 01:11:12,161 | INFO | train_grpo_train | metrics_logged mode=train step=1964

                                                       
{'loss': 0.0047, 'grad_norm': 2.332364320755005, 'learning_rate': 4.0515151515151516e-06, 'num_tokens': 4426910.0, 'completions/mean_length': 181.0, 'completions/min_length': 180.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 181.0, 'completions/min_terminated_length': 180.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.9980403780937195, 'rewards/meter/std': 0.00041446235263720155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8472222685813904, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.8455538749694824, 'rewards/total_composite/std': 0.08242159336805344, 'reward': 0.8455538749694824, 'reward_std': 0.08242160081863403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021872445940971375, 'sampling/sampling_logp_difference/max': 1.738257884979248, 'sampling/importance_sampling_ratio/min': 0.17582646012306213, 'sampling/importance_sampling_ratio/mean': 1.0034029483795166, 'sampling/importance_sampling_ratio/max': 1.5945336818695068, 'entropy': 0.1502722892910242, 'clip_ratio/low_mean': 0.001377504551783204, 'clip_ratio/low_min': 0.001377504551783204, 'clip_ratio/high_mean': 0.009661172516644001, 'clip_ratio/high_max': 0.009661172516644001, 'clip_ratio/region_mean': 0.011038677068427205, 'reward_total_mean': 0.8455538749694824, 'reward_meter_mean': 0.9980403780937195, 'reward_meter_std': 0.00041446235263720155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8472222685813904, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.8455538749694824, 'reward_total_composite_std': 0.08242159336805344, 'epoch': 0.08}

 60%|█████▉    | 1964/3300 [2:34:21<2:00:28,  5.41s/it]INFO 04-12 01:11:12 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1965/3300 [2:34:26<1:57:31,  5.28s/it]2026-04-12 01:11:17,143 | INFO | train_grpo_train | metrics_logged mode=train step=1965

                                                       
{'loss': -0.0106, 'grad_norm': 7.793933391571045, 'learning_rate': 4.048484848484849e-06, 'num_tokens': 4428746.0, 'completions/mean_length': 74.5, 'completions/min_length': 71.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9973430037498474, 'rewards/meter/std': 0.001286112586967647, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973430037498474, 'rewards/total_composite/std': 0.001286112586967647, 'reward': 0.9973430037498474, 'reward_std': 0.0012860975693911314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05123698338866234, 'sampling/sampling_logp_difference/max': 2.1309890747070312, 'sampling/importance_sampling_ratio/min': 0.11871980875730515, 'sampling/importance_sampling_ratio/mean': 1.0060445070266724, 'sampling/importance_sampling_ratio/max': 1.865578293800354, 'entropy': 0.43255916610360146, 'clip_ratio/low_mean': 0.007042253389954567, 'clip_ratio/low_min': 0.007042253389954567, 'clip_ratio/high_mean': 0.024971136124804616, 'clip_ratio/high_max': 0.024971136124804616, 'clip_ratio/region_mean': 0.03201338951475918, 'reward_total_mean': 0.9973430037498474, 'reward_meter_mean': 0.9973430037498474, 'reward_meter_std': 0.001286112586967647, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973430037498474, 'reward_total_composite_std': 0.001286112586967647, 'epoch': 0.08}

 60%|█████▉    | 1965/3300 [2:34:26<1:57:31,  5.28s/it]INFO 04-12 01:11:17 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1966/3300 [2:34:31<1:57:25,  5.28s/it]2026-04-12 01:11:22,420 | INFO | train_grpo_train | metrics_logged mode=train step=1966

                                                       
{'loss': 0.0066, 'grad_norm': 2.1039681434631348, 'learning_rate': 4.045454545454546e-06, 'num_tokens': 4430956.0, 'completions/mean_length': 109.25, 'completions/min_length': 103.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.25, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.993117094039917, 'rewards/meter/std': 0.010319688357412815, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9682018160820007, 'rewards/total_composite/std': 0.06981157511472702, 'reward': 0.9682018160820007, 'reward_std': 0.06981158256530762, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04413502663373947, 'sampling/sampling_logp_difference/max': 1.36668062210083, 'sampling/importance_sampling_ratio/min': 0.2549518346786499, 'sampling/importance_sampling_ratio/mean': 1.0064747333526611, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.38045662455260754, 'clip_ratio/low_mean': 0.00909090880304575, 'clip_ratio/low_min': 0.00909090880304575, 'clip_ratio/high_mean': 0.027550982777029276, 'clip_ratio/high_max': 0.027550982777029276, 'clip_ratio/region_mean': 0.036641891580075026, 'reward_total_mean': 0.9682018160820007, 'reward_meter_mean': 0.993117094039917, 'reward_meter_std': 0.010319688357412815, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9682018160820007, 'reward_total_composite_std': 0.06981157511472702, 'epoch': 0.08}

 60%|█████▉    | 1966/3300 [2:34:31<1:57:25,  5.28s/it]INFO 04-12 01:11:22 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1967/3300 [2:34:36<1:53:51,  5.13s/it]2026-04-12 01:11:27,182 | INFO | train_grpo_train | metrics_logged mode=train step=1967

                                                       
{'loss': 0.0108, 'grad_norm': 2.3863680362701416, 'learning_rate': 4.0424242424242425e-06, 'num_tokens': 4432773.0, 'completions/mean_length': 71.125, 'completions/min_length': 68.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9883999824523926, 'rewards/meter/std': 0.005968651734292507, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9883999824523926, 'rewards/total_composite/std': 0.005968651734292507, 'reward': 0.9883999824523926, 'reward_std': 0.005968641955405474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04225626587867737, 'sampling/sampling_logp_difference/max': 1.4220304489135742, 'sampling/importance_sampling_ratio/min': 0.24122373759746552, 'sampling/importance_sampling_ratio/mean': 1.0037516355514526, 'sampling/importance_sampling_ratio/max': 1.7200582027435303, 'entropy': 0.2595768254250288, 'clip_ratio/low_mean': 0.022320898715406656, 'clip_ratio/low_min': 0.022320898715406656, 'clip_ratio/high_mean': 0.019368293695151806, 'clip_ratio/high_max': 0.019368293695151806, 'clip_ratio/region_mean': 0.04168919241055846, 'reward_total_mean': 0.9883999824523926, 'reward_meter_mean': 0.9883999824523926, 'reward_meter_std': 0.005968651734292507, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9883999824523926, 'reward_total_composite_std': 0.005968651734292507, 'epoch': 0.08}

 60%|█████▉    | 1967/3300 [2:34:36<1:53:51,  5.13s/it]INFO 04-12 01:11:27 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1968/3300 [2:34:40<1:48:42,  4.90s/it]2026-04-12 01:11:31,549 | INFO | train_grpo_train | metrics_logged mode=train step=1968

                                                       
{'loss': 0.0097, 'grad_norm': 3.711296319961548, 'learning_rate': 4.03939393939394e-06, 'num_tokens': 4434286.0, 'completions/mean_length': 33.125, 'completions/min_length': 33.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9655854105949402, 'rewards/meter/std': 0.0028627016581594944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9655854105949402, 'rewards/total_composite/std': 0.0028627016581594944, 'reward': 0.9655854105949402, 'reward_std': 0.0028626781422644854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012964806519448757, 'sampling/sampling_logp_difference/max': 0.5334515571594238, 'sampling/importance_sampling_ratio/min': 0.5865768194198608, 'sampling/importance_sampling_ratio/mean': 1.0038316249847412, 'sampling/importance_sampling_ratio/max': 1.5242719650268555, 'entropy': 0.09357678238302469, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.007575757801532745, 'clip_ratio/high_max': 0.007575757801532745, 'clip_ratio/region_mean': 0.011363636702299118, 'reward_total_mean': 0.9655854105949402, 'reward_meter_mean': 0.9655854105949402, 'reward_meter_std': 0.0028627016581594944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9655854105949402, 'reward_total_composite_std': 0.0028627016581594944, 'epoch': 0.08}

 60%|█████▉    | 1968/3300 [2:34:40<1:48:42,  4.90s/it]INFO 04-12 01:11:31 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1969/3300 [2:34:45<1:47:31,  4.85s/it]2026-04-12 01:11:36,284 | INFO | train_grpo_train | metrics_logged mode=train step=1969

                                                       
{'loss': 0.0, 'grad_norm': 0.08461344242095947, 'learning_rate': 4.036363636363637e-06, 'num_tokens': 4436078.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9970186352729797, 'rewards/meter/std': 3.89859178540064e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970186352729797, 'rewards/total_composite/std': 3.89859178540064e-06, 'reward': 0.9970186352729797, 'reward_std': 3.889565959980246e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002396326744928956, 'sampling/sampling_logp_difference/max': 0.5296306610107422, 'sampling/importance_sampling_ratio/min': 0.5888224244117737, 'sampling/importance_sampling_ratio/mean': 1.0005252361297607, 'sampling/importance_sampling_ratio/max': 1.0470123291015625, 'entropy': 0.012546915793791413, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9970186352729797, 'reward_meter_mean': 0.9970186352729797, 'reward_meter_std': 3.89859178540064e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970186352729797, 'reward_total_composite_std': 3.89859178540064e-06, 'epoch': 0.08}

 60%|█████▉    | 1969/3300 [2:34:45<1:47:31,  4.85s/it]INFO 04-12 01:11:36 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1970/3300 [2:34:50<1:48:11,  4.88s/it]2026-04-12 01:11:41,232 | INFO | train_grpo_train | metrics_logged mode=train step=1970

                                                       
{'loss': 0.0007, 'grad_norm': 4.984862327575684, 'learning_rate': 4.033333333333333e-06, 'num_tokens': 4437774.0, 'completions/mean_length': 33.0, 'completions/min_length': 33.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9650729894638062, 'rewards/meter/std': 0.0013392050750553608, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9650729894638062, 'rewards/total_composite/std': 0.0013392050750553608, 'reward': 0.9650729894638062, 'reward_std': 0.0013392041437327862, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010843876749277115, 'sampling/sampling_logp_difference/max': 0.6145038604736328, 'sampling/importance_sampling_ratio/min': 0.5409092307090759, 'sampling/importance_sampling_ratio/mean': 1.0015711784362793, 'sampling/importance_sampling_ratio/max': 1.5502804517745972, 'entropy': 0.08069896046072245, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.011363636702299118, 'reward_total_mean': 0.9650729894638062, 'reward_meter_mean': 0.9650729894638062, 'reward_meter_std': 0.0013392050750553608, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9650729894638062, 'reward_total_composite_std': 0.0013392050750553608, 'epoch': 0.08}

 60%|█████▉    | 1970/3300 [2:34:50<1:48:11,  4.88s/it]INFO 04-12 01:11:41 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1971/3300 [2:34:54<1:46:38,  4.81s/it]2026-04-12 01:11:45,898 | INFO | train_grpo_train | metrics_logged mode=train step=1971

                                                       
{'loss': 0.0063, 'grad_norm': 2.97381854057312, 'learning_rate': 4.030303030303031e-06, 'num_tokens': 4439592.0, 'completions/mean_length': 72.25, 'completions/min_length': 70.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9976744651794434, 'rewards/meter/std': 0.0007276128162629902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976744651794434, 'rewards/total_composite/std': 0.0007276128162629902, 'reward': 0.9976744651794434, 'reward_std': 0.0007276106043718755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03132731840014458, 'sampling/sampling_logp_difference/max': 1.332169532775879, 'sampling/importance_sampling_ratio/min': 0.2639040946960449, 'sampling/importance_sampling_ratio/mean': 1.0109328031539917, 'sampling/importance_sampling_ratio/max': 1.839429497718811, 'entropy': 0.24298556335270405, 'clip_ratio/low_mean': 0.008561643771827221, 'clip_ratio/low_min': 0.008561643771827221, 'clip_ratio/high_mean': 0.01056614622939378, 'clip_ratio/high_max': 0.01056614622939378, 'clip_ratio/region_mean': 0.019127790001221, 'reward_total_mean': 0.9976744651794434, 'reward_meter_mean': 0.9976744651794434, 'reward_meter_std': 0.0007276128162629902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976744651794434, 'reward_total_composite_std': 0.0007276128162629902, 'epoch': 0.08}

 60%|█████▉    | 1971/3300 [2:34:54<1:46:38,  4.81s/it]INFO 04-12 01:11:46 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1972/3300 [2:34:59<1:47:40,  4.86s/it]2026-04-12 01:11:50,879 | INFO | train_grpo_train | metrics_logged mode=train step=1972

                                                       
{'loss': -0.0272, 'grad_norm': 5.1056904792785645, 'learning_rate': 4.027272727272727e-06, 'num_tokens': 4441521.0, 'completions/mean_length': 75.125, 'completions/min_length': 70.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9973230361938477, 'rewards/meter/std': 0.0019267015159130096, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973230361938477, 'rewards/total_composite/std': 0.0019267015159130096, 'reward': 0.9973230361938477, 'reward_std': 0.0019267020979896188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04758862033486366, 'sampling/sampling_logp_difference/max': 1.3674206733703613, 'sampling/importance_sampling_ratio/min': 0.25476324558258057, 'sampling/importance_sampling_ratio/mean': 1.0059734582901, 'sampling/importance_sampling_ratio/max': 1.9101423025131226, 'entropy': 0.47906696051359177, 'clip_ratio/low_mean': 0.015499898232519627, 'clip_ratio/low_min': 0.015499898232519627, 'clip_ratio/high_mean': 0.02989264251664281, 'clip_ratio/high_max': 0.02989264251664281, 'clip_ratio/region_mean': 0.045392540749162436, 'reward_total_mean': 0.9973230361938477, 'reward_meter_mean': 0.9973230361938477, 'reward_meter_std': 0.0019267015159130096, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973230361938477, 'reward_total_composite_std': 0.0019267015159130096, 'epoch': 0.08}

 60%|█████▉    | 1972/3300 [2:34:59<1:47:40,  4.86s/it]INFO 04-12 01:11:51 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1973/3300 [2:35:05<1:51:08,  5.03s/it]2026-04-12 01:11:56,277 | INFO | train_grpo_train | metrics_logged mode=train step=1973

                                                       
{'loss': -0.0034, 'grad_norm': 1.4667316675186157, 'learning_rate': 4.024242424242424e-06, 'num_tokens': 4443935.0, 'completions/mean_length': 134.75, 'completions/min_length': 133.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.75, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.8778820037841797, 'rewards/meter/std': 0.01677864044904709, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7524703145027161, 'rewards/total_composite/std': 0.014381703920662403, 'reward': 0.7524703145027161, 'reward_std': 0.014381689950823784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009153633378446102, 'sampling/sampling_logp_difference/max': 1.524185061454773, 'sampling/importance_sampling_ratio/min': 0.21779848635196686, 'sampling/importance_sampling_ratio/mean': 1.0035121440887451, 'sampling/importance_sampling_ratio/max': 1.7678629159927368, 'entropy': 0.0543739995919168, 'clip_ratio/low_mean': 0.0028056252049282193, 'clip_ratio/low_min': 0.0028056252049282193, 'clip_ratio/high_mean': 0.003703703638166189, 'clip_ratio/high_max': 0.003703703638166189, 'clip_ratio/region_mean': 0.0065093288430944085, 'reward_total_mean': 0.7524703145027161, 'reward_meter_mean': 0.8778820037841797, 'reward_meter_std': 0.01677864044904709, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7524703145027161, 'reward_total_composite_std': 0.014381703920662403, 'epoch': 0.08}

 60%|█████▉    | 1973/3300 [2:35:05<1:51:08,  5.03s/it]INFO 04-12 01:11:56 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1974/3300 [2:35:10<1:52:32,  5.09s/it]2026-04-12 01:12:01,525 | INFO | train_grpo_train | metrics_logged mode=train step=1974

                                                       
{'loss': -0.0005, 'grad_norm': 1.7984775304794312, 'learning_rate': 4.0212121212121216e-06, 'num_tokens': 4446120.0, 'completions/mean_length': 100.125, 'completions/min_length': 99.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9985292553901672, 'rewards/meter/std': 0.000194300344446674, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9985292553901672, 'rewards/total_composite/std': 0.000194300344446674, 'reward': 0.9985292553901672, 'reward_std': 0.00019429647363722324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01695932447910309, 'sampling/sampling_logp_difference/max': 1.2210431098937988, 'sampling/importance_sampling_ratio/min': 0.2949223816394806, 'sampling/importance_sampling_ratio/mean': 1.0018852949142456, 'sampling/importance_sampling_ratio/max': 1.6485373973846436, 'entropy': 0.10710672754794359, 'clip_ratio/low_mean': 0.007525252411141992, 'clip_ratio/low_min': 0.007525252411141992, 'clip_ratio/high_mean': 0.006263126386329532, 'clip_ratio/high_max': 0.006263126386329532, 'clip_ratio/region_mean': 0.013788378797471523, 'reward_total_mean': 0.9985292553901672, 'reward_meter_mean': 0.9985292553901672, 'reward_meter_std': 0.000194300344446674, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9985292553901672, 'reward_total_composite_std': 0.000194300344446674, 'epoch': 0.08}

 60%|█████▉    | 1974/3300 [2:35:10<1:52:32,  5.09s/it]INFO 04-12 01:12:01 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1975/3300 [2:35:15<1:50:13,  4.99s/it]2026-04-12 01:12:06,278 | INFO | train_grpo_train | metrics_logged mode=train step=1975

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.018181818181818e-06, 'num_tokens': 4447928.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9970200061798096, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970200061798096, 'rewards/total_composite/std': 0.0, 'reward': 0.9970200061798096, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0013655778020620346, 'sampling/sampling_logp_difference/max': 0.04307138919830322, 'sampling/importance_sampling_ratio/min': 0.9628649353981018, 'sampling/importance_sampling_ratio/mean': 1.0012253522872925, 'sampling/importance_sampling_ratio/max': 1.044012427330017, 'entropy': 0.011537018930539489, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9970200061798096, 'reward_meter_mean': 0.9970200061798096, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970200061798096, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 60%|█████▉    | 1975/3300 [2:35:15<1:50:13,  4.99s/it]INFO 04-12 01:12:06 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1976/3300 [2:35:21<2:00:44,  5.47s/it]2026-04-12 01:12:12,873 | INFO | train_grpo_train | metrics_logged mode=train step=1976

                                                       
{'loss': 0.0074, 'grad_norm': 3.668858766555786, 'learning_rate': 4.015151515151515e-06, 'num_tokens': 4451433.0, 'completions/mean_length': 222.125, 'completions/min_length': 216.0, 'completions/max_length': 229.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 222.125, 'completions/min_terminated_length': 216.0, 'completions/max_terminated_length': 229.0, 'rewards/meter/mean': 0.9412789940834045, 'rewards/meter/std': 0.15677009522914886, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9695513248443604, 'rewards/repeat_penalty/std': 0.042069755494594574, 'rewards/total_composite/mean': 0.8963793516159058, 'rewards/total_composite/std': 0.17684437334537506, 'reward': 0.8963793516159058, 'reward_std': 0.17684435844421387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04191061109304428, 'sampling/sampling_logp_difference/max': 5.00631856918335, 'sampling/importance_sampling_ratio/min': 0.0066955070942640305, 'sampling/importance_sampling_ratio/mean': 0.9989370107650757, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18782638758420944, 'clip_ratio/low_mean': 0.0077665450517088175, 'clip_ratio/low_min': 0.0077665450517088175, 'clip_ratio/high_mean': 0.022601958131417632, 'clip_ratio/high_max': 0.022601958131417632, 'clip_ratio/region_mean': 0.03036850318312645, 'reward_total_mean': 0.8963793516159058, 'reward_meter_mean': 0.9412789940834045, 'reward_meter_std': 0.15677009522914886, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9695513248443604, 'reward_repeat_penalty_std': 0.042069755494594574, 'reward_total_composite_mean': 0.8963793516159058, 'reward_total_composite_std': 0.17684437334537506, 'epoch': 0.08}

 60%|█████▉    | 1976/3300 [2:35:21<2:00:44,  5.47s/it]INFO 04-12 01:12:13 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1977/3300 [2:35:28<2:05:35,  5.70s/it]2026-04-12 01:12:19,093 | INFO | train_grpo_train | metrics_logged mode=train step=1977

                                                       
{'loss': 0.0098, 'grad_norm': 2.525111436843872, 'learning_rate': 4.0121212121212125e-06, 'num_tokens': 4454371.0, 'completions/mean_length': 184.25, 'completions/min_length': 179.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 184.25, 'completions/min_terminated_length': 179.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.993905782699585, 'rewards/meter/std': 0.00516868568956852, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.930555522441864, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.9247466325759888, 'rewards/total_composite/std': 0.08044659346342087, 'reward': 0.9247466325759888, 'reward_std': 0.08044657856225967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049122218042612076, 'sampling/sampling_logp_difference/max': 1.7033472061157227, 'sampling/importance_sampling_ratio/min': 0.18207307159900665, 'sampling/importance_sampling_ratio/mean': 1.0052932500839233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37936991080641747, 'clip_ratio/low_mean': 0.011562932399101555, 'clip_ratio/low_min': 0.011562932399101555, 'clip_ratio/high_mean': 0.02449705172330141, 'clip_ratio/high_max': 0.02449705172330141, 'clip_ratio/region_mean': 0.036059984122402966, 'reward_total_mean': 0.9247466325759888, 'reward_meter_mean': 0.993905782699585, 'reward_meter_std': 0.00516868568956852, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.930555522441864, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.9247466325759888, 'reward_total_composite_std': 0.08044659346342087, 'epoch': 0.08}

 60%|█████▉    | 1977/3300 [2:35:28<2:05:35,  5.70s/it]INFO 04-12 01:12:19 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1978/3300 [2:35:32<2:00:06,  5.45s/it]2026-04-12 01:12:23,975 | INFO | train_grpo_train | metrics_logged mode=train step=1978

                                                       
{'loss': 0.0005, 'grad_norm': 5.933767318725586, 'learning_rate': 4.009090909090909e-06, 'num_tokens': 4456189.0, 'completions/mean_length': 76.25, 'completions/min_length': 74.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.944914698600769, 'rewards/meter/std': 0.14462126791477203, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.944914698600769, 'rewards/total_composite/std': 0.14462126791477203, 'reward': 0.944914698600769, 'reward_std': 0.14462128281593323, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0594903789460659, 'sampling/sampling_logp_difference/max': 1.3728089332580566, 'sampling/importance_sampling_ratio/min': 0.2533941864967346, 'sampling/importance_sampling_ratio/mean': 1.005175232887268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4438166059553623, 'clip_ratio/low_mean': 0.009999999776482582, 'clip_ratio/low_min': 0.009999999776482582, 'clip_ratio/high_mean': 0.03106398310046643, 'clip_ratio/high_max': 0.03106398310046643, 'clip_ratio/region_mean': 0.04106398287694901, 'reward_total_mean': 0.944914698600769, 'reward_meter_mean': 0.944914698600769, 'reward_meter_std': 0.14462126791477203, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.944914698600769, 'reward_total_composite_std': 0.14462126791477203, 'epoch': 0.08}

 60%|█████▉    | 1978/3300 [2:35:32<2:00:06,  5.45s/it]INFO 04-12 01:12:24 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1979/3300 [2:35:37<1:55:25,  5.24s/it]2026-04-12 01:12:28,741 | INFO | train_grpo_train | metrics_logged mode=train step=1979

                                                       
{'loss': 0.0023, 'grad_norm': 1.3233561515808105, 'learning_rate': 4.006060606060607e-06, 'num_tokens': 4457998.0, 'completions/mean_length': 66.125, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9989874362945557, 'rewards/meter/std': 5.0926621042890474e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989874362945557, 'rewards/total_composite/std': 5.0926621042890474e-05, 'reward': 0.9989874362945557, 'reward_std': 5.0924856623169035e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010206947103142738, 'sampling/sampling_logp_difference/max': 0.8685154914855957, 'sampling/importance_sampling_ratio/min': 0.41957396268844604, 'sampling/importance_sampling_ratio/mean': 0.9985904693603516, 'sampling/importance_sampling_ratio/max': 1.4895697832107544, 'entropy': 0.03965302975848317, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.0056535504991188645, 'reward_total_mean': 0.9989874362945557, 'reward_meter_mean': 0.9989874362945557, 'reward_meter_std': 5.0926621042890474e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989874362945557, 'reward_total_composite_std': 5.0926621042890474e-05, 'epoch': 0.08}

 60%|█████▉    | 1979/3300 [2:35:37<1:55:25,  5.24s/it]INFO 04-12 01:12:28 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1980/3300 [2:35:44<2:04:22,  5.65s/it]2026-04-12 01:12:35,341 | INFO | train_grpo_train | metrics_logged mode=train step=1980

                                                       
{'loss': 0.087, 'grad_norm': 2.8739843368530273, 'learning_rate': 4.003030303030303e-06, 'num_tokens': 4461140.0, 'completions/mean_length': 199.75, 'completions/min_length': 169.0, 'completions/max_length': 222.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 199.75, 'completions/min_terminated_length': 169.0, 'completions/max_terminated_length': 222.0, 'rewards/meter/mean': 0.9881318211555481, 'rewards/meter/std': 0.012296332977712154, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8166666626930237, 'rewards/repeat_penalty/std': 0.05781134217977524, 'rewards/total_composite/mean': 0.7281157374382019, 'rewards/total_composite/std': 0.11554723232984543, 'reward': 0.7281157374382019, 'reward_std': 0.11554723978042603, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031819190829992294, 'sampling/sampling_logp_difference/max': 4.230698585510254, 'sampling/importance_sampling_ratio/min': 0.014542227610945702, 'sampling/importance_sampling_ratio/mean': 0.9996204972267151, 'sampling/importance_sampling_ratio/max': 1.922302007675171, 'entropy': 0.16436910070478916, 'clip_ratio/low_mean': 0.010806537233293056, 'clip_ratio/low_min': 0.010806537233293056, 'clip_ratio/high_mean': 0.013719010865315795, 'clip_ratio/high_max': 0.013719010865315795, 'clip_ratio/region_mean': 0.02452554809860885, 'reward_total_mean': 0.7281157374382019, 'reward_meter_mean': 0.9881318211555481, 'reward_meter_std': 0.012296332977712154, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8166666626930237, 'reward_repeat_penalty_std': 0.05781134217977524, 'reward_total_composite_mean': 0.7281157374382019, 'reward_total_composite_std': 0.11554723232984543, 'epoch': 0.08}

 60%|██████    | 1980/3300 [2:35:44<2:04:22,  5.65s/it]INFO 04-12 01:12:35 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1981/3300 [2:35:49<1:59:16,  5.43s/it]2026-04-12 01:12:40,237 | INFO | train_grpo_train | metrics_logged mode=train step=1981

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.000000000000001e-06, 'num_tokens': 4463220.0, 'completions/mean_length': 90.0, 'completions/min_length': 90.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.0, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9955950975418091, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955950975418091, 'rewards/total_composite/std': 0.0, 'reward': 0.9955950975418091, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0013981742085888982, 'sampling/sampling_logp_difference/max': 0.12119700014591217, 'sampling/importance_sampling_ratio/min': 0.937319278717041, 'sampling/importance_sampling_ratio/mean': 1.0010671615600586, 'sampling/importance_sampling_ratio/max': 1.1288472414016724, 'entropy': 0.011815825360827148, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9955950975418091, 'reward_meter_mean': 0.9955950975418091, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955950975418091, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 60%|██████    | 1981/3300 [2:35:49<1:59:16,  5.43s/it]INFO 04-12 01:12:40 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1982/3300 [2:35:54<1:56:45,  5.31s/it]2026-04-12 01:12:45,291 | INFO | train_grpo_train | metrics_logged mode=train step=1982

                                                       
{'loss': 0.0058, 'grad_norm': 6.157790660858154, 'learning_rate': 3.996969696969698e-06, 'num_tokens': 4465131.0, 'completions/mean_length': 76.875, 'completions/min_length': 73.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.875, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9915469884872437, 'rewards/meter/std': 0.0046110316179692745, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9915469884872437, 'rewards/total_composite/std': 0.0046110316179692745, 'reward': 0.9915469884872437, 'reward_std': 0.004611044656485319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03136689215898514, 'sampling/sampling_logp_difference/max': 1.0577783584594727, 'sampling/importance_sampling_ratio/min': 0.3472263514995575, 'sampling/importance_sampling_ratio/mean': 1.0080691576004028, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.357822060585022, 'clip_ratio/low_mean': 0.008034686907194555, 'clip_ratio/low_min': 0.008034686907194555, 'clip_ratio/high_mean': 0.01946915965527296, 'clip_ratio/high_max': 0.01946915965527296, 'clip_ratio/region_mean': 0.027503846562467515, 'reward_total_mean': 0.9915469884872437, 'reward_meter_mean': 0.9915469884872437, 'reward_meter_std': 0.0046110316179692745, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9915469884872437, 'reward_total_composite_std': 0.0046110316179692745, 'epoch': 0.08}

 60%|██████    | 1982/3300 [2:35:54<1:56:45,  5.31s/it]INFO 04-12 01:12:45 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1983/3300 [2:35:58<1:49:55,  5.01s/it]2026-04-12 01:12:49,588 | INFO | train_grpo_train | metrics_logged mode=train step=1983

                                                       
{'loss': -0.0135, 'grad_norm': 4.942656993865967, 'learning_rate': 3.993939393939394e-06, 'num_tokens': 4466706.0, 'completions/mean_length': 32.875, 'completions/min_length': 32.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9034555554389954, 'rewards/meter/std': 0.1775951385498047, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9034555554389954, 'rewards/total_composite/std': 0.1775951385498047, 'reward': 0.9034555554389954, 'reward_std': 0.1775951385498047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0072293514385819435, 'sampling/sampling_logp_difference/max': 0.4028921127319336, 'sampling/importance_sampling_ratio/min': 0.6683841943740845, 'sampling/importance_sampling_ratio/mean': 1.0033267736434937, 'sampling/importance_sampling_ratio/max': 1.1638069152832031, 'entropy': 0.06188688660040498, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.9034555554389954, 'reward_meter_mean': 0.9034555554389954, 'reward_meter_std': 0.1775951385498047, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9034555554389954, 'reward_total_composite_std': 0.1775951385498047, 'epoch': 0.08}

 60%|██████    | 1983/3300 [2:35:58<1:49:55,  5.01s/it]INFO 04-12 01:12:49 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1984/3300 [2:36:03<1:48:45,  4.96s/it]2026-04-12 01:12:54,462 | INFO | train_grpo_train | metrics_logged mode=train step=1984

                                                       
{'loss': -0.0081, 'grad_norm': 0.42240381240844727, 'learning_rate': 3.990909090909092e-06, 'num_tokens': 4468579.0, 'completions/mean_length': 67.125, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9162052869796753, 'rewards/meter/std': 0.0074835000559687614, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9162052869796753, 'rewards/total_composite/std': 0.0074835000559687614, 'reward': 0.9162052869796753, 'reward_std': 0.007483497262001038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004512776155024767, 'sampling/sampling_logp_difference/max': 0.38767099380493164, 'sampling/importance_sampling_ratio/min': 0.7210829854011536, 'sampling/importance_sampling_ratio/mean': 1.0025385618209839, 'sampling/importance_sampling_ratio/max': 1.473544955253601, 'entropy': 0.03432013071142137, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.003703906899318099, 'reward_total_mean': 0.9162052869796753, 'reward_meter_mean': 0.9162052869796753, 'reward_meter_std': 0.0074835000559687614, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9162052869796753, 'reward_total_composite_std': 0.0074835000559687614, 'epoch': 0.08}

 60%|██████    | 1984/3300 [2:36:03<1:48:45,  4.96s/it]INFO 04-12 01:12:54 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1985/3300 [2:36:08<1:46:58,  4.88s/it]2026-04-12 01:12:59,133 | INFO | train_grpo_train | metrics_logged mode=train step=1985

                                                       
{'loss': 0.0002, 'grad_norm': 1.1522923707962036, 'learning_rate': 3.987878787878788e-06, 'num_tokens': 4470379.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.994735836982727, 'rewards/meter/std': 6.81514575262554e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.994735836982727, 'rewards/total_composite/std': 6.81514575262554e-05, 'reward': 0.994735836982727, 'reward_std': 6.814543303335086e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0007778764702379704, 'sampling/sampling_logp_difference/max': 0.019494354724884033, 'sampling/importance_sampling_ratio/min': 0.9891203045845032, 'sampling/importance_sampling_ratio/mean': 1.0006917715072632, 'sampling/importance_sampling_ratio/max': 1.0196856260299683, 'entropy': 0.007963738986290991, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.994735836982727, 'reward_meter_mean': 0.994735836982727, 'reward_meter_std': 6.81514575262554e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.994735836982727, 'reward_total_composite_std': 6.81514575262554e-05, 'epoch': 0.08}

 60%|██████    | 1985/3300 [2:36:08<1:46:58,  4.88s/it]INFO 04-12 01:12:59 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1986/3300 [2:36:12<1:44:19,  4.76s/it]2026-04-12 01:13:03,613 | INFO | train_grpo_train | metrics_logged mode=train step=1986

                                                       
{'loss': 0.0087, 'grad_norm': 13.343822479248047, 'learning_rate': 3.984848484848485e-06, 'num_tokens': 4471880.0, 'completions/mean_length': 40.625, 'completions/min_length': 37.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5094391703605652, 'rewards/meter/std': 0.4160431921482086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.47001466155052185, 'rewards/total_composite/std': 0.39801424741744995, 'reward': 0.47001466155052185, 'reward_std': 0.39801424741744995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10109325498342514, 'sampling/sampling_logp_difference/max': 1.5112323760986328, 'sampling/importance_sampling_ratio/min': 0.22063790261745453, 'sampling/importance_sampling_ratio/mean': 1.0084450244903564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6491754315793514, 'clip_ratio/low_mean': 0.05110658518970013, 'clip_ratio/low_min': 0.05110658518970013, 'clip_ratio/high_mean': 0.03596452111378312, 'clip_ratio/high_max': 0.03596452111378312, 'clip_ratio/region_mean': 0.08707110630348325, 'reward_total_mean': 0.47001466155052185, 'reward_meter_mean': 0.5094391703605652, 'reward_meter_std': 0.4160431921482086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.47001466155052185, 'reward_total_composite_std': 0.39801424741744995, 'epoch': 0.08}

 60%|██████    | 1986/3300 [2:36:12<1:44:19,  4.76s/it]INFO 04-12 01:13:03 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1987/3300 [2:36:18<1:50:34,  5.05s/it]2026-04-12 01:13:09,346 | INFO | train_grpo_train | metrics_logged mode=train step=1987

                                                       
{'loss': -0.0103, 'grad_norm': 4.379970073699951, 'learning_rate': 3.9818181818181825e-06, 'num_tokens': 4474379.0, 'completions/mean_length': 149.375, 'completions/min_length': 137.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 149.375, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9875733852386475, 'rewards/meter/std': 0.02084759995341301, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9405426383018494, 'rewards/total_composite/std': 0.10836569964885712, 'reward': 0.9405426383018494, 'reward_std': 0.10836568474769592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04869993403553963, 'sampling/sampling_logp_difference/max': 1.6888227462768555, 'sampling/importance_sampling_ratio/min': 0.18473687767982483, 'sampling/importance_sampling_ratio/mean': 1.0045157670974731, 'sampling/importance_sampling_ratio/max': 1.912219524383545, 'entropy': 0.4626317396759987, 'clip_ratio/low_mean': 0.010287561686709523, 'clip_ratio/low_min': 0.010287561686709523, 'clip_ratio/high_mean': 0.02709868340753019, 'clip_ratio/high_max': 0.02709868340753019, 'clip_ratio/region_mean': 0.03738624509423971, 'reward_total_mean': 0.9405426383018494, 'reward_meter_mean': 0.9875733852386475, 'reward_meter_std': 0.02084759995341301, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9405426383018494, 'reward_total_composite_std': 0.10836569964885712, 'epoch': 0.08}

 60%|██████    | 1987/3300 [2:36:18<1:50:34,  5.05s/it]INFO 04-12 01:13:09 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1988/3300 [2:36:23<1:49:15,  5.00s/it]2026-04-12 01:13:14,206 | INFO | train_grpo_train | metrics_logged mode=train step=1988

                                                       
{'loss': 0.0169, 'grad_norm': 6.39749002456665, 'learning_rate': 3.978787878787879e-06, 'num_tokens': 4476330.0, 'completions/mean_length': 76.875, 'completions/min_length': 75.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.988801121711731, 'rewards/meter/std': 0.014979206025600433, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.988801121711731, 'rewards/total_composite/std': 0.014979206025600433, 'reward': 0.988801121711731, 'reward_std': 0.014979207888245583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05862836539745331, 'sampling/sampling_logp_difference/max': 1.5416898727416992, 'sampling/importance_sampling_ratio/min': 0.21401913464069366, 'sampling/importance_sampling_ratio/mean': 1.0134505033493042, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4746681675314903, 'clip_ratio/low_mean': 0.008121267077513039, 'clip_ratio/low_min': 0.008121267077513039, 'clip_ratio/high_mean': 0.022676826687529683, 'clip_ratio/high_max': 0.022676826687529683, 'clip_ratio/region_mean': 0.030798093765042722, 'reward_total_mean': 0.988801121711731, 'reward_meter_mean': 0.988801121711731, 'reward_meter_std': 0.014979206025600433, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.988801121711731, 'reward_total_composite_std': 0.014979206025600433, 'epoch': 0.08}

 60%|██████    | 1988/3300 [2:36:23<1:49:15,  5.00s/it]INFO 04-12 01:13:14 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1989/3300 [2:36:27<1:47:50,  4.94s/it]2026-04-12 01:13:19,000 | INFO | train_grpo_train | metrics_logged mode=train step=1989

                                                       
{'loss': -0.0063, 'grad_norm': 5.730335712432861, 'learning_rate': 3.975757575757576e-06, 'num_tokens': 4478142.0, 'completions/mean_length': 67.5, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9175406694412231, 'rewards/meter/std': 0.021055592224001884, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9175406694412231, 'rewards/total_composite/std': 0.021055592224001884, 'reward': 0.9175406694412231, 'reward_std': 0.021055590361356735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014552988111972809, 'sampling/sampling_logp_difference/max': 0.96074378490448, 'sampling/importance_sampling_ratio/min': 0.38260820508003235, 'sampling/importance_sampling_ratio/mean': 1.0017634630203247, 'sampling/importance_sampling_ratio/max': 1.4384418725967407, 'entropy': 0.08089068345725536, 'clip_ratio/low_mean': 0.005597014795057476, 'clip_ratio/low_min': 0.005597014795057476, 'clip_ratio/high_mean': 0.0036764706019312143, 'clip_ratio/high_max': 0.0036764706019312143, 'clip_ratio/region_mean': 0.00927348539698869, 'reward_total_mean': 0.9175406694412231, 'reward_meter_mean': 0.9175406694412231, 'reward_meter_std': 0.021055592224001884, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9175406694412231, 'reward_total_composite_std': 0.021055592224001884, 'epoch': 0.08}

 60%|██████    | 1989/3300 [2:36:27<1:47:50,  4.94s/it]INFO 04-12 01:13:19 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1990/3300 [2:36:34<1:58:04,  5.41s/it]2026-04-12 01:13:25,508 | INFO | train_grpo_train | metrics_logged mode=train step=1990

                                                       
{'loss': -0.0004, 'grad_norm': 2.3331921100616455, 'learning_rate': 3.972727272727273e-06, 'num_tokens': 4481575.0, 'completions/mean_length': 216.125, 'completions/min_length': 206.0, 'completions/max_length': 222.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 216.125, 'completions/min_terminated_length': 206.0, 'completions/max_terminated_length': 222.0, 'rewards/meter/mean': 0.998195230960846, 'rewards/meter/std': 0.0007794296252541244, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9043561220169067, 'rewards/repeat_penalty/std': 0.053097911179065704, 'rewards/total_composite/mean': 0.9027365446090698, 'rewards/total_composite/std': 0.053245700895786285, 'reward': 0.9027365446090698, 'reward_std': 0.05324570834636688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023263096809387207, 'sampling/sampling_logp_difference/max': 1.5214588642120361, 'sampling/importance_sampling_ratio/min': 0.21839307248592377, 'sampling/importance_sampling_ratio/mean': 1.0023928880691528, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09882919304072857, 'clip_ratio/low_mean': 0.005168477655388415, 'clip_ratio/low_min': 0.005168477655388415, 'clip_ratio/high_mean': 0.016820423072203994, 'clip_ratio/high_max': 0.016820423072203994, 'clip_ratio/region_mean': 0.02198890072759241, 'reward_total_mean': 0.9027365446090698, 'reward_meter_mean': 0.998195230960846, 'reward_meter_std': 0.0007794296252541244, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9043561220169067, 'reward_repeat_penalty_std': 0.053097911179065704, 'reward_total_composite_mean': 0.9027365446090698, 'reward_total_composite_std': 0.053245700895786285, 'epoch': 0.08}

 60%|██████    | 1990/3300 [2:36:34<1:58:04,  5.41s/it]INFO 04-12 01:13:25 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1991/3300 [2:36:42<2:12:27,  6.07s/it]2026-04-12 01:13:33,138 | INFO | train_grpo_train | metrics_logged mode=train step=1991

                                                       
{'loss': 0.0014, 'grad_norm': 1.446068286895752, 'learning_rate': 3.96969696969697e-06, 'num_tokens': 4485554.0, 'completions/mean_length': 288.375, 'completions/min_length': 287.0, 'completions/max_length': 290.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 288.375, 'completions/min_terminated_length': 287.0, 'completions/max_terminated_length': 290.0, 'rewards/meter/mean': 0.9093496203422546, 'rewards/meter/std': 0.01770671270787716, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6796875, 'rewards/repeat_penalty/std': 0.022097086533904076, 'rewards/total_composite/mean': 0.411979079246521, 'rewards/total_composite/std': 0.013528194278478622, 'reward': 0.411979079246521, 'reward_std': 0.013528193347156048, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02342934161424637, 'sampling/sampling_logp_difference/max': 13.880716323852539, 'sampling/importance_sampling_ratio/min': 9.368746418658702e-07, 'sampling/importance_sampling_ratio/mean': 0.9994128942489624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.059712667018175125, 'clip_ratio/low_mean': 0.001727131224470213, 'clip_ratio/low_min': 0.001727131224470213, 'clip_ratio/high_mean': 0.007373977248789743, 'clip_ratio/high_max': 0.007373977248789743, 'clip_ratio/region_mean': 0.009101108473259956, 'reward_total_mean': 0.411979079246521, 'reward_meter_mean': 0.9093496203422546, 'reward_meter_std': 0.01770671270787716, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6796875, 'reward_repeat_penalty_std': 0.022097086533904076, 'reward_total_composite_mean': 0.411979079246521, 'reward_total_composite_std': 0.013528194278478622, 'epoch': 0.08}

 60%|██████    | 1991/3300 [2:36:42<2:12:27,  6.07s/it]INFO 04-12 01:13:33 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1992/3300 [2:36:47<2:10:12,  5.97s/it]2026-04-12 01:13:38,874 | INFO | train_grpo_train | metrics_logged mode=train step=1992

                                                       
{'loss': -0.0152, 'grad_norm': 1.818966031074524, 'learning_rate': 3.966666666666667e-06, 'num_tokens': 4488257.0, 'completions/mean_length': 157.875, 'completions/min_length': 150.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.875, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.9970660209655762, 'rewards/meter/std': 0.00133254355750978, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7754957675933838, 'rewards/total_composite/std': 0.0010364169720560312, 'reward': 0.7754957675933838, 'reward_std': 0.001036411034874618, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002477371832355857, 'sampling/sampling_logp_difference/max': 0.8110690116882324, 'sampling/importance_sampling_ratio/min': 0.44438278675079346, 'sampling/importance_sampling_ratio/mean': 1.0000977516174316, 'sampling/importance_sampling_ratio/max': 1.1387183666229248, 'entropy': 0.011172075115609914, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7754957675933838, 'reward_meter_mean': 0.9970660209655762, 'reward_meter_std': 0.00133254355750978, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7754957675933838, 'reward_total_composite_std': 0.0010364169720560312, 'epoch': 0.08}

 60%|██████    | 1992/3300 [2:36:47<2:10:12,  5.97s/it]INFO 04-12 01:13:39 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1993/3300 [2:36:52<2:02:08,  5.61s/it]2026-04-12 01:13:43,628 | INFO | train_grpo_train | metrics_logged mode=train step=1993

                                                       
{'loss': -0.0053, 'grad_norm': 3.636791467666626, 'learning_rate': 3.963636363636364e-06, 'num_tokens': 4490233.0, 'completions/mean_length': 73.0, 'completions/min_length': 72.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.0, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9986278414726257, 'rewards/meter/std': 0.0007868955726735294, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9986278414726257, 'rewards/total_composite/std': 0.0007868955726735294, 'reward': 0.9986278414726257, 'reward_std': 0.0007868955144658685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03113717958331108, 'sampling/sampling_logp_difference/max': 1.0237507820129395, 'sampling/importance_sampling_ratio/min': 0.3592449724674225, 'sampling/importance_sampling_ratio/mean': 1.0071430206298828, 'sampling/importance_sampling_ratio/max': 1.5565049648284912, 'entropy': 0.21745091304183006, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.011917525203898549, 'clip_ratio/high_max': 0.011917525203898549, 'clip_ratio/region_mean': 0.015389747451990843, 'reward_total_mean': 0.9986278414726257, 'reward_meter_mean': 0.9986278414726257, 'reward_meter_std': 0.0007868955726735294, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9986278414726257, 'reward_total_composite_std': 0.0007868955726735294, 'epoch': 0.08}

 60%|██████    | 1993/3300 [2:36:52<2:02:08,  5.61s/it]INFO 04-12 01:13:43 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1994/3300 [2:36:58<2:03:19,  5.67s/it]2026-04-12 01:13:49,430 | INFO | train_grpo_train | metrics_logged mode=train step=1994

                                                       
{'loss': -0.0042, 'grad_norm': 3.498656749725342, 'learning_rate': 3.960606060606061e-06, 'num_tokens': 4492892.0, 'completions/mean_length': 147.375, 'completions/min_length': 144.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.375, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9948342442512512, 'rewards/meter/std': 0.006030126474797726, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948342442512512, 'rewards/total_composite/std': 0.006030126474797726, 'reward': 0.9948342442512512, 'reward_std': 0.006030108779668808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06232795864343643, 'sampling/sampling_logp_difference/max': 1.819002628326416, 'sampling/importance_sampling_ratio/min': 0.16218742728233337, 'sampling/importance_sampling_ratio/mean': 1.0064527988433838, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4670833945274353, 'clip_ratio/low_mean': 0.006076388992369175, 'clip_ratio/low_min': 0.006076388992369175, 'clip_ratio/high_mean': 0.03634089510887861, 'clip_ratio/high_max': 0.03634089510887861, 'clip_ratio/region_mean': 0.04241728410124779, 'reward_total_mean': 0.9948342442512512, 'reward_meter_mean': 0.9948342442512512, 'reward_meter_std': 0.006030126474797726, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948342442512512, 'reward_total_composite_std': 0.006030126474797726, 'epoch': 0.08}

 60%|██████    | 1994/3300 [2:36:58<2:03:19,  5.67s/it]INFO 04-12 01:13:49 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1995/3300 [2:37:02<1:53:52,  5.24s/it]2026-04-12 01:13:53,663 | INFO | train_grpo_train | metrics_logged mode=train step=1995

                                                       
{'loss': 0.0012, 'grad_norm': 0.7663399577140808, 'learning_rate': 3.957575757575758e-06, 'num_tokens': 4494276.0, 'completions/mean_length': 31.0, 'completions/min_length': 31.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.9957044124603271, 'rewards/meter/std': 3.0345732739078812e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957044124603271, 'rewards/total_composite/std': 3.0345732739078812e-05, 'reward': 0.9957044124603271, 'reward_std': 3.034573092008941e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02179253287613392, 'sampling/sampling_logp_difference/max': 4.0259270668029785, 'sampling/importance_sampling_ratio/min': 0.01784687116742134, 'sampling/importance_sampling_ratio/mean': 0.9947730898857117, 'sampling/importance_sampling_ratio/max': 1.0392725467681885, 'entropy': 0.03334390290547162, 'clip_ratio/low_mean': 0.012096773833036423, 'clip_ratio/low_min': 0.012096773833036423, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.012096773833036423, 'reward_total_mean': 0.9957044124603271, 'reward_meter_mean': 0.9957044124603271, 'reward_meter_std': 3.0345732739078812e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957044124603271, 'reward_total_composite_std': 3.0345732739078812e-05, 'epoch': 0.08}

 60%|██████    | 1995/3300 [2:37:02<1:53:52,  5.24s/it]INFO 04-12 01:13:53 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1996/3300 [2:37:06<1:47:42,  4.96s/it]2026-04-12 01:13:57,966 | INFO | train_grpo_train | metrics_logged mode=train step=1996

                                                       
{'loss': -0.0084, 'grad_norm': 11.98099422454834, 'learning_rate': 3.954545454545454e-06, 'num_tokens': 4495750.0, 'completions/mean_length': 34.25, 'completions/min_length': 34.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9950810074806213, 'rewards/meter/std': 0.0017581552965566516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9950810074806213, 'rewards/total_composite/std': 0.0017581552965566516, 'reward': 0.9950810074806213, 'reward_std': 0.0017581689171493053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015304782427847385, 'sampling/sampling_logp_difference/max': 1.0786250829696655, 'sampling/importance_sampling_ratio/min': 0.549746036529541, 'sampling/importance_sampling_ratio/mean': 1.008811354637146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.046473859809339046, 'clip_ratio/low_mean': 0.011029411805793643, 'clip_ratio/low_min': 0.011029411805793643, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.011029411805793643, 'reward_total_mean': 0.9950810074806213, 'reward_meter_mean': 0.9950810074806213, 'reward_meter_std': 0.0017581552965566516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9950810074806213, 'reward_total_composite_std': 0.0017581552965566516, 'epoch': 0.08}

 60%|██████    | 1996/3300 [2:37:06<1:47:42,  4.96s/it]INFO 04-12 01:13:58 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 1997/3300 [2:37:11<1:45:34,  4.86s/it]2026-04-12 01:14:02,610 | INFO | train_grpo_train | metrics_logged mode=train step=1997

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.951515151515152e-06, 'num_tokens': 4497582.0, 'completions/mean_length': 69.0, 'completions/min_length': 69.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9355690479278564, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9355690479278564, 'rewards/total_composite/std': 0.0, 'reward': 0.9355690479278564, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0048114219680428505, 'sampling/sampling_logp_difference/max': 0.08930085599422455, 'sampling/importance_sampling_ratio/min': 0.9659483432769775, 'sampling/importance_sampling_ratio/mean': 1.004404067993164, 'sampling/importance_sampling_ratio/max': 1.093409538269043, 'entropy': 0.048666974529623985, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9355690479278564, 'reward_meter_mean': 0.9355690479278564, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9355690479278564, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 61%|██████    | 1997/3300 [2:37:11<1:45:34,  4.86s/it]INFO 04-12 01:14:02 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 1998/3300 [2:37:15<1:41:12,  4.66s/it]2026-04-12 01:14:06,810 | INFO | train_grpo_train | metrics_logged mode=train step=1998

                                                       
{'loss': 0.0006, 'grad_norm': 1.003632664680481, 'learning_rate': 3.948484848484849e-06, 'num_tokens': 4499102.0, 'completions/mean_length': 30.0, 'completions/min_length': 30.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9927566647529602, 'rewards/meter/std': 3.355137232574634e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9927566647529602, 'rewards/total_composite/std': 3.355137232574634e-05, 'reward': 0.9927566647529602, 'reward_std': 3.355137232574634e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010052908211946487, 'sampling/sampling_logp_difference/max': 0.4464629590511322, 'sampling/importance_sampling_ratio/min': 0.6398874521255493, 'sampling/importance_sampling_ratio/mean': 1.0059435367584229, 'sampling/importance_sampling_ratio/max': 1.4227073192596436, 'entropy': 0.05545296333730221, 'clip_ratio/low_mean': 0.01666666753590107, 'clip_ratio/low_min': 0.01666666753590107, 'clip_ratio/high_mean': 0.004166666883975267, 'clip_ratio/high_max': 0.004166666883975267, 'clip_ratio/region_mean': 0.020833334419876337, 'reward_total_mean': 0.9927566647529602, 'reward_meter_mean': 0.9927566647529602, 'reward_meter_std': 3.355137232574634e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9927566647529602, 'reward_total_composite_std': 3.355137232574634e-05, 'epoch': 0.08}

 61%|██████    | 1998/3300 [2:37:15<1:41:12,  4.66s/it]INFO 04-12 01:14:07 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 1999/3300 [2:37:20<1:41:53,  4.70s/it]2026-04-12 01:14:11,592 | INFO | train_grpo_train | metrics_logged mode=train step=1999

                                                       
{'loss': -0.0049, 'grad_norm': 2.3003454208374023, 'learning_rate': 3.945454545454545e-06, 'num_tokens': 4501021.0, 'completions/mean_length': 73.875, 'completions/min_length': 71.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9951367378234863, 'rewards/meter/std': 0.003929613158106804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9951367378234863, 'rewards/total_composite/std': 0.003929613158106804, 'reward': 0.9951367378234863, 'reward_std': 0.003929623868316412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04188167303800583, 'sampling/sampling_logp_difference/max': 1.4766731262207031, 'sampling/importance_sampling_ratio/min': 0.22839628159999847, 'sampling/importance_sampling_ratio/mean': 1.0134656429290771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40391664765775204, 'clip_ratio/low_mean': 0.008561643771827221, 'clip_ratio/low_min': 0.008561643771827221, 'clip_ratio/high_mean': 0.02339910331647843, 'clip_ratio/high_max': 0.02339910331647843, 'clip_ratio/region_mean': 0.03196074708830565, 'reward_total_mean': 0.9951367378234863, 'reward_meter_mean': 0.9951367378234863, 'reward_meter_std': 0.003929613158106804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9951367378234863, 'reward_total_composite_std': 0.003929613158106804, 'epoch': 0.08}

 61%|██████    | 1999/3300 [2:37:20<1:41:53,  4.70s/it]INFO 04-12 01:14:11 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2000/3300 [2:37:26<1:49:23,  5.05s/it]2026-04-12 01:14:17,496 | INFO | train_grpo_train | metrics_logged mode=train step=2000

                                                       
{'loss': 0.0069, 'grad_norm': 3.8840928077697754, 'learning_rate': 3.942424242424243e-06, 'num_tokens': 4503974.0, 'completions/mean_length': 147.125, 'completions/min_length': 141.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.125, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.9887276887893677, 'rewards/meter/std': 0.01330604124814272, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9359963536262512, 'rewards/total_composite/std': 0.07730759680271149, 'reward': 0.9359963536262512, 'reward_std': 0.07730759680271149, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05714353546500206, 'sampling/sampling_logp_difference/max': 2.683351516723633, 'sampling/importance_sampling_ratio/min': 0.06833374500274658, 'sampling/importance_sampling_ratio/mean': 1.0066624879837036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3696784619241953, 'clip_ratio/low_mean': 0.005059222981799394, 'clip_ratio/low_min': 0.005059222981799394, 'clip_ratio/high_mean': 0.026401946786791086, 'clip_ratio/high_max': 0.026401946786791086, 'clip_ratio/region_mean': 0.03146116976859048, 'reward_total_mean': 0.9359963536262512, 'reward_meter_mean': 0.9887276887893677, 'reward_meter_std': 0.01330604124814272, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9359963536262512, 'reward_total_composite_std': 0.07730759680271149, 'epoch': 0.08}

 61%|██████    | 2000/3300 [2:37:26<1:49:23,  5.05s/it]INFO 04-12 01:14:17 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:31,  2.87s/it]

 23%|██▎       | 3/13 [00:09<00:32,  3.25s/it]

 31%|███       | 4/13 [00:14<00:35,  3.90s/it]

 38%|███▊      | 5/13 [00:19<00:34,  4.29s/it]

 46%|████▌     | 6/13 [00:25<00:32,  4.71s/it]

 54%|█████▍    | 7/13 [00:30<00:28,  4.80s/it]

 62%|██████▏   | 8/13 [00:35<00:25,  5.05s/it]

 69%|██████▉   | 9/13 [00:41<00:21,  5.39s/it]

 77%|███████▋  | 10/13 [00:46<00:15,  5.17s/it]

 85%|████████▍ | 11/13 [00:51<00:10,  5.14s/it]

 92%|█████████▏| 12/13 [00:56<00:05,  5.16s/it]

100%|██████████| 13/13 [01:01<00:00,  5.13s/it]2026-04-12 01:15:25,463 | INFO | train_grpo_train | metrics_logged mode=eval step=2000

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 67.8915, 'eval_samples_per_second': 1.532, 'eval_steps_per_second': 0.191, 'eval_num_tokens': 4503974.0, 'eval_completions/mean_length': 205.2596153846154, 'eval_completions/min_length': 62.92307692307692, 'eval_completions/max_length': 358.15384615384613, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 205.2596153846154, 'eval_completions/min_terminated_length': 62.92307692307692, 'eval_completions/max_terminated_length': 358.15384615384613, 'eval_rewards/meter/mean': 0.7367691305967478, 'eval_rewards/meter/std': 0.3722034446322001, 'eval_rewards/count_adherence/mean': 0.9229050003565274, 'eval_rewards/count_adherence/std': 0.11920174153951499, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8324548143606919, 'eval_rewards/repeat_penalty/std': 0.1328287388269718, 'eval_rewards/total_composite/mean': 0.568704937513058, 'eval_rewards/total_composite/std': 0.34218758459274584, 'eval_reward': 0.568704937513058, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.017686504655732557, 'eval_sampling/sampling_logp_difference/max': 1.1488754015702467, 'eval_sampling/importance_sampling_ratio/min': 0.33040788540473354, 'eval_sampling/importance_sampling_ratio/mean': 1.0038928618797889, 'eval_sampling/importance_sampling_ratio/max': 1.4441173993624175, 'eval_entropy': 0.16280974046542093, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.568704937513058, 'eval_reward_meter_mean': 0.7367691305967478, 'eval_reward_meter_std': 0.3722034446322001, 'eval_reward_count_adherence_mean': 0.9229050003565274, 'eval_reward_count_adherence_std': 0.11920174153951499, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8324548143606919, 'eval_reward_repeat_penalty_std': 0.1328287388269718, 'eval_reward_total_composite_mean': 0.568704937513058, 'eval_reward_total_composite_std': 0.34218758459274584, 'epoch': 0.08}

 61%|██████    | 2000/3300 [2:38:34<1:49:23,  5.05s/it]

100%|██████████| 13/13 [01:01<00:00,  5.13s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 61%|██████    | 2001/3300 [2:38:42<9:28:29, 26.26s/it]2026-04-12 01:15:33,199 | INFO | train_grpo_train | metrics_logged mode=train step=2001

                                                       
{'loss': 0.0013, 'grad_norm': 1.4718478918075562, 'learning_rate': 3.93939393939394e-06, 'num_tokens': 4506302.0, 'completions/mean_length': 122.0, 'completions/min_length': 122.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.0, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9962645173072815, 'rewards/meter/std': 4.681191057898104e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8717328310012817, 'rewards/total_composite/std': 0.050349388271570206, 'reward': 0.8717328310012817, 'reward_std': 0.05034938082098961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007403408642858267, 'sampling/sampling_logp_difference/max': 1.1984543800354004, 'sampling/importance_sampling_ratio/min': 0.30166009068489075, 'sampling/importance_sampling_ratio/mean': 1.0027551651000977, 'sampling/importance_sampling_ratio/max': 1.2980602979660034, 'entropy': 0.043996233493089676, 'clip_ratio/low_mean': 0.005122950533404946, 'clip_ratio/low_min': 0.005122950533404946, 'clip_ratio/high_mean': 0.0010245901066809893, 'clip_ratio/high_max': 0.0010245901066809893, 'clip_ratio/region_mean': 0.006147540640085936, 'reward_total_mean': 0.8717328310012817, 'reward_meter_mean': 0.9962645173072815, 'reward_meter_std': 4.681191057898104e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8717328310012817, 'reward_total_composite_std': 0.050349388271570206, 'epoch': 0.08}

 61%|██████    | 2001/3300 [2:38:42<9:28:29, 26.26s/it]INFO 04-12 01:15:33 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2002/3300 [2:38:46<7:06:45, 19.73s/it]2026-04-12 01:15:37,687 | INFO | train_grpo_train | metrics_logged mode=train step=2002

                                                       
{'loss': -0.0194, 'grad_norm': 5.45490837097168, 'learning_rate': 3.936363636363636e-06, 'num_tokens': 4507873.0, 'completions/mean_length': 37.375, 'completions/min_length': 36.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9985835552215576, 'rewards/meter/std': 0.00047840786282904446, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9985835552215576, 'rewards/total_composite/std': 0.00047840786282904446, 'reward': 0.9985835552215576, 'reward_std': 0.0004784138291142881, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017835740000009537, 'sampling/sampling_logp_difference/max': 0.7125928401947021, 'sampling/importance_sampling_ratio/min': 0.49037110805511475, 'sampling/importance_sampling_ratio/mean': 1.0080413818359375, 'sampling/importance_sampling_ratio/max': 1.2333381175994873, 'entropy': 0.1442318456247449, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0032051282469183207, 'clip_ratio/high_max': 0.0032051282469183207, 'clip_ratio/region_mean': 0.0032051282469183207, 'reward_total_mean': 0.9985835552215576, 'reward_meter_mean': 0.9985835552215576, 'reward_meter_std': 0.00047840786282904446, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9985835552215576, 'reward_total_composite_std': 0.00047840786282904446, 'epoch': 0.08}

 61%|██████    | 2002/3300 [2:38:46<7:06:45, 19.73s/it]INFO 04-12 01:15:37 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2003/3300 [2:38:51<5:32:32, 15.38s/it]2026-04-12 01:15:42,936 | INFO | train_grpo_train | metrics_logged mode=train step=2003

                                                       
{'loss': -0.0002, 'grad_norm': 0.0041707647033035755, 'learning_rate': 3.9333333333333335e-06, 'num_tokens': 4510281.0, 'completions/mean_length': 127.0, 'completions/min_length': 127.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.0, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.9974512457847595, 'rewards/meter/std': 1.9482802599668503e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.854958176612854, 'rewards/total_composite/std': 1.66965983225964e-05, 'reward': 0.854958176612854, 'reward_std': 1.669656739977654e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0020607816986739635, 'sampling/sampling_logp_difference/max': 1.2046804428100586, 'sampling/importance_sampling_ratio/min': 0.2997877895832062, 'sampling/importance_sampling_ratio/mean': 1.000118374824524, 'sampling/importance_sampling_ratio/max': 1.0359206199645996, 'entropy': 0.007686095021199435, 'clip_ratio/low_mean': 0.0009842519648373127, 'clip_ratio/low_min': 0.0009842519648373127, 'clip_ratio/high_mean': 0.0009842519648373127, 'clip_ratio/high_max': 0.0009842519648373127, 'clip_ratio/region_mean': 0.0019685039296746254, 'reward_total_mean': 0.854958176612854, 'reward_meter_mean': 0.9974512457847595, 'reward_meter_std': 1.9482802599668503e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.854958176612854, 'reward_total_composite_std': 1.66965983225964e-05, 'epoch': 0.08}

 61%|██████    | 2003/3300 [2:38:51<5:32:32, 15.38s/it]INFO 04-12 01:15:43 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2004/3300 [2:38:56<4:20:49, 12.07s/it]2026-04-12 01:15:47,292 | INFO | train_grpo_train | metrics_logged mode=train step=2004

                                                       
{'loss': 0.0022, 'grad_norm': 4.7584662437438965, 'learning_rate': 3.930303030303031e-06, 'num_tokens': 4511890.0, 'completions/mean_length': 38.125, 'completions/min_length': 35.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.999361515045166, 'rewards/meter/std': 0.00022483796055894345, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8744252920150757, 'rewards/total_composite/std': 0.35332125425338745, 'reward': 0.8744252920150757, 'reward_std': 0.35332125425338745, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044290993362665176, 'sampling/sampling_logp_difference/max': 1.2952303886413574, 'sampling/importance_sampling_ratio/min': 0.273834764957428, 'sampling/importance_sampling_ratio/mean': 1.0098013877868652, 'sampling/importance_sampling_ratio/max': 1.5466128587722778, 'entropy': 0.2946350034326315, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03295032726600766, 'clip_ratio/high_max': 0.03295032726600766, 'clip_ratio/region_mean': 0.03295032726600766, 'reward_total_mean': 0.8744252920150757, 'reward_meter_mean': 0.999361515045166, 'reward_meter_std': 0.00022483796055894345, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8744252920150757, 'reward_total_composite_std': 0.35332125425338745, 'epoch': 0.08}

 61%|██████    | 2004/3300 [2:38:56<4:20:49, 12.07s/it]INFO 04-12 01:15:47 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2005/3300 [2:39:01<3:36:56, 10.05s/it]2026-04-12 01:15:52,627 | INFO | train_grpo_train | metrics_logged mode=train step=2005

                                                       
{'loss': 0.0085, 'grad_norm': 2.326425552368164, 'learning_rate': 3.927272727272727e-06, 'num_tokens': 4514367.0, 'completions/mean_length': 121.625, 'completions/min_length': 119.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.625, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9959496259689331, 'rewards/meter/std': 0.0007494749734178185, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8892085552215576, 'rewards/total_composite/std': 0.06532733142375946, 'reward': 0.8892085552215576, 'reward_std': 0.06532733887434006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011545258574187756, 'sampling/sampling_logp_difference/max': 1.276261329650879, 'sampling/importance_sampling_ratio/min': 0.27907872200012207, 'sampling/importance_sampling_ratio/mean': 1.0025670528411865, 'sampling/importance_sampling_ratio/max': 1.928596019744873, 'entropy': 0.07582355523481965, 'clip_ratio/low_mean': 0.00512295076623559, 'clip_ratio/low_min': 0.00512295076623559, 'clip_ratio/high_mean': 0.0072237912099808455, 'clip_ratio/high_max': 0.0072237912099808455, 'clip_ratio/region_mean': 0.012346741976216435, 'reward_total_mean': 0.8892085552215576, 'reward_meter_mean': 0.9959496259689331, 'reward_meter_std': 0.0007494749734178185, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8892085552215576, 'reward_total_composite_std': 0.06532733142375946, 'epoch': 0.08}

 61%|██████    | 2005/3300 [2:39:01<3:36:56, 10.05s/it]INFO 04-12 01:15:52 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2006/3300 [2:39:05<2:59:01,  8.30s/it]2026-04-12 01:15:56,845 | INFO | train_grpo_train | metrics_logged mode=train step=2006

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.9242424242424244e-06, 'num_tokens': 4515983.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9979252815246582, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979252815246582, 'rewards/total_composite/std': 0.0, 'reward': 0.9979252815246582, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.004444928839802742, 'sampling/sampling_logp_difference/max': 0.16487574577331543, 'sampling/importance_sampling_ratio/min': 0.9554311633110046, 'sampling/importance_sampling_ratio/mean': 1.0037531852722168, 'sampling/importance_sampling_ratio/max': 1.1792465448379517, 'entropy': 0.04299367079511285, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9979252815246582, 'reward_meter_mean': 0.9979252815246582, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979252815246582, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 61%|██████    | 2006/3300 [2:39:05<2:59:01,  8.30s/it]INFO 04-12 01:15:57 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2007/3300 [2:39:10<2:36:31,  7.26s/it]2026-04-12 01:16:01,682 | INFO | train_grpo_train | metrics_logged mode=train step=2007

                                                       
{'loss': 0.0055, 'grad_norm': 3.622284412384033, 'learning_rate': 3.921212121212122e-06, 'num_tokens': 4518013.0, 'completions/mean_length': 74.75, 'completions/min_length': 71.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9929636716842651, 'rewards/meter/std': 0.008374504745006561, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9929636716842651, 'rewards/total_composite/std': 0.008374504745006561, 'reward': 0.9929636716842651, 'reward_std': 0.008374514989554882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04201868921518326, 'sampling/sampling_logp_difference/max': 1.1939210891723633, 'sampling/importance_sampling_ratio/min': 0.30303072929382324, 'sampling/importance_sampling_ratio/mean': 1.006500244140625, 'sampling/importance_sampling_ratio/max': 1.6419928073883057, 'entropy': 0.378201762214303, 'clip_ratio/low_mean': 0.00826923071872443, 'clip_ratio/low_min': 0.00826923071872443, 'clip_ratio/high_mean': 0.026706379372626543, 'clip_ratio/high_max': 0.026706379372626543, 'clip_ratio/region_mean': 0.03497561009135097, 'reward_total_mean': 0.9929636716842651, 'reward_meter_mean': 0.9929636716842651, 'reward_meter_std': 0.008374504745006561, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9929636716842651, 'reward_total_composite_std': 0.008374504745006561, 'epoch': 0.08}

 61%|██████    | 2007/3300 [2:39:10<2:36:31,  7.26s/it]INFO 04-12 01:16:01 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2008/3300 [2:39:18<2:41:00,  7.48s/it]2026-04-12 01:16:09,656 | INFO | train_grpo_train | metrics_logged mode=train step=2008

                                                       
{'loss': 0.0072, 'grad_norm': 1.4125186204910278, 'learning_rate': 3.918181818181819e-06, 'num_tokens': 4522347.0, 'completions/mean_length': 324.75, 'completions/min_length': 309.0, 'completions/max_length': 337.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 324.75, 'completions/min_terminated_length': 309.0, 'completions/max_terminated_length': 337.0, 'rewards/meter/mean': 0.7302596569061279, 'rewards/meter/std': 0.25050464272499084, 'rewards/count_adherence/mean': 0.7272727489471436, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7244791984558105, 'rewards/repeat_penalty/std': 0.03113570623099804, 'rewards/total_composite/mean': 0.38363176584243774, 'rewards/total_composite/std': 0.13244113326072693, 'reward': 0.38363176584243774, 'reward_std': 0.13244113326072693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03116171434521675, 'sampling/sampling_logp_difference/max': 13.405669212341309, 'sampling/importance_sampling_ratio/min': 1.5065788829815574e-06, 'sampling/importance_sampling_ratio/mean': 1.0025417804718018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1241519721224904, 'clip_ratio/low_mean': 0.009641719050705433, 'clip_ratio/low_min': 0.009641719050705433, 'clip_ratio/high_mean': 0.007035914051812142, 'clip_ratio/high_max': 0.007035914051812142, 'clip_ratio/region_mean': 0.016677633102517575, 'reward_total_mean': 0.38363176584243774, 'reward_meter_mean': 0.7302596569061279, 'reward_meter_std': 0.25050464272499084, 'reward_count_adherence_mean': 0.7272727489471436, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7244791984558105, 'reward_repeat_penalty_std': 0.03113570623099804, 'reward_total_composite_mean': 0.38363176584243774, 'reward_total_composite_std': 0.13244113326072693, 'epoch': 0.08}

 61%|██████    | 2008/3300 [2:39:18<2:41:00,  7.48s/it]INFO 04-12 01:16:09 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2009/3300 [2:39:23<2:21:54,  6.60s/it]2026-04-12 01:16:14,199 | INFO | train_grpo_train | metrics_logged mode=train step=2009

                                                       
{'loss': -0.0129, 'grad_norm': 2.9392364025115967, 'learning_rate': 3.915151515151515e-06, 'num_tokens': 4524064.0, 'completions/mean_length': 57.625, 'completions/min_length': 56.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.625, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9944944977760315, 'rewards/meter/std': 0.0007702266448177397, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9944944977760315, 'rewards/total_composite/std': 0.0007702266448177397, 'reward': 0.9944944977760315, 'reward_std': 0.0007702277507632971, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011414974927902222, 'sampling/sampling_logp_difference/max': 0.5186891555786133, 'sampling/importance_sampling_ratio/min': 0.5953003764152527, 'sampling/importance_sampling_ratio/mean': 1.0033068656921387, 'sampling/importance_sampling_ratio/max': 1.263137936592102, 'entropy': 0.08374510239809752, 'clip_ratio/low_mean': 0.0022321429569274187, 'clip_ratio/low_min': 0.0022321429569274187, 'clip_ratio/high_mean': 0.006465517217293382, 'clip_ratio/high_max': 0.006465517217293382, 'clip_ratio/region_mean': 0.0086976601742208, 'reward_total_mean': 0.9944944977760315, 'reward_meter_mean': 0.9944944977760315, 'reward_meter_std': 0.0007702266448177397, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9944944977760315, 'reward_total_composite_std': 0.0007702266448177397, 'epoch': 0.08}

 61%|██████    | 2009/3300 [2:39:23<2:21:54,  6.60s/it]INFO 04-12 01:16:14 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2010/3300 [2:39:28<2:13:05,  6.19s/it]2026-04-12 01:16:19,442 | INFO | train_grpo_train | metrics_logged mode=train step=2010

                                                       
{'loss': 0.008, 'grad_norm': 2.544851303100586, 'learning_rate': 3.912121212121213e-06, 'num_tokens': 4526204.0, 'completions/mean_length': 103.5, 'completions/min_length': 103.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.5, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.8996114730834961, 'rewards/meter/std': 0.042233821004629135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8996114730834961, 'rewards/total_composite/std': 0.042233821004629135, 'reward': 0.8996114730834961, 'reward_std': 0.04223383218050003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016247009858489037, 'sampling/sampling_logp_difference/max': 1.265535831451416, 'sampling/importance_sampling_ratio/min': 0.28208810091018677, 'sampling/importance_sampling_ratio/mean': 0.9998170137405396, 'sampling/importance_sampling_ratio/max': 1.8066370487213135, 'entropy': 0.08177287224680185, 'clip_ratio/low_mean': 0.005918904324062169, 'clip_ratio/low_min': 0.005918904324062169, 'clip_ratio/high_mean': 0.00849514571018517, 'clip_ratio/high_max': 0.00849514571018517, 'clip_ratio/region_mean': 0.014414050034247339, 'reward_total_mean': 0.8996114730834961, 'reward_meter_mean': 0.8996114730834961, 'reward_meter_std': 0.042233821004629135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8996114730834961, 'reward_total_composite_std': 0.042233821004629135, 'epoch': 0.08}

 61%|██████    | 2010/3300 [2:39:28<2:13:05,  6.19s/it]INFO 04-12 01:16:19 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2011/3300 [2:39:33<2:04:09,  5.78s/it]2026-04-12 01:16:24,262 | INFO | train_grpo_train | metrics_logged mode=train step=2011

                                                       
{'loss': 0.0002, 'grad_norm': 6.366806983947754, 'learning_rate': 3.90909090909091e-06, 'num_tokens': 4528044.0, 'completions/mean_length': 74.0, 'completions/min_length': 71.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9870051145553589, 'rewards/meter/std': 0.02932342328131199, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9870051145553589, 'rewards/total_composite/std': 0.02932342328131199, 'reward': 0.9870051145553589, 'reward_std': 0.029323413968086243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04717674106359482, 'sampling/sampling_logp_difference/max': 1.317805290222168, 'sampling/importance_sampling_ratio/min': 0.2677222192287445, 'sampling/importance_sampling_ratio/mean': 1.0154587030410767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35354609973728657, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.028711115941405296, 'clip_ratio/high_max': 0.028711115941405296, 'clip_ratio/region_mean': 0.032089494401589036, 'reward_total_mean': 0.9870051145553589, 'reward_meter_mean': 0.9870051145553589, 'reward_meter_std': 0.02932342328131199, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9870051145553589, 'reward_total_composite_std': 0.02932342328131199, 'epoch': 0.08}

 61%|██████    | 2011/3300 [2:39:33<2:04:09,  5.78s/it]INFO 04-12 01:16:24 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2012/3300 [2:39:37<1:54:20,  5.33s/it]2026-04-12 01:16:28,533 | INFO | train_grpo_train | metrics_logged mode=train step=2012

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.906060606060606e-06, 'num_tokens': 4529676.0, 'completions/mean_length': 33.0, 'completions/min_length': 33.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9990787506103516, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990787506103516, 'rewards/total_composite/std': 0.0, 'reward': 0.9990787506103516, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00452659884467721, 'sampling/sampling_logp_difference/max': 0.21105718612670898, 'sampling/importance_sampling_ratio/min': 0.8097277879714966, 'sampling/importance_sampling_ratio/mean': 1.0008872747421265, 'sampling/importance_sampling_ratio/max': 1.086820363998413, 'entropy': 0.025131846428848803, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990787506103516, 'reward_meter_mean': 0.9990787506103516, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990787506103516, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 61%|██████    | 2012/3300 [2:39:37<1:54:20,  5.33s/it]INFO 04-12 01:16:28 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2013/3300 [2:39:42<1:52:01,  5.22s/it]2026-04-12 01:16:33,510 | INFO | train_grpo_train | metrics_logged mode=train step=2013

                                                       
{'loss': 0.0012, 'grad_norm': 1.6898747682571411, 'learning_rate': 3.9030303030303035e-06, 'num_tokens': 4531806.0, 'completions/mean_length': 89.25, 'completions/min_length': 88.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.25, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.995819091796875, 'rewards/meter/std': 0.0001791357935871929, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.995819091796875, 'rewards/total_composite/std': 0.0001791357935871929, 'reward': 0.995819091796875, 'reward_std': 0.00017913819465320557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015769556164741516, 'sampling/sampling_logp_difference/max': 1.6130731105804443, 'sampling/importance_sampling_ratio/min': 0.19927427172660828, 'sampling/importance_sampling_ratio/mean': 1.0000059604644775, 'sampling/importance_sampling_ratio/max': 1.4373438358306885, 'entropy': 0.09196147602051497, 'clip_ratio/low_mean': 0.009800249827094376, 'clip_ratio/low_min': 0.009800249827094376, 'clip_ratio/high_mean': 0.005494505632668734, 'clip_ratio/high_max': 0.005494505632668734, 'clip_ratio/region_mean': 0.01529475545976311, 'reward_total_mean': 0.995819091796875, 'reward_meter_mean': 0.995819091796875, 'reward_meter_std': 0.0001791357935871929, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.995819091796875, 'reward_total_composite_std': 0.0001791357935871929, 'epoch': 0.08}

 61%|██████    | 2013/3300 [2:39:42<1:52:01,  5.22s/it]INFO 04-12 01:16:33 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2014/3300 [2:39:46<1:47:18,  5.01s/it]2026-04-12 01:16:38,018 | INFO | train_grpo_train | metrics_logged mode=train step=2014

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.900000000000001e-06, 'num_tokens': 4533534.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9970200061798096, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970200061798096, 'rewards/total_composite/std': 0.0, 'reward': 0.9970200061798096, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.002094635274261236, 'sampling/sampling_logp_difference/max': 0.07730516046285629, 'sampling/importance_sampling_ratio/min': 0.9277377724647522, 'sampling/importance_sampling_ratio/mean': 1.0017708539962769, 'sampling/importance_sampling_ratio/max': 1.0803717374801636, 'entropy': 0.015853506047278643, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9970200061798096, 'reward_meter_mean': 0.9970200061798096, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970200061798096, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 61%|██████    | 2014/3300 [2:39:47<1:47:18,  5.01s/it]INFO 04-12 01:16:38 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2015/3300 [2:39:51<1:47:01,  5.00s/it]2026-04-12 01:16:42,990 | INFO | train_grpo_train | metrics_logged mode=train step=2015

                                                       
{'loss': 0.0009, 'grad_norm': 1.428508996963501, 'learning_rate': 3.896969696969697e-06, 'num_tokens': 4535555.0, 'completions/mean_length': 96.625, 'completions/min_length': 96.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.625, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.9975524544715881, 'rewards/meter/std': 0.0001665082381805405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975524544715881, 'rewards/total_composite/std': 0.0001665082381805405, 'reward': 0.9975524544715881, 'reward_std': 0.00016649799363221973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018747342750430107, 'sampling/sampling_logp_difference/max': 0.7523272037506104, 'sampling/importance_sampling_ratio/min': 0.47126853466033936, 'sampling/importance_sampling_ratio/mean': 1.0018181800842285, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11758578661829233, 'clip_ratio/low_mean': 0.0026041667442768812, 'clip_ratio/low_min': 0.0026041667442768812, 'clip_ratio/high_mean': 0.015477340901270509, 'clip_ratio/high_max': 0.015477340901270509, 'clip_ratio/region_mean': 0.01808150764554739, 'reward_total_mean': 0.9975524544715881, 'reward_meter_mean': 0.9975524544715881, 'reward_meter_std': 0.0001665082381805405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975524544715881, 'reward_total_composite_std': 0.0001665082381805405, 'epoch': 0.08}

 61%|██████    | 2015/3300 [2:39:51<1:47:01,  5.00s/it]INFO 04-12 01:16:43 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2016/3300 [2:39:56<1:42:40,  4.80s/it]2026-04-12 01:16:47,334 | INFO | train_grpo_train | metrics_logged mode=train step=2016

                                                       
{'loss': -0.035, 'grad_norm': 5.452454566955566, 'learning_rate': 3.8939393939393944e-06, 'num_tokens': 4536992.0, 'completions/mean_length': 32.625, 'completions/min_length': 30.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9971754550933838, 'rewards/meter/std': 0.004386701621115208, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971754550933838, 'rewards/total_composite/std': 0.004386701621115208, 'reward': 0.9971754550933838, 'reward_std': 0.0043867104686796665, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013498242013156414, 'sampling/sampling_logp_difference/max': 1.0278760194778442, 'sampling/importance_sampling_ratio/min': 0.3577660620212555, 'sampling/importance_sampling_ratio/mean': 0.9991469979286194, 'sampling/importance_sampling_ratio/max': 1.2878613471984863, 'entropy': 0.028148290934041142, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.007575757801532745, 'clip_ratio/high_max': 0.007575757801532745, 'clip_ratio/region_mean': 0.011363636702299118, 'reward_total_mean': 0.9971754550933838, 'reward_meter_mean': 0.9971754550933838, 'reward_meter_std': 0.004386701621115208, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971754550933838, 'reward_total_composite_std': 0.004386701621115208, 'epoch': 0.08}

 61%|██████    | 2016/3300 [2:39:56<1:42:40,  4.80s/it]INFO 04-12 01:16:47 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2017/3300 [2:40:01<1:45:22,  4.93s/it]2026-04-12 01:16:52,557 | INFO | train_grpo_train | metrics_logged mode=train step=2017

                                                       
{'loss': 0.0048, 'grad_norm': 4.091221332550049, 'learning_rate': 3.890909090909092e-06, 'num_tokens': 4538962.0, 'completions/mean_length': 66.25, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9984210729598999, 'rewards/meter/std': 0.0010966768022626638, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984210729598999, 'rewards/total_composite/std': 0.0010966768022626638, 'reward': 0.9984210729598999, 'reward_std': 0.0010966742411255836, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02255828306078911, 'sampling/sampling_logp_difference/max': 1.67164945602417, 'sampling/importance_sampling_ratio/min': 0.18793681263923645, 'sampling/importance_sampling_ratio/mean': 0.9979748129844666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06328104855492711, 'clip_ratio/low_mean': 0.005681818351149559, 'clip_ratio/low_min': 0.005681818351149559, 'clip_ratio/high_mean': 0.007604895276017487, 'clip_ratio/high_max': 0.007604895276017487, 'clip_ratio/region_mean': 0.013286713627167046, 'reward_total_mean': 0.9984210729598999, 'reward_meter_mean': 0.9984210729598999, 'reward_meter_std': 0.0010966768022626638, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984210729598999, 'reward_total_composite_std': 0.0010966768022626638, 'epoch': 0.08}

 61%|██████    | 2017/3300 [2:40:01<1:45:22,  4.93s/it]INFO 04-12 01:16:52 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2018/3300 [2:40:06<1:45:07,  4.92s/it]2026-04-12 01:16:57,461 | INFO | train_grpo_train | metrics_logged mode=train step=2018

                                                       
{'loss': 0.0544, 'grad_norm': 9.685022354125977, 'learning_rate': 3.887878787878788e-06, 'num_tokens': 4540736.0, 'completions/mean_length': 62.75, 'completions/min_length': 57.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7036042213439941, 'rewards/meter/std': 0.2752836048603058, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.800000011920929, 'rewards/repeat_penalty/std': 0.1511857807636261, 'rewards/total_composite/mean': 0.5434250831604004, 'rewards/total_composite/std': 0.20307792723178864, 'reward': 0.5434250831604004, 'reward_std': 0.20307794213294983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0515943206846714, 'sampling/sampling_logp_difference/max': 1.9400761127471924, 'sampling/importance_sampling_ratio/min': 0.1436930149793625, 'sampling/importance_sampling_ratio/mean': 1.0063507556915283, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28926723077893257, 'clip_ratio/low_mean': 0.011172067141160369, 'clip_ratio/low_min': 0.011172067141160369, 'clip_ratio/high_mean': 0.0331978602334857, 'clip_ratio/high_max': 0.0331978602334857, 'clip_ratio/region_mean': 0.04436992737464607, 'reward_total_mean': 0.5434250831604004, 'reward_meter_mean': 0.7036042213439941, 'reward_meter_std': 0.2752836048603058, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.800000011920929, 'reward_repeat_penalty_std': 0.1511857807636261, 'reward_total_composite_mean': 0.5434250831604004, 'reward_total_composite_std': 0.20307792723178864, 'epoch': 0.08}

 61%|██████    | 2018/3300 [2:40:06<1:45:07,  4.92s/it]INFO 04-12 01:16:57 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2019/3300 [2:40:11<1:48:28,  5.08s/it]2026-04-12 01:17:02,909 | INFO | train_grpo_train | metrics_logged mode=train step=2019

                                                       
{'loss': -0.0182, 'grad_norm': 3.233832359313965, 'learning_rate': 3.884848484848485e-06, 'num_tokens': 4542988.0, 'completions/mean_length': 106.5, 'completions/min_length': 100.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.5, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9943101406097412, 'rewards/meter/std': 0.0025435402058064938, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943101406097412, 'rewards/total_composite/std': 0.0025435402058064938, 'reward': 0.9943101406097412, 'reward_std': 0.0025435429997742176, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03694016858935356, 'sampling/sampling_logp_difference/max': 1.4739065170288086, 'sampling/importance_sampling_ratio/min': 0.22902902960777283, 'sampling/importance_sampling_ratio/mean': 1.010913610458374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3407053891569376, 'clip_ratio/low_mean': 0.014337376691401005, 'clip_ratio/low_min': 0.014337376691401005, 'clip_ratio/high_mean': 0.013733877101913095, 'clip_ratio/high_max': 0.013733877101913095, 'clip_ratio/region_mean': 0.0280712537933141, 'reward_total_mean': 0.9943101406097412, 'reward_meter_mean': 0.9943101406097412, 'reward_meter_std': 0.0025435402058064938, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943101406097412, 'reward_total_composite_std': 0.0025435402058064938, 'epoch': 0.08}

 61%|██████    | 2019/3300 [2:40:11<1:48:28,  5.08s/it]INFO 04-12 01:17:03 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2020/3300 [2:40:17<1:53:49,  5.34s/it]2026-04-12 01:17:08,875 | INFO | train_grpo_train | metrics_logged mode=train step=2020

                                                       
{'loss': 0.0002, 'grad_norm': 2.484889507293701, 'learning_rate': 3.881818181818182e-06, 'num_tokens': 4545802.0, 'completions/mean_length': 173.75, 'completions/min_length': 173.0, 'completions/max_length': 174.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.75, 'completions/min_terminated_length': 173.0, 'completions/max_terminated_length': 174.0, 'rewards/meter/mean': 0.9751757383346558, 'rewards/meter/std': 0.06318987905979156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7124999761581421, 'rewards/repeat_penalty/std': 0.0353553481400013, 'rewards/total_composite/mean': 0.6928579211235046, 'rewards/total_composite/std': 0.015284620225429535, 'reward': 0.6928579211235046, 'reward_std': 0.015284637920558453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013370201922953129, 'sampling/sampling_logp_difference/max': 1.3443384170532227, 'sampling/importance_sampling_ratio/min': 0.2607121467590332, 'sampling/importance_sampling_ratio/mean': 1.0012462139129639, 'sampling/importance_sampling_ratio/max': 1.7495958805084229, 'entropy': 0.07485386310145259, 'clip_ratio/low_mean': 0.0007225433364510536, 'clip_ratio/low_min': 0.0007225433364510536, 'clip_ratio/high_mean': 0.01079247216694057, 'clip_ratio/high_max': 0.01079247216694057, 'clip_ratio/region_mean': 0.011515015503391623, 'reward_total_mean': 0.6928579211235046, 'reward_meter_mean': 0.9751757383346558, 'reward_meter_std': 0.06318987905979156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7124999761581421, 'reward_repeat_penalty_std': 0.0353553481400013, 'reward_total_composite_mean': 0.6928579211235046, 'reward_total_composite_std': 0.015284620225429535, 'epoch': 0.08}

 61%|██████    | 2020/3300 [2:40:17<1:53:49,  5.34s/it]INFO 04-12 01:17:09 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2021/3300 [2:40:23<1:54:52,  5.39s/it]2026-04-12 01:17:14,357 | INFO | train_grpo_train | metrics_logged mode=train step=2021

                                                       
{'loss': 0.0001, 'grad_norm': 1.3713654279708862, 'learning_rate': 3.878787878787879e-06, 'num_tokens': 4548138.0, 'completions/mean_length': 137.0, 'completions/min_length': 137.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.0, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.904876172542572, 'rewards/meter/std': 0.0019419160671532154, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7756081223487854, 'rewards/total_composite/std': 0.0016644843854010105, 'reward': 0.7756081223487854, 'reward_std': 0.0016644754214212298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.001481961109675467, 'sampling/sampling_logp_difference/max': 0.2336844801902771, 'sampling/importance_sampling_ratio/min': 0.9551554918289185, 'sampling/importance_sampling_ratio/mean': 1.0013093948364258, 'sampling/importance_sampling_ratio/max': 1.263245940208435, 'entropy': 0.011890202295035124, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7756081223487854, 'reward_meter_mean': 0.904876172542572, 'reward_meter_std': 0.0019419160671532154, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7756081223487854, 'reward_total_composite_std': 0.0016644843854010105, 'epoch': 0.08}

 61%|██████    | 2021/3300 [2:40:23<1:54:52,  5.39s/it]INFO 04-12 01:17:14 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2022/3300 [2:40:27<1:48:15,  5.08s/it]2026-04-12 01:17:18,724 | INFO | train_grpo_train | metrics_logged mode=train step=2022

                                                       
{'loss': -0.0341, 'grad_norm': 6.894834518432617, 'learning_rate': 3.875757575757576e-06, 'num_tokens': 4549759.0, 'completions/mean_length': 34.625, 'completions/min_length': 32.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9927211403846741, 'rewards/meter/std': 0.009707700461149216, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9927211403846741, 'rewards/total_composite/std': 0.009707700461149216, 'reward': 0.9927211403846741, 'reward_std': 0.009707717224955559, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02147146500647068, 'sampling/sampling_logp_difference/max': 1.2381019592285156, 'sampling/importance_sampling_ratio/min': 0.2899340093135834, 'sampling/importance_sampling_ratio/mean': 0.9955629110336304, 'sampling/importance_sampling_ratio/max': 1.2991583347320557, 'entropy': 0.07363486429676414, 'clip_ratio/low_mean': 0.011029412038624287, 'clip_ratio/low_min': 0.011029412038624287, 'clip_ratio/high_mean': 0.014087301678955555, 'clip_ratio/high_max': 0.014087301678955555, 'clip_ratio/region_mean': 0.02511671371757984, 'reward_total_mean': 0.9927211403846741, 'reward_meter_mean': 0.9927211403846741, 'reward_meter_std': 0.009707700461149216, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9927211403846741, 'reward_total_composite_std': 0.009707700461149216, 'epoch': 0.08}

 61%|██████▏   | 2022/3300 [2:40:27<1:48:15,  5.08s/it]INFO 04-12 01:17:18 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2023/3300 [2:40:32<1:46:12,  4.99s/it]2026-04-12 01:17:23,510 | INFO | train_grpo_train | metrics_logged mode=train step=2023

                                                       
{'loss': 0.0061, 'grad_norm': 2.7678604125976562, 'learning_rate': 3.872727272727273e-06, 'num_tokens': 4551755.0, 'completions/mean_length': 72.5, 'completions/min_length': 72.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.5, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9986792802810669, 'rewards/meter/std': 0.000637872377410531, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9986792802810669, 'rewards/total_composite/std': 0.000637872377410531, 'reward': 0.9986792802810669, 'reward_std': 0.0006378723192028701, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03179734945297241, 'sampling/sampling_logp_difference/max': 0.8867998123168945, 'sampling/importance_sampling_ratio/min': 0.4119720160961151, 'sampling/importance_sampling_ratio/mean': 1.0089229345321655, 'sampling/importance_sampling_ratio/max': 1.7916842699050903, 'entropy': 0.24161814339458942, 'clip_ratio/low_mean': 0.005184551002457738, 'clip_ratio/low_min': 0.005184551002457738, 'clip_ratio/high_mean': 0.006896879756823182, 'clip_ratio/high_max': 0.006896879756823182, 'clip_ratio/region_mean': 0.01208143075928092, 'reward_total_mean': 0.9986792802810669, 'reward_meter_mean': 0.9986792802810669, 'reward_meter_std': 0.000637872377410531, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9986792802810669, 'reward_total_composite_std': 0.000637872377410531, 'epoch': 0.08}

 61%|██████▏   | 2023/3300 [2:40:32<1:46:12,  4.99s/it]INFO 04-12 01:17:23 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2024/3300 [2:40:37<1:47:07,  5.04s/it]2026-04-12 01:17:28,646 | INFO | train_grpo_train | metrics_logged mode=train step=2024

                                                       
{'loss': 0.0087, 'grad_norm': 1.6175122261047363, 'learning_rate': 3.86969696969697e-06, 'num_tokens': 4553953.0, 'completions/mean_length': 102.75, 'completions/min_length': 100.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.75, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9968816041946411, 'rewards/meter/std': 0.0006491380045190454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9968816041946411, 'rewards/total_composite/std': 0.0006491380045190454, 'reward': 0.9968816041946411, 'reward_std': 0.0006491419626399875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010959350503981113, 'sampling/sampling_logp_difference/max': 0.9986443519592285, 'sampling/importance_sampling_ratio/min': 0.36837852001190186, 'sampling/importance_sampling_ratio/mean': 0.9999393820762634, 'sampling/importance_sampling_ratio/max': 1.40573251247406, 'entropy': 0.05600249394774437, 'clip_ratio/low_mean': 0.0024038462433964014, 'clip_ratio/low_min': 0.0024038462433964014, 'clip_ratio/high_mean': 0.006140776677057147, 'clip_ratio/high_max': 0.006140776677057147, 'clip_ratio/region_mean': 0.008544622920453548, 'reward_total_mean': 0.9968816041946411, 'reward_meter_mean': 0.9968816041946411, 'reward_meter_std': 0.0006491380045190454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9968816041946411, 'reward_total_composite_std': 0.0006491380045190454, 'epoch': 0.08}

 61%|██████▏   | 2024/3300 [2:40:37<1:47:07,  5.04s/it]INFO 04-12 01:17:28 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2025/3300 [2:40:43<1:51:52,  5.26s/it]2026-04-12 01:17:34,446 | INFO | train_grpo_train | metrics_logged mode=train step=2025

                                                       
{'loss': 0.0171, 'grad_norm': 3.409259557723999, 'learning_rate': 3.866666666666667e-06, 'num_tokens': 4556516.0, 'completions/mean_length': 150.375, 'completions/min_length': 139.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.375, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.890097975730896, 'rewards/meter/std': 0.1463252604007721, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.8427009582519531, 'rewards/total_composite/std': 0.15601050853729248, 'reward': 0.8427009582519531, 'reward_std': 0.1560104936361313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05061601847410202, 'sampling/sampling_logp_difference/max': 3.8258116245269775, 'sampling/importance_sampling_ratio/min': 0.021800734102725983, 'sampling/importance_sampling_ratio/mean': 0.9980630278587341, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2885005250573158, 'clip_ratio/low_mean': 0.017962860874831676, 'clip_ratio/low_min': 0.017962860874831676, 'clip_ratio/high_mean': 0.025219141854904592, 'clip_ratio/high_max': 0.025219141854904592, 'clip_ratio/region_mean': 0.04318200272973627, 'reward_total_mean': 0.8427009582519531, 'reward_meter_mean': 0.890097975730896, 'reward_meter_std': 0.1463252604007721, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.8427009582519531, 'reward_total_composite_std': 0.15601050853729248, 'epoch': 0.08}

 61%|██████▏   | 2025/3300 [2:40:43<1:51:52,  5.26s/it]INFO 04-12 01:17:34 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2026/3300 [2:40:48<1:47:55,  5.08s/it]2026-04-12 01:17:39,101 | INFO | train_grpo_train | metrics_logged mode=train step=2026

                                                       
{'loss': 0.0053, 'grad_norm': 2.6563684940338135, 'learning_rate': 3.863636363636364e-06, 'num_tokens': 4558243.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9982448220252991, 'rewards/meter/std': 0.0003502867475617677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982448220252991, 'rewards/total_composite/std': 0.0003502867475617677, 'reward': 0.9982448220252991, 'reward_std': 0.00035030534490942955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01347498781979084, 'sampling/sampling_logp_difference/max': 0.7937655448913574, 'sampling/importance_sampling_ratio/min': 0.4521390497684479, 'sampling/importance_sampling_ratio/mean': 0.9976450800895691, 'sampling/importance_sampling_ratio/max': 1.3493366241455078, 'entropy': 0.038185699842870235, 'clip_ratio/low_mean': 0.0037313431967049837, 'clip_ratio/low_min': 0.0037313431967049837, 'clip_ratio/high_mean': 0.003759611048735678, 'clip_ratio/high_max': 0.003759611048735678, 'clip_ratio/region_mean': 0.007490954245440662, 'reward_total_mean': 0.9982448220252991, 'reward_meter_mean': 0.9982448220252991, 'reward_meter_std': 0.0003502867475617677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982448220252991, 'reward_total_composite_std': 0.0003502867475617677, 'epoch': 0.08}

 61%|██████▏   | 2026/3300 [2:40:48<1:47:55,  5.08s/it]INFO 04-12 01:17:39 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2027/3300 [2:40:56<2:09:55,  6.12s/it]2026-04-12 01:17:47,654 | INFO | train_grpo_train | metrics_logged mode=train step=2027

                                                       
{'loss': 0.0071, 'grad_norm': 2.305088758468628, 'learning_rate': 3.860606060606061e-06, 'num_tokens': 4562915.0, 'completions/mean_length': 340.0, 'completions/min_length': 333.0, 'completions/max_length': 342.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 340.0, 'completions/min_terminated_length': 333.0, 'completions/max_terminated_length': 342.0, 'rewards/meter/mean': 0.9870275259017944, 'rewards/meter/std': 0.0033544274047017097, 'rewards/count_adherence/mean': 0.701923131942749, 'rewards/count_adherence/std': 0.027196412906050682, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7379385828971863, 'rewards/repeat_penalty/std': 0.02510136552155018, 'rewards/total_composite/mean': 0.5113018751144409, 'rewards/total_composite/std': 0.02731543779373169, 'reward': 0.5113018751144409, 'reward_std': 0.027315424755215645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013775442726910114, 'sampling/sampling_logp_difference/max': 1.7148866653442383, 'sampling/importance_sampling_ratio/min': 0.17998412251472473, 'sampling/importance_sampling_ratio/mean': 1.0029667615890503, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06602455815300345, 'clip_ratio/low_mean': 0.005135256564244628, 'clip_ratio/low_min': 0.005135256564244628, 'clip_ratio/high_mean': 0.0037086496595293283, 'clip_ratio/high_max': 0.0037086496595293283, 'clip_ratio/region_mean': 0.008843906223773956, 'reward_total_mean': 0.5113018751144409, 'reward_meter_mean': 0.9870275259017944, 'reward_meter_std': 0.0033544274047017097, 'reward_count_adherence_mean': 0.701923131942749, 'reward_count_adherence_std': 0.027196412906050682, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7379385828971863, 'reward_repeat_penalty_std': 0.02510136552155018, 'reward_total_composite_mean': 0.5113018751144409, 'reward_total_composite_std': 0.02731543779373169, 'epoch': 0.08}

 61%|██████▏   | 2027/3300 [2:40:56<2:09:55,  6.12s/it]INFO 04-12 01:17:47 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2028/3300 [2:41:01<2:01:10,  5.72s/it]2026-04-12 01:17:52,419 | INFO | train_grpo_train | metrics_logged mode=train step=2028

                                                       
{'loss': 0.0109, 'grad_norm': 4.434993743896484, 'learning_rate': 3.857575757575758e-06, 'num_tokens': 4564619.0, 'completions/mean_length': 69.0, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9325088262557983, 'rewards/meter/std': 0.02886304259300232, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9325088262557983, 'rewards/total_composite/std': 0.02886304259300232, 'reward': 0.9325088262557983, 'reward_std': 0.02886303886771202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018320947885513306, 'sampling/sampling_logp_difference/max': 0.9960315227508545, 'sampling/importance_sampling_ratio/min': 0.36934226751327515, 'sampling/importance_sampling_ratio/mean': 0.9968265295028687, 'sampling/importance_sampling_ratio/max': 1.2660948038101196, 'entropy': 0.06549654342234135, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.007299659075215459, 'clip_ratio/high_max': 0.007299659075215459, 'clip_ratio/region_mean': 0.0090853733709082, 'reward_total_mean': 0.9325088262557983, 'reward_meter_mean': 0.9325088262557983, 'reward_meter_std': 0.02886304259300232, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9325088262557983, 'reward_total_composite_std': 0.02886304259300232, 'epoch': 0.08}

 61%|██████▏   | 2028/3300 [2:41:01<2:01:10,  5.72s/it]INFO 04-12 01:17:52 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2029/3300 [2:41:06<1:56:58,  5.52s/it]2026-04-12 01:17:57,492 | INFO | train_grpo_train | metrics_logged mode=train step=2029

                                                       
{'loss': -0.0021, 'grad_norm': 0.44382914900779724, 'learning_rate': 3.8545454545454545e-06, 'num_tokens': 4566681.0, 'completions/mean_length': 102.75, 'completions/min_length': 102.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.75, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9243794083595276, 'rewards/meter/std': 0.014914168044924736, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9243794083595276, 'rewards/total_composite/std': 0.014914168044924736, 'reward': 0.9243794083595276, 'reward_std': 0.01491414662450552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004690106958150864, 'sampling/sampling_logp_difference/max': 0.31058406829833984, 'sampling/importance_sampling_ratio/min': 0.733018696308136, 'sampling/importance_sampling_ratio/mean': 1.002354621887207, 'sampling/importance_sampling_ratio/max': 1.119174838066101, 'entropy': 0.04349301569163799, 'clip_ratio/low_mean': 0.0036526747280731797, 'clip_ratio/low_min': 0.0036526747280731797, 'clip_ratio/high_mean': 0.0012135922443121672, 'clip_ratio/high_max': 0.0012135922443121672, 'clip_ratio/region_mean': 0.004866266972385347, 'reward_total_mean': 0.9243794083595276, 'reward_meter_mean': 0.9243794083595276, 'reward_meter_std': 0.014914168044924736, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9243794083595276, 'reward_total_composite_std': 0.014914168044924736, 'epoch': 0.08}

 61%|██████▏   | 2029/3300 [2:41:06<1:56:58,  5.52s/it]INFO 04-12 01:17:57 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2030/3300 [2:41:11<1:52:34,  5.32s/it]2026-04-12 01:18:02,331 | INFO | train_grpo_train | metrics_logged mode=train step=2030

                                                       
{'loss': 0.3073, 'grad_norm': 10.0611572265625, 'learning_rate': 3.851515151515152e-06, 'num_tokens': 4568259.0, 'completions/mean_length': 47.25, 'completions/min_length': 37.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9984946846961975, 'rewards/meter/std': 0.0014516202500090003, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7492817640304565, 'rewards/total_composite/std': 0.46246692538261414, 'reward': 0.7492817640304565, 'reward_std': 0.46246692538261414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.057400476187467575, 'sampling/sampling_logp_difference/max': 2.0996389389038086, 'sampling/importance_sampling_ratio/min': 0.12250065058469772, 'sampling/importance_sampling_ratio/mean': 1.003790259361267, 'sampling/importance_sampling_ratio/max': 1.675675392150879, 'entropy': 0.3015878004953265, 'clip_ratio/low_mean': 0.013769977260380983, 'clip_ratio/low_min': 0.013769977260380983, 'clip_ratio/high_mean': 0.019332326017320156, 'clip_ratio/high_max': 0.019332326017320156, 'clip_ratio/region_mean': 0.03310230327770114, 'reward_total_mean': 0.7492817640304565, 'reward_meter_mean': 0.9984946846961975, 'reward_meter_std': 0.0014516202500090003, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7492817640304565, 'reward_total_composite_std': 0.46246692538261414, 'epoch': 0.08}

 62%|██████▏   | 2030/3300 [2:41:11<1:52:34,  5.32s/it]INFO 04-12 01:18:02 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2031/3300 [2:41:16<1:53:39,  5.37s/it]2026-04-12 01:18:07,833 | INFO | train_grpo_train | metrics_logged mode=train step=2031

                                                       
{'loss': -0.004, 'grad_norm': 3.1920166015625, 'learning_rate': 3.848484848484848e-06, 'num_tokens': 4570662.0, 'completions/mean_length': 129.375, 'completions/min_length': 128.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.375, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9977312684059143, 'rewards/meter/std': 0.00020651114755310118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8908217549324036, 'rewards/total_composite/std': 0.06581787765026093, 'reward': 0.8908217549324036, 'reward_std': 0.06581789255142212, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01242055743932724, 'sampling/sampling_logp_difference/max': 1.437847375869751, 'sampling/importance_sampling_ratio/min': 0.23743832111358643, 'sampling/importance_sampling_ratio/mean': 1.0028420686721802, 'sampling/importance_sampling_ratio/max': 1.6425533294677734, 'entropy': 0.06218431191518903, 'clip_ratio/low_mean': 0.011571898590773344, 'clip_ratio/low_min': 0.011571898590773344, 'clip_ratio/high_mean': 0.004751632455736399, 'clip_ratio/high_max': 0.004751632455736399, 'clip_ratio/region_mean': 0.016323531046509743, 'reward_total_mean': 0.8908217549324036, 'reward_meter_mean': 0.9977312684059143, 'reward_meter_std': 0.00020651114755310118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8908217549324036, 'reward_total_composite_std': 0.06581787765026093, 'epoch': 0.08}

 62%|██████▏   | 2031/3300 [2:41:16<1:53:39,  5.37s/it]INFO 04-12 01:18:08 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2032/3300 [2:41:21<1:50:35,  5.23s/it]2026-04-12 01:18:12,741 | INFO | train_grpo_train | metrics_logged mode=train step=2032

                                                       
{'loss': 0.0009, 'grad_norm': 1.8542088270187378, 'learning_rate': 3.8454545454545454e-06, 'num_tokens': 4572734.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9978283643722534, 'rewards/meter/std': 0.00016764015890657902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978283643722534, 'rewards/total_composite/std': 0.00016764015890657902, 'reward': 0.9978283643722534, 'reward_std': 0.00016763212624937296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011215658858418465, 'sampling/sampling_logp_difference/max': 0.35844850540161133, 'sampling/importance_sampling_ratio/min': 0.6987596154212952, 'sampling/importance_sampling_ratio/mean': 1.0011532306671143, 'sampling/importance_sampling_ratio/max': 1.2213056087493896, 'entropy': 0.06596994632855058, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.009469697251915932, 'clip_ratio/high_max': 0.009469697251915932, 'clip_ratio/region_mean': 0.013257576152682304, 'reward_total_mean': 0.9978283643722534, 'reward_meter_mean': 0.9978283643722534, 'reward_meter_std': 0.00016764015890657902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978283643722534, 'reward_total_composite_std': 0.00016764015890657902, 'epoch': 0.08}

 62%|██████▏   | 2032/3300 [2:41:21<1:50:35,  5.23s/it]INFO 04-12 01:18:12 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2033/3300 [2:41:27<1:54:58,  5.44s/it]2026-04-12 01:18:18,705 | INFO | train_grpo_train | metrics_logged mode=train step=2033

                                                       
{'loss': 0.0323, 'grad_norm': 3.462294578552246, 'learning_rate': 3.842424242424243e-06, 'num_tokens': 4575273.0, 'completions/mean_length': 148.375, 'completions/min_length': 136.0, 'completions/max_length': 163.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.375, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 163.0, 'rewards/meter/mean': 0.91922527551651, 'rewards/meter/std': 0.09434106945991516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.8339765071868896, 'rewards/total_composite/std': 0.08177480101585388, 'reward': 0.8339765071868896, 'reward_std': 0.08177479356527328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04417654499411583, 'sampling/sampling_logp_difference/max': 1.6957603693008423, 'sampling/importance_sampling_ratio/min': 0.18345968425273895, 'sampling/importance_sampling_ratio/mean': 1.0021495819091797, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24012545123696327, 'clip_ratio/low_mean': 0.010510046035051346, 'clip_ratio/low_min': 0.010510046035051346, 'clip_ratio/high_mean': 0.023466870421543717, 'clip_ratio/high_max': 0.023466870421543717, 'clip_ratio/region_mean': 0.03397691645659506, 'reward_total_mean': 0.8339765071868896, 'reward_meter_mean': 0.91922527551651, 'reward_meter_std': 0.09434106945991516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.8339765071868896, 'reward_total_composite_std': 0.08177480101585388, 'epoch': 0.08}

 62%|██████▏   | 2033/3300 [2:41:27<1:54:58,  5.44s/it]INFO 04-12 01:18:18 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2034/3300 [2:41:32<1:49:48,  5.20s/it]2026-04-12 01:18:23,328 | INFO | train_grpo_train | metrics_logged mode=train step=2034

                                                       
{'loss': 0.0002, 'grad_norm': 0.040686190128326416, 'learning_rate': 3.839393939393939e-06, 'num_tokens': 4576945.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.997020959854126, 'rewards/meter/std': 2.6552993404038716e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997020959854126, 'rewards/total_composite/std': 2.6552993404038716e-06, 'reward': 0.997020959854126, 'reward_std': 2.661313374119345e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0038556025829166174, 'sampling/sampling_logp_difference/max': 0.41426944732666016, 'sampling/importance_sampling_ratio/min': 0.6608229279518127, 'sampling/importance_sampling_ratio/mean': 1.0010058879852295, 'sampling/importance_sampling_ratio/max': 1.0738173723220825, 'entropy': 0.024745611241087317, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.997020959854126, 'reward_meter_mean': 0.997020959854126, 'reward_meter_std': 2.6552993404038716e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997020959854126, 'reward_total_composite_std': 2.6552993404038716e-06, 'epoch': 0.08}

 62%|██████▏   | 2034/3300 [2:41:32<1:49:48,  5.20s/it]INFO 04-12 01:18:23 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2035/3300 [2:41:37<1:48:09,  5.13s/it]2026-04-12 01:18:28,295 | INFO | train_grpo_train | metrics_logged mode=train step=2035

                                                       
{'loss': 0.0137, 'grad_norm': 5.467068195343018, 'learning_rate': 3.836363636363636e-06, 'num_tokens': 4578843.0, 'completions/mean_length': 74.25, 'completions/min_length': 73.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.25, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8532536625862122, 'rewards/meter/std': 0.21336820721626282, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7550032138824463, 'rewards/total_composite/std': 0.3712865710258484, 'reward': 0.7550032138824463, 'reward_std': 0.3712865710258484, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04074222594499588, 'sampling/sampling_logp_difference/max': 1.6433086395263672, 'sampling/importance_sampling_ratio/min': 0.1933393031358719, 'sampling/importance_sampling_ratio/mean': 1.005318522453308, 'sampling/importance_sampling_ratio/max': 1.8733973503112793, 'entropy': 0.2636965289711952, 'clip_ratio/low_mean': 0.008355855825357139, 'clip_ratio/low_min': 0.008355855825357139, 'clip_ratio/high_mean': 0.0169863011687994, 'clip_ratio/high_max': 0.0169863011687994, 'clip_ratio/region_mean': 0.02534215699415654, 'reward_total_mean': 0.7550032138824463, 'reward_meter_mean': 0.8532536625862122, 'reward_meter_std': 0.21336820721626282, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7550032138824463, 'reward_total_composite_std': 0.3712865710258484, 'epoch': 0.08}

 62%|██████▏   | 2035/3300 [2:41:37<1:48:09,  5.13s/it]INFO 04-12 01:18:28 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2036/3300 [2:41:42<1:49:14,  5.19s/it]2026-04-12 01:18:33,609 | INFO | train_grpo_train | metrics_logged mode=train step=2036

                                                       
{'loss': 0.003, 'grad_norm': 1.754205584526062, 'learning_rate': 3.833333333333334e-06, 'num_tokens': 4581108.0, 'completions/mean_length': 108.125, 'completions/min_length': 105.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.125, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9986991882324219, 'rewards/meter/std': 0.0005537345423363149, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9737250804901123, 'rewards/total_composite/std': 0.07053209096193314, 'reward': 0.9737250804901123, 'reward_std': 0.07053209841251373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02059764787554741, 'sampling/sampling_logp_difference/max': 0.9686670303344727, 'sampling/importance_sampling_ratio/min': 0.37958869338035583, 'sampling/importance_sampling_ratio/mean': 1.00703763961792, 'sampling/importance_sampling_ratio/max': 1.7996537685394287, 'entropy': 0.21733380481600761, 'clip_ratio/low_mean': 0.0022935778833925724, 'clip_ratio/low_min': 0.0022935778833925724, 'clip_ratio/high_mean': 0.016197497956454754, 'clip_ratio/high_max': 0.016197497956454754, 'clip_ratio/region_mean': 0.018491075839847326, 'reward_total_mean': 0.9737250804901123, 'reward_meter_mean': 0.9986991882324219, 'reward_meter_std': 0.0005537345423363149, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9737250804901123, 'reward_total_composite_std': 0.07053209096193314, 'epoch': 0.08}

 62%|██████▏   | 2036/3300 [2:41:42<1:49:14,  5.19s/it]INFO 04-12 01:18:33 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2037/3300 [2:41:47<1:45:48,  5.03s/it]2026-04-12 01:18:38,265 | INFO | train_grpo_train | metrics_logged mode=train step=2037

                                                       
{'loss': 0.0, 'grad_norm': 0.042626433074474335, 'learning_rate': 3.830303030303031e-06, 'num_tokens': 4582996.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9979811906814575, 'rewards/meter/std': 3.9898877730593085e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979811906814575, 'rewards/total_composite/std': 3.9898877730593085e-06, 'reward': 0.9979811906814575, 'reward_std': 4.0090494621836115e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00694712670519948, 'sampling/sampling_logp_difference/max': 0.40966981649398804, 'sampling/importance_sampling_ratio/min': 0.6638693809509277, 'sampling/importance_sampling_ratio/mean': 1.0025206804275513, 'sampling/importance_sampling_ratio/max': 1.3170571327209473, 'entropy': 0.040048055816441774, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.005681818351149559, 'reward_total_mean': 0.9979811906814575, 'reward_meter_mean': 0.9979811906814575, 'reward_meter_std': 3.9898877730593085e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979811906814575, 'reward_total_composite_std': 3.9898877730593085e-06, 'epoch': 0.08}

 62%|██████▏   | 2037/3300 [2:41:47<1:45:48,  5.03s/it]INFO 04-12 01:18:38 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2038/3300 [2:41:52<1:46:19,  5.06s/it]2026-04-12 01:18:43,367 | INFO | train_grpo_train | metrics_logged mode=train step=2038

                                                       
{'loss': 0.0002, 'grad_norm': 0.05830654874444008, 'learning_rate': 3.827272727272728e-06, 'num_tokens': 4585172.0, 'completions/mean_length': 97.0, 'completions/min_length': 97.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.9979192614555359, 'rewards/meter/std': 5.3105509323359e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979192614555359, 'rewards/total_composite/std': 5.3105509323359e-06, 'reward': 0.9979192614555359, 'reward_std': 5.296794824971585e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0037312970962375402, 'sampling/sampling_logp_difference/max': 0.49465620517730713, 'sampling/importance_sampling_ratio/min': 0.6097805500030518, 'sampling/importance_sampling_ratio/mean': 1.0008537769317627, 'sampling/importance_sampling_ratio/max': 1.242266297340393, 'entropy': 0.025324456859380007, 'clip_ratio/low_mean': 0.0012886597542092204, 'clip_ratio/low_min': 0.0012886597542092204, 'clip_ratio/high_mean': 0.0038659792626276612, 'clip_ratio/high_max': 0.0038659792626276612, 'clip_ratio/region_mean': 0.005154639016836882, 'reward_total_mean': 0.9979192614555359, 'reward_meter_mean': 0.9979192614555359, 'reward_meter_std': 5.3105509323359e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979192614555359, 'reward_total_composite_std': 5.3105509323359e-06, 'epoch': 0.08}

 62%|██████▏   | 2038/3300 [2:41:52<1:46:19,  5.06s/it]INFO 04-12 01:18:43 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2039/3300 [2:41:57<1:47:54,  5.13s/it]2026-04-12 01:18:48,701 | INFO | train_grpo_train | metrics_logged mode=train step=2039

                                                       
{'loss': 0.0239, 'grad_norm': 10.090517044067383, 'learning_rate': 3.8242424242424245e-06, 'num_tokens': 4587310.0, 'completions/mean_length': 110.25, 'completions/min_length': 104.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.25, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9942349195480347, 'rewards/meter/std': 0.0053013949654996395, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9696316123008728, 'rewards/total_composite/std': 0.07374817132949829, 'reward': 0.9696316123008728, 'reward_std': 0.07374817132949829, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038861531764268875, 'sampling/sampling_logp_difference/max': 1.0281339883804321, 'sampling/importance_sampling_ratio/min': 0.3576737940311432, 'sampling/importance_sampling_ratio/mean': 1.0104196071624756, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32149537093937397, 'clip_ratio/low_mean': 0.004385964944958687, 'clip_ratio/low_min': 0.004385964944958687, 'clip_ratio/high_mean': 0.02182656608056277, 'clip_ratio/high_max': 0.02182656608056277, 'clip_ratio/region_mean': 0.026212531025521457, 'reward_total_mean': 0.9696316123008728, 'reward_meter_mean': 0.9942349195480347, 'reward_meter_std': 0.0053013949654996395, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9696316123008728, 'reward_total_composite_std': 0.07374817132949829, 'epoch': 0.08}

 62%|██████▏   | 2039/3300 [2:41:57<1:47:54,  5.13s/it]INFO 04-12 01:18:48 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2040/3300 [2:42:02<1:45:12,  5.01s/it]2026-04-12 01:18:53,417 | INFO | train_grpo_train | metrics_logged mode=train step=2040

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.821212121212122e-06, 'num_tokens': 4589110.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990598559379578, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990598559379578, 'rewards/total_composite/std': 0.0, 'reward': 0.9990598559379578, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0009253994794562459, 'sampling/sampling_logp_difference/max': 0.022413522005081177, 'sampling/importance_sampling_ratio/min': 0.9778358340263367, 'sampling/importance_sampling_ratio/mean': 1.0006673336029053, 'sampling/importance_sampling_ratio/max': 1.0154409408569336, 'entropy': 0.009250407747458667, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990598559379578, 'reward_meter_mean': 0.9990598559379578, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990598559379578, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 62%|██████▏   | 2040/3300 [2:42:02<1:45:12,  5.01s/it]INFO 04-12 01:18:53 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2041/3300 [2:42:06<1:42:41,  4.89s/it]2026-04-12 01:18:58,046 | INFO | train_grpo_train | metrics_logged mode=train step=2041

                                                       
{'loss': -0.0008, 'grad_norm': 4.351480007171631, 'learning_rate': 3.818181818181819e-06, 'num_tokens': 4590654.0, 'completions/mean_length': 39.0, 'completions/min_length': 38.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9944432973861694, 'rewards/meter/std': 0.0023516512010246515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9944432973861694, 'rewards/total_composite/std': 0.0023516512010246515, 'reward': 0.9944432973861694, 'reward_std': 0.0023516479413956404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022262930870056152, 'sampling/sampling_logp_difference/max': 0.5608975887298584, 'sampling/importance_sampling_ratio/min': 0.5875173807144165, 'sampling/importance_sampling_ratio/mean': 1.0104680061340332, 'sampling/importance_sampling_ratio/max': 1.7522445917129517, 'entropy': 0.21922382153570652, 'clip_ratio/low_mean': 0.006330128293484449, 'clip_ratio/low_min': 0.006330128293484449, 'clip_ratio/high_mean': 0.009615384740754962, 'clip_ratio/high_max': 0.009615384740754962, 'clip_ratio/region_mean': 0.01594551303423941, 'reward_total_mean': 0.9944432973861694, 'reward_meter_mean': 0.9944432973861694, 'reward_meter_std': 0.0023516512010246515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9944432973861694, 'reward_total_composite_std': 0.0023516512010246515, 'epoch': 0.08}

 62%|██████▏   | 2041/3300 [2:42:07<1:42:41,  4.89s/it]INFO 04-12 01:18:58 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2042/3300 [2:42:11<1:40:29,  4.79s/it]2026-04-12 01:19:02,595 | INFO | train_grpo_train | metrics_logged mode=train step=2042

                                                       
{'loss': 0.0002, 'grad_norm': 0.49870744347572327, 'learning_rate': 3.8151515151515155e-06, 'num_tokens': 4592318.0, 'completions/mean_length': 58.0, 'completions/min_length': 58.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9948880672454834, 'rewards/meter/std': 3.073411789955571e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948880672454834, 'rewards/total_composite/std': 3.073411789955571e-05, 'reward': 0.9948880672454834, 'reward_std': 3.0727183911949396e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004786766599863768, 'sampling/sampling_logp_difference/max': 0.8091294765472412, 'sampling/importance_sampling_ratio/min': 0.44524553418159485, 'sampling/importance_sampling_ratio/mean': 1.001407265663147, 'sampling/importance_sampling_ratio/max': 1.819435477256775, 'entropy': 0.023268206976354122, 'clip_ratio/low_mean': 0.0021551724057644606, 'clip_ratio/low_min': 0.0021551724057644606, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0021551724057644606, 'reward_total_mean': 0.9948880672454834, 'reward_meter_mean': 0.9948880672454834, 'reward_meter_std': 3.073411789955571e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948880672454834, 'reward_total_composite_std': 3.073411789955571e-05, 'epoch': 0.08}

 62%|██████▏   | 2042/3300 [2:42:11<1:40:29,  4.79s/it]INFO 04-12 01:19:02 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2043/3300 [2:42:20<2:04:56,  5.96s/it]2026-04-12 01:19:11,293 | INFO | train_grpo_train | metrics_logged mode=train step=2043

                                                       
{'loss': -0.0015, 'grad_norm': 2.7885191440582275, 'learning_rate': 3.8121212121212127e-06, 'num_tokens': 4597255.0, 'completions/mean_length': 382.125, 'completions/min_length': 379.0, 'completions/max_length': 392.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 382.125, 'completions/min_terminated_length': 379.0, 'completions/max_terminated_length': 392.0, 'rewards/meter/mean': 0.9941146373748779, 'rewards/meter/std': 0.0013542358065024018, 'rewards/count_adherence/mean': 0.6397058963775635, 'rewards/count_adherence/std': 0.020797256380319595, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6714285612106323, 'rewards/repeat_penalty/std': 0.06172133609652519, 'rewards/total_composite/mean': 0.42605897784233093, 'rewards/total_composite/std': 0.027482789009809494, 'reward': 0.42605897784233093, 'reward_std': 0.02748279646039009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017567651346325874, 'sampling/sampling_logp_difference/max': 5.018847465515137, 'sampling/importance_sampling_ratio/min': 0.006612143013626337, 'sampling/importance_sampling_ratio/mean': 1.0006448030471802, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07445414271205664, 'clip_ratio/low_mean': 0.0013192612677812576, 'clip_ratio/low_min': 0.0013192612677812576, 'clip_ratio/high_mean': 0.01007624133490026, 'clip_ratio/high_max': 0.01007624133490026, 'clip_ratio/region_mean': 0.011395502602681518, 'reward_total_mean': 0.42605897784233093, 'reward_meter_mean': 0.9941146373748779, 'reward_meter_std': 0.0013542358065024018, 'reward_count_adherence_mean': 0.6397058963775635, 'reward_count_adherence_std': 0.020797256380319595, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6714285612106323, 'reward_repeat_penalty_std': 0.06172133609652519, 'reward_total_composite_mean': 0.42605897784233093, 'reward_total_composite_std': 0.027482789009809494, 'epoch': 0.08}

 62%|██████▏   | 2043/3300 [2:42:20<2:04:56,  5.96s/it]INFO 04-12 01:19:11 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2044/3300 [2:42:25<1:57:53,  5.63s/it]2026-04-12 01:19:16,156 | INFO | train_grpo_train | metrics_logged mode=train step=2044

                                                       
{'loss': -0.018, 'grad_norm': 3.922247886657715, 'learning_rate': 3.8090909090909095e-06, 'num_tokens': 4599133.0, 'completions/mean_length': 71.75, 'completions/min_length': 69.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9981293082237244, 'rewards/meter/std': 0.0010699051199480891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981293082237244, 'rewards/total_composite/std': 0.0010699051199480891, 'reward': 0.9981293082237244, 'reward_std': 0.0010698941769078374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03494240716099739, 'sampling/sampling_logp_difference/max': 1.515822410583496, 'sampling/importance_sampling_ratio/min': 0.2196274846792221, 'sampling/importance_sampling_ratio/mean': 1.008084774017334, 'sampling/importance_sampling_ratio/max': 1.6578000783920288, 'entropy': 0.23516237549483776, 'clip_ratio/low_mean': 0.012303744442760944, 'clip_ratio/low_min': 0.012303744442760944, 'clip_ratio/high_mean': 0.00866253802087158, 'clip_ratio/high_max': 0.00866253802087158, 'clip_ratio/region_mean': 0.020966282463632524, 'reward_total_mean': 0.9981293082237244, 'reward_meter_mean': 0.9981293082237244, 'reward_meter_std': 0.0010699051199480891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981293082237244, 'reward_total_composite_std': 0.0010699051199480891, 'epoch': 0.08}

 62%|██████▏   | 2044/3300 [2:42:25<1:57:53,  5.63s/it]INFO 04-12 01:19:16 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2045/3300 [2:42:30<1:57:30,  5.62s/it]2026-04-12 01:19:21,740 | INFO | train_grpo_train | metrics_logged mode=train step=2045

                                                       
{'loss': -0.0019, 'grad_norm': 3.759932279586792, 'learning_rate': 3.8060606060606064e-06, 'num_tokens': 4601598.0, 'completions/mean_length': 128.125, 'completions/min_length': 127.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.125, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9989614486694336, 'rewards/meter/std': 9.313080954598263e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8919284343719482, 'rewards/total_composite/std': 0.06603793799877167, 'reward': 0.8919284343719482, 'reward_std': 0.06603794544935226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013363703154027462, 'sampling/sampling_logp_difference/max': 1.402510166168213, 'sampling/importance_sampling_ratio/min': 0.24597874283790588, 'sampling/importance_sampling_ratio/mean': 0.9983217120170593, 'sampling/importance_sampling_ratio/max': 1.4929101467132568, 'entropy': 0.04036114830523729, 'clip_ratio/low_mean': 0.004853463382460177, 'clip_ratio/low_min': 0.004853463382460177, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004853463382460177, 'reward_total_mean': 0.8919284343719482, 'reward_meter_mean': 0.9989614486694336, 'reward_meter_std': 9.313080954598263e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8919284343719482, 'reward_total_composite_std': 0.06603793799877167, 'epoch': 0.08}

 62%|██████▏   | 2045/3300 [2:42:30<1:57:30,  5.62s/it]INFO 04-12 01:19:21 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2046/3300 [2:42:35<1:51:32,  5.34s/it]2026-04-12 01:19:26,424 | INFO | train_grpo_train | metrics_logged mode=train step=2046

                                                       
{'loss': -0.0005, 'grad_norm': 0.4616295397281647, 'learning_rate': 3.803030303030303e-06, 'num_tokens': 4603398.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.999051570892334, 'rewards/meter/std': 2.334937744308263e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999051570892334, 'rewards/total_composite/std': 2.334937744308263e-05, 'reward': 0.999051570892334, 'reward_std': 2.3349353796220385e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0014573887456208467, 'sampling/sampling_logp_difference/max': 0.3329801559448242, 'sampling/importance_sampling_ratio/min': 0.7167844176292419, 'sampling/importance_sampling_ratio/mean': 1.0000971555709839, 'sampling/importance_sampling_ratio/max': 1.0310380458831787, 'entropy': 0.007882928592152894, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.999051570892334, 'reward_meter_mean': 0.999051570892334, 'reward_meter_std': 2.334937744308263e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999051570892334, 'reward_total_composite_std': 2.334937744308263e-05, 'epoch': 0.08}

 62%|██████▏   | 2046/3300 [2:42:35<1:51:32,  5.34s/it]INFO 04-12 01:19:26 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2047/3300 [2:42:40<1:47:24,  5.14s/it]2026-04-12 01:19:31,115 | INFO | train_grpo_train | metrics_logged mode=train step=2047

                                                       
{'loss': -0.0001, 'grad_norm': 0.1668306142091751, 'learning_rate': 3.8000000000000005e-06, 'num_tokens': 4605142.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9979839324951172, 'rewards/meter/std': 4.366625034890603e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979839324951172, 'rewards/total_composite/std': 4.366625034890603e-06, 'reward': 0.9979839324951172, 'reward_std': 4.369384441815782e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005823140498250723, 'sampling/sampling_logp_difference/max': 0.598487377166748, 'sampling/importance_sampling_ratio/min': 0.5496423840522766, 'sampling/importance_sampling_ratio/mean': 1.0016472339630127, 'sampling/importance_sampling_ratio/max': 1.3960132598876953, 'entropy': 0.03522463422268629, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.0037878789007663727, 'reward_total_mean': 0.9979839324951172, 'reward_meter_mean': 0.9979839324951172, 'reward_meter_std': 4.366625034890603e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979839324951172, 'reward_total_composite_std': 4.366625034890603e-06, 'epoch': 0.08}

 62%|██████▏   | 2047/3300 [2:42:40<1:47:24,  5.14s/it]INFO 04-12 01:19:31 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2048/3300 [2:42:46<1:57:01,  5.61s/it]2026-04-12 01:19:37,807 | INFO | train_grpo_train | metrics_logged mode=train step=2048

                                                       
{'loss': -0.0143, 'grad_norm': 1.707572102546692, 'learning_rate': 3.7969696969696973e-06, 'num_tokens': 4608320.0, 'completions/mean_length': 213.25, 'completions/min_length': 187.0, 'completions/max_length': 221.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 213.25, 'completions/min_terminated_length': 187.0, 'completions/max_terminated_length': 221.0, 'rewards/meter/mean': 0.9987424612045288, 'rewards/meter/std': 0.0006651075091212988, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8977272510528564, 'rewards/repeat_penalty/std': 0.09009374678134918, 'rewards/total_composite/mean': 0.8758239150047302, 'rewards/total_composite/std': 0.08225142955780029, 'reward': 0.8758239150047302, 'reward_std': 0.08225142955780029, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02781507931649685, 'sampling/sampling_logp_difference/max': 1.3763513565063477, 'sampling/importance_sampling_ratio/min': 0.25249814987182617, 'sampling/importance_sampling_ratio/mean': 1.006170392036438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2369670756161213, 'clip_ratio/low_mean': 0.012748151202686131, 'clip_ratio/low_min': 0.012748151202686131, 'clip_ratio/high_mean': 0.018360439455136657, 'clip_ratio/high_max': 0.018360439455136657, 'clip_ratio/region_mean': 0.031108590657822788, 'reward_total_mean': 0.8758239150047302, 'reward_meter_mean': 0.9987424612045288, 'reward_meter_std': 0.0006651075091212988, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8977272510528564, 'reward_repeat_penalty_std': 0.09009374678134918, 'reward_total_composite_mean': 0.8758239150047302, 'reward_total_composite_std': 0.08225142955780029, 'epoch': 0.08}

 62%|██████▏   | 2048/3300 [2:42:46<1:57:01,  5.61s/it]INFO 04-12 01:19:38 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2049/3300 [2:42:51<1:52:06,  5.38s/it]2026-04-12 01:19:42,641 | INFO | train_grpo_train | metrics_logged mode=train step=2049

                                                       
{'loss': 0.0002, 'grad_norm': 3.169921636581421, 'learning_rate': 3.793939393939394e-06, 'num_tokens': 4610199.0, 'completions/mean_length': 72.875, 'completions/min_length': 71.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.875, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9960318207740784, 'rewards/meter/std': 0.0027530721854418516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9960318207740784, 'rewards/total_composite/std': 0.0027530721854418516, 'reward': 0.9960318207740784, 'reward_std': 0.0027530903462320566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02808314375579357, 'sampling/sampling_logp_difference/max': 1.0131583213806152, 'sampling/importance_sampling_ratio/min': 0.3630704879760742, 'sampling/importance_sampling_ratio/mean': 1.0030927658081055, 'sampling/importance_sampling_ratio/max': 1.5734020471572876, 'entropy': 0.19552704505622387, 'clip_ratio/low_mean': 0.0017123287543654442, 'clip_ratio/low_min': 0.0017123287543654442, 'clip_ratio/high_mean': 0.00854364933911711, 'clip_ratio/high_max': 0.00854364933911711, 'clip_ratio/region_mean': 0.010255978093482554, 'reward_total_mean': 0.9960318207740784, 'reward_meter_mean': 0.9960318207740784, 'reward_meter_std': 0.0027530721854418516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9960318207740784, 'reward_total_composite_std': 0.0027530721854418516, 'epoch': 0.08}

 62%|██████▏   | 2049/3300 [2:42:51<1:52:06,  5.38s/it]INFO 04-12 01:19:42 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2050/3300 [2:42:56<1:47:51,  5.18s/it]2026-04-12 01:19:47,355 | INFO | train_grpo_train | metrics_logged mode=train step=2050

                                                       
{'loss': -0.0035, 'grad_norm': 2.9366605281829834, 'learning_rate': 3.7909090909090914e-06, 'num_tokens': 4612079.0, 'completions/mean_length': 63.0, 'completions/min_length': 62.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8283913731575012, 'rewards/meter/std': 0.1856846958398819, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8283913731575012, 'rewards/total_composite/std': 0.1856846958398819, 'reward': 0.8283913731575012, 'reward_std': 0.1856846958398819, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02338075079023838, 'sampling/sampling_logp_difference/max': 2.3884801864624023, 'sampling/importance_sampling_ratio/min': 0.09176904708147049, 'sampling/importance_sampling_ratio/mean': 1.000585913658142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15631293877959251, 'clip_ratio/low_mean': 0.009891632944345474, 'clip_ratio/low_min': 0.009891632944345474, 'clip_ratio/high_mean': 0.007843502098694444, 'clip_ratio/high_max': 0.007843502098694444, 'clip_ratio/region_mean': 0.017735135043039918, 'reward_total_mean': 0.8283913731575012, 'reward_meter_mean': 0.8283913731575012, 'reward_meter_std': 0.1856846958398819, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8283913731575012, 'reward_total_composite_std': 0.1856846958398819, 'epoch': 0.08}

 62%|██████▏   | 2050/3300 [2:42:56<1:47:51,  5.18s/it]INFO 04-12 01:19:47 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:31,  2.87s/it]

 23%|██▎       | 3/13 [00:09<00:32,  3.26s/it]

 31%|███       | 4/13 [00:14<00:35,  3.93s/it]

 38%|███▊      | 5/13 [00:19<00:34,  4.26s/it]

 46%|████▌     | 6/13 [00:25<00:32,  4.69s/it]

 54%|█████▍    | 7/13 [00:29<00:28,  4.70s/it]

 62%|██████▏   | 8/13 [00:35<00:25,  5.08s/it]

 69%|██████▉   | 9/13 [00:41<00:21,  5.29s/it]

 77%|███████▋  | 10/13 [00:46<00:15,  5.07s/it]

 85%|████████▍ | 11/13 [00:51<00:10,  5.15s/it]

 92%|█████████▏| 12/13 [00:56<00:05,  5.17s/it]

100%|██████████| 13/13 [01:01<00:00,  5.06s/it]2026-04-12 01:20:55,009 | INFO | train_grpo_train | metrics_logged mode=eval step=2050

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 67.5758, 'eval_samples_per_second': 1.539, 'eval_steps_per_second': 0.192, 'eval_num_tokens': 4612079.0, 'eval_completions/mean_length': 201.2403846153846, 'eval_completions/min_length': 63.30769230769231, 'eval_completions/max_length': 354.6923076923077, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 201.2403846153846, 'eval_completions/min_terminated_length': 63.30769230769231, 'eval_completions/max_terminated_length': 354.6923076923077, 'eval_rewards/meter/mean': 0.7046224291508014, 'eval_rewards/meter/std': 0.42778917917838466, 'eval_rewards/count_adherence/mean': 0.9102904154704168, 'eval_rewards/count_adherence/std': 0.12148115927210221, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.80578757249392, 'eval_rewards/repeat_penalty/std': 0.18289457318874505, 'eval_rewards/total_composite/mean': 0.5263216747687414, 'eval_rewards/total_composite/std': 0.37407297583726734, 'eval_reward': 0.5263216747687414, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.013937483756588055, 'eval_sampling/sampling_logp_difference/max': 1.54466306246244, 'eval_sampling/importance_sampling_ratio/min': 0.2842731796778165, 'eval_sampling/importance_sampling_ratio/mean': 1.0030618355824397, 'eval_sampling/importance_sampling_ratio/max': 1.4798904473964984, 'eval_entropy': 0.12492357309047993, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5263216747687414, 'eval_reward_meter_mean': 0.7046224291508014, 'eval_reward_meter_std': 0.42778917917838466, 'eval_reward_count_adherence_mean': 0.9102904154704168, 'eval_reward_count_adherence_std': 0.12148115927210221, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.80578757249392, 'eval_reward_repeat_penalty_std': 0.18289457318874505, 'eval_reward_total_composite_mean': 0.5263216747687414, 'eval_reward_total_composite_std': 0.37407297583726734, 'epoch': 0.08}

 62%|██████▏   | 2050/3300 [2:44:04<1:47:51,  5.18s/it]

100%|██████████| 13/13 [01:01<00:00,  5.06s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 62%|██████▏   | 2051/3300 [2:44:14<9:25:56, 27.19s/it]2026-04-12 01:21:05,885 | INFO | train_grpo_train | metrics_logged mode=train step=2051

                                                       
{'loss': -0.0002, 'grad_norm': 0.02642267569899559, 'learning_rate': 3.7878787878787882e-06, 'num_tokens': 4616222.0, 'completions/mean_length': 312.875, 'completions/min_length': 312.0, 'completions/max_length': 313.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 312.875, 'completions/min_terminated_length': 312.0, 'completions/max_terminated_length': 313.0, 'rewards/meter/mean': 0.9973797798156738, 'rewards/meter/std': 8.498986062477343e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.5789473652839661, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5774303674697876, 'rewards/total_composite/std': 4.927758254780201e-06, 'reward': 0.5774303674697876, 'reward_std': 4.94351661473047e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.001822733087465167, 'sampling/sampling_logp_difference/max': 0.5258277654647827, 'sampling/importance_sampling_ratio/min': 0.5910658836364746, 'sampling/importance_sampling_ratio/mean': 1.0007059574127197, 'sampling/importance_sampling_ratio/max': 1.331079125404358, 'entropy': 0.01078800146933645, 'clip_ratio/low_mean': 0.0003993610152974725, 'clip_ratio/low_min': 0.0003993610152974725, 'clip_ratio/high_mean': 0.0008000020461622626, 'clip_ratio/high_max': 0.0008000020461622626, 'clip_ratio/region_mean': 0.001199363061459735, 'reward_total_mean': 0.5774303674697876, 'reward_meter_mean': 0.9973797798156738, 'reward_meter_std': 8.498986062477343e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.5789473652839661, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5774303674697876, 'reward_total_composite_std': 4.927758254780201e-06, 'epoch': 0.08}

 62%|██████▏   | 2051/3300 [2:44:14<9:25:56, 27.19s/it]INFO 04-12 01:21:06 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2052/3300 [2:44:21<7:17:43, 21.04s/it]2026-04-12 01:21:12,595 | INFO | train_grpo_train | metrics_logged mode=train step=2052

                                                       
{'loss': -0.0126, 'grad_norm': 1.5723775625228882, 'learning_rate': 3.784848484848485e-06, 'num_tokens': 4619661.0, 'completions/mean_length': 214.875, 'completions/min_length': 208.0, 'completions/max_length': 221.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 214.875, 'completions/min_terminated_length': 208.0, 'completions/max_terminated_length': 221.0, 'rewards/meter/mean': 0.9847596883773804, 'rewards/meter/std': 0.03142433241009712, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.625, 'rewards/repeat_penalty/std': 0.2591308057308197, 'rewards/total_composite/mean': 0.6123065948486328, 'rewards/total_composite/std': 0.24669423699378967, 'reward': 0.6123065948486328, 'reward_std': 0.24669425189495087, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023984558880329132, 'sampling/sampling_logp_difference/max': 1.9047002792358398, 'sampling/importance_sampling_ratio/min': 0.14886726438999176, 'sampling/importance_sampling_ratio/mean': 1.0024995803833008, 'sampling/importance_sampling_ratio/max': 1.9927300214767456, 'entropy': 0.18392128869891167, 'clip_ratio/low_mean': 0.0017639786819927394, 'clip_ratio/low_min': 0.0017639786819927394, 'clip_ratio/high_mean': 0.014248464838601649, 'clip_ratio/high_max': 0.014248464838601649, 'clip_ratio/region_mean': 0.016012443520594388, 'reward_total_mean': 0.6123065948486328, 'reward_meter_mean': 0.9847596883773804, 'reward_meter_std': 0.03142433241009712, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.625, 'reward_repeat_penalty_std': 0.2591308057308197, 'reward_total_composite_mean': 0.6123065948486328, 'reward_total_composite_std': 0.24669423699378967, 'epoch': 0.08}

 62%|██████▏   | 2052/3300 [2:44:21<7:17:43, 21.04s/it]INFO 04-12 01:21:12 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2053/3300 [2:44:25<5:33:57, 16.07s/it]2026-04-12 01:21:17,059 | INFO | train_grpo_train | metrics_logged mode=train step=2053

                                                       
{'loss': 0.0254, 'grad_norm': 7.329649448394775, 'learning_rate': 3.781818181818182e-06, 'num_tokens': 4621203.0, 'completions/mean_length': 38.75, 'completions/min_length': 36.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9991681575775146, 'rewards/meter/std': 0.0005326797836460173, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991681575775146, 'rewards/total_composite/std': 0.0005326797836460173, 'reward': 0.9991681575775146, 'reward_std': 0.0005326886312104762, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02681351639330387, 'sampling/sampling_logp_difference/max': 0.8889703750610352, 'sampling/importance_sampling_ratio/min': 0.4110788106918335, 'sampling/importance_sampling_ratio/mean': 1.006339430809021, 'sampling/importance_sampling_ratio/max': 1.5577260255813599, 'entropy': 0.16460048034787178, 'clip_ratio/low_mean': 0.006253908621147275, 'clip_ratio/low_min': 0.006253908621147275, 'clip_ratio/high_mean': 0.0065972222946584225, 'clip_ratio/high_max': 0.0065972222946584225, 'clip_ratio/region_mean': 0.012851130915805697, 'reward_total_mean': 0.9991681575775146, 'reward_meter_mean': 0.9991681575775146, 'reward_meter_std': 0.0005326797836460173, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991681575775146, 'reward_total_composite_std': 0.0005326797836460173, 'epoch': 0.08}

 62%|██████▏   | 2053/3300 [2:44:26<5:33:57, 16.07s/it]INFO 04-12 01:21:17 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2054/3300 [2:44:30<4:23:27, 12.69s/it]2026-04-12 01:21:21,871 | INFO | train_grpo_train | metrics_logged mode=train step=2054

                                                       
{'loss': 0.0004, 'grad_norm': 0.35728779435157776, 'learning_rate': 3.778787878787879e-06, 'num_tokens': 4623051.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.997982382774353, 'rewards/meter/std': 1.786020766303409e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997982382774353, 'rewards/total_composite/std': 1.786020766303409e-05, 'reward': 0.997982382774353, 'reward_std': 1.7854295947472565e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009083742275834084, 'sampling/sampling_logp_difference/max': 0.5304985046386719, 'sampling/importance_sampling_ratio/min': 0.5883116722106934, 'sampling/importance_sampling_ratio/mean': 0.999168872833252, 'sampling/importance_sampling_ratio/max': 1.3692970275878906, 'entropy': 0.03993460722267628, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.007575757801532745, 'clip_ratio/high_max': 0.007575757801532745, 'clip_ratio/region_mean': 0.01515151560306549, 'reward_total_mean': 0.997982382774353, 'reward_meter_mean': 0.997982382774353, 'reward_meter_std': 1.786020766303409e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997982382774353, 'reward_total_composite_std': 1.786020766303409e-05, 'epoch': 0.08}

 62%|██████▏   | 2054/3300 [2:44:30<4:23:27, 12.69s/it]INFO 04-12 01:21:22 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2055/3300 [2:44:38<3:54:40, 11.31s/it]2026-04-12 01:21:29,952 | INFO | train_grpo_train | metrics_logged mode=train step=2055

                                                       
{'loss': -0.0003, 'grad_norm': 2.26758074760437, 'learning_rate': 3.775757575757576e-06, 'num_tokens': 4627144.0, 'completions/mean_length': 315.625, 'completions/min_length': 305.0, 'completions/max_length': 330.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 315.625, 'completions/min_terminated_length': 305.0, 'completions/max_terminated_length': 330.0, 'rewards/meter/mean': 0.9283133149147034, 'rewards/meter/std': 0.12819351255893707, 'rewards/count_adherence/mean': 0.862500011920929, 'rewards/count_adherence/std': 0.0517548993229866, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8436580896377563, 'rewards/repeat_penalty/std': 0.11612330377101898, 'rewards/total_composite/mean': 0.6760411262512207, 'rewards/total_composite/std': 0.1399042010307312, 'reward': 0.6760411262512207, 'reward_std': 0.1399042010307312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0354769341647625, 'sampling/sampling_logp_difference/max': 2.784627914428711, 'sampling/importance_sampling_ratio/min': 0.061752062290906906, 'sampling/importance_sampling_ratio/mean': 1.002432942390442, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22647083643823862, 'clip_ratio/low_mean': 0.006354167824611068, 'clip_ratio/low_min': 0.006354167824611068, 'clip_ratio/high_mean': 0.026429779594764113, 'clip_ratio/high_max': 0.026429779594764113, 'clip_ratio/region_mean': 0.03278394741937518, 'reward_total_mean': 0.6760411262512207, 'reward_meter_mean': 0.9283133149147034, 'reward_meter_std': 0.12819351255893707, 'reward_count_adherence_mean': 0.862500011920929, 'reward_count_adherence_std': 0.0517548993229866, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8436580896377563, 'reward_repeat_penalty_std': 0.11612330377101898, 'reward_total_composite_mean': 0.6760411262512207, 'reward_total_composite_std': 0.1399042010307312, 'epoch': 0.08}

 62%|██████▏   | 2055/3300 [2:44:38<3:54:40, 11.31s/it]INFO 04-12 01:21:30 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2056/3300 [2:44:43<3:13:06,  9.31s/it]2026-04-12 01:21:34,603 | INFO | train_grpo_train | metrics_logged mode=train step=2056

                                                       
{'loss': -0.0133, 'grad_norm': 8.520014762878418, 'learning_rate': 3.772727272727273e-06, 'num_tokens': 4628924.0, 'completions/mean_length': 61.5, 'completions/min_length': 42.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6756709814071655, 'rewards/meter/std': 0.25546279549598694, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6134400963783264, 'rewards/total_composite/std': 0.22521547973155975, 'reward': 0.6134400963783264, 'reward_std': 0.22521549463272095, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031044768169522285, 'sampling/sampling_logp_difference/max': 2.2168161869049072, 'sampling/importance_sampling_ratio/min': 0.10895545780658722, 'sampling/importance_sampling_ratio/mean': 1.0012702941894531, 'sampling/importance_sampling_ratio/max': 1.7200733423233032, 'entropy': 0.2539948094636202, 'clip_ratio/low_mean': 0.007753314450383186, 'clip_ratio/low_min': 0.007753314450383186, 'clip_ratio/high_mean': 0.013611778849735856, 'clip_ratio/high_max': 0.013611778849735856, 'clip_ratio/region_mean': 0.021365093300119042, 'reward_total_mean': 0.6134400963783264, 'reward_meter_mean': 0.6756709814071655, 'reward_meter_std': 0.25546279549598694, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6134400963783264, 'reward_total_composite_std': 0.22521547973155975, 'epoch': 0.08}

 62%|██████▏   | 2056/3300 [2:44:43<3:13:06,  9.31s/it]INFO 04-12 01:21:34 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2057/3300 [2:44:48<2:43:57,  7.91s/it]2026-04-12 01:21:39,255 | INFO | train_grpo_train | metrics_logged mode=train step=2057

                                                       
{'loss': -0.001, 'grad_norm': 0.4782918691635132, 'learning_rate': 3.76969696969697e-06, 'num_tokens': 4630892.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990358352661133, 'rewards/meter/std': 6.781429692637175e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990358352661133, 'rewards/total_composite/std': 6.781429692637175e-05, 'reward': 0.9990358352661133, 'reward_std': 6.782029959140345e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0013893973082304, 'sampling/sampling_logp_difference/max': 0.41433000564575195, 'sampling/importance_sampling_ratio/min': 0.6607828736305237, 'sampling/importance_sampling_ratio/mean': 0.9997783303260803, 'sampling/importance_sampling_ratio/max': 1.0163952112197876, 'entropy': 0.00615574762923643, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990358352661133, 'reward_meter_mean': 0.9990358352661133, 'reward_meter_std': 6.781429692637175e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990358352661133, 'reward_total_composite_std': 6.781429692637175e-05, 'epoch': 0.08}

 62%|██████▏   | 2057/3300 [2:44:48<2:43:57,  7.91s/it]INFO 04-12 01:21:39 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2058/3300 [2:44:52<2:23:27,  6.93s/it]2026-04-12 01:21:43,894 | INFO | train_grpo_train | metrics_logged mode=train step=2058

                                                       
{'loss': 0.0, 'grad_norm': 0.47224369645118713, 'learning_rate': 3.766666666666667e-06, 'num_tokens': 4632844.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9979783296585083, 'rewards/meter/std': 4.64778822788503e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979783296585083, 'rewards/total_composite/std': 4.64778822788503e-05, 'reward': 0.9979783296585083, 'reward_std': 4.6468860091408715e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004593671765178442, 'sampling/sampling_logp_difference/max': 0.4634366035461426, 'sampling/importance_sampling_ratio/min': 0.6291179060935974, 'sampling/importance_sampling_ratio/mean': 1.0018833875656128, 'sampling/importance_sampling_ratio/max': 1.5034801959991455, 'entropy': 0.029094983357936144, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9979783296585083, 'reward_meter_mean': 0.9979783296585083, 'reward_meter_std': 4.64778822788503e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979783296585083, 'reward_total_composite_std': 4.64778822788503e-05, 'epoch': 0.08}

 62%|██████▏   | 2058/3300 [2:44:52<2:23:27,  6.93s/it]INFO 04-12 01:21:44 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2059/3300 [2:44:57<2:09:17,  6.25s/it]2026-04-12 01:21:48,561 | INFO | train_grpo_train | metrics_logged mode=train step=2059

                                                       
{'loss': -0.0039, 'grad_norm': 3.3012750148773193, 'learning_rate': 3.7636363636363637e-06, 'num_tokens': 4634771.0, 'completions/mean_length': 65.875, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9989724159240723, 'rewards/meter/std': 0.00024742307141423225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989724159240723, 'rewards/total_composite/std': 0.00024742307141423225, 'reward': 0.9989724159240723, 'reward_std': 0.0002474260691087693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0033799612428992987, 'sampling/sampling_logp_difference/max': 1.322075366973877, 'sampling/importance_sampling_ratio/min': 0.2665814757347107, 'sampling/importance_sampling_ratio/mean': 0.9990096688270569, 'sampling/importance_sampling_ratio/max': 1.0218313932418823, 'entropy': 0.0075465834233909845, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9989724159240723, 'reward_meter_mean': 0.9989724159240723, 'reward_meter_std': 0.00024742307141423225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989724159240723, 'reward_total_composite_std': 0.00024742307141423225, 'epoch': 0.08}

 62%|██████▏   | 2059/3300 [2:44:57<2:09:17,  6.25s/it]INFO 04-12 01:21:48 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2060/3300 [2:45:02<2:02:04,  5.91s/it]2026-04-12 01:21:53,661 | INFO | train_grpo_train | metrics_logged mode=train step=2060

                                                       
{'loss': -0.0013, 'grad_norm': 0.7633002400398254, 'learning_rate': 3.7606060606060605e-06, 'num_tokens': 4636937.0, 'completions/mean_length': 98.75, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.75, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.99910569190979, 'rewards/meter/std': 4.124943006900139e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99910569190979, 'rewards/total_composite/std': 4.124943006900139e-05, 'reward': 0.99910569190979, 'reward_std': 4.125862324144691e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0010279221460223198, 'sampling/sampling_logp_difference/max': 0.05605363845825195, 'sampling/importance_sampling_ratio/min': 0.9609670042991638, 'sampling/importance_sampling_ratio/mean': 1.0008070468902588, 'sampling/importance_sampling_ratio/max': 1.0576543807983398, 'entropy': 0.011058209463953972, 'clip_ratio/low_mean': 0.0025510203558951616, 'clip_ratio/low_min': 0.0025510203558951616, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0025510203558951616, 'reward_total_mean': 0.99910569190979, 'reward_meter_mean': 0.99910569190979, 'reward_meter_std': 4.124943006900139e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99910569190979, 'reward_total_composite_std': 4.124943006900139e-05, 'epoch': 0.08}

 62%|██████▏   | 2060/3300 [2:45:02<2:02:04,  5.91s/it]INFO 04-12 01:21:53 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2061/3300 [2:45:07<1:58:44,  5.75s/it]2026-04-12 01:21:59,046 | INFO | train_grpo_train | metrics_logged mode=train step=2061

                                                       
{'loss': 0.0046, 'grad_norm': 2.503458023071289, 'learning_rate': 3.757575757575758e-06, 'num_tokens': 4638510.0, 'completions/mean_length': 38.625, 'completions/min_length': 36.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9993641972541809, 'rewards/meter/std': 0.0002709543623495847, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993641972541809, 'rewards/total_composite/std': 0.0002709543623495847, 'reward': 0.9993641972541809, 'reward_std': 0.0002709643158596009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026116767898201942, 'sampling/sampling_logp_difference/max': 1.114222526550293, 'sampling/importance_sampling_ratio/min': 0.3281702995300293, 'sampling/importance_sampling_ratio/mean': 1.0021777153015137, 'sampling/importance_sampling_ratio/max': 1.691472053527832, 'entropy': 0.1621959926560521, 'clip_ratio/low_mean': 0.012820512987673283, 'clip_ratio/low_min': 0.012820512987673283, 'clip_ratio/high_mean': 0.016136696096509695, 'clip_ratio/high_max': 0.016136696096509695, 'clip_ratio/region_mean': 0.028957209084182978, 'reward_total_mean': 0.9993641972541809, 'reward_meter_mean': 0.9993641972541809, 'reward_meter_std': 0.0002709543623495847, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993641972541809, 'reward_total_composite_std': 0.0002709543623495847, 'epoch': 0.08}

 62%|██████▏   | 2061/3300 [2:45:08<1:58:44,  5.75s/it]INFO 04-12 01:21:59 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2062/3300 [2:45:13<1:56:00,  5.62s/it]2026-04-12 01:22:04,374 | INFO | train_grpo_train | metrics_logged mode=train step=2062

                                                       
{'loss': 0.0007, 'grad_norm': 0.12022405117750168, 'learning_rate': 3.7545454545454546e-06, 'num_tokens': 4640965.0, 'completions/mean_length': 127.875, 'completions/min_length': 127.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.875, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9978995323181152, 'rewards/meter/std': 1.4925647519703489e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8553425073623657, 'rewards/total_composite/std': 1.2786828847310971e-05, 'reward': 0.8553425073623657, 'reward_std': 1.2798592251783703e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004213304258882999, 'sampling/sampling_logp_difference/max': 0.8202402591705322, 'sampling/importance_sampling_ratio/min': 0.44032585620880127, 'sampling/importance_sampling_ratio/mean': 0.9991137981414795, 'sampling/importance_sampling_ratio/max': 1.108210802078247, 'entropy': 0.019355892902240157, 'clip_ratio/low_mean': 0.0009765625, 'clip_ratio/low_min': 0.0009765625, 'clip_ratio/high_mean': 0.0029450664296746254, 'clip_ratio/high_max': 0.0029450664296746254, 'clip_ratio/region_mean': 0.003921628929674625, 'reward_total_mean': 0.8553425073623657, 'reward_meter_mean': 0.9978995323181152, 'reward_meter_std': 1.4925647519703489e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8553425073623657, 'reward_total_composite_std': 1.2786828847310971e-05, 'epoch': 0.08}

 62%|██████▏   | 2062/3300 [2:45:13<1:56:00,  5.62s/it]INFO 04-12 01:22:04 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2063/3300 [2:45:18<1:51:58,  5.43s/it]2026-04-12 01:22:09,355 | INFO | train_grpo_train | metrics_logged mode=train step=2063

                                                       
{'loss': -0.0003, 'grad_norm': 0.1908394694328308, 'learning_rate': 3.7515151515151515e-06, 'num_tokens': 4643037.0, 'completions/mean_length': 97.0, 'completions/min_length': 97.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.9979283809661865, 'rewards/meter/std': 1.0387539987277705e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979283809661865, 'rewards/total_composite/std': 1.0387539987277705e-05, 'reward': 0.9979283809661865, 'reward_std': 1.038749087456381e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0028961896896362305, 'sampling/sampling_logp_difference/max': 0.6855463981628418, 'sampling/importance_sampling_ratio/min': 0.5038148760795593, 'sampling/importance_sampling_ratio/mean': 1.001339077949524, 'sampling/importance_sampling_ratio/max': 1.0733686685562134, 'entropy': 0.020256070652976632, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9979283809661865, 'reward_meter_mean': 0.9979283809661865, 'reward_meter_std': 1.0387539987277705e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979283809661865, 'reward_total_composite_std': 1.0387539987277705e-05, 'epoch': 0.08}

 63%|██████▎   | 2063/3300 [2:45:18<1:51:58,  5.43s/it]INFO 04-12 01:22:09 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2064/3300 [2:45:23<1:50:27,  5.36s/it]2026-04-12 01:22:14,549 | INFO | train_grpo_train | metrics_logged mode=train step=2064

                                                       
{'loss': -0.0028, 'grad_norm': 0.6916006207466125, 'learning_rate': 3.748484848484849e-06, 'num_tokens': 4645538.0, 'completions/mean_length': 119.625, 'completions/min_length': 118.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.625, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.996715784072876, 'rewards/meter/std': 0.00013060594210401177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8543277978897095, 'rewards/total_composite/std': 0.00011193905083928257, 'reward': 0.8543277978897095, 'reward_std': 0.00011193443788215518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004635249730199575, 'sampling/sampling_logp_difference/max': 0.4385652542114258, 'sampling/importance_sampling_ratio/min': 0.6449611186981201, 'sampling/importance_sampling_ratio/mean': 1.0017560720443726, 'sampling/importance_sampling_ratio/max': 1.5008399486541748, 'entropy': 0.02747240220196545, 'clip_ratio/low_mean': 0.004228386213071644, 'clip_ratio/low_min': 0.004228386213071644, 'clip_ratio/high_mean': 0.0010416667209938169, 'clip_ratio/high_max': 0.0010416667209938169, 'clip_ratio/region_mean': 0.005270052934065461, 'reward_total_mean': 0.8543277978897095, 'reward_meter_mean': 0.996715784072876, 'reward_meter_std': 0.00013060594210401177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8543277978897095, 'reward_total_composite_std': 0.00011193905083928257, 'epoch': 0.08}

 63%|██████▎   | 2064/3300 [2:45:23<1:50:27,  5.36s/it]INFO 04-12 01:22:14 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2065/3300 [2:45:28<1:46:43,  5.18s/it]2026-04-12 01:22:19,323 | INFO | train_grpo_train | metrics_logged mode=train step=2065

                                                       
{'loss': -0.0076, 'grad_norm': 3.2012574672698975, 'learning_rate': 3.745454545454546e-06, 'num_tokens': 4647558.0, 'completions/mean_length': 74.5, 'completions/min_length': 72.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.5, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9918162822723389, 'rewards/meter/std': 0.009919598698616028, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9918162822723389, 'rewards/total_composite/std': 0.009919598698616028, 'reward': 0.9918162822723389, 'reward_std': 0.009919599629938602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023357173427939415, 'sampling/sampling_logp_difference/max': 1.084247350692749, 'sampling/importance_sampling_ratio/min': 0.33815622329711914, 'sampling/importance_sampling_ratio/mean': 1.004847526550293, 'sampling/importance_sampling_ratio/max': 1.9737117290496826, 'entropy': 0.1982443816959858, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.016551562468521297, 'clip_ratio/high_max': 0.016551562468521297, 'clip_ratio/region_mean': 0.02002378471661359, 'reward_total_mean': 0.9918162822723389, 'reward_meter_mean': 0.9918162822723389, 'reward_meter_std': 0.009919598698616028, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9918162822723389, 'reward_total_composite_std': 0.009919598698616028, 'epoch': 0.08}

 63%|██████▎   | 2065/3300 [2:45:28<1:46:43,  5.18s/it]INFO 04-12 01:22:19 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2066/3300 [2:45:32<1:42:39,  4.99s/it]2026-04-12 01:22:23,870 | INFO | train_grpo_train | metrics_logged mode=train step=2066

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.742424242424243e-06, 'num_tokens': 4649430.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990598559379578, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990598559379578, 'rewards/total_composite/std': 0.0, 'reward': 0.9990598559379578, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002762091171462089, 'sampling/sampling_logp_difference/max': 0.004217715933918953, 'sampling/importance_sampling_ratio/min': 0.997229278087616, 'sampling/importance_sampling_ratio/mean': 1.0002522468566895, 'sampling/importance_sampling_ratio/max': 1.0042266845703125, 'entropy': 0.002455963665852323, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990598559379578, 'reward_meter_mean': 0.9990598559379578, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990598559379578, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 63%|██████▎   | 2066/3300 [2:45:32<1:42:39,  4.99s/it]INFO 04-12 01:22:24 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2067/3300 [2:45:41<2:04:34,  6.06s/it]2026-04-12 01:22:32,437 | INFO | train_grpo_train | metrics_logged mode=train step=2067

                                                       
{'loss': -0.0169, 'grad_norm': 1.5572625398635864, 'learning_rate': 3.73939393939394e-06, 'num_tokens': 4654027.0, 'completions/mean_length': 381.625, 'completions/min_length': 362.0, 'completions/max_length': 398.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 381.625, 'completions/min_terminated_length': 362.0, 'completions/max_terminated_length': 398.0, 'rewards/meter/mean': 0.9989120960235596, 'rewards/meter/std': 0.0005744565860368311, 'rewards/count_adherence/mean': 0.65625, 'rewards/count_adherence/std': 0.033407654613256454, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7585839033126831, 'rewards/repeat_penalty/std': 0.06556817144155502, 'rewards/total_composite/mean': 0.4973776340484619, 'rewards/total_composite/std': 0.04983275756239891, 'reward': 0.4973776340484619, 'reward_std': 0.04983276501297951, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034968629479408264, 'sampling/sampling_logp_difference/max': 14.583293914794922, 'sampling/importance_sampling_ratio/min': 4.6404053932747047e-07, 'sampling/importance_sampling_ratio/mean': 1.0025993585586548, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2074881698936224, 'clip_ratio/low_mean': 0.009894335642457008, 'clip_ratio/low_min': 0.009894335642457008, 'clip_ratio/high_mean': 0.012662693159654737, 'clip_ratio/high_max': 0.012662693159654737, 'clip_ratio/region_mean': 0.022557028802111745, 'reward_total_mean': 0.4973776340484619, 'reward_meter_mean': 0.9989120960235596, 'reward_meter_std': 0.0005744565860368311, 'reward_count_adherence_mean': 0.65625, 'reward_count_adherence_std': 0.033407654613256454, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7585839033126831, 'reward_repeat_penalty_std': 0.06556817144155502, 'reward_total_composite_mean': 0.4973776340484619, 'reward_total_composite_std': 0.04983275756239891, 'epoch': 0.08}

 63%|██████▎   | 2067/3300 [2:45:41<2:04:34,  6.06s/it]INFO 04-12 01:22:32 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2068/3300 [2:45:46<1:56:14,  5.66s/it]2026-04-12 01:22:37,160 | INFO | train_grpo_train | metrics_logged mode=train step=2068

                                                       
{'loss': -0.0084, 'grad_norm': 5.365983009338379, 'learning_rate': 3.736363636363637e-06, 'num_tokens': 4655794.0, 'completions/mean_length': 74.875, 'completions/min_length': 70.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9989867210388184, 'rewards/meter/std': 0.0009042008896358311, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989867210388184, 'rewards/total_composite/std': 0.0009042008896358311, 'reward': 0.9989867210388184, 'reward_std': 0.000904214452020824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027786457911133766, 'sampling/sampling_logp_difference/max': 0.7996888160705566, 'sampling/importance_sampling_ratio/min': 0.44946879148483276, 'sampling/importance_sampling_ratio/mean': 1.0074291229248047, 'sampling/importance_sampling_ratio/max': 1.8772411346435547, 'entropy': 0.22592910565435886, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.01815969729796052, 'clip_ratio/high_max': 0.01815969729796052, 'clip_ratio/region_mean': 0.021731125889346004, 'reward_total_mean': 0.9989867210388184, 'reward_meter_mean': 0.9989867210388184, 'reward_meter_std': 0.0009042008896358311, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989867210388184, 'reward_total_composite_std': 0.0009042008896358311, 'epoch': 0.08}

 63%|██████▎   | 2068/3300 [2:45:46<1:56:14,  5.66s/it]INFO 04-12 01:22:37 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2069/3300 [2:45:50<1:47:04,  5.22s/it]2026-04-12 01:22:41,343 | INFO | train_grpo_train | metrics_logged mode=train step=2069

                                                       
{'loss': 0.008, 'grad_norm': 3.824120283126831, 'learning_rate': 3.7333333333333337e-06, 'num_tokens': 4657320.0, 'completions/mean_length': 29.75, 'completions/min_length': 29.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9928488731384277, 'rewards/meter/std': 7.0027461333666e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9928488731384277, 'rewards/total_composite/std': 7.0027461333666e-05, 'reward': 0.9928488731384277, 'reward_std': 7.002745405770838e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0071687763556838036, 'sampling/sampling_logp_difference/max': 0.6952643394470215, 'sampling/importance_sampling_ratio/min': 0.9426794052124023, 'sampling/importance_sampling_ratio/mean': 1.008081316947937, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03380009834654629, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9928488731384277, 'reward_meter_mean': 0.9928488731384277, 'reward_meter_std': 7.0027461333666e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9928488731384277, 'reward_total_composite_std': 7.0027461333666e-05, 'epoch': 0.08}

 63%|██████▎   | 2069/3300 [2:45:50<1:47:04,  5.22s/it]INFO 04-12 01:22:41 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2070/3300 [2:45:54<1:42:30,  5.00s/it]2026-04-12 01:22:45,831 | INFO | train_grpo_train | metrics_logged mode=train step=2070

                                                       
{'loss': 0.0041, 'grad_norm': 2.178863286972046, 'learning_rate': 3.7303030303030306e-06, 'num_tokens': 4659028.0, 'completions/mean_length': 57.5, 'completions/min_length': 56.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9949300289154053, 'rewards/meter/std': 4.729199281428009e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949300289154053, 'rewards/total_composite/std': 4.729199281428009e-05, 'reward': 0.9949300289154053, 'reward_std': 4.729198190034367e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006639817729592323, 'sampling/sampling_logp_difference/max': 0.6818382740020752, 'sampling/importance_sampling_ratio/min': 0.5056865215301514, 'sampling/importance_sampling_ratio/mean': 1.0030012130737305, 'sampling/importance_sampling_ratio/max': 1.951530933380127, 'entropy': 0.03393339877948165, 'clip_ratio/low_mean': 0.008620689623057842, 'clip_ratio/low_min': 0.008620689623057842, 'clip_ratio/high_mean': 0.0022321429569274187, 'clip_ratio/high_max': 0.0022321429569274187, 'clip_ratio/region_mean': 0.010852832579985261, 'reward_total_mean': 0.9949300289154053, 'reward_meter_mean': 0.9949300289154053, 'reward_meter_std': 4.729199281428009e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949300289154053, 'reward_total_composite_std': 4.729199281428009e-05, 'epoch': 0.08}

 63%|██████▎   | 2070/3300 [2:45:54<1:42:30,  5.00s/it]INFO 04-12 01:22:46 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2071/3300 [2:45:59<1:41:46,  4.97s/it]2026-04-12 01:22:50,726 | INFO | train_grpo_train | metrics_logged mode=train step=2071

                                                       
{'loss': 0.0012, 'grad_norm': 2.837266206741333, 'learning_rate': 3.727272727272728e-06, 'num_tokens': 4661038.0, 'completions/mean_length': 87.25, 'completions/min_length': 86.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.25, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.996016263961792, 'rewards/meter/std': 0.00032200998975895345, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.996016263961792, 'rewards/total_composite/std': 0.00032200998975895345, 'reward': 0.996016263961792, 'reward_std': 0.0003220097569283098, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015046549029648304, 'sampling/sampling_logp_difference/max': 1.0829687118530273, 'sampling/importance_sampling_ratio/min': 0.3910037875175476, 'sampling/importance_sampling_ratio/mean': 1.0012880563735962, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06696707848459482, 'clip_ratio/low_mean': 0.0028572361916303635, 'clip_ratio/low_min': 0.0028572361916303635, 'clip_ratio/high_mean': 0.011511339340358973, 'clip_ratio/high_max': 0.011511339340358973, 'clip_ratio/region_mean': 0.014368575531989336, 'reward_total_mean': 0.996016263961792, 'reward_meter_mean': 0.996016263961792, 'reward_meter_std': 0.00032200998975895345, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.996016263961792, 'reward_total_composite_std': 0.00032200998975895345, 'epoch': 0.08}

 63%|██████▎   | 2071/3300 [2:45:59<1:41:46,  4.97s/it]INFO 04-12 01:22:50 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2072/3300 [2:46:05<1:46:26,  5.20s/it]2026-04-12 01:22:56,471 | INFO | train_grpo_train | metrics_logged mode=train step=2072

                                                       
{'loss': 0.003, 'grad_norm': 3.793785572052002, 'learning_rate': 3.7242424242424246e-06, 'num_tokens': 4663829.0, 'completions/mean_length': 151.875, 'completions/min_length': 149.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.875, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9847034215927124, 'rewards/meter/std': 0.01465265266597271, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8616158962249756, 'rewards/total_composite/std': 0.05130418762564659, 'reward': 0.8616158962249756, 'reward_std': 0.05130419507622719, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02519957348704338, 'sampling/sampling_logp_difference/max': 2.1398165225982666, 'sampling/importance_sampling_ratio/min': 0.11767642945051193, 'sampling/importance_sampling_ratio/mean': 1.002585530281067, 'sampling/importance_sampling_ratio/max': 1.8727295398712158, 'entropy': 0.11779375281184912, 'clip_ratio/low_mean': 0.014854058739729226, 'clip_ratio/low_min': 0.014854058739729226, 'clip_ratio/high_mean': 0.003246753243729472, 'clip_ratio/high_max': 0.003246753243729472, 'clip_ratio/region_mean': 0.018100811983458698, 'reward_total_mean': 0.8616158962249756, 'reward_meter_mean': 0.9847034215927124, 'reward_meter_std': 0.01465265266597271, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8616158962249756, 'reward_total_composite_std': 0.05130418762564659, 'epoch': 0.08}

 63%|██████▎   | 2072/3300 [2:46:05<1:46:26,  5.20s/it]INFO 04-12 01:22:56 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2073/3300 [2:46:10<1:46:17,  5.20s/it]2026-04-12 01:23:01,660 | INFO | train_grpo_train | metrics_logged mode=train step=2073

                                                       
{'loss': -0.0078, 'grad_norm': 3.991828203201294, 'learning_rate': 3.7212121212121215e-06, 'num_tokens': 4665887.0, 'completions/mean_length': 112.25, 'completions/min_length': 108.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.25, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.99208664894104, 'rewards/meter/std': 0.012512140907347202, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99208664894104, 'rewards/total_composite/std': 0.012512140907347202, 'reward': 0.99208664894104, 'reward_std': 0.012512128800153732, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04193767532706261, 'sampling/sampling_logp_difference/max': 1.4444141387939453, 'sampling/importance_sampling_ratio/min': 0.23588424921035767, 'sampling/importance_sampling_ratio/mean': 1.0028138160705566, 'sampling/importance_sampling_ratio/max': 1.9161767959594727, 'entropy': 0.3353299479931593, 'clip_ratio/low_mean': 0.009154040599241853, 'clip_ratio/low_min': 0.009154040599241853, 'clip_ratio/high_mean': 0.023355348501354456, 'clip_ratio/high_max': 0.023355348501354456, 'clip_ratio/region_mean': 0.03250938910059631, 'reward_total_mean': 0.99208664894104, 'reward_meter_mean': 0.99208664894104, 'reward_meter_std': 0.012512140907347202, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99208664894104, 'reward_total_composite_std': 0.012512140907347202, 'epoch': 0.08}

 63%|██████▎   | 2073/3300 [2:46:10<1:46:17,  5.20s/it]INFO 04-12 01:23:01 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2074/3300 [2:46:15<1:41:37,  4.97s/it]2026-04-12 01:23:06,108 | INFO | train_grpo_train | metrics_logged mode=train step=2074

                                                       
{'loss': 0.0016, 'grad_norm': 1.4791640043258667, 'learning_rate': 3.7181818181818187e-06, 'num_tokens': 4667687.0, 'completions/mean_length': 63.0, 'completions/min_length': 63.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9969918131828308, 'rewards/meter/std': 0.00010383435437688604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969918131828308, 'rewards/total_composite/std': 0.00010383435437688604, 'reward': 0.9969918131828308, 'reward_std': 0.00010383195331087336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005412998143583536, 'sampling/sampling_logp_difference/max': 0.561959981918335, 'sampling/importance_sampling_ratio/min': 0.7295204997062683, 'sampling/importance_sampling_ratio/mean': 1.0029563903808594, 'sampling/importance_sampling_ratio/max': 1.7541072368621826, 'entropy': 0.02836236241273582, 'clip_ratio/low_mean': 0.0019841270986944437, 'clip_ratio/low_min': 0.0019841270986944437, 'clip_ratio/high_mean': 0.0019841270986944437, 'clip_ratio/high_max': 0.0019841270986944437, 'clip_ratio/region_mean': 0.003968254197388887, 'reward_total_mean': 0.9969918131828308, 'reward_meter_mean': 0.9969918131828308, 'reward_meter_std': 0.00010383435437688604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969918131828308, 'reward_total_composite_std': 0.00010383435437688604, 'epoch': 0.08}

 63%|██████▎   | 2074/3300 [2:46:15<1:41:37,  4.97s/it]INFO 04-12 01:23:06 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2075/3300 [2:46:19<1:38:44,  4.84s/it]2026-04-12 01:23:10,625 | INFO | train_grpo_train | metrics_logged mode=train step=2075

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.7151515151515156e-06, 'num_tokens': 4669527.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990598559379578, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990598559379578, 'rewards/total_composite/std': 0.0, 'reward': 0.9990598559379578, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002672372793313116, 'sampling/sampling_logp_difference/max': 0.008290700614452362, 'sampling/importance_sampling_ratio/min': 0.9917435646057129, 'sampling/importance_sampling_ratio/mean': 1.0002044439315796, 'sampling/importance_sampling_ratio/max': 1.0042898654937744, 'entropy': 0.0028630376618821174, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990598559379578, 'reward_meter_mean': 0.9990598559379578, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990598559379578, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 63%|██████▎   | 2075/3300 [2:46:19<1:38:44,  4.84s/it]INFO 04-12 01:23:10 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2076/3300 [2:46:24<1:41:12,  4.96s/it]2026-04-12 01:23:15,880 | INFO | train_grpo_train | metrics_logged mode=train step=2076

                                                       
{'loss': -0.0008, 'grad_norm': 2.183760404586792, 'learning_rate': 3.7121212121212124e-06, 'num_tokens': 4671885.0, 'completions/mean_length': 116.75, 'completions/min_length': 116.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.75, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.996731698513031, 'rewards/meter/std': 3.716555511346087e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8721410036087036, 'rewards/total_composite/std': 0.050359878689050674, 'reward': 0.8721410036087036, 'reward_std': 0.05035989359021187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008479885756969452, 'sampling/sampling_logp_difference/max': 0.6767293214797974, 'sampling/importance_sampling_ratio/min': 0.5082767009735107, 'sampling/importance_sampling_ratio/mean': 1.000087022781372, 'sampling/importance_sampling_ratio/max': 1.463352918624878, 'entropy': 0.04457000829279423, 'clip_ratio/low_mean': 0.005342192598618567, 'clip_ratio/low_min': 0.005342192598618567, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.005342192598618567, 'reward_total_mean': 0.8721410036087036, 'reward_meter_mean': 0.996731698513031, 'reward_meter_std': 3.716555511346087e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8721410036087036, 'reward_total_composite_std': 0.050359878689050674, 'epoch': 0.08}

 63%|██████▎   | 2076/3300 [2:46:24<1:41:12,  4.96s/it]INFO 04-12 01:23:16 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2077/3300 [2:46:29<1:39:15,  4.87s/it]2026-04-12 01:23:20,534 | INFO | train_grpo_train | metrics_logged mode=train step=2077

                                                       
{'loss': -0.005, 'grad_norm': 5.309196949005127, 'learning_rate': 3.7090909090909092e-06, 'num_tokens': 4673787.0, 'completions/mean_length': 68.75, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9131183624267578, 'rewards/meter/std': 0.07230029255151749, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9131183624267578, 'rewards/total_composite/std': 0.07230029255151749, 'reward': 0.9131183624267578, 'reward_std': 0.07230029255151749, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007340815383940935, 'sampling/sampling_logp_difference/max': 0.4846305847167969, 'sampling/importance_sampling_ratio/min': 0.6702300310134888, 'sampling/importance_sampling_ratio/mean': 1.0044362545013428, 'sampling/importance_sampling_ratio/max': 1.6235750913619995, 'entropy': 0.05782870203256607, 'clip_ratio/low_mean': 0.0037313431967049837, 'clip_ratio/low_min': 0.0037313431967049837, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0037313431967049837, 'reward_total_mean': 0.9131183624267578, 'reward_meter_mean': 0.9131183624267578, 'reward_meter_std': 0.07230029255151749, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9131183624267578, 'reward_total_composite_std': 0.07230029255151749, 'epoch': 0.08}

 63%|██████▎   | 2077/3300 [2:46:29<1:39:15,  4.87s/it]INFO 04-12 01:23:20 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2078/3300 [2:46:35<1:43:58,  5.11s/it]2026-04-12 01:23:26,188 | INFO | train_grpo_train | metrics_logged mode=train step=2078

                                                       
{'loss': -0.0024, 'grad_norm': 2.607077121734619, 'learning_rate': 3.7060606060606065e-06, 'num_tokens': 4676302.0, 'completions/mean_length': 148.375, 'completions/min_length': 146.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.375, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.9990102648735046, 'rewards/meter/std': 0.00021416397066786885, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.909817099571228, 'rewards/total_composite/std': 0.0739244893193245, 'reward': 0.909817099571228, 'reward_std': 0.0739244818687439, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023411214351654053, 'sampling/sampling_logp_difference/max': 1.31553316116333, 'sampling/importance_sampling_ratio/min': 0.26833122968673706, 'sampling/importance_sampling_ratio/mean': 1.0051114559173584, 'sampling/importance_sampling_ratio/max': 1.6944489479064941, 'entropy': 0.20128737576305866, 'clip_ratio/low_mean': 0.006791699095629156, 'clip_ratio/low_min': 0.006791699095629156, 'clip_ratio/high_mean': 0.00663501548115164, 'clip_ratio/high_max': 0.00663501548115164, 'clip_ratio/region_mean': 0.013426714576780796, 'reward_total_mean': 0.909817099571228, 'reward_meter_mean': 0.9990102648735046, 'reward_meter_std': 0.00021416397066786885, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.909817099571228, 'reward_total_composite_std': 0.0739244893193245, 'epoch': 0.08}

 63%|██████▎   | 2078/3300 [2:46:35<1:43:58,  5.11s/it]INFO 04-12 01:23:26 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2079/3300 [2:46:39<1:41:03,  4.97s/it]2026-04-12 01:23:30,830 | INFO | train_grpo_train | metrics_logged mode=train step=2079

                                                       
{'loss': -0.0002, 'grad_norm': 0.3224610984325409, 'learning_rate': 3.7030303030303033e-06, 'num_tokens': 4678159.0, 'completions/mean_length': 66.125, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9980389475822449, 'rewards/meter/std': 2.6020999939646572e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980389475822449, 'rewards/total_composite/std': 2.6020999939646572e-05, 'reward': 0.9980389475822449, 'reward_std': 2.602726817713119e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007125611882656813, 'sampling/sampling_logp_difference/max': 1.095078468322754, 'sampling/importance_sampling_ratio/min': 0.3345133364200592, 'sampling/importance_sampling_ratio/mean': 0.9994921088218689, 'sampling/importance_sampling_ratio/max': 1.2167119979858398, 'entropy': 0.026706934673711658, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.00562528264708817, 'reward_total_mean': 0.9980389475822449, 'reward_meter_mean': 0.9980389475822449, 'reward_meter_std': 2.6020999939646572e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980389475822449, 'reward_total_composite_std': 2.6020999939646572e-05, 'epoch': 0.08}

 63%|██████▎   | 2079/3300 [2:46:39<1:41:03,  4.97s/it]INFO 04-12 01:23:31 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2080/3300 [2:46:44<1:36:46,  4.76s/it]2026-04-12 01:23:35,103 | INFO | train_grpo_train | metrics_logged mode=train step=2080

                                                       
{'loss': -0.0081, 'grad_norm': 9.717095375061035, 'learning_rate': 3.7e-06, 'num_tokens': 4679892.0, 'completions/mean_length': 39.625, 'completions/min_length': 38.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9970946311950684, 'rewards/meter/std': 0.003818577155470848, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970946311950684, 'rewards/total_composite/std': 0.003818577155470848, 'reward': 0.9970946311950684, 'reward_std': 0.0038185729645192623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058814503252506256, 'sampling/sampling_logp_difference/max': 1.591008186340332, 'sampling/importance_sampling_ratio/min': 0.2037201076745987, 'sampling/importance_sampling_ratio/mean': 0.9983224868774414, 'sampling/importance_sampling_ratio/max': 1.6322083473205566, 'entropy': 0.395177373662591, 'clip_ratio/low_mean': 0.02227564202621579, 'clip_ratio/low_min': 0.02227564202621579, 'clip_ratio/high_mean': 0.03190367738716304, 'clip_ratio/high_max': 0.03190367738716304, 'clip_ratio/region_mean': 0.054179319413378835, 'reward_total_mean': 0.9970946311950684, 'reward_meter_mean': 0.9970946311950684, 'reward_meter_std': 0.003818577155470848, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970946311950684, 'reward_total_composite_std': 0.003818577155470848, 'epoch': 0.08}

 63%|██████▎   | 2080/3300 [2:46:44<1:36:46,  4.76s/it]INFO 04-12 01:23:35 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2081/3300 [2:46:48<1:36:21,  4.74s/it]2026-04-12 01:23:39,812 | INFO | train_grpo_train | metrics_logged mode=train step=2081

                                                       
{'loss': -0.0053, 'grad_norm': 7.125683784484863, 'learning_rate': 3.6969696969696974e-06, 'num_tokens': 4681889.0, 'completions/mean_length': 75.625, 'completions/min_length': 74.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.625, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.8755629062652588, 'rewards/meter/std': 0.3493680953979492, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8755629062652588, 'rewards/total_composite/std': 0.3493680953979492, 'reward': 0.8755629062652588, 'reward_std': 0.3493681252002716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032194141298532486, 'sampling/sampling_logp_difference/max': 0.9960083961486816, 'sampling/importance_sampling_ratio/min': 0.3693508207798004, 'sampling/importance_sampling_ratio/mean': 1.0054106712341309, 'sampling/importance_sampling_ratio/max': 1.5550665855407715, 'entropy': 0.22988875024020672, 'clip_ratio/low_mean': 0.00844594556838274, 'clip_ratio/low_min': 0.00844594556838274, 'clip_ratio/high_mean': 0.009914012742228806, 'clip_ratio/high_max': 0.009914012742228806, 'clip_ratio/region_mean': 0.018359958310611546, 'reward_total_mean': 0.8755629062652588, 'reward_meter_mean': 0.8755629062652588, 'reward_meter_std': 0.3493680953979492, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8755629062652588, 'reward_total_composite_std': 0.3493680953979492, 'epoch': 0.08}

 63%|██████▎   | 2081/3300 [2:46:48<1:36:21,  4.74s/it]INFO 04-12 01:23:40 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2082/3300 [2:46:53<1:34:52,  4.67s/it]2026-04-12 01:23:44,331 | INFO | train_grpo_train | metrics_logged mode=train step=2082

                                                       
{'loss': -0.0013, 'grad_norm': 0.7917746305465698, 'learning_rate': 3.6939393939393942e-06, 'num_tokens': 4683848.0, 'completions/mean_length': 65.875, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9979181289672852, 'rewards/meter/std': 0.00028089084662497044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979181289672852, 'rewards/total_composite/std': 0.00028089084662497044, 'reward': 0.9979181289672852, 'reward_std': 0.00028088997351005673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008248316124081612, 'sampling/sampling_logp_difference/max': 1.082160472869873, 'sampling/importance_sampling_ratio/min': 0.3388626277446747, 'sampling/importance_sampling_ratio/mean': 0.9990794658660889, 'sampling/importance_sampling_ratio/max': 1.159103274345398, 'entropy': 0.026136466301977634, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.001923076924867928, 'clip_ratio/high_max': 0.001923076924867928, 'clip_ratio/region_mean': 0.001923076924867928, 'reward_total_mean': 0.9979181289672852, 'reward_meter_mean': 0.9979181289672852, 'reward_meter_std': 0.00028089084662497044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979181289672852, 'reward_total_composite_std': 0.00028089084662497044, 'epoch': 0.08}

 63%|██████▎   | 2082/3300 [2:46:53<1:34:52,  4.67s/it]INFO 04-12 01:23:44 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2083/3300 [2:46:58<1:40:29,  4.95s/it]2026-04-12 01:23:49,930 | INFO | train_grpo_train | metrics_logged mode=train step=2083

                                                       
{'loss': 0.0021, 'grad_norm': 3.155012845993042, 'learning_rate': 3.690909090909091e-06, 'num_tokens': 4686452.0, 'completions/mean_length': 147.5, 'completions/min_length': 144.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.5, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.9979666471481323, 'rewards/meter/std': 0.0005141000729054213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9801425337791443, 'rewards/total_composite/std': 0.05034194886684418, 'reward': 0.9801425337791443, 'reward_std': 0.050341930240392685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039336614310741425, 'sampling/sampling_logp_difference/max': 1.2754813432693481, 'sampling/importance_sampling_ratio/min': 0.27929648756980896, 'sampling/importance_sampling_ratio/mean': 1.0034173727035522, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2962766755372286, 'clip_ratio/low_mean': 0.0051369862630963326, 'clip_ratio/low_min': 0.0051369862630963326, 'clip_ratio/high_mean': 0.03633964783512056, 'clip_ratio/high_max': 0.03633964783512056, 'clip_ratio/region_mean': 0.04147663409821689, 'reward_total_mean': 0.9801425337791443, 'reward_meter_mean': 0.9979666471481323, 'reward_meter_std': 0.0005141000729054213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9801425337791443, 'reward_total_composite_std': 0.05034194886684418, 'epoch': 0.08}

 63%|██████▎   | 2083/3300 [2:46:58<1:40:29,  4.95s/it]INFO 04-12 01:23:50 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2084/3300 [2:47:04<1:44:53,  5.18s/it]2026-04-12 01:23:55,625 | INFO | train_grpo_train | metrics_logged mode=train step=2084

                                                       
{'loss': 0.0021, 'grad_norm': 5.227519989013672, 'learning_rate': 3.687878787878788e-06, 'num_tokens': 4689035.0, 'completions/mean_length': 151.875, 'completions/min_length': 142.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.875, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.8921954035758972, 'rewards/meter/std': 0.1858879029750824, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.7994784116744995, 'rewards/total_composite/std': 0.18709427118301392, 'reward': 0.7994784116744995, 'reward_std': 0.18709427118301392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031672801822423935, 'sampling/sampling_logp_difference/max': 2.7413718700408936, 'sampling/importance_sampling_ratio/min': 0.06448182463645935, 'sampling/importance_sampling_ratio/mean': 1.0021029710769653, 'sampling/importance_sampling_ratio/max': 1.7797260284423828, 'entropy': 0.14115788787603378, 'clip_ratio/low_mean': 0.004023639601655304, 'clip_ratio/low_min': 0.004023639601655304, 'clip_ratio/high_mean': 0.020977655542083085, 'clip_ratio/high_max': 0.020977655542083085, 'clip_ratio/region_mean': 0.02500129514373839, 'reward_total_mean': 0.7994784116744995, 'reward_meter_mean': 0.8921954035758972, 'reward_meter_std': 0.1858879029750824, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.7994784116744995, 'reward_total_composite_std': 0.18709427118301392, 'epoch': 0.08}

 63%|██████▎   | 2084/3300 [2:47:04<1:44:53,  5.18s/it]INFO 04-12 01:23:55 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2085/3300 [2:47:12<1:59:09,  5.88s/it]2026-04-12 01:24:03,159 | INFO | train_grpo_train | metrics_logged mode=train step=2085

                                                       
{'loss': 0.0076, 'grad_norm': 1.964224934577942, 'learning_rate': 3.684848484848485e-06, 'num_tokens': 4693004.0, 'completions/mean_length': 294.125, 'completions/min_length': 288.0, 'completions/max_length': 310.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 294.125, 'completions/min_terminated_length': 288.0, 'completions/max_terminated_length': 310.0, 'rewards/meter/mean': 0.998891294002533, 'rewards/meter/std': 0.0004529697762336582, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8760416507720947, 'rewards/repeat_penalty/std': 0.023332269862294197, 'rewards/total_composite/mean': 0.8750678896903992, 'rewards/total_composite/std': 0.02319573611021042, 'reward': 0.8750678896903992, 'reward_std': 0.023195745423436165, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025770394131541252, 'sampling/sampling_logp_difference/max': 2.1700761318206787, 'sampling/importance_sampling_ratio/min': 0.1141689196228981, 'sampling/importance_sampling_ratio/mean': 1.0067555904388428, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2046606820076704, 'clip_ratio/low_mean': 0.018270427826792, 'clip_ratio/low_min': 0.018270427826792, 'clip_ratio/high_mean': 0.003448275849223137, 'clip_ratio/high_max': 0.003448275849223137, 'clip_ratio/region_mean': 0.02171870367601514, 'reward_total_mean': 0.8750678896903992, 'reward_meter_mean': 0.998891294002533, 'reward_meter_std': 0.0004529697762336582, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8760416507720947, 'reward_repeat_penalty_std': 0.023332269862294197, 'reward_total_composite_mean': 0.8750678896903992, 'reward_total_composite_std': 0.02319573611021042, 'epoch': 0.08}

 63%|██████▎   | 2085/3300 [2:47:12<1:59:09,  5.88s/it]INFO 04-12 01:24:03 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2086/3300 [2:47:19<2:08:34,  6.35s/it]2026-04-12 01:24:10,614 | INFO | train_grpo_train | metrics_logged mode=train step=2086

                                                       
{'loss': -0.0218, 'grad_norm': 1.5824189186096191, 'learning_rate': 3.681818181818182e-06, 'num_tokens': 4696909.0, 'completions/mean_length': 291.125, 'completions/min_length': 270.0, 'completions/max_length': 306.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 291.125, 'completions/min_terminated_length': 270.0, 'completions/max_terminated_length': 306.0, 'rewards/meter/mean': 0.9986486434936523, 'rewards/meter/std': 0.0003730578755494207, 'rewards/count_adherence/mean': 0.984375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8737351298332214, 'rewards/repeat_penalty/std': 0.06906846165657043, 'rewards/total_composite/mean': 0.8602873086929321, 'rewards/total_composite/std': 0.09172172844409943, 'reward': 0.8602873086929321, 'reward_std': 0.09172171354293823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0357806496322155, 'sampling/sampling_logp_difference/max': 15.902763366699219, 'sampling/importance_sampling_ratio/min': 1.240273945768422e-07, 'sampling/importance_sampling_ratio/mean': 1.003360390663147, 'sampling/importance_sampling_ratio/max': 1.9022753238677979, 'entropy': 0.23322945460677147, 'clip_ratio/low_mean': 0.007962123490869999, 'clip_ratio/low_min': 0.007962123490869999, 'clip_ratio/high_mean': 0.020660461392253637, 'clip_ratio/high_max': 0.020660461392253637, 'clip_ratio/region_mean': 0.028622584883123636, 'reward_total_mean': 0.8602873086929321, 'reward_meter_mean': 0.9986486434936523, 'reward_meter_std': 0.0003730578755494207, 'reward_count_adherence_mean': 0.984375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8737351298332214, 'reward_repeat_penalty_std': 0.06906846165657043, 'reward_total_composite_mean': 0.8602873086929321, 'reward_total_composite_std': 0.09172172844409943, 'epoch': 0.08}

 63%|██████▎   | 2086/3300 [2:47:19<2:08:34,  6.35s/it]INFO 04-12 01:24:10 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2087/3300 [2:47:23<1:56:51,  5.78s/it]2026-04-12 01:24:15,047 | INFO | train_grpo_train | metrics_logged mode=train step=2087

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.678787878787879e-06, 'num_tokens': 4698765.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990598559379578, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990598559379578, 'rewards/total_composite/std': 0.0, 'reward': 0.9990598559379578, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.000278584222542122, 'sampling/sampling_logp_difference/max': 0.006727563217282295, 'sampling/importance_sampling_ratio/min': 0.9977940917015076, 'sampling/importance_sampling_ratio/mean': 1.0002596378326416, 'sampling/importance_sampling_ratio/max': 1.0067503452301025, 'entropy': 0.0027300505607854575, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990598559379578, 'reward_meter_mean': 0.9990598559379578, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990598559379578, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 63%|██████▎   | 2087/3300 [2:47:24<1:56:51,  5.78s/it]INFO 04-12 01:24:15 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2088/3300 [2:47:29<1:52:52,  5.59s/it]2026-04-12 01:24:20,188 | INFO | train_grpo_train | metrics_logged mode=train step=2088

                                                       
{'loss': -0.0178, 'grad_norm': 3.470649480819702, 'learning_rate': 3.6757575757575757e-06, 'num_tokens': 4701006.0, 'completions/mean_length': 112.125, 'completions/min_length': 102.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.125, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9967917203903198, 'rewards/meter/std': 0.002291632117703557, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9967917203903198, 'rewards/total_composite/std': 0.002291632117703557, 'reward': 0.9967917203903198, 'reward_std': 0.002291641663759947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02775193564593792, 'sampling/sampling_logp_difference/max': 1.1847543716430664, 'sampling/importance_sampling_ratio/min': 0.3058212697505951, 'sampling/importance_sampling_ratio/mean': 1.0074383020401, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18240004498511553, 'clip_ratio/low_mean': 0.007895297603681684, 'clip_ratio/low_min': 0.007895297603681684, 'clip_ratio/high_mean': 0.013125053141266108, 'clip_ratio/high_max': 0.013125053141266108, 'clip_ratio/region_mean': 0.02102035074494779, 'reward_total_mean': 0.9967917203903198, 'reward_meter_mean': 0.9967917203903198, 'reward_meter_std': 0.002291632117703557, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9967917203903198, 'reward_total_composite_std': 0.002291632117703557, 'epoch': 0.08}

 63%|██████▎   | 2088/3300 [2:47:29<1:52:52,  5.59s/it]INFO 04-12 01:24:20 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2089/3300 [2:47:33<1:46:01,  5.25s/it]2026-04-12 01:24:24,663 | INFO | train_grpo_train | metrics_logged mode=train step=2089

                                                       
{'loss': -0.0005, 'grad_norm': 1.3091340065002441, 'learning_rate': 3.672727272727273e-06, 'num_tokens': 4702766.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990187883377075, 'rewards/meter/std': 0.00011603027814999223, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990187883377075, 'rewards/total_composite/std': 0.00011603027814999223, 'reward': 0.9990187883377075, 'reward_std': 0.00011603629536693916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0023684005718678236, 'sampling/sampling_logp_difference/max': 1.0964341163635254, 'sampling/importance_sampling_ratio/min': 0.3340601921081543, 'sampling/importance_sampling_ratio/mean': 0.9990017414093018, 'sampling/importance_sampling_ratio/max': 1.0052270889282227, 'entropy': 0.0029864944226574153, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990187883377075, 'reward_meter_mean': 0.9990187883377075, 'reward_meter_std': 0.00011603027814999223, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990187883377075, 'reward_total_composite_std': 0.00011603027814999223, 'epoch': 0.08}

 63%|██████▎   | 2089/3300 [2:47:33<1:46:01,  5.25s/it]INFO 04-12 01:24:24 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2090/3300 [2:47:38<1:42:56,  5.10s/it]2026-04-12 01:24:29,418 | INFO | train_grpo_train | metrics_logged mode=train step=2090

                                                       
{'loss': -0.0218, 'grad_norm': 2.0700643062591553, 'learning_rate': 3.6696969696969697e-06, 'num_tokens': 4704838.0, 'completions/mean_length': 86.0, 'completions/min_length': 80.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9953513741493225, 'rewards/meter/std': 0.0020426749251782894, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953513741493225, 'rewards/total_composite/std': 0.0020426749251782894, 'reward': 0.9953513741493225, 'reward_std': 0.0020426807459443808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013093557208776474, 'sampling/sampling_logp_difference/max': 1.2429170608520508, 'sampling/importance_sampling_ratio/min': 0.288541316986084, 'sampling/importance_sampling_ratio/mean': 1.0013517141342163, 'sampling/importance_sampling_ratio/max': 1.8694008588790894, 'entropy': 0.05126679269596934, 'clip_ratio/low_mean': 0.0015625000232830644, 'clip_ratio/low_min': 0.0015625000232830644, 'clip_ratio/high_mean': 0.0100741779897362, 'clip_ratio/high_max': 0.0100741779897362, 'clip_ratio/region_mean': 0.011636678013019264, 'reward_total_mean': 0.9953513741493225, 'reward_meter_mean': 0.9953513741493225, 'reward_meter_std': 0.0020426749251782894, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953513741493225, 'reward_total_composite_std': 0.0020426749251782894, 'epoch': 0.08}

 63%|██████▎   | 2090/3300 [2:47:38<1:42:56,  5.10s/it]INFO 04-12 01:24:29 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2091/3300 [2:47:43<1:45:14,  5.22s/it]2026-04-12 01:24:34,916 | INFO | train_grpo_train | metrics_logged mode=train step=2091

                                                       
{'loss': 0.0105, 'grad_norm': 2.3087821006774902, 'learning_rate': 3.6666666666666666e-06, 'num_tokens': 4707486.0, 'completions/mean_length': 147.0, 'completions/min_length': 144.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.0, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.9949081540107727, 'rewards/meter/std': 0.004385494627058506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.959259033203125, 'rewards/total_composite/std': 0.0641113817691803, 'reward': 0.959259033203125, 'reward_std': 0.0641113668680191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027278585359454155, 'sampling/sampling_logp_difference/max': 1.1111698150634766, 'sampling/importance_sampling_ratio/min': 0.3291736841201782, 'sampling/importance_sampling_ratio/mean': 1.0080565214157104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19478315208107233, 'clip_ratio/low_mean': 0.0033445945591665804, 'clip_ratio/low_min': 0.0033445945591665804, 'clip_ratio/high_mean': 0.017056229757145047, 'clip_ratio/high_max': 0.017056229757145047, 'clip_ratio/region_mean': 0.020400824316311628, 'reward_total_mean': 0.959259033203125, 'reward_meter_mean': 0.9949081540107727, 'reward_meter_std': 0.004385494627058506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.959259033203125, 'reward_total_composite_std': 0.0641113817691803, 'epoch': 0.08}

 63%|██████▎   | 2091/3300 [2:47:43<1:45:14,  5.22s/it]INFO 04-12 01:24:35 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2092/3300 [2:47:48<1:39:14,  4.93s/it]2026-04-12 01:24:39,163 | INFO | train_grpo_train | metrics_logged mode=train step=2092

                                                       
{'loss': 0.0061, 'grad_norm': 6.707172393798828, 'learning_rate': 3.6636363636363643e-06, 'num_tokens': 4709052.0, 'completions/mean_length': 38.75, 'completions/min_length': 37.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9994734525680542, 'rewards/meter/std': 0.0001241376594407484, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994734525680542, 'rewards/total_composite/std': 0.0001241376594407484, 'reward': 0.9994734525680542, 'reward_std': 0.00012412367505021393, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031195173040032387, 'sampling/sampling_logp_difference/max': 0.9616420269012451, 'sampling/importance_sampling_ratio/min': 0.3822647035121918, 'sampling/importance_sampling_ratio/mean': 0.9975427389144897, 'sampling/importance_sampling_ratio/max': 1.5661958456039429, 'entropy': 0.1498966608196497, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.012833506800234318, 'clip_ratio/high_max': 0.012833506800234318, 'clip_ratio/region_mean': 0.016211885260418057, 'reward_total_mean': 0.9994734525680542, 'reward_meter_mean': 0.9994734525680542, 'reward_meter_std': 0.0001241376594407484, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994734525680542, 'reward_total_composite_std': 0.0001241376594407484, 'epoch': 0.08}

 63%|██████▎   | 2092/3300 [2:47:48<1:39:14,  4.93s/it]INFO 04-12 01:24:39 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2093/3300 [2:47:52<1:34:56,  4.72s/it]2026-04-12 01:24:43,398 | INFO | train_grpo_train | metrics_logged mode=train step=2093

                                                       
{'loss': 0.0237, 'grad_norm': 7.211493015289307, 'learning_rate': 3.660606060606061e-06, 'num_tokens': 4710458.0, 'completions/mean_length': 37.75, 'completions/min_length': 37.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9790675640106201, 'rewards/meter/std': 0.02823478914797306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9790675640106201, 'rewards/total_composite/std': 0.02823478914797306, 'reward': 0.9790675640106201, 'reward_std': 0.02823478914797306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014203748665750027, 'sampling/sampling_logp_difference/max': 0.29107916355133057, 'sampling/importance_sampling_ratio/min': 0.7474565505981445, 'sampling/importance_sampling_ratio/mean': 1.0054656267166138, 'sampling/importance_sampling_ratio/max': 1.3320231437683105, 'entropy': 0.12450700253248215, 'clip_ratio/low_mean': 0.009615384740754962, 'clip_ratio/low_min': 0.009615384740754962, 'clip_ratio/high_mean': 0.006756756920367479, 'clip_ratio/high_max': 0.006756756920367479, 'clip_ratio/region_mean': 0.01637214166112244, 'reward_total_mean': 0.9790675640106201, 'reward_meter_mean': 0.9790675640106201, 'reward_meter_std': 0.02823478914797306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9790675640106201, 'reward_total_composite_std': 0.02823478914797306, 'epoch': 0.08}

 63%|██████▎   | 2093/3300 [2:47:52<1:34:56,  4.72s/it]INFO 04-12 01:24:43 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2094/3300 [2:47:56<1:33:29,  4.65s/it]2026-04-12 01:24:47,882 | INFO | train_grpo_train | metrics_logged mode=train step=2094

                                                       
{'loss': 0.0017, 'grad_norm': 0.3633219003677368, 'learning_rate': 3.657575757575758e-06, 'num_tokens': 4712186.0, 'completions/mean_length': 69.0, 'completions/min_length': 69.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.938814103603363, 'rewards/meter/std': 0.0002475241490174085, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.938814103603363, 'rewards/total_composite/std': 0.0002475241490174085, 'reward': 0.938814103603363, 'reward_std': 0.0002475333458278328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0060532353818416595, 'sampling/sampling_logp_difference/max': 0.7302188873291016, 'sampling/importance_sampling_ratio/min': 0.4818035662174225, 'sampling/importance_sampling_ratio/mean': 0.999586820602417, 'sampling/importance_sampling_ratio/max': 1.0647722482681274, 'entropy': 0.04171428643167019, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0036231884732842445, 'clip_ratio/high_max': 0.0036231884732842445, 'clip_ratio/region_mean': 0.0036231884732842445, 'reward_total_mean': 0.938814103603363, 'reward_meter_mean': 0.938814103603363, 'reward_meter_std': 0.0002475241490174085, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.938814103603363, 'reward_total_composite_std': 0.0002475241490174085, 'epoch': 0.08}

 63%|██████▎   | 2094/3300 [2:47:56<1:33:29,  4.65s/it]INFO 04-12 01:24:48 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2095/3300 [2:48:01<1:32:38,  4.61s/it]2026-04-12 01:24:52,411 | INFO | train_grpo_train | metrics_logged mode=train step=2095

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.654545454545455e-06, 'num_tokens': 4714042.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990598559379578, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990598559379578, 'rewards/total_composite/std': 0.0, 'reward': 0.9990598559379578, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005561705911532044, 'sampling/sampling_logp_difference/max': 0.016185477375984192, 'sampling/importance_sampling_ratio/min': 0.9876941442489624, 'sampling/importance_sampling_ratio/mean': 1.0003678798675537, 'sampling/importance_sampling_ratio/max': 1.0163171291351318, 'entropy': 0.00596828549169004, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9990598559379578, 'reward_meter_mean': 0.9990598559379578, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990598559379578, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 63%|██████▎   | 2095/3300 [2:48:01<1:32:38,  4.61s/it]INFO 04-12 01:24:52 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2096/3300 [2:48:05<1:32:30,  4.61s/it]2026-04-12 01:24:57,011 | INFO | train_grpo_train | metrics_logged mode=train step=2096

                                                       
{'loss': 0.006, 'grad_norm': 3.099059581756592, 'learning_rate': 3.651515151515152e-06, 'num_tokens': 4716098.0, 'completions/mean_length': 69.0, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9360407590866089, 'rewards/meter/std': 0.005241453181952238, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9360407590866089, 'rewards/total_composite/std': 0.005241453181952238, 'reward': 0.9360407590866089, 'reward_std': 0.0052414704114198685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011832116171717644, 'sampling/sampling_logp_difference/max': 0.7093310356140137, 'sampling/importance_sampling_ratio/min': 0.49197322130203247, 'sampling/importance_sampling_ratio/mean': 1.001249074935913, 'sampling/importance_sampling_ratio/max': 1.8257942199707031, 'entropy': 0.07018918683752418, 'clip_ratio/low_mean': 0.0018115942366421223, 'clip_ratio/low_min': 0.0018115942366421223, 'clip_ratio/high_mean': 0.0036231884732842445, 'clip_ratio/high_max': 0.0036231884732842445, 'clip_ratio/region_mean': 0.005434782709926367, 'reward_total_mean': 0.9360407590866089, 'reward_meter_mean': 0.9360407590866089, 'reward_meter_std': 0.005241453181952238, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9360407590866089, 'reward_total_composite_std': 0.005241453181952238, 'epoch': 0.08}

 64%|██████▎   | 2096/3300 [2:48:06<1:32:30,  4.61s/it]INFO 04-12 01:24:57 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2097/3300 [2:48:09<1:28:59,  4.44s/it]2026-04-12 01:25:01,048 | INFO | train_grpo_train | metrics_logged mode=train step=2097

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.648484848484849e-06, 'num_tokens': 4717538.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9942078590393066, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9942078590393066, 'rewards/total_composite/std': 0.0, 'reward': 0.9942078590393066, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0031366986222565174, 'sampling/sampling_logp_difference/max': 0.15313977003097534, 'sampling/importance_sampling_ratio/min': 0.8580098152160645, 'sampling/importance_sampling_ratio/mean': 1.0009886026382446, 'sampling/importance_sampling_ratio/max': 1.0437800884246826, 'entropy': 0.030742917326278985, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9942078590393066, 'reward_meter_mean': 0.9942078590393066, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9942078590393066, 'reward_total_composite_std': 0.0, 'epoch': 0.08}

 64%|██████▎   | 2097/3300 [2:48:10<1:28:59,  4.44s/it]INFO 04-12 01:25:01 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2098/3300 [2:48:14<1:30:42,  4.53s/it]2026-04-12 01:25:05,784 | INFO | train_grpo_train | metrics_logged mode=train step=2098

                                                       
{'loss': 0.0272, 'grad_norm': 2.8510286808013916, 'learning_rate': 3.645454545454546e-06, 'num_tokens': 4719239.0, 'completions/mean_length': 75.625, 'completions/min_length': 71.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.625, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9827600717544556, 'rewards/meter/std': 0.0294841006398201, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9827600717544556, 'rewards/total_composite/std': 0.0294841006398201, 'reward': 0.9827600717544556, 'reward_std': 0.029484104365110397, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033739130944013596, 'sampling/sampling_logp_difference/max': 1.151628017425537, 'sampling/importance_sampling_ratio/min': 0.3161216974258423, 'sampling/importance_sampling_ratio/mean': 0.9956676363945007, 'sampling/importance_sampling_ratio/max': 1.6294512748718262, 'entropy': 0.16028152592480183, 'clip_ratio/low_mean': 0.0015625000232830644, 'clip_ratio/low_min': 0.0015625000232830644, 'clip_ratio/high_mean': 0.028136852895841002, 'clip_ratio/high_max': 0.028136852895841002, 'clip_ratio/region_mean': 0.029699352919124067, 'reward_total_mean': 0.9827600717544556, 'reward_meter_mean': 0.9827600717544556, 'reward_meter_std': 0.0294841006398201, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9827600717544556, 'reward_total_composite_std': 0.0294841006398201, 'epoch': 0.08}

 64%|██████▎   | 2098/3300 [2:48:14<1:30:42,  4.53s/it]INFO 04-12 01:25:06 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2099/3300 [2:48:19<1:33:34,  4.67s/it]2026-04-12 01:25:10,803 | INFO | train_grpo_train | metrics_logged mode=train step=2099

                                                       
{'loss': -0.0241, 'grad_norm': 4.330075740814209, 'learning_rate': 3.642424242424243e-06, 'num_tokens': 4721395.0, 'completions/mean_length': 105.5, 'completions/min_length': 96.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.5, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.8654639720916748, 'rewards/meter/std': 0.23051653802394867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8491353988647461, 'rewards/total_composite/std': 0.25138652324676514, 'reward': 0.8491353988647461, 'reward_std': 0.25138649344444275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03153692185878754, 'sampling/sampling_logp_difference/max': 1.3270950317382812, 'sampling/importance_sampling_ratio/min': 0.3179101049900055, 'sampling/importance_sampling_ratio/mean': 1.007623553276062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29543587751686573, 'clip_ratio/low_mean': 0.008760618977248669, 'clip_ratio/low_min': 0.008760618977248669, 'clip_ratio/high_mean': 0.016345139825716615, 'clip_ratio/high_max': 0.016345139825716615, 'clip_ratio/region_mean': 0.025105758802965283, 'reward_total_mean': 0.8491353988647461, 'reward_meter_mean': 0.8654639720916748, 'reward_meter_std': 0.23051653802394867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8491353988647461, 'reward_total_composite_std': 0.25138652324676514, 'epoch': 0.08}

 64%|██████▎   | 2099/3300 [2:48:19<1:33:34,  4.67s/it]INFO 04-12 01:25:11 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2100/3300 [2:48:24<1:33:59,  4.70s/it]2026-04-12 01:25:15,571 | INFO | train_grpo_train | metrics_logged mode=train step=2100

                                                       
{'loss': 0.0052, 'grad_norm': 7.931367874145508, 'learning_rate': 3.6393939393939398e-06, 'num_tokens': 4723535.0, 'completions/mean_length': 86.5, 'completions/min_length': 83.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.5, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9882360100746155, 'rewards/meter/std': 0.022600755095481873, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9882360100746155, 'rewards/total_composite/std': 0.022600755095481873, 'reward': 0.9882360100746155, 'reward_std': 0.022600755095481873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022525055333971977, 'sampling/sampling_logp_difference/max': 1.4979562759399414, 'sampling/importance_sampling_ratio/min': 0.22358666360378265, 'sampling/importance_sampling_ratio/mean': 1.0013149976730347, 'sampling/importance_sampling_ratio/max': 1.8274431228637695, 'entropy': 0.20394995296373963, 'clip_ratio/low_mean': 0.0060240961611270905, 'clip_ratio/low_min': 0.0060240961611270905, 'clip_ratio/high_mean': 0.008620689623057842, 'clip_ratio/high_max': 0.008620689623057842, 'clip_ratio/region_mean': 0.014644785784184933, 'reward_total_mean': 0.9882360100746155, 'reward_meter_mean': 0.9882360100746155, 'reward_meter_std': 0.022600755095481873, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9882360100746155, 'reward_total_composite_std': 0.022600755095481873, 'epoch': 0.08}

 64%|██████▎   | 2100/3300 [2:48:24<1:33:59,  4.70s/it]INFO 04-12 01:25:15 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:31,  2.86s/it]

 23%|██▎       | 3/13 [00:09<00:32,  3.24s/it]

 31%|███       | 4/13 [00:14<00:35,  3.95s/it]

 38%|███▊      | 5/13 [00:19<00:34,  4.37s/it]

 46%|████▌     | 6/13 [00:25<00:33,  4.74s/it]

 54%|█████▍    | 7/13 [00:30<00:28,  4.75s/it]

 62%|██████▏   | 8/13 [00:35<00:25,  5.09s/it]

 69%|██████▉   | 9/13 [00:41<00:21,  5.38s/it]

 77%|███████▋  | 10/13 [00:46<00:15,  5.10s/it]

 85%|████████▍ | 11/13 [00:51<00:10,  5.18s/it]

 92%|█████████▏| 12/13 [00:57<00:05,  5.29s/it]

100%|██████████| 13/13 [01:02<00:00,  5.20s/it]2026-04-12 01:26:24,192 | INFO | train_grpo_train | metrics_logged mode=eval step=2100

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 68.5013, 'eval_samples_per_second': 1.518, 'eval_steps_per_second': 0.19, 'eval_num_tokens': 4723535.0, 'eval_completions/mean_length': 201.65384615384616, 'eval_completions/min_length': 61.92307692307692, 'eval_completions/max_length': 356.53846153846155, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 201.65384615384616, 'eval_completions/min_terminated_length': 61.92307692307692, 'eval_completions/max_terminated_length': 356.53846153846155, 'eval_rewards/meter/mean': 0.7131731968659621, 'eval_rewards/meter/std': 0.4128540616769057, 'eval_rewards/count_adherence/mean': 0.9138026375036973, 'eval_rewards/count_adherence/std': 0.12064289301633835, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8164694079985986, 'eval_rewards/repeat_penalty/std': 0.1486260168827497, 'eval_rewards/total_composite/mean': 0.542255353469115, 'eval_rewards/total_composite/std': 0.3623279837461618, 'eval_reward': 0.542255353469115, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.013091834047092842, 'eval_sampling/sampling_logp_difference/max': 0.9667405440257146, 'eval_sampling/importance_sampling_ratio/min': 0.38750078357183015, 'eval_sampling/importance_sampling_ratio/mean': 1.0036665751383855, 'eval_sampling/importance_sampling_ratio/max': 1.501889733167795, 'eval_entropy': 0.12711333655394041, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.542255353469115, 'eval_reward_meter_mean': 0.7131731968659621, 'eval_reward_meter_std': 0.4128540616769057, 'eval_reward_count_adherence_mean': 0.9138026375036973, 'eval_reward_count_adherence_std': 0.12064289301633835, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8164694079985986, 'eval_reward_repeat_penalty_std': 0.1486260168827497, 'eval_reward_total_composite_mean': 0.542255353469115, 'eval_reward_total_composite_std': 0.3623279837461618, 'epoch': 0.08}

 64%|██████▎   | 2100/3300 [2:49:33<1:33:59,  4.70s/it]

100%|██████████| 13/13 [01:02<00:00,  5.20s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 64%|██████▎   | 2101/3300 [2:49:42<8:54:31, 26.75s/it]2026-04-12 01:26:33,771 | INFO | train_grpo_train | metrics_logged mode=train step=2101

                                                       
{'loss': -0.0478, 'grad_norm': 1.2342872619628906, 'learning_rate': 3.6363636363636366e-06, 'num_tokens': 4726322.0, 'completions/mean_length': 176.375, 'completions/min_length': 160.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 176.375, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.9766891598701477, 'rewards/meter/std': 0.022879159078001976, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.761904776096344, 'rewards/repeat_penalty/std': 0.07493293285369873, 'rewards/total_composite/mean': 0.6860010623931885, 'rewards/total_composite/std': 0.08363299816846848, 'reward': 0.6860010623931885, 'reward_std': 0.08363301306962967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01796703413128853, 'sampling/sampling_logp_difference/max': 1.0117740631103516, 'sampling/importance_sampling_ratio/min': 0.36357343196868896, 'sampling/importance_sampling_ratio/mean': 1.0021946430206299, 'sampling/importance_sampling_ratio/max': 1.9307383298873901, 'entropy': 0.11573350336402655, 'clip_ratio/low_mean': 0.007700952875893563, 'clip_ratio/low_min': 0.007700952875893563, 'clip_ratio/high_mean': 0.00595723238075152, 'clip_ratio/high_max': 0.00595723238075152, 'clip_ratio/region_mean': 0.013658185256645083, 'reward_total_mean': 0.6860010623931885, 'reward_meter_mean': 0.9766891598701477, 'reward_meter_std': 0.022879159078001976, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.761904776096344, 'reward_repeat_penalty_std': 0.07493293285369873, 'reward_total_composite_mean': 0.6860010623931885, 'reward_total_composite_std': 0.08363299816846848, 'epoch': 0.08}

 64%|██████▎   | 2101/3300 [2:49:42<8:54:31, 26.75s/it]INFO 04-12 01:26:34 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2102/3300 [2:49:46<6:39:32, 20.01s/it]2026-04-12 01:26:38,054 | INFO | train_grpo_train | metrics_logged mode=train step=2102

                                                       
{'loss': 0.0086, 'grad_norm': 14.514649391174316, 'learning_rate': 3.633333333333334e-06, 'num_tokens': 4727945.0, 'completions/mean_length': 28.875, 'completions/min_length': 28.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9928563833236694, 'rewards/meter/std': 0.0002995376707985997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9928563833236694, 'rewards/total_composite/std': 0.0002995376707985997, 'reward': 0.9928563833236694, 'reward_std': 0.0002995316463056952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0030352873727679253, 'sampling/sampling_logp_difference/max': 0.14670616388320923, 'sampling/importance_sampling_ratio/min': 0.8847241401672363, 'sampling/importance_sampling_ratio/mean': 1.0009772777557373, 'sampling/importance_sampling_ratio/max': 1.1580135822296143, 'entropy': 0.03053366276435554, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9928563833236694, 'reward_meter_mean': 0.9928563833236694, 'reward_meter_std': 0.0002995376707985997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9928563833236694, 'reward_total_composite_std': 0.0002995376707985997, 'epoch': 0.08}

 64%|██████▎   | 2102/3300 [2:49:47<6:39:32, 20.01s/it]INFO 04-12 01:26:38 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2103/3300 [2:49:54<5:23:28, 16.21s/it]2026-04-12 01:26:45,411 | INFO | train_grpo_train | metrics_logged mode=train step=2103

                                                       
{'loss': -0.0264, 'grad_norm': 1.2050065994262695, 'learning_rate': 3.6303030303030307e-06, 'num_tokens': 4731675.0, 'completions/mean_length': 279.25, 'completions/min_length': 257.0, 'completions/max_length': 286.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 279.25, 'completions/min_terminated_length': 257.0, 'completions/max_terminated_length': 286.0, 'rewards/meter/mean': 0.9298574924468994, 'rewards/meter/std': 0.1897430717945099, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8583333492279053, 'rewards/repeat_penalty/std': 0.19002924859523773, 'rewards/total_composite/mean': 0.8288908004760742, 'rewards/total_composite/std': 0.26403236389160156, 'reward': 0.8288908004760742, 'reward_std': 0.2640323340892792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023323198780417442, 'sampling/sampling_logp_difference/max': 1.9472055435180664, 'sampling/importance_sampling_ratio/min': 0.14267219603061676, 'sampling/importance_sampling_ratio/mean': 1.0039503574371338, 'sampling/importance_sampling_ratio/max': 1.7007434368133545, 'entropy': 0.272721977904439, 'clip_ratio/low_mean': 0.00048638132284395397, 'clip_ratio/low_min': 0.00048638132284395397, 'clip_ratio/high_mean': 0.02126871724613011, 'clip_ratio/high_max': 0.02126871724613011, 'clip_ratio/region_mean': 0.021755098568974063, 'reward_total_mean': 0.8288908004760742, 'reward_meter_mean': 0.9298574924468994, 'reward_meter_std': 0.1897430717945099, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8583333492279053, 'reward_repeat_penalty_std': 0.19002924859523773, 'reward_total_composite_mean': 0.8288908004760742, 'reward_total_composite_std': 0.26403236389160156, 'epoch': 0.08}

 64%|██████▎   | 2103/3300 [2:49:54<5:23:28, 16.21s/it]INFO 04-12 01:26:45 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2104/3300 [2:50:02<4:32:39, 13.68s/it]2026-04-12 01:26:53,168 | INFO | train_grpo_train | metrics_logged mode=train step=2104

                                                       
{'loss': -0.0122, 'grad_norm': 1.8657711744308472, 'learning_rate': 3.6272727272727275e-06, 'num_tokens': 4735591.0, 'completions/mean_length': 296.5, 'completions/min_length': 284.0, 'completions/max_length': 315.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 296.5, 'completions/min_terminated_length': 284.0, 'completions/max_terminated_length': 315.0, 'rewards/meter/mean': 0.9669773578643799, 'rewards/meter/std': 0.07153221219778061, 'rewards/count_adherence/mean': 0.984375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7205882668495178, 'rewards/repeat_penalty/std': 0.061355408281087875, 'rewards/total_composite/mean': 0.6847283840179443, 'rewards/total_composite/std': 0.07457956671714783, 'reward': 0.6847283840179443, 'reward_std': 0.07457955926656723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021201852709054947, 'sampling/sampling_logp_difference/max': 1.1567082405090332, 'sampling/importance_sampling_ratio/min': 0.3145197927951813, 'sampling/importance_sampling_ratio/mean': 1.0043342113494873, 'sampling/importance_sampling_ratio/max': 1.9632225036621094, 'entropy': 0.17900856165215373, 'clip_ratio/low_mean': 0.004951660754159093, 'clip_ratio/low_min': 0.004951660754159093, 'clip_ratio/high_mean': 0.009039964294061065, 'clip_ratio/high_max': 0.009039964294061065, 'clip_ratio/region_mean': 0.013991625048220158, 'reward_total_mean': 0.6847283840179443, 'reward_meter_mean': 0.9669773578643799, 'reward_meter_std': 0.07153221219778061, 'reward_count_adherence_mean': 0.984375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7205882668495178, 'reward_repeat_penalty_std': 0.061355408281087875, 'reward_total_composite_mean': 0.6847283840179443, 'reward_total_composite_std': 0.07457956671714783, 'epoch': 0.08}

 64%|██████▍   | 2104/3300 [2:50:02<4:32:39, 13.68s/it]INFO 04-12 01:26:53 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2105/3300 [2:50:07<3:44:44, 11.28s/it]2026-04-12 01:26:58,863 | INFO | train_grpo_train | metrics_logged mode=train step=2105

                                                       
{'loss': -0.0002, 'grad_norm': 4.257723331451416, 'learning_rate': 3.6242424242424248e-06, 'num_tokens': 4738461.0, 'completions/mean_length': 164.75, 'completions/min_length': 161.0, 'completions/max_length': 166.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 164.75, 'completions/min_terminated_length': 161.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.9989630579948425, 'rewards/meter/std': 0.00037122031790204346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7769712209701538, 'rewards/total_composite/std': 0.00028871477115899324, 'reward': 0.7769712209701538, 'reward_std': 0.0002887141308747232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016738269478082657, 'sampling/sampling_logp_difference/max': 2.3132505416870117, 'sampling/importance_sampling_ratio/min': 0.09893912822008133, 'sampling/importance_sampling_ratio/mean': 1.0012632608413696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0402666125446558, 'clip_ratio/low_mean': 0.004629980307072401, 'clip_ratio/low_min': 0.004629980307072401, 'clip_ratio/high_mean': 0.009068090235814452, 'clip_ratio/high_max': 0.009068090235814452, 'clip_ratio/region_mean': 0.013698070542886853, 'reward_total_mean': 0.7769712209701538, 'reward_meter_mean': 0.9989630579948425, 'reward_meter_std': 0.00037122031790204346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7769712209701538, 'reward_total_composite_std': 0.00028871477115899324, 'epoch': 0.08}

 64%|██████▍   | 2105/3300 [2:50:07<3:44:44, 11.28s/it]INFO 04-12 01:26:59 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2106/3300 [2:50:15<3:26:00, 10.35s/it]2026-04-12 01:27:07,041 | INFO | train_grpo_train | metrics_logged mode=train step=2106

                                                       
{'loss': -0.0003, 'grad_norm': 1.7383067607879639, 'learning_rate': 3.6212121212121216e-06, 'num_tokens': 4742393.0, 'completions/mean_length': 279.5, 'completions/min_length': 276.0, 'completions/max_length': 282.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 279.5, 'completions/min_terminated_length': 276.0, 'completions/max_terminated_length': 282.0, 'rewards/meter/mean': 0.9945741891860962, 'rewards/meter/std': 0.0010695032542571425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7578125, 'rewards/repeat_penalty/std': 0.07790146768093109, 'rewards/total_composite/mean': 0.7536386847496033, 'rewards/total_composite/std': 0.07672014087438583, 'reward': 0.7536386847496033, 'reward_std': 0.07672013342380524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010709014721214771, 'sampling/sampling_logp_difference/max': 1.3207725286483765, 'sampling/importance_sampling_ratio/min': 0.2669290006160736, 'sampling/importance_sampling_ratio/mean': 0.999245285987854, 'sampling/importance_sampling_ratio/max': 1.8650121688842773, 'entropy': 0.045266281347721815, 'clip_ratio/low_mean': 0.00402287530596368, 'clip_ratio/low_min': 0.00402287530596368, 'clip_ratio/high_mean': 0.00443893379997462, 'clip_ratio/high_max': 0.00443893379997462, 'clip_ratio/region_mean': 0.0084618091059383, 'reward_total_mean': 0.7536386847496033, 'reward_meter_mean': 0.9945741891860962, 'reward_meter_std': 0.0010695032542571425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7578125, 'reward_repeat_penalty_std': 0.07790146768093109, 'reward_total_composite_mean': 0.7536386847496033, 'reward_total_composite_std': 0.07672014087438583, 'epoch': 0.08}

 64%|██████▍   | 2106/3300 [2:50:16<3:26:00, 10.35s/it]INFO 04-12 01:27:07 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2107/3300 [2:50:21<2:56:16,  8.87s/it]2026-04-12 01:27:12,447 | INFO | train_grpo_train | metrics_logged mode=train step=2107

                                                       
{'loss': -0.0, 'grad_norm': 0.0878043994307518, 'learning_rate': 3.6181818181818184e-06, 'num_tokens': 4744793.0, 'completions/mean_length': 118.0, 'completions/min_length': 118.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.0, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9968082904815674, 'rewards/meter/std': 4.3226536945439875e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8544071912765503, 'rewards/total_composite/std': 3.7037829315522686e-05, 'reward': 0.8544071912765503, 'reward_std': 3.7032186810392886e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005184296518564224, 'sampling/sampling_logp_difference/max': 1.0821123123168945, 'sampling/importance_sampling_ratio/min': 0.33887895941734314, 'sampling/importance_sampling_ratio/mean': 0.9996362924575806, 'sampling/importance_sampling_ratio/max': 1.1646240949630737, 'entropy': 0.026277633383870125, 'clip_ratio/low_mean': 0.0021186440717428923, 'clip_ratio/low_min': 0.0021186440717428923, 'clip_ratio/high_mean': 0.0010593220358714461, 'clip_ratio/high_max': 0.0010593220358714461, 'clip_ratio/region_mean': 0.0031779661076143384, 'reward_total_mean': 0.8544071912765503, 'reward_meter_mean': 0.9968082904815674, 'reward_meter_std': 4.3226536945439875e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8544071912765503, 'reward_total_composite_std': 3.7037829315522686e-05, 'epoch': 0.08}

 64%|██████▍   | 2107/3300 [2:50:21<2:56:16,  8.87s/it]INFO 04-12 01:27:12 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2108/3300 [2:50:26<2:35:00,  7.80s/it]2026-04-12 01:27:17,764 | INFO | train_grpo_train | metrics_logged mode=train step=2108

                                                       
{'loss': -0.0069, 'grad_norm': 1.9401229619979858, 'learning_rate': 3.6151515151515153e-06, 'num_tokens': 4747165.0, 'completions/mean_length': 120.5, 'completions/min_length': 120.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.5, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9796810746192932, 'rewards/meter/std': 0.003423919202759862, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8397266268730164, 'rewards/total_composite/std': 0.002934773452579975, 'reward': 0.8397266268730164, 'reward_std': 0.00293477950617671, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004073810297995806, 'sampling/sampling_logp_difference/max': 0.2728769779205322, 'sampling/importance_sampling_ratio/min': 0.761186420917511, 'sampling/importance_sampling_ratio/mean': 1.0006637573242188, 'sampling/importance_sampling_ratio/max': 1.215042233467102, 'entropy': 0.029683739179745317, 'clip_ratio/low_mean': 0.004166666883975267, 'clip_ratio/low_min': 0.004166666883975267, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004166666883975267, 'reward_total_mean': 0.8397266268730164, 'reward_meter_mean': 0.9796810746192932, 'reward_meter_std': 0.003423919202759862, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8397266268730164, 'reward_total_composite_std': 0.002934773452579975, 'epoch': 0.08}

 64%|██████▍   | 2108/3300 [2:50:26<2:35:00,  7.80s/it]INFO 04-12 01:27:18 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2109/3300 [2:50:31<2:15:44,  6.84s/it]2026-04-12 01:27:22,351 | INFO | train_grpo_train | metrics_logged mode=train step=2109

                                                       
{'loss': 0.0148, 'grad_norm': 14.469276428222656, 'learning_rate': 3.6121212121212125e-06, 'num_tokens': 4748903.0, 'completions/mean_length': 59.25, 'completions/min_length': 57.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.675230860710144, 'rewards/meter/std': 0.3926205635070801, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.675230860710144, 'rewards/total_composite/std': 0.3926205635070801, 'reward': 0.675230860710144, 'reward_std': 0.3926205635070801, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032343991100788116, 'sampling/sampling_logp_difference/max': 1.3949341773986816, 'sampling/importance_sampling_ratio/min': 0.35194310545921326, 'sampling/importance_sampling_ratio/mean': 1.008945107460022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24129995983093977, 'clip_ratio/low_mean': 0.021334989927709103, 'clip_ratio/low_min': 0.021334989927709103, 'clip_ratio/high_mean': 0.006250000325962901, 'clip_ratio/high_max': 0.006250000325962901, 'clip_ratio/region_mean': 0.027584990253672004, 'reward_total_mean': 0.675230860710144, 'reward_meter_mean': 0.675230860710144, 'reward_meter_std': 0.3926205635070801, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.675230860710144, 'reward_total_composite_std': 0.3926205635070801, 'epoch': 0.08}

 64%|██████▍   | 2109/3300 [2:50:31<2:15:44,  6.84s/it]INFO 04-12 01:27:22 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2110/3300 [2:50:35<2:01:50,  6.14s/it]2026-04-12 01:27:26,872 | INFO | train_grpo_train | metrics_logged mode=train step=2110

                                                       
{'loss': -0.0044, 'grad_norm': 0.8487328290939331, 'learning_rate': 3.6090909090909093e-06, 'num_tokens': 4750550.0, 'completions/mean_length': 55.875, 'completions/min_length': 55.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9948891401290894, 'rewards/meter/std': 0.0002349100832361728, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948891401290894, 'rewards/total_composite/std': 0.0002349100832361728, 'reward': 0.9948891401290894, 'reward_std': 0.00023490135208703578, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005522139370441437, 'sampling/sampling_logp_difference/max': 0.4381282329559326, 'sampling/importance_sampling_ratio/min': 0.7602133750915527, 'sampling/importance_sampling_ratio/mean': 1.0012332201004028, 'sampling/importance_sampling_ratio/max': 1.5498037338256836, 'entropy': 0.031467124819755554, 'clip_ratio/low_mean': 0.0022727272007614374, 'clip_ratio/low_min': 0.0022727272007614374, 'clip_ratio/high_mean': 0.0022321429569274187, 'clip_ratio/high_max': 0.0022321429569274187, 'clip_ratio/region_mean': 0.004504870157688856, 'reward_total_mean': 0.9948891401290894, 'reward_meter_mean': 0.9948891401290894, 'reward_meter_std': 0.0002349100832361728, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948891401290894, 'reward_total_composite_std': 0.0002349100832361728, 'epoch': 0.08}

 64%|██████▍   | 2110/3300 [2:50:35<2:01:50,  6.14s/it]INFO 04-12 01:27:27 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2111/3300 [2:50:40<1:53:33,  5.73s/it]2026-04-12 01:27:31,642 | INFO | train_grpo_train | metrics_logged mode=train step=2111

                                                       
{'loss': 0.0071, 'grad_norm': 4.360041618347168, 'learning_rate': 3.606060606060606e-06, 'num_tokens': 4752319.0, 'completions/mean_length': 72.125, 'completions/min_length': 71.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9952445030212402, 'rewards/meter/std': 0.0019698443356901407, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9952445030212402, 'rewards/total_composite/std': 0.0019698443356901407, 'reward': 0.9952445030212402, 'reward_std': 0.00196986086666584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039312709122896194, 'sampling/sampling_logp_difference/max': 1.4191043376922607, 'sampling/importance_sampling_ratio/min': 0.24193060398101807, 'sampling/importance_sampling_ratio/mean': 1.0026034116744995, 'sampling/importance_sampling_ratio/max': 1.8677667379379272, 'entropy': 0.2009260579943657, 'clip_ratio/low_mean': 0.012152777868323028, 'clip_ratio/low_min': 0.012152777868323028, 'clip_ratio/high_mean': 0.022355403285473585, 'clip_ratio/high_max': 0.022355403285473585, 'clip_ratio/region_mean': 0.03450818115379661, 'reward_total_mean': 0.9952445030212402, 'reward_meter_mean': 0.9952445030212402, 'reward_meter_std': 0.0019698443356901407, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9952445030212402, 'reward_total_composite_std': 0.0019698443356901407, 'epoch': 0.08}

 64%|██████▍   | 2111/3300 [2:50:40<1:53:33,  5.73s/it]INFO 04-12 01:27:31 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2112/3300 [2:50:45<1:48:07,  5.46s/it]2026-04-12 01:27:36,478 | INFO | train_grpo_train | metrics_logged mode=train step=2112

                                                       
{'loss': 0.0141, 'grad_norm': 3.6561806201934814, 'learning_rate': 3.603030303030303e-06, 'num_tokens': 4754067.0, 'completions/mean_length': 75.5, 'completions/min_length': 74.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9953248500823975, 'rewards/meter/std': 0.002693862421438098, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953248500823975, 'rewards/total_composite/std': 0.002693862421438098, 'reward': 0.9953248500823975, 'reward_std': 0.002693843562155962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0394962802529335, 'sampling/sampling_logp_difference/max': 1.5090522766113281, 'sampling/importance_sampling_ratio/min': 0.2211194485425949, 'sampling/importance_sampling_ratio/mean': 1.0075361728668213, 'sampling/importance_sampling_ratio/max': 1.8056482076644897, 'entropy': 0.28017595037817955, 'clip_ratio/low_mean': 0.009980987291783094, 'clip_ratio/low_min': 0.009980987291783094, 'clip_ratio/high_mean': 0.013099490897729993, 'clip_ratio/high_max': 0.013099490897729993, 'clip_ratio/region_mean': 0.023080478189513087, 'reward_total_mean': 0.9953248500823975, 'reward_meter_mean': 0.9953248500823975, 'reward_meter_std': 0.002693862421438098, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953248500823975, 'reward_total_composite_std': 0.002693862421438098, 'epoch': 0.08}

 64%|██████▍   | 2112/3300 [2:50:45<1:48:07,  5.46s/it]INFO 04-12 01:27:36 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2113/3300 [2:50:54<2:06:49,  6.41s/it]2026-04-12 01:27:45,156 | INFO | train_grpo_train | metrics_logged mode=train step=2113

                                                       
{'loss': 0.0042, 'grad_norm': 1.0902267694473267, 'learning_rate': 3.6000000000000003e-06, 'num_tokens': 4758194.0, 'completions/mean_length': 283.875, 'completions/min_length': 254.0, 'completions/max_length': 289.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 283.875, 'completions/min_terminated_length': 254.0, 'completions/max_terminated_length': 289.0, 'rewards/meter/mean': 0.9955472946166992, 'rewards/meter/std': 0.0001822304038796574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6993464231491089, 'rewards/repeat_penalty/std': 0.05026431009173393, 'rewards/total_composite/mean': 0.6962326169013977, 'rewards/total_composite/std': 0.05004315823316574, 'reward': 0.6962326169013977, 'reward_std': 0.05004315450787544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0065383305773139, 'sampling/sampling_logp_difference/max': 0.8383898735046387, 'sampling/importance_sampling_ratio/min': 0.4324061870574951, 'sampling/importance_sampling_ratio/mean': 0.9997674822807312, 'sampling/importance_sampling_ratio/max': 1.2696807384490967, 'entropy': 0.03367242659442127, 'clip_ratio/low_mean': 0.0008680555620230734, 'clip_ratio/low_min': 0.0008680555620230734, 'clip_ratio/high_mean': 0.0022863353369757533, 'clip_ratio/high_max': 0.0022863353369757533, 'clip_ratio/region_mean': 0.0031543908989988267, 'reward_total_mean': 0.6962326169013977, 'reward_meter_mean': 0.9955472946166992, 'reward_meter_std': 0.0001822304038796574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6993464231491089, 'reward_repeat_penalty_std': 0.05026431009173393, 'reward_total_composite_mean': 0.6962326169013977, 'reward_total_composite_std': 0.05004315823316574, 'epoch': 0.08}

 64%|██████▍   | 2113/3300 [2:50:54<2:06:49,  6.41s/it]INFO 04-12 01:27:45 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2114/3300 [2:51:00<2:08:56,  6.52s/it]2026-04-12 01:27:51,911 | INFO | train_grpo_train | metrics_logged mode=train step=2114

                                                       
{'loss': 0.005, 'grad_norm': 2.4666974544525146, 'learning_rate': 3.596969696969697e-06, 'num_tokens': 4761069.0, 'completions/mean_length': 184.375, 'completions/min_length': 176.0, 'completions/max_length': 193.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 184.375, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 193.0, 'rewards/meter/mean': 0.9969959855079651, 'rewards/meter/std': 0.0019897068850696087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.07856741547584534, 'rewards/total_composite/mean': 0.8585879802703857, 'rewards/total_composite/std': 0.0791376605629921, 'reward': 0.8585879802703857, 'reward_std': 0.0791376456618309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0264764241874218, 'sampling/sampling_logp_difference/max': 0.995415210723877, 'sampling/importance_sampling_ratio/min': 0.3695699870586395, 'sampling/importance_sampling_ratio/mean': 1.0035723447799683, 'sampling/importance_sampling_ratio/max': 1.9149926900863647, 'entropy': 0.16935125924646854, 'clip_ratio/low_mean': 0.003332580265123397, 'clip_ratio/low_min': 0.003332580265123397, 'clip_ratio/high_mean': 0.01766940689412877, 'clip_ratio/high_max': 0.01766940689412877, 'clip_ratio/region_mean': 0.021001987159252167, 'reward_total_mean': 0.8585879802703857, 'reward_meter_mean': 0.9969959855079651, 'reward_meter_std': 0.0019897068850696087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.07856741547584534, 'reward_total_composite_mean': 0.8585879802703857, 'reward_total_composite_std': 0.0791376605629921, 'epoch': 0.08}

 64%|██████▍   | 2114/3300 [2:51:00<2:08:56,  6.52s/it]INFO 04-12 01:27:52 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2115/3300 [2:51:10<2:26:01,  7.39s/it]2026-04-12 01:28:01,311 | INFO | train_grpo_train | metrics_logged mode=train step=2115

                                                       
{'loss': -0.0201, 'grad_norm': 1.3121997117996216, 'learning_rate': 3.593939393939394e-06, 'num_tokens': 4765867.0, 'completions/mean_length': 371.75, 'completions/min_length': 346.0, 'completions/max_length': 399.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 371.75, 'completions/min_terminated_length': 346.0, 'completions/max_terminated_length': 399.0, 'rewards/meter/mean': 0.9931106567382812, 'rewards/meter/std': 0.003387609263882041, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7990131378173828, 'rewards/repeat_penalty/std': 0.08971592038869858, 'rewards/total_composite/mean': 0.5668459534645081, 'rewards/total_composite/std': 0.06429950892925262, 'reward': 0.5668459534645081, 'reward_std': 0.06429950892925262, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03195931762456894, 'sampling/sampling_logp_difference/max': 2.8946592807769775, 'sampling/importance_sampling_ratio/min': 0.055317867547273636, 'sampling/importance_sampling_ratio/mean': 1.0030711889266968, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23509259428828955, 'clip_ratio/low_mean': 0.008480485761538148, 'clip_ratio/low_min': 0.008480485761538148, 'clip_ratio/high_mean': 0.013179366709664464, 'clip_ratio/high_max': 0.013179366709664464, 'clip_ratio/region_mean': 0.021659852471202612, 'reward_total_mean': 0.5668459534645081, 'reward_meter_mean': 0.9931106567382812, 'reward_meter_std': 0.003387609263882041, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7990131378173828, 'reward_repeat_penalty_std': 0.08971592038869858, 'reward_total_composite_mean': 0.5668459534645081, 'reward_total_composite_std': 0.06429950892925262, 'epoch': 0.08}

 64%|██████▍   | 2115/3300 [2:51:10<2:26:01,  7.39s/it]INFO 04-12 01:28:01 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2116/3300 [2:51:16<2:19:59,  7.09s/it]2026-04-12 01:28:07,749 | INFO | train_grpo_train | metrics_logged mode=train step=2116

                                                       
{'loss': 0.0004, 'grad_norm': 1.9204028844833374, 'learning_rate': 3.590909090909091e-06, 'num_tokens': 4768740.0, 'completions/mean_length': 165.125, 'completions/min_length': 165.0, 'completions/max_length': 166.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.125, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.9991816878318787, 'rewards/meter/std': 9.469691576668993e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.819444477558136, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.8187738060951233, 'rewards/total_composite/std': 0.057456936687231064, 'reward': 0.8187738060951233, 'reward_std': 0.05745694413781166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009189879521727562, 'sampling/sampling_logp_difference/max': 1.0871639251708984, 'sampling/importance_sampling_ratio/min': 0.33717140555381775, 'sampling/importance_sampling_ratio/mean': 0.9997456669807434, 'sampling/importance_sampling_ratio/max': 1.796520471572876, 'entropy': 0.04427940887399018, 'clip_ratio/low_mean': 0.0015151514671742916, 'clip_ratio/low_min': 0.0015151514671742916, 'clip_ratio/high_mean': 0.0015060240402817726, 'clip_ratio/high_max': 0.0015060240402817726, 'clip_ratio/region_mean': 0.0030211755074560642, 'reward_total_mean': 0.8187738060951233, 'reward_meter_mean': 0.9991816878318787, 'reward_meter_std': 9.469691576668993e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.819444477558136, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.8187738060951233, 'reward_total_composite_std': 0.057456936687231064, 'epoch': 0.08}

 64%|██████▍   | 2116/3300 [2:51:16<2:19:59,  7.09s/it]INFO 04-12 01:28:08 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2117/3300 [2:51:23<2:19:02,  7.05s/it]2026-04-12 01:28:14,712 | INFO | train_grpo_train | metrics_logged mode=train step=2117

                                                       
{'loss': 0.0088, 'grad_norm': 1.891755223274231, 'learning_rate': 3.587878787878788e-06, 'num_tokens': 4771637.0, 'completions/mean_length': 182.125, 'completions/min_length': 178.0, 'completions/max_length': 186.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 182.125, 'completions/min_terminated_length': 178.0, 'completions/max_terminated_length': 186.0, 'rewards/meter/mean': 0.9905871152877808, 'rewards/meter/std': 0.00902599561959505, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8055555820465088, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.7980640530586243, 'rewards/total_composite/std': 0.05313723534345627, 'reward': 0.7980640530586243, 'reward_std': 0.053137242794036865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018866930156946182, 'sampling/sampling_logp_difference/max': 1.2814984321594238, 'sampling/importance_sampling_ratio/min': 0.27762100100517273, 'sampling/importance_sampling_ratio/mean': 1.0019904375076294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09395178686827421, 'clip_ratio/low_mean': 0.010289846803061664, 'clip_ratio/low_min': 0.010289846803061664, 'clip_ratio/high_mean': 0.006211843807250261, 'clip_ratio/high_max': 0.006211843807250261, 'clip_ratio/region_mean': 0.016501690610311925, 'reward_total_mean': 0.7980640530586243, 'reward_meter_mean': 0.9905871152877808, 'reward_meter_std': 0.00902599561959505, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8055555820465088, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.7980640530586243, 'reward_total_composite_std': 0.05313723534345627, 'epoch': 0.09}

 64%|██████▍   | 2117/3300 [2:51:23<2:19:02,  7.05s/it]INFO 04-12 01:28:15 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2118/3300 [2:51:31<2:25:57,  7.41s/it]2026-04-12 01:28:22,899 | INFO | train_grpo_train | metrics_logged mode=train step=2118

                                                       
{'loss': -0.0196, 'grad_norm': 1.509689211845398, 'learning_rate': 3.584848484848485e-06, 'num_tokens': 4775851.0, 'completions/mean_length': 297.75, 'completions/min_length': 279.0, 'completions/max_length': 312.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 297.75, 'completions/min_terminated_length': 279.0, 'completions/max_terminated_length': 312.0, 'rewards/meter/mean': 0.99297034740448, 'rewards/meter/std': 0.005611373111605644, 'rewards/count_adherence/mean': 0.8888888955116272, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8921875357627869, 'rewards/repeat_penalty/std': 0.1552397757768631, 'rewards/total_composite/mean': 0.7875304222106934, 'rewards/total_composite/std': 0.13774561882019043, 'reward': 0.7875304222106934, 'reward_std': 0.13774563372135162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04715384542942047, 'sampling/sampling_logp_difference/max': 1.5841054916381836, 'sampling/importance_sampling_ratio/min': 0.20513120293617249, 'sampling/importance_sampling_ratio/mean': 1.00875985622406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.424762025475502, 'clip_ratio/low_mean': 0.00718731596134603, 'clip_ratio/low_min': 0.00718731596134603, 'clip_ratio/high_mean': 0.02530877012759447, 'clip_ratio/high_max': 0.02530877012759447, 'clip_ratio/region_mean': 0.0324960860889405, 'reward_total_mean': 0.7875304222106934, 'reward_meter_mean': 0.99297034740448, 'reward_meter_std': 0.005611373111605644, 'reward_count_adherence_mean': 0.8888888955116272, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8921875357627869, 'reward_repeat_penalty_std': 0.1552397757768631, 'reward_total_composite_mean': 0.7875304222106934, 'reward_total_composite_std': 0.13774561882019043, 'epoch': 0.09}

 64%|██████▍   | 2118/3300 [2:51:31<2:25:57,  7.41s/it]INFO 04-12 01:28:23 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2119/3300 [2:51:36<2:08:28,  6.53s/it]2026-04-12 01:28:27,386 | INFO | train_grpo_train | metrics_logged mode=train step=2119

                                                       
{'loss': -0.004, 'grad_norm': 3.0825119018554688, 'learning_rate': 3.5818181818181817e-06, 'num_tokens': 4777544.0, 'completions/mean_length': 55.625, 'completions/min_length': 54.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9949027895927429, 'rewards/meter/std': 0.00035745359491556883, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949027895927429, 'rewards/total_composite/std': 0.00035745359491556883, 'reward': 0.9949027895927429, 'reward_std': 0.00035746910725720227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01183076947927475, 'sampling/sampling_logp_difference/max': 0.6624326705932617, 'sampling/importance_sampling_ratio/min': 0.515595555305481, 'sampling/importance_sampling_ratio/mean': 0.9991430640220642, 'sampling/importance_sampling_ratio/max': 1.3233060836791992, 'entropy': 0.08348313788883388, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.004545454401522875, 'clip_ratio/high_max': 0.004545454401522875, 'clip_ratio/region_mean': 0.004545454401522875, 'reward_total_mean': 0.9949027895927429, 'reward_meter_mean': 0.9949027895927429, 'reward_meter_std': 0.00035745359491556883, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949027895927429, 'reward_total_composite_std': 0.00035745359491556883, 'epoch': 0.09}

 64%|██████▍   | 2119/3300 [2:51:36<2:08:28,  6.53s/it]INFO 04-12 01:28:27 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2120/3300 [2:51:41<1:57:42,  5.98s/it]2026-04-12 01:28:32,085 | INFO | train_grpo_train | metrics_logged mode=train step=2120

                                                       
{'loss': 0.0047, 'grad_norm': 3.3534421920776367, 'learning_rate': 3.578787878787879e-06, 'num_tokens': 4779365.0, 'completions/mean_length': 73.625, 'completions/min_length': 72.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9964693784713745, 'rewards/meter/std': 0.002390751615166664, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9964693784713745, 'rewards/total_composite/std': 0.002390751615166664, 'reward': 0.9964693784713745, 'reward_std': 0.002390736248344183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03976722061634064, 'sampling/sampling_logp_difference/max': 1.3943734169006348, 'sampling/importance_sampling_ratio/min': 0.24798838794231415, 'sampling/importance_sampling_ratio/mean': 0.9964186549186707, 'sampling/importance_sampling_ratio/max': 1.3884204626083374, 'entropy': 0.3053772822022438, 'clip_ratio/low_mean': 0.015343417064286768, 'clip_ratio/low_min': 0.015343417064286768, 'clip_ratio/high_mean': 0.020273972768336535, 'clip_ratio/high_max': 0.020273972768336535, 'clip_ratio/region_mean': 0.0356173898326233, 'reward_total_mean': 0.9964693784713745, 'reward_meter_mean': 0.9964693784713745, 'reward_meter_std': 0.002390751615166664, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9964693784713745, 'reward_total_composite_std': 0.002390751615166664, 'epoch': 0.09}

 64%|██████▍   | 2120/3300 [2:51:41<1:57:42,  5.98s/it]INFO 04-12 01:28:32 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2121/3300 [2:51:45<1:48:18,  5.51s/it]2026-04-12 01:28:36,493 | INFO | train_grpo_train | metrics_logged mode=train step=2121

                                                       
{'loss': -0.0199, 'grad_norm': 4.196082592010498, 'learning_rate': 3.575757575757576e-06, 'num_tokens': 4781079.0, 'completions/mean_length': 55.25, 'completions/min_length': 52.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.25, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.994369626045227, 'rewards/meter/std': 0.0011783665977418423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.994369626045227, 'rewards/total_composite/std': 0.0011783665977418423, 'reward': 0.994369626045227, 'reward_std': 0.0011783792870119214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00968268234282732, 'sampling/sampling_logp_difference/max': 0.9451119899749756, 'sampling/importance_sampling_ratio/min': 0.7468781471252441, 'sampling/importance_sampling_ratio/mean': 1.0018255710601807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.053725593723356724, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0022321429569274187, 'clip_ratio/high_max': 0.0022321429569274187, 'clip_ratio/region_mean': 0.004546957788988948, 'reward_total_mean': 0.994369626045227, 'reward_meter_mean': 0.994369626045227, 'reward_meter_std': 0.0011783665977418423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.994369626045227, 'reward_total_composite_std': 0.0011783665977418423, 'epoch': 0.09}

 64%|██████▍   | 2121/3300 [2:51:45<1:48:18,  5.51s/it]INFO 04-12 01:28:36 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2122/3300 [2:51:54<2:09:48,  6.61s/it]2026-04-12 01:28:45,678 | INFO | train_grpo_train | metrics_logged mode=train step=2122

                                                       
{'loss': -0.0042, 'grad_norm': 1.6565197706222534, 'learning_rate': 3.5727272727272734e-06, 'num_tokens': 4786520.0, 'completions/mean_length': 410.125, 'completions/min_length': 383.0, 'completions/max_length': 429.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 410.125, 'completions/min_terminated_length': 383.0, 'completions/max_terminated_length': 429.0, 'rewards/meter/mean': 0.9933267831802368, 'rewards/meter/std': 0.007835081778466702, 'rewards/count_adherence/mean': 0.671875, 'rewards/count_adherence/std': 0.0289318785071373, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.82259202003479, 'rewards/repeat_penalty/std': 0.032492659986019135, 'rewards/total_composite/mean': 0.5489572286605835, 'rewards/total_composite/std': 0.031959354877471924, 'reward': 0.5489572286605835, 'reward_std': 0.031959328800439835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030476173385977745, 'sampling/sampling_logp_difference/max': 4.228629112243652, 'sampling/importance_sampling_ratio/min': 0.014572354033589363, 'sampling/importance_sampling_ratio/mean': 1.0041075944900513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21682057529687881, 'clip_ratio/low_mean': 0.009774382109753788, 'clip_ratio/low_min': 0.009774382109753788, 'clip_ratio/high_mean': 0.01417637791018933, 'clip_ratio/high_max': 0.01417637791018933, 'clip_ratio/region_mean': 0.023950760019943118, 'reward_total_mean': 0.5489572286605835, 'reward_meter_mean': 0.9933267831802368, 'reward_meter_std': 0.007835081778466702, 'reward_count_adherence_mean': 0.671875, 'reward_count_adherence_std': 0.0289318785071373, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.82259202003479, 'reward_repeat_penalty_std': 0.032492659986019135, 'reward_total_composite_mean': 0.5489572286605835, 'reward_total_composite_std': 0.031959354877471924, 'epoch': 0.09}

 64%|██████▍   | 2122/3300 [2:51:54<2:09:48,  6.61s/it]INFO 04-12 01:28:45 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2123/3300 [2:51:59<1:57:19,  5.98s/it]2026-04-12 01:28:50,187 | INFO | train_grpo_train | metrics_logged mode=train step=2123

                                                       
{'loss': 0.0001, 'grad_norm': 0.5491177439689636, 'learning_rate': 3.5696969696969703e-06, 'num_tokens': 4788264.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.999085009098053, 'rewards/meter/std': 1.2789209904440213e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999085009098053, 'rewards/total_composite/std': 1.2789209904440213e-05, 'reward': 0.999085009098053, 'reward_std': 1.2776405128533952e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004210304468870163, 'sampling/sampling_logp_difference/max': 0.254666805267334, 'sampling/importance_sampling_ratio/min': 0.8502018451690674, 'sampling/importance_sampling_ratio/mean': 1.0031498670578003, 'sampling/importance_sampling_ratio/max': 1.2900317907333374, 'entropy': 0.04304229188710451, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.999085009098053, 'reward_meter_mean': 0.999085009098053, 'reward_meter_std': 1.2789209904440213e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999085009098053, 'reward_total_composite_std': 1.2789209904440213e-05, 'epoch': 0.09}

 64%|██████▍   | 2123/3300 [2:51:59<1:57:19,  5.98s/it]INFO 04-12 01:28:50 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2124/3300 [2:52:04<1:51:14,  5.68s/it]2026-04-12 01:28:55,180 | INFO | train_grpo_train | metrics_logged mode=train step=2124

                                                       
{'loss': 0.0436, 'grad_norm': 5.764961242675781, 'learning_rate': 3.566666666666667e-06, 'num_tokens': 4790353.0, 'completions/mean_length': 82.125, 'completions/min_length': 80.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.125, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.4380970001220703, 'rewards/meter/std': 0.3100660443305969, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4380970001220703, 'rewards/total_composite/std': 0.3100660443305969, 'reward': 0.4380970001220703, 'reward_std': 0.3100660443305969, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030523981899023056, 'sampling/sampling_logp_difference/max': 1.2927680015563965, 'sampling/importance_sampling_ratio/min': 0.27450987696647644, 'sampling/importance_sampling_ratio/mean': 1.0140416622161865, 'sampling/importance_sampling_ratio/max': 1.603717565536499, 'entropy': 0.26322684064507484, 'clip_ratio/low_mean': 0.017462147399783134, 'clip_ratio/low_min': 0.017462147399783134, 'clip_ratio/high_mean': 0.0031250000465661287, 'clip_ratio/high_max': 0.0031250000465661287, 'clip_ratio/region_mean': 0.020587147446349263, 'reward_total_mean': 0.4380970001220703, 'reward_meter_mean': 0.4380970001220703, 'reward_meter_std': 0.3100660443305969, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4380970001220703, 'reward_total_composite_std': 0.3100660443305969, 'epoch': 0.09}

 64%|██████▍   | 2124/3300 [2:52:04<1:51:14,  5.68s/it]INFO 04-12 01:28:55 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2125/3300 [2:52:09<1:52:04,  5.72s/it]2026-04-12 01:29:00,979 | INFO | train_grpo_train | metrics_logged mode=train step=2125

                                                       
{'loss': 0.0064, 'grad_norm': 1.9213858842849731, 'learning_rate': 3.563636363636364e-06, 'num_tokens': 4793170.0, 'completions/mean_length': 156.125, 'completions/min_length': 154.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 156.125, 'completions/min_terminated_length': 154.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.9884169101715088, 'rewards/meter/std': 0.011561714112758636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7687686681747437, 'rewards/total_composite/std': 0.008992438204586506, 'reward': 0.7687686681747437, 'reward_std': 0.008992421440780163, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008227693848311901, 'sampling/sampling_logp_difference/max': 1.741776704788208, 'sampling/importance_sampling_ratio/min': 0.17520882189273834, 'sampling/importance_sampling_ratio/mean': 1.000543475151062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.03148621576838195, 'clip_ratio/low_mean': 0.0007861634949222207, 'clip_ratio/low_min': 0.0007861634949222207, 'clip_ratio/high_mean': 0.0072432131273671985, 'clip_ratio/high_max': 0.0072432131273671985, 'clip_ratio/region_mean': 0.00802937662228942, 'reward_total_mean': 0.7687686681747437, 'reward_meter_mean': 0.9884169101715088, 'reward_meter_std': 0.011561714112758636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7687686681747437, 'reward_total_composite_std': 0.008992438204586506, 'epoch': 0.09}

 64%|██████▍   | 2125/3300 [2:52:09<1:52:04,  5.72s/it]INFO 04-12 01:29:01 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2126/3300 [2:52:14<1:43:34,  5.29s/it]2026-04-12 01:29:05,272 | INFO | train_grpo_train | metrics_logged mode=train step=2126

                                                       
{'loss': -0.0063, 'grad_norm': 6.276111125946045, 'learning_rate': 3.560606060606061e-06, 'num_tokens': 4794531.0, 'completions/mean_length': 33.125, 'completions/min_length': 33.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9689929485321045, 'rewards/meter/std': 0.0023667889181524515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9689929485321045, 'rewards/total_composite/std': 0.0023667889181524515, 'reward': 0.9689929485321045, 'reward_std': 0.0023667817004024982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0230213962495327, 'sampling/sampling_logp_difference/max': 0.7616815567016602, 'sampling/importance_sampling_ratio/min': 0.4668806791305542, 'sampling/importance_sampling_ratio/mean': 1.0059295892715454, 'sampling/importance_sampling_ratio/max': 1.4770609140396118, 'entropy': 0.14932595752179623, 'clip_ratio/low_mean': 0.018939394503831863, 'clip_ratio/low_min': 0.018939394503831863, 'clip_ratio/high_mean': 0.018605169840157032, 'clip_ratio/high_max': 0.018605169840157032, 'clip_ratio/region_mean': 0.037544564343988895, 'reward_total_mean': 0.9689929485321045, 'reward_meter_mean': 0.9689929485321045, 'reward_meter_std': 0.0023667889181524515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9689929485321045, 'reward_total_composite_std': 0.0023667889181524515, 'epoch': 0.09}

 64%|██████▍   | 2126/3300 [2:52:14<1:43:34,  5.29s/it]INFO 04-12 01:29:05 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2127/3300 [2:52:18<1:39:02,  5.07s/it]2026-04-12 01:29:09,808 | INFO | train_grpo_train | metrics_logged mode=train step=2127

                                                       
{'loss': 0.0053, 'grad_norm': 3.2434964179992676, 'learning_rate': 3.557575757575758e-06, 'num_tokens': 4796260.0, 'completions/mean_length': 54.125, 'completions/min_length': 54.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7054018974304199, 'rewards/meter/std': 0.08442744612693787, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7054018974304199, 'rewards/total_composite/std': 0.08442744612693787, 'reward': 0.7054018974304199, 'reward_std': 0.08442744612693787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009495479986071587, 'sampling/sampling_logp_difference/max': 0.5380792617797852, 'sampling/importance_sampling_ratio/min': 0.5838686227798462, 'sampling/importance_sampling_ratio/mean': 1.0031954050064087, 'sampling/importance_sampling_ratio/max': 1.4377515316009521, 'entropy': 0.049557043705135584, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0069444444961845875, 'clip_ratio/high_max': 0.0069444444961845875, 'clip_ratio/region_mean': 0.0069444444961845875, 'reward_total_mean': 0.7054018974304199, 'reward_meter_mean': 0.7054018974304199, 'reward_meter_std': 0.08442744612693787, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7054018974304199, 'reward_total_composite_std': 0.08442744612693787, 'epoch': 0.09}

 64%|██████▍   | 2127/3300 [2:52:18<1:39:02,  5.07s/it]INFO 04-12 01:29:10 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2128/3300 [2:52:25<1:50:39,  5.67s/it]2026-04-12 01:29:16,872 | INFO | train_grpo_train | metrics_logged mode=train step=2128

                                                       
{'loss': -0.0058, 'grad_norm': 2.1206531524658203, 'learning_rate': 3.554545454545455e-06, 'num_tokens': 4799957.0, 'completions/mean_length': 261.125, 'completions/min_length': 248.0, 'completions/max_length': 268.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 261.125, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 268.0, 'rewards/meter/mean': 0.9929981231689453, 'rewards/meter/std': 0.0028132039587944746, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8942307829856873, 'rewards/repeat_penalty/std': 0.05723259970545769, 'rewards/total_composite/mean': 0.8880020976066589, 'rewards/total_composite/std': 0.05752360075712204, 'reward': 0.8880020976066589, 'reward_std': 0.05752360448241234, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04205208271741867, 'sampling/sampling_logp_difference/max': 0.9664134979248047, 'sampling/importance_sampling_ratio/min': 0.38044506311416626, 'sampling/importance_sampling_ratio/mean': 1.0112462043762207, 'sampling/importance_sampling_ratio/max': 1.9891395568847656, 'entropy': 0.43808240443468094, 'clip_ratio/low_mean': 0.009215709753334522, 'clip_ratio/low_min': 0.009215709753334522, 'clip_ratio/high_mean': 0.02282354235649109, 'clip_ratio/high_max': 0.02282354235649109, 'clip_ratio/region_mean': 0.03203925210982561, 'reward_total_mean': 0.8880020976066589, 'reward_meter_mean': 0.9929981231689453, 'reward_meter_std': 0.0028132039587944746, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8942307829856873, 'reward_repeat_penalty_std': 0.05723259970545769, 'reward_total_composite_mean': 0.8880020976066589, 'reward_total_composite_std': 0.05752360075712204, 'epoch': 0.09}

 64%|██████▍   | 2128/3300 [2:52:25<1:50:39,  5.67s/it]INFO 04-12 01:29:17 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2129/3300 [2:52:32<1:59:27,  6.12s/it]2026-04-12 01:29:24,053 | INFO | train_grpo_train | metrics_logged mode=train step=2129

                                                       
{'loss': -0.0205, 'grad_norm': 2.319450855255127, 'learning_rate': 3.551515151515152e-06, 'num_tokens': 4803610.0, 'completions/mean_length': 264.625, 'completions/min_length': 252.0, 'completions/max_length': 281.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 264.625, 'completions/min_terminated_length': 252.0, 'completions/max_terminated_length': 281.0, 'rewards/meter/mean': 0.9950700998306274, 'rewards/meter/std': 0.0019416395807638764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8949176073074341, 'rewards/repeat_penalty/std': 0.04042290523648262, 'rewards/total_composite/mean': 0.8904905915260315, 'rewards/total_composite/std': 0.039885781705379486, 'reward': 0.8904905915260315, 'reward_std': 0.039885781705379486, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040030911564826965, 'sampling/sampling_logp_difference/max': 3.957237720489502, 'sampling/importance_sampling_ratio/min': 0.01911584474146366, 'sampling/importance_sampling_ratio/mean': 1.0098717212677002, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35909293219447136, 'clip_ratio/low_mean': 0.01029538200236857, 'clip_ratio/low_min': 0.01029538200236857, 'clip_ratio/high_mean': 0.02008880244102329, 'clip_ratio/high_max': 0.02008880244102329, 'clip_ratio/region_mean': 0.03038418444339186, 'reward_total_mean': 0.8904905915260315, 'reward_meter_mean': 0.9950700998306274, 'reward_meter_std': 0.0019416395807638764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8949176073074341, 'reward_repeat_penalty_std': 0.04042290523648262, 'reward_total_composite_mean': 0.8904905915260315, 'reward_total_composite_std': 0.039885781705379486, 'epoch': 0.09}

 65%|██████▍   | 2129/3300 [2:52:33<1:59:27,  6.12s/it]INFO 04-12 01:29:24 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2130/3300 [2:52:37<1:51:09,  5.70s/it]2026-04-12 01:29:28,778 | INFO | train_grpo_train | metrics_logged mode=train step=2130

                                                       
{'loss': 0.0131, 'grad_norm': 4.817035675048828, 'learning_rate': 3.548484848484849e-06, 'num_tokens': 4805553.0, 'completions/mean_length': 70.875, 'completions/min_length': 68.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9955647587776184, 'rewards/meter/std': 0.0008866624557413161, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955647587776184, 'rewards/total_composite/std': 0.0008866624557413161, 'reward': 0.9955647587776184, 'reward_std': 0.0008866624557413161, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.037295129150152206, 'sampling/sampling_logp_difference/max': 1.1011011600494385, 'sampling/importance_sampling_ratio/min': 0.3325047492980957, 'sampling/importance_sampling_ratio/mean': 1.006104826927185, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20211379043757915, 'clip_ratio/low_mean': 0.017461657291278243, 'clip_ratio/low_min': 0.017461657291278243, 'clip_ratio/high_mean': 0.01770029927138239, 'clip_ratio/high_max': 0.01770029927138239, 'clip_ratio/region_mean': 0.035161956562660635, 'reward_total_mean': 0.9955647587776184, 'reward_meter_mean': 0.9955647587776184, 'reward_meter_std': 0.0008866624557413161, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955647587776184, 'reward_total_composite_std': 0.0008866624557413161, 'epoch': 0.09}

 65%|██████▍   | 2130/3300 [2:52:37<1:51:09,  5.70s/it]INFO 04-12 01:29:29 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2131/3300 [2:52:43<1:50:04,  5.65s/it]2026-04-12 01:29:34,305 | INFO | train_grpo_train | metrics_logged mode=train step=2131

                                                       
{'loss': -0.0121, 'grad_norm': 2.6760542392730713, 'learning_rate': 3.5454545454545458e-06, 'num_tokens': 4808185.0, 'completions/mean_length': 142.0, 'completions/min_length': 138.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.0, 'completions/min_terminated_length': 138.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.9969066381454468, 'rewards/meter/std': 0.0009874114766716957, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969066381454468, 'rewards/total_composite/std': 0.0009874114766716957, 'reward': 0.9969066381454468, 'reward_std': 0.0009874175302684307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03481653332710266, 'sampling/sampling_logp_difference/max': 1.2930517196655273, 'sampling/importance_sampling_ratio/min': 0.2744320034980774, 'sampling/importance_sampling_ratio/mean': 1.0037314891815186, 'sampling/importance_sampling_ratio/max': 1.913059115409851, 'entropy': 0.28764577955007553, 'clip_ratio/low_mean': 0.015936652896925807, 'clip_ratio/low_min': 0.015936652896925807, 'clip_ratio/high_mean': 0.014961764682084322, 'clip_ratio/high_max': 0.014961764682084322, 'clip_ratio/region_mean': 0.03089841757901013, 'reward_total_mean': 0.9969066381454468, 'reward_meter_mean': 0.9969066381454468, 'reward_meter_std': 0.0009874114766716957, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969066381454468, 'reward_total_composite_std': 0.0009874114766716957, 'epoch': 0.09}

 65%|██████▍   | 2131/3300 [2:52:43<1:50:04,  5.65s/it]INFO 04-12 01:29:34 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2132/3300 [2:52:48<1:47:16,  5.51s/it]2026-04-12 01:29:39,496 | INFO | train_grpo_train | metrics_logged mode=train step=2132

                                                       
{'loss': 0.0002, 'grad_norm': 3.024273633956909, 'learning_rate': 3.5424242424242426e-06, 'num_tokens': 4810360.0, 'completions/mean_length': 107.875, 'completions/min_length': 106.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.875, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9939501285552979, 'rewards/meter/std': 0.003658889327198267, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9690351486206055, 'rewards/total_composite/std': 0.06948735564947128, 'reward': 0.9690351486206055, 'reward_std': 0.06948734074831009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03451121971011162, 'sampling/sampling_logp_difference/max': 1.3759374618530273, 'sampling/importance_sampling_ratio/min': 0.2526026964187622, 'sampling/importance_sampling_ratio/mean': 1.0037057399749756, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17346674110740423, 'clip_ratio/low_mean': 0.004672897048294544, 'clip_ratio/low_min': 0.004672897048294544, 'clip_ratio/high_mean': 0.026603603502735496, 'clip_ratio/high_max': 0.026603603502735496, 'clip_ratio/region_mean': 0.03127650055103004, 'reward_total_mean': 0.9690351486206055, 'reward_meter_mean': 0.9939501285552979, 'reward_meter_std': 0.003658889327198267, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9690351486206055, 'reward_total_composite_std': 0.06948735564947128, 'epoch': 0.09}

 65%|██████▍   | 2132/3300 [2:52:48<1:47:16,  5.51s/it]INFO 04-12 01:29:39 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2133/3300 [2:52:54<1:48:35,  5.58s/it]2026-04-12 01:29:45,246 | INFO | train_grpo_train | metrics_logged mode=train step=2133

                                                       
{'loss': 0.001, 'grad_norm': 0.6826779246330261, 'learning_rate': 3.53939393939394e-06, 'num_tokens': 4813068.0, 'completions/mean_length': 165.5, 'completions/min_length': 165.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.5, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.999198317527771, 'rewards/meter/std': 3.6679448385257274e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7916666865348816, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.7910317182540894, 'rewards/total_composite/std': 0.039245039224624634, 'reward': 0.7910317182540894, 'reward_std': 0.03924502432346344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008064919151365757, 'sampling/sampling_logp_difference/max': 1.644291639328003, 'sampling/importance_sampling_ratio/min': 0.19314932823181152, 'sampling/importance_sampling_ratio/mean': 1.0013805627822876, 'sampling/importance_sampling_ratio/max': 1.5342899560928345, 'entropy': 0.03646064782515168, 'clip_ratio/low_mean': 0.0030303029343485832, 'clip_ratio/low_min': 0.0030303029343485832, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0030303029343485832, 'reward_total_mean': 0.7910317182540894, 'reward_meter_mean': 0.999198317527771, 'reward_meter_std': 3.6679448385257274e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7916666865348816, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.7910317182540894, 'reward_total_composite_std': 0.039245039224624634, 'epoch': 0.09}

 65%|██████▍   | 2133/3300 [2:52:54<1:48:35,  5.58s/it]INFO 04-12 01:29:45 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2134/3300 [2:52:58<1:41:27,  5.22s/it]2026-04-12 01:29:49,637 | INFO | train_grpo_train | metrics_logged mode=train step=2134

                                                       
{'loss': 0.0003, 'grad_norm': 0.44613105058670044, 'learning_rate': 3.5363636363636367e-06, 'num_tokens': 4814740.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7352948188781738, 'rewards/meter/std': 3.051780367968604e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7352948188781738, 'rewards/total_composite/std': 3.051780367968604e-05, 'reward': 0.7352948188781738, 'reward_std': 3.0517016057274304e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005636686459183693, 'sampling/sampling_logp_difference/max': 0.6817889213562012, 'sampling/importance_sampling_ratio/min': 0.7626789808273315, 'sampling/importance_sampling_ratio/mean': 1.0043625831604004, 'sampling/importance_sampling_ratio/max': 1.9774119853973389, 'entropy': 0.037100087851285934, 'clip_ratio/low_mean': 0.004629629664123058, 'clip_ratio/low_min': 0.004629629664123058, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004629629664123058, 'reward_total_mean': 0.7352948188781738, 'reward_meter_mean': 0.7352948188781738, 'reward_meter_std': 3.051780367968604e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7352948188781738, 'reward_total_composite_std': 3.051780367968604e-05, 'epoch': 0.09}

 65%|██████▍   | 2134/3300 [2:52:58<1:41:27,  5.22s/it]INFO 04-12 01:29:49 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2135/3300 [2:53:05<1:51:20,  5.73s/it]2026-04-12 01:29:56,555 | INFO | train_grpo_train | metrics_logged mode=train step=2135

                                                       
{'loss': 0.0222, 'grad_norm': 4.601060390472412, 'learning_rate': 3.5333333333333335e-06, 'num_tokens': 4818135.0, 'completions/mean_length': 229.375, 'completions/min_length': 224.0, 'completions/max_length': 240.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 229.375, 'completions/min_terminated_length': 224.0, 'completions/max_terminated_length': 240.0, 'rewards/meter/mean': 0.971871018409729, 'rewards/meter/std': 0.06841260194778442, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9431818723678589, 'rewards/repeat_penalty/std': 0.047049909830093384, 'rewards/total_composite/mean': 0.9174596071243286, 'rewards/total_composite/std': 0.08795089274644852, 'reward': 0.9174596071243286, 'reward_std': 0.08795087039470673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041487276554107666, 'sampling/sampling_logp_difference/max': 1.7106046676635742, 'sampling/importance_sampling_ratio/min': 0.18075646460056305, 'sampling/importance_sampling_ratio/mean': 1.0070955753326416, 'sampling/importance_sampling_ratio/max': 1.8775804042816162, 'entropy': 0.4052262455224991, 'clip_ratio/low_mean': 0.016251615015789866, 'clip_ratio/low_min': 0.016251615015789866, 'clip_ratio/high_mean': 0.014304422307759523, 'clip_ratio/high_max': 0.014304422307759523, 'clip_ratio/region_mean': 0.03055603732354939, 'reward_total_mean': 0.9174596071243286, 'reward_meter_mean': 0.971871018409729, 'reward_meter_std': 0.06841260194778442, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9431818723678589, 'reward_repeat_penalty_std': 0.047049909830093384, 'reward_total_composite_mean': 0.9174596071243286, 'reward_total_composite_std': 0.08795089274644852, 'epoch': 0.09}

 65%|██████▍   | 2135/3300 [2:53:05<1:51:20,  5.73s/it]INFO 04-12 01:29:56 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2136/3300 [2:53:11<1:53:45,  5.86s/it]2026-04-12 01:30:02,737 | INFO | train_grpo_train | metrics_logged mode=train step=2136

                                                       
{'loss': -0.0019, 'grad_norm': 1.9712090492248535, 'learning_rate': 3.5303030303030304e-06, 'num_tokens': 4820878.0, 'completions/mean_length': 165.875, 'completions/min_length': 165.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.875, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9991864562034607, 'rewards/meter/std': 2.615616722323466e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.07856741547584534, 'rewards/total_composite/mean': 0.8604106903076172, 'rewards/total_composite/std': 0.07850649952888489, 'reward': 0.8604106903076172, 'reward_std': 0.07850649952888489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009452839381992817, 'sampling/sampling_logp_difference/max': 0.7467837333679199, 'sampling/importance_sampling_ratio/min': 0.5129733085632324, 'sampling/importance_sampling_ratio/mean': 1.0031864643096924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06440420588478446, 'clip_ratio/low_mean': 0.004527199547737837, 'clip_ratio/low_min': 0.004527199547737837, 'clip_ratio/high_mean': 0.0015015150420367718, 'clip_ratio/high_max': 0.0015015150420367718, 'clip_ratio/region_mean': 0.006028714589774609, 'reward_total_mean': 0.8604106903076172, 'reward_meter_mean': 0.9991864562034607, 'reward_meter_std': 2.615616722323466e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.07856741547584534, 'reward_total_composite_mean': 0.8604106903076172, 'reward_total_composite_std': 0.07850649952888489, 'epoch': 0.09}

 65%|██████▍   | 2136/3300 [2:53:11<1:53:45,  5.86s/it]INFO 04-12 01:30:03 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2137/3300 [2:53:17<1:53:16,  5.84s/it]2026-04-12 01:30:08,514 | INFO | train_grpo_train | metrics_logged mode=train step=2137

                                                       
{'loss': -0.0007, 'grad_norm': 1.4629026651382446, 'learning_rate': 3.5272727272727276e-06, 'num_tokens': 4823547.0, 'completions/mean_length': 145.625, 'completions/min_length': 145.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 145.625, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.9696428775787354, 'rewards/meter/std': 0.0028882811311632395, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7541667222976685, 'rewards/total_composite/std': 0.0022464508656412363, 'reward': 0.7541667222976685, 'reward_std': 0.0022464566864073277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011477367021143436, 'sampling/sampling_logp_difference/max': 1.31986665725708, 'sampling/importance_sampling_ratio/min': 0.26717090606689453, 'sampling/importance_sampling_ratio/mean': 1.0017269849777222, 'sampling/importance_sampling_ratio/max': 1.7878315448760986, 'entropy': 0.04937975388020277, 'clip_ratio/low_mean': 0.006010864395648241, 'clip_ratio/low_min': 0.006010864395648241, 'clip_ratio/high_mean': 0.0025862068869173527, 'clip_ratio/high_max': 0.0025862068869173527, 'clip_ratio/region_mean': 0.008597071282565594, 'reward_total_mean': 0.7541667222976685, 'reward_meter_mean': 0.9696428775787354, 'reward_meter_std': 0.0028882811311632395, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7541667222976685, 'reward_total_composite_std': 0.0022464508656412363, 'epoch': 0.09}

 65%|██████▍   | 2137/3300 [2:53:17<1:53:16,  5.84s/it]INFO 04-12 01:30:08 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2138/3300 [2:53:24<2:00:48,  6.24s/it]2026-04-12 01:30:15,695 | INFO | train_grpo_train | metrics_logged mode=train step=2138

                                                       
{'loss': 0.0015, 'grad_norm': 2.5688741207122803, 'learning_rate': 3.5242424242424244e-06, 'num_tokens': 4826897.0, 'completions/mean_length': 228.75, 'completions/min_length': 225.0, 'completions/max_length': 233.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 228.75, 'completions/min_terminated_length': 225.0, 'completions/max_terminated_length': 233.0, 'rewards/meter/mean': 0.9738757610321045, 'rewards/meter/std': 0.055281247943639755, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9659091234207153, 'rewards/repeat_penalty/std': 0.047049909830093384, 'rewards/total_composite/mean': 0.9418025016784668, 'rewards/total_composite/std': 0.08390164375305176, 'reward': 0.9418025016784668, 'reward_std': 0.08390163630247116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04527745395898819, 'sampling/sampling_logp_difference/max': 1.5867786407470703, 'sampling/importance_sampling_ratio/min': 0.20458358526229858, 'sampling/importance_sampling_ratio/mean': 1.0137181282043457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40121684968471527, 'clip_ratio/low_mean': 0.01095851231366396, 'clip_ratio/low_min': 0.01095851231366396, 'clip_ratio/high_mean': 0.02402331749908626, 'clip_ratio/high_max': 0.02402331749908626, 'clip_ratio/region_mean': 0.03498182981275022, 'reward_total_mean': 0.9418025016784668, 'reward_meter_mean': 0.9738757610321045, 'reward_meter_std': 0.055281247943639755, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9659091234207153, 'reward_repeat_penalty_std': 0.047049909830093384, 'reward_total_composite_mean': 0.9418025016784668, 'reward_total_composite_std': 0.08390164375305176, 'epoch': 0.09}

 65%|██████▍   | 2138/3300 [2:53:24<2:00:48,  6.24s/it]INFO 04-12 01:30:15 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2139/3300 [2:53:29<1:51:16,  5.75s/it]2026-04-12 01:30:20,306 | INFO | train_grpo_train | metrics_logged mode=train step=2139

                                                       
{'loss': 0.0018, 'grad_norm': 1.6966145038604736, 'learning_rate': 3.5212121212121213e-06, 'num_tokens': 4828577.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7483930587768555, 'rewards/meter/std': 0.02422301471233368, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7483930587768555, 'rewards/total_composite/std': 0.02422301471233368, 'reward': 0.7483930587768555, 'reward_std': 0.024223024025559425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005149809177964926, 'sampling/sampling_logp_difference/max': 0.4711747169494629, 'sampling/importance_sampling_ratio/min': 0.6358863711357117, 'sampling/importance_sampling_ratio/mean': 1.0017740726470947, 'sampling/importance_sampling_ratio/max': 1.6018749475479126, 'entropy': 0.020970646291971207, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.002314814832061529, 'clip_ratio/high_max': 0.002314814832061529, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.7483930587768555, 'reward_meter_mean': 0.7483930587768555, 'reward_meter_std': 0.02422301471233368, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7483930587768555, 'reward_total_composite_std': 0.02422301471233368, 'epoch': 0.09}

 65%|██████▍   | 2139/3300 [2:53:29<1:51:16,  5.75s/it]INFO 04-12 01:30:20 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2140/3300 [2:53:34<1:47:23,  5.55s/it]2026-04-12 01:30:25,403 | INFO | train_grpo_train | metrics_logged mode=train step=2140

                                                       
{'loss': -0.0033, 'grad_norm': 3.241760015487671, 'learning_rate': 3.5181818181818185e-06, 'num_tokens': 4830436.0, 'completions/mean_length': 76.375, 'completions/min_length': 73.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9935569763183594, 'rewards/meter/std': 0.004046422429382801, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9935569763183594, 'rewards/total_composite/std': 0.004046422429382801, 'reward': 0.9935569763183594, 'reward_std': 0.004046411253511906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03363337367773056, 'sampling/sampling_logp_difference/max': 1.207101821899414, 'sampling/importance_sampling_ratio/min': 0.2990627586841583, 'sampling/importance_sampling_ratio/mean': 1.008273720741272, 'sampling/importance_sampling_ratio/max': 1.7292462587356567, 'entropy': 0.3370419256389141, 'clip_ratio/low_mean': 0.004959081998094916, 'clip_ratio/low_min': 0.004959081998094916, 'clip_ratio/high_mean': 0.02094770153053105, 'clip_ratio/high_max': 0.02094770153053105, 'clip_ratio/region_mean': 0.025906783528625965, 'reward_total_mean': 0.9935569763183594, 'reward_meter_mean': 0.9935569763183594, 'reward_meter_std': 0.004046422429382801, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9935569763183594, 'reward_total_composite_std': 0.004046422429382801, 'epoch': 0.09}

 65%|██████▍   | 2140/3300 [2:53:34<1:47:23,  5.55s/it]INFO 04-12 01:30:25 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2141/3300 [2:53:38<1:41:07,  5.24s/it]2026-04-12 01:30:29,879 | INFO | train_grpo_train | metrics_logged mode=train step=2141

                                                       
{'loss': 0.0014, 'grad_norm': 4.896651268005371, 'learning_rate': 3.5151515151515154e-06, 'num_tokens': 4832051.0, 'completions/mean_length': 55.875, 'completions/min_length': 55.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9941670894622803, 'rewards/meter/std': 0.002000106731429696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9941670894622803, 'rewards/total_composite/std': 0.002000106731429696, 'reward': 0.9941670894622803, 'reward_std': 0.0020000922959297895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011897793039679527, 'sampling/sampling_logp_difference/max': 0.8048620223999023, 'sampling/importance_sampling_ratio/min': 0.4471496343612671, 'sampling/importance_sampling_ratio/mean': 1.0028151273727417, 'sampling/importance_sampling_ratio/max': 1.2935843467712402, 'entropy': 0.0797986383549869, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.004464285913854837, 'clip_ratio/high_max': 0.004464285913854837, 'clip_ratio/region_mean': 0.004464285913854837, 'reward_total_mean': 0.9941670894622803, 'reward_meter_mean': 0.9941670894622803, 'reward_meter_std': 0.002000106731429696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9941670894622803, 'reward_total_composite_std': 0.002000106731429696, 'epoch': 0.09}

 65%|██████▍   | 2141/3300 [2:53:38<1:41:07,  5.24s/it]INFO 04-12 01:30:30 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2142/3300 [2:53:48<2:04:06,  6.43s/it]2026-04-12 01:30:39,091 | INFO | train_grpo_train | metrics_logged mode=train step=2142

                                                       
{'loss': 0.013, 'grad_norm': 1.6217327117919922, 'learning_rate': 3.512121212121212e-06, 'num_tokens': 4837339.0, 'completions/mean_length': 428.0, 'completions/min_length': 411.0, 'completions/max_length': 448.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 428.0, 'completions/min_terminated_length': 411.0, 'completions/max_terminated_length': 448.0, 'rewards/meter/mean': 0.9931625127792358, 'rewards/meter/std': 0.0038301812019199133, 'rewards/count_adherence/mean': 0.6953125, 'rewards/count_adherence/std': 0.022097086533904076, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8884575366973877, 'rewards/repeat_penalty/std': 0.061322376132011414, 'rewards/total_composite/mean': 0.6134706139564514, 'rewards/total_composite/std': 0.04515310749411583, 'reward': 0.6134706139564514, 'reward_std': 0.04515310376882553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042900450527668, 'sampling/sampling_logp_difference/max': 2.926726818084717, 'sampling/importance_sampling_ratio/min': 0.05357210338115692, 'sampling/importance_sampling_ratio/mean': 1.0085182189941406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3651280999183655, 'clip_ratio/low_mean': 0.005864104256033897, 'clip_ratio/low_min': 0.005864104256033897, 'clip_ratio/high_mean': 0.027495570946484804, 'clip_ratio/high_max': 0.027495570946484804, 'clip_ratio/region_mean': 0.0333596752025187, 'reward_total_mean': 0.6134706139564514, 'reward_meter_mean': 0.9931625127792358, 'reward_meter_std': 0.0038301812019199133, 'reward_count_adherence_mean': 0.6953125, 'reward_count_adherence_std': 0.022097086533904076, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8884575366973877, 'reward_repeat_penalty_std': 0.061322376132011414, 'reward_total_composite_mean': 0.6134706139564514, 'reward_total_composite_std': 0.04515310749411583, 'epoch': 0.09}

 65%|██████▍   | 2142/3300 [2:53:48<2:04:06,  6.43s/it]INFO 04-12 01:30:39 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2143/3300 [2:53:52<1:53:54,  5.91s/it]2026-04-12 01:30:43,788 | INFO | train_grpo_train | metrics_logged mode=train step=2143

                                                       
{'loss': 0.0064, 'grad_norm': 5.659821510314941, 'learning_rate': 3.509090909090909e-06, 'num_tokens': 4839153.0, 'completions/mean_length': 74.75, 'completions/min_length': 72.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9972976446151733, 'rewards/meter/std': 0.0019415906863287091, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972976446151733, 'rewards/total_composite/std': 0.0019415906863287091, 'reward': 0.9972976446151733, 'reward_std': 0.0019415918504819274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06912153214216232, 'sampling/sampling_logp_difference/max': 1.5144741535186768, 'sampling/importance_sampling_ratio/min': 0.4317888617515564, 'sampling/importance_sampling_ratio/mean': 1.0229203701019287, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6021371968090534, 'clip_ratio/low_mean': 0.040365297347307205, 'clip_ratio/low_min': 0.040365297347307205, 'clip_ratio/high_mean': 0.024497864302247763, 'clip_ratio/high_max': 0.024497864302247763, 'clip_ratio/region_mean': 0.06486316164955497, 'reward_total_mean': 0.9972976446151733, 'reward_meter_mean': 0.9972976446151733, 'reward_meter_std': 0.0019415906863287091, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972976446151733, 'reward_total_composite_std': 0.0019415906863287091, 'epoch': 0.09}

 65%|██████▍   | 2143/3300 [2:53:52<1:53:54,  5.91s/it]INFO 04-12 01:30:44 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2144/3300 [2:53:57<1:46:43,  5.54s/it]2026-04-12 01:30:48,468 | INFO | train_grpo_train | metrics_logged mode=train step=2144

                                                       
{'loss': -0.0009, 'grad_norm': 5.293501853942871, 'learning_rate': 3.5060606060606063e-06, 'num_tokens': 4841063.0, 'completions/mean_length': 74.75, 'completions/min_length': 72.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9942959547042847, 'rewards/meter/std': 0.0059119779616594315, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9942959547042847, 'rewards/total_composite/std': 0.0059119779616594315, 'reward': 0.9942959547042847, 'reward_std': 0.005911969114094973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06987594068050385, 'sampling/sampling_logp_difference/max': 2.395887851715088, 'sampling/importance_sampling_ratio/min': 0.09109176695346832, 'sampling/importance_sampling_ratio/mean': 1.0201735496520996, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48777300119400024, 'clip_ratio/low_mean': 0.021914413664489985, 'clip_ratio/low_min': 0.021914413664489985, 'clip_ratio/high_mean': 0.025277130538597703, 'clip_ratio/high_max': 0.025277130538597703, 'clip_ratio/region_mean': 0.04719154420308769, 'reward_total_mean': 0.9942959547042847, 'reward_meter_mean': 0.9942959547042847, 'reward_meter_std': 0.0059119779616594315, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9942959547042847, 'reward_total_composite_std': 0.0059119779616594315, 'epoch': 0.09}

 65%|██████▍   | 2144/3300 [2:53:57<1:46:43,  5.54s/it]INFO 04-12 01:30:48 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2145/3300 [2:54:01<1:38:50,  5.13s/it]2026-04-12 01:30:52,654 | INFO | train_grpo_train | metrics_logged mode=train step=2145

                                                       
{'loss': 0.0132, 'grad_norm': 7.162275314331055, 'learning_rate': 3.503030303030303e-06, 'num_tokens': 4842409.0, 'completions/mean_length': 29.25, 'completions/min_length': 29.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9901993274688721, 'rewards/meter/std': 0.004359308164566755, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9901993274688721, 'rewards/total_composite/std': 0.004359308164566755, 'reward': 0.9901993274688721, 'reward_std': 0.0043593174777925014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01579299196600914, 'sampling/sampling_logp_difference/max': 0.4930543899536133, 'sampling/importance_sampling_ratio/min': 0.6107580661773682, 'sampling/importance_sampling_ratio/mean': 0.9988601207733154, 'sampling/importance_sampling_ratio/max': 1.2712383270263672, 'entropy': 0.09351515863090754, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.01278735650703311, 'clip_ratio/high_max': 0.01278735650703311, 'clip_ratio/region_mean': 0.01278735650703311, 'reward_total_mean': 0.9901993274688721, 'reward_meter_mean': 0.9901993274688721, 'reward_meter_std': 0.004359308164566755, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9901993274688721, 'reward_total_composite_std': 0.004359308164566755, 'epoch': 0.09}

 65%|██████▌   | 2145/3300 [2:54:01<1:38:50,  5.13s/it]INFO 04-12 01:30:52 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2146/3300 [2:54:10<2:02:49,  6.39s/it]2026-04-12 01:31:01,959 | INFO | train_grpo_train | metrics_logged mode=train step=2146

                                                       
{'loss': 0.5018, 'grad_norm': 5.25761604309082, 'learning_rate': 3.5e-06, 'num_tokens': 4845311.0, 'completions/mean_length': 198.75, 'completions/min_length': 143.0, 'completions/max_length': 488.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 198.75, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 488.0, 'rewards/meter/mean': 0.8164126873016357, 'rewards/meter/std': 0.32973745465278625, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9434523582458496, 'rewards/repeat_penalty/std': 0.0783882662653923, 'rewards/total_composite/mean': 0.7369348406791687, 'rewards/total_composite/std': 0.32373327016830444, 'reward': 0.7369348406791687, 'reward_std': 0.32373327016830444, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09025763720273972, 'sampling/sampling_logp_difference/max': 1.6007208824157715, 'sampling/importance_sampling_ratio/min': 0.20175102353096008, 'sampling/importance_sampling_ratio/mean': 1.013696312904358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2716873697936535, 'clip_ratio/low_mean': 0.008386766072362661, 'clip_ratio/low_min': 0.008386766072362661, 'clip_ratio/high_mean': 0.02845356403850019, 'clip_ratio/high_max': 0.02845356403850019, 'clip_ratio/region_mean': 0.03684033011086285, 'reward_total_mean': 0.7369348406791687, 'reward_meter_mean': 0.8164126873016357, 'reward_meter_std': 0.32973745465278625, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9434523582458496, 'reward_repeat_penalty_std': 0.0783882662653923, 'reward_total_composite_mean': 0.7369348406791687, 'reward_total_composite_std': 0.32373327016830444, 'epoch': 0.09}

 65%|██████▌   | 2146/3300 [2:54:10<2:02:49,  6.39s/it]INFO 04-12 01:31:02 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2147/3300 [2:54:15<1:53:09,  5.89s/it]2026-04-12 01:31:06,686 | INFO | train_grpo_train | metrics_logged mode=train step=2147

                                                       
{'loss': -0.0056, 'grad_norm': 7.887787342071533, 'learning_rate': 3.496969696969697e-06, 'num_tokens': 4847185.0, 'completions/mean_length': 73.25, 'completions/min_length': 69.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8275381922721863, 'rewards/meter/std': 0.3271496891975403, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8275381922721863, 'rewards/total_composite/std': 0.3271496891975403, 'reward': 0.8275381922721863, 'reward_std': 0.3271496891975403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0584087148308754, 'sampling/sampling_logp_difference/max': 1.7163329124450684, 'sampling/importance_sampling_ratio/min': 0.17972400784492493, 'sampling/importance_sampling_ratio/mean': 1.0108306407928467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41511600464582443, 'clip_ratio/low_mean': 0.00345633109100163, 'clip_ratio/low_min': 0.00345633109100163, 'clip_ratio/high_mean': 0.029611698118969798, 'clip_ratio/high_max': 0.029611698118969798, 'clip_ratio/region_mean': 0.03306802920997143, 'reward_total_mean': 0.8275381922721863, 'reward_meter_mean': 0.8275381922721863, 'reward_meter_std': 0.3271496891975403, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8275381922721863, 'reward_total_composite_std': 0.3271496891975403, 'epoch': 0.09}

 65%|██████▌   | 2147/3300 [2:54:15<1:53:09,  5.89s/it]INFO 04-12 01:31:06 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2148/3300 [2:54:19<1:42:26,  5.34s/it]2026-04-12 01:31:10,732 | INFO | train_grpo_train | metrics_logged mode=train step=2148

                                                       
{'loss': -0.0013, 'grad_norm': 6.055057525634766, 'learning_rate': 3.493939393939394e-06, 'num_tokens': 4848601.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9926831722259521, 'rewards/meter/std': 0.0014186145272105932, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9926831722259521, 'rewards/total_composite/std': 0.0014186145272105932, 'reward': 0.9926831722259521, 'reward_std': 0.0014186184853315353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018664227798581123, 'sampling/sampling_logp_difference/max': 0.8764760494232178, 'sampling/importance_sampling_ratio/min': 0.4162471890449524, 'sampling/importance_sampling_ratio/mean': 0.9984536170959473, 'sampling/importance_sampling_ratio/max': 1.2208077907562256, 'entropy': 0.0592546658590436, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.012931034434586763, 'clip_ratio/high_max': 0.012931034434586763, 'clip_ratio/region_mean': 0.017241379246115685, 'reward_total_mean': 0.9926831722259521, 'reward_meter_mean': 0.9926831722259521, 'reward_meter_std': 0.0014186145272105932, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9926831722259521, 'reward_total_composite_std': 0.0014186145272105932, 'epoch': 0.09}

 65%|██████▌   | 2148/3300 [2:54:19<1:42:26,  5.34s/it]INFO 04-12 01:31:10 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2149/3300 [2:54:25<1:46:43,  5.56s/it]2026-04-12 01:31:16,826 | INFO | train_grpo_train | metrics_logged mode=train step=2149

                                                       
{'loss': 0.012, 'grad_norm': 3.5529558658599854, 'learning_rate': 3.4909090909090913e-06, 'num_tokens': 4851588.0, 'completions/mean_length': 186.375, 'completions/min_length': 181.0, 'completions/max_length': 196.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.375, 'completions/min_terminated_length': 181.0, 'completions/max_terminated_length': 196.0, 'rewards/meter/mean': 0.9706813097000122, 'rewards/meter/std': 0.06584999710321426, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666269302368, 'rewards/repeat_penalty/std': 0.07856741547584534, 'rewards/total_composite/mean': 0.8901457786560059, 'rewards/total_composite/std': 0.10105938464403152, 'reward': 0.8901457786560059, 'reward_std': 0.10105939954519272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05906498804688454, 'sampling/sampling_logp_difference/max': 1.6412153244018555, 'sampling/importance_sampling_ratio/min': 0.19374443590641022, 'sampling/importance_sampling_ratio/mean': 1.0136346817016602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5572481229901314, 'clip_ratio/low_mean': 0.021443208097480237, 'clip_ratio/low_min': 0.021443208097480237, 'clip_ratio/high_mean': 0.02710459241643548, 'clip_ratio/high_max': 0.02710459241643548, 'clip_ratio/region_mean': 0.04854780051391572, 'reward_total_mean': 0.8901457786560059, 'reward_meter_mean': 0.9706813097000122, 'reward_meter_std': 0.06584999710321426, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666269302368, 'reward_repeat_penalty_std': 0.07856741547584534, 'reward_total_composite_mean': 0.8901457786560059, 'reward_total_composite_std': 0.10105938464403152, 'epoch': 0.09}

 65%|██████▌   | 2149/3300 [2:54:25<1:46:43,  5.56s/it]INFO 04-12 01:31:17 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2150/3300 [2:54:31<1:48:35,  5.67s/it]2026-04-12 01:31:22,731 | INFO | train_grpo_train | metrics_logged mode=train step=2150

                                                       
{'loss': 0.0075, 'grad_norm': 2.0675745010375977, 'learning_rate': 3.4878787878787885e-06, 'num_tokens': 4854574.0, 'completions/mean_length': 179.25, 'completions/min_length': 174.0, 'completions/max_length': 182.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 179.25, 'completions/min_terminated_length': 174.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.9972676634788513, 'rewards/meter/std': 0.0014413135359063745, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.08399210125207901, 'rewards/total_composite/mean': 0.9418361783027649, 'rewards/total_composite/std': 0.08344186097383499, 'reward': 0.9418361783027649, 'reward_std': 0.0834418535232544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02511134371161461, 'sampling/sampling_logp_difference/max': 1.1651017665863037, 'sampling/importance_sampling_ratio/min': 0.3118909001350403, 'sampling/importance_sampling_ratio/mean': 1.0084565877914429, 'sampling/importance_sampling_ratio/max': 1.8844407796859741, 'entropy': 0.2794443480670452, 'clip_ratio/low_mean': 0.006343001441564411, 'clip_ratio/low_min': 0.006343001441564411, 'clip_ratio/high_mean': 0.009757570689544082, 'clip_ratio/high_max': 0.009757570689544082, 'clip_ratio/region_mean': 0.016100572131108493, 'reward_total_mean': 0.9418361783027649, 'reward_meter_mean': 0.9972676634788513, 'reward_meter_std': 0.0014413135359063745, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.08399210125207901, 'reward_total_composite_mean': 0.9418361783027649, 'reward_total_composite_std': 0.08344186097383499, 'epoch': 0.09}

 65%|██████▌   | 2150/3300 [2:54:31<1:48:35,  5.67s/it]INFO 04-12 01:31:22 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:34,  3.09s/it]

 23%|██▎       | 3/13 [00:09<00:33,  3.39s/it]

 31%|███       | 4/13 [00:15<00:37,  4.12s/it]

 38%|███▊      | 5/13 [00:20<00:35,  4.40s/it]

 46%|████▌     | 6/13 [00:25<00:33,  4.74s/it]

 54%|█████▍    | 7/13 [00:30<00:29,  4.91s/it]

 62%|██████▏   | 8/13 [00:37<00:26,  5.37s/it]

 69%|██████▉   | 9/13 [00:43<00:22,  5.69s/it]

 77%|███████▋  | 10/13 [00:48<00:16,  5.37s/it]

 85%|████████▍ | 11/13 [00:53<00:10,  5.40s/it]

 92%|█████████▏| 12/13 [00:59<00:05,  5.47s/it]

100%|██████████| 13/13 [01:04<00:00,  5.38s/it]2026-04-12 01:32:33,849 | INFO | train_grpo_train | metrics_logged mode=eval step=2150

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 71.0335, 'eval_samples_per_second': 1.464, 'eval_steps_per_second': 0.183, 'eval_num_tokens': 4854574.0, 'eval_completions/mean_length': 208.15384615384616, 'eval_completions/min_length': 61.23076923076923, 'eval_completions/max_length': 377.61538461538464, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 208.15384615384616, 'eval_completions/min_terminated_length': 61.23076923076923, 'eval_completions/max_terminated_length': 377.61538461538464, 'eval_rewards/meter/mean': 0.7046629740641668, 'eval_rewards/meter/std': 0.41940173277488124, 'eval_rewards/count_adherence/mean': 0.9253275165191064, 'eval_rewards/count_adherence/std': 0.10161341182314433, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.843710142832536, 'eval_rewards/repeat_penalty/std': 0.15254983076682457, 'eval_rewards/total_composite/mean': 0.5750110378632178, 'eval_rewards/total_composite/std': 0.38267656473013073, 'eval_reward': 0.5750110378632178, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.02115864851153814, 'eval_sampling/sampling_logp_difference/max': 1.116743463736314, 'eval_sampling/importance_sampling_ratio/min': 0.3407066028851729, 'eval_sampling/importance_sampling_ratio/mean': 1.0052896371254554, 'eval_sampling/importance_sampling_ratio/max': 1.5623481090252216, 'eval_entropy': 0.20996456879835862, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5750110378632178, 'eval_reward_meter_mean': 0.7046629740641668, 'eval_reward_meter_std': 0.41940173277488124, 'eval_reward_count_adherence_mean': 0.9253275165191064, 'eval_reward_count_adherence_std': 0.10161341182314433, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.843710142832536, 'eval_reward_repeat_penalty_std': 0.15254983076682457, 'eval_reward_total_composite_mean': 0.5750110378632178, 'eval_reward_total_composite_std': 0.38267656473013073, 'epoch': 0.09}

 65%|██████▌   | 2150/3300 [2:55:42<1:48:35,  5.67s/it]

100%|██████████| 13/13 [01:04<00:00,  5.38s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 65%|██████▌   | 2151/3300 [2:55:55<9:15:08, 28.99s/it]2026-04-12 01:32:46,143 | INFO | train_grpo_train | metrics_logged mode=train step=2151

                                                       
{'loss': -0.0636, 'grad_norm': 3.3067500591278076, 'learning_rate': 3.4848484848484854e-06, 'num_tokens': 4856155.0, 'completions/mean_length': 101.625, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.000003814697266, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.4682069420814514, 'rewards/meter/std': 0.35492372512817383, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.4681951105594635, 'rewards/total_composite/std': 0.35494157671928406, 'reward': 0.4681951105594635, 'reward_std': 0.35494154691696167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11403389275074005, 'sampling/sampling_logp_difference/max': 2.580704689025879, 'sampling/importance_sampling_ratio/min': 0.07572062313556671, 'sampling/importance_sampling_ratio/mean': 1.0368061065673828, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9084643498063087, 'clip_ratio/low_mean': 0.06512659136205912, 'clip_ratio/low_min': 0.06512659136205912, 'clip_ratio/high_mean': 0.023466601967811584, 'clip_ratio/high_max': 0.023466601967811584, 'clip_ratio/region_mean': 0.0885931933298707, 'reward_total_mean': 0.4681951105594635, 'reward_meter_mean': 0.4682069420814514, 'reward_meter_std': 0.35492372512817383, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.4681951105594635, 'reward_total_composite_std': 0.35494157671928406, 'epoch': 0.09}

 65%|██████▌   | 2151/3300 [2:55:55<9:15:08, 28.99s/it]INFO 04-12 01:32:46 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2152/3300 [2:56:02<7:09:21, 22.44s/it]2026-04-12 01:32:53,304 | INFO | train_grpo_train | metrics_logged mode=train step=2152

                                                       
{'loss': 0.0533, 'grad_norm': 6.426714897155762, 'learning_rate': 3.481818181818182e-06, 'num_tokens': 4859871.0, 'completions/mean_length': 247.5, 'completions/min_length': 238.0, 'completions/max_length': 276.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 247.5, 'completions/min_terminated_length': 238.0, 'completions/max_terminated_length': 276.0, 'rewards/meter/mean': 0.9982793927192688, 'rewards/meter/std': 0.00048778695054352283, 'rewards/count_adherence/mean': 0.9642857313156128, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8910256624221802, 'rewards/repeat_penalty/std': 0.06309091299772263, 'rewards/total_composite/mean': 0.8597774505615234, 'rewards/total_composite/std': 0.10398055613040924, 'reward': 0.8597774505615234, 'reward_std': 0.10398054867982864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028371067717671394, 'sampling/sampling_logp_difference/max': 3.265172243118286, 'sampling/importance_sampling_ratio/min': 0.03819035738706589, 'sampling/importance_sampling_ratio/mean': 1.0031156539916992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1678062528371811, 'clip_ratio/low_mean': 0.010588002507574856, 'clip_ratio/low_min': 0.010588002507574856, 'clip_ratio/high_mean': 0.0109593840315938, 'clip_ratio/high_max': 0.0109593840315938, 'clip_ratio/region_mean': 0.021547386539168656, 'reward_total_mean': 0.8597774505615234, 'reward_meter_mean': 0.9982793927192688, 'reward_meter_std': 0.00048778695054352283, 'reward_count_adherence_mean': 0.9642857313156128, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8910256624221802, 'reward_repeat_penalty_std': 0.06309091299772263, 'reward_total_composite_mean': 0.8597774505615234, 'reward_total_composite_std': 0.10398055613040924, 'epoch': 0.09}

 65%|██████▌   | 2152/3300 [2:56:02<7:09:21, 22.44s/it]INFO 04-12 01:32:53 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2153/3300 [2:56:06<5:26:27, 17.08s/it]2026-04-12 01:32:57,880 | INFO | train_grpo_train | metrics_logged mode=train step=2153

                                                       
{'loss': 0.0, 'grad_norm': 0.24814273416996002, 'learning_rate': 3.4787878787878795e-06, 'num_tokens': 4861871.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990808367729187, 'rewards/meter/std': 1.3548809874919243e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990808367729187, 'rewards/total_composite/std': 1.3548809874919243e-05, 'reward': 0.9990808367729187, 'reward_std': 1.3551140909839887e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011720891110599041, 'sampling/sampling_logp_difference/max': 1.327371597290039, 'sampling/importance_sampling_ratio/min': 0.2651733458042145, 'sampling/importance_sampling_ratio/mean': 1.0004948377609253, 'sampling/importance_sampling_ratio/max': 1.5080136060714722, 'entropy': 0.07956603448837996, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.007575757801532745, 'reward_total_mean': 0.9990808367729187, 'reward_meter_mean': 0.9990808367729187, 'reward_meter_std': 1.3548809874919243e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990808367729187, 'reward_total_composite_std': 1.3548809874919243e-05, 'epoch': 0.09}

 65%|██████▌   | 2153/3300 [2:56:06<5:26:27, 17.08s/it]INFO 04-12 01:32:58 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2154/3300 [2:56:11<4:15:13, 13.36s/it]2026-04-12 01:33:02,556 | INFO | train_grpo_train | metrics_logged mode=train step=2154

                                                       
{'loss': 0.0058, 'grad_norm': 3.136981248855591, 'learning_rate': 3.4757575757575763e-06, 'num_tokens': 4863701.0, 'completions/mean_length': 72.75, 'completions/min_length': 71.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.75, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9944970607757568, 'rewards/meter/std': 0.004257019143551588, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9944970607757568, 'rewards/total_composite/std': 0.004257019143551588, 'reward': 0.9944970607757568, 'reward_std': 0.004257019609212875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03498362377285957, 'sampling/sampling_logp_difference/max': 1.4066009521484375, 'sampling/importance_sampling_ratio/min': 0.2449745386838913, 'sampling/importance_sampling_ratio/mean': 1.0077894926071167, 'sampling/importance_sampling_ratio/max': 1.6701947450637817, 'entropy': 0.23252216633409262, 'clip_ratio/low_mean': 0.011893743183463812, 'clip_ratio/low_min': 0.011893743183463812, 'clip_ratio/high_mean': 0.022406263276934624, 'clip_ratio/high_max': 0.022406263276934624, 'clip_ratio/region_mean': 0.034300006460398436, 'reward_total_mean': 0.9944970607757568, 'reward_meter_mean': 0.9944970607757568, 'reward_meter_std': 0.004257019143551588, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9944970607757568, 'reward_total_composite_std': 0.004257019143551588, 'epoch': 0.09}

 65%|██████▌   | 2154/3300 [2:56:11<4:15:13, 13.36s/it]INFO 04-12 01:33:02 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2155/3300 [2:56:15<3:24:16, 10.70s/it]2026-04-12 01:33:07,054 | INFO | train_grpo_train | metrics_logged mode=train step=2155

                                                       
{'loss': -0.0003, 'grad_norm': 0.40986379981040955, 'learning_rate': 3.472727272727273e-06, 'num_tokens': 4865461.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990643262863159, 'rewards/meter/std': 2.489090729795862e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990643262863159, 'rewards/total_composite/std': 2.489090729795862e-05, 'reward': 0.9990643262863159, 'reward_std': 2.4881244826246984e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005097586195915937, 'sampling/sampling_logp_difference/max': 0.27621328830718994, 'sampling/importance_sampling_ratio/min': 0.8046210408210754, 'sampling/importance_sampling_ratio/mean': 1.0042916536331177, 'sampling/importance_sampling_ratio/max': 1.318129062652588, 'entropy': 0.07073036208748817, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.005681818351149559, 'reward_total_mean': 0.9990643262863159, 'reward_meter_mean': 0.9990643262863159, 'reward_meter_std': 2.489090729795862e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990643262863159, 'reward_total_composite_std': 2.489090729795862e-05, 'epoch': 0.09}

 65%|██████▌   | 2155/3300 [2:56:16<3:24:16, 10.70s/it]INFO 04-12 01:33:07 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2156/3300 [2:56:21<2:51:49,  9.01s/it]2026-04-12 01:33:12,120 | INFO | train_grpo_train | metrics_logged mode=train step=2156

                                                       
{'loss': -0.0024, 'grad_norm': 1.6892346143722534, 'learning_rate': 3.46969696969697e-06, 'num_tokens': 4867682.0, 'completions/mean_length': 117.625, 'completions/min_length': 116.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.625, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.9963239431381226, 'rewards/meter/std': 0.0006225909455679357, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9429587125778198, 'rewards/total_composite/std': 0.07379542291164398, 'reward': 0.9429587125778198, 'reward_std': 0.07379541546106339, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020759226754307747, 'sampling/sampling_logp_difference/max': 1.4962215423583984, 'sampling/importance_sampling_ratio/min': 0.22397485375404358, 'sampling/importance_sampling_ratio/mean': 1.0034620761871338, 'sampling/importance_sampling_ratio/max': 1.3534270524978638, 'entropy': 0.13127075601369143, 'clip_ratio/low_mean': 0.0032144944416359067, 'clip_ratio/low_min': 0.0032144944416359067, 'clip_ratio/high_mean': 0.009525301051326096, 'clip_ratio/high_max': 0.009525301051326096, 'clip_ratio/region_mean': 0.012739795492962003, 'reward_total_mean': 0.9429587125778198, 'reward_meter_mean': 0.9963239431381226, 'reward_meter_std': 0.0006225909455679357, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9429587125778198, 'reward_total_composite_std': 0.07379542291164398, 'epoch': 0.09}

 65%|██████▌   | 2156/3300 [2:56:21<2:51:49,  9.01s/it]INFO 04-12 01:33:12 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2157/3300 [2:56:25<2:24:30,  7.59s/it]2026-04-12 01:33:16,381 | INFO | train_grpo_train | metrics_logged mode=train step=2157

                                                       
{'loss': 0.0158, 'grad_norm': 1.6599547863006592, 'learning_rate': 3.4666666666666672e-06, 'num_tokens': 4869399.0, 'completions/mean_length': 56.625, 'completions/min_length': 54.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.725243330001831, 'rewards/meter/std': 0.03141331672668457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.725243330001831, 'rewards/total_composite/std': 0.03141331672668457, 'reward': 0.725243330001831, 'reward_std': 0.03141331672668457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01150722336024046, 'sampling/sampling_logp_difference/max': 0.5583809614181519, 'sampling/importance_sampling_ratio/min': 0.5721346139907837, 'sampling/importance_sampling_ratio/mean': 1.0009208917617798, 'sampling/importance_sampling_ratio/max': 1.4941579103469849, 'entropy': 0.054770099464803934, 'clip_ratio/low_mean': 0.004385964944958687, 'clip_ratio/low_min': 0.004385964944958687, 'clip_ratio/high_mean': 0.00889376224949956, 'clip_ratio/high_max': 0.00889376224949956, 'clip_ratio/region_mean': 0.013279727194458246, 'reward_total_mean': 0.725243330001831, 'reward_meter_mean': 0.725243330001831, 'reward_meter_std': 0.03141331672668457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.725243330001831, 'reward_total_composite_std': 0.03141331672668457, 'epoch': 0.09}

 65%|██████▌   | 2157/3300 [2:56:25<2:24:30,  7.59s/it]INFO 04-12 01:33:16 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2158/3300 [2:56:29<2:06:23,  6.64s/it]2026-04-12 01:33:20,812 | INFO | train_grpo_train | metrics_logged mode=train step=2158

                                                       
{'loss': 0.0002, 'grad_norm': 0.1878003627061844, 'learning_rate': 3.463636363636364e-06, 'num_tokens': 4871311.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990821480751038, 'rewards/meter/std': 1.2787881132680923e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990821480751038, 'rewards/total_composite/std': 1.2787881132680923e-05, 'reward': 0.9990821480751038, 'reward_std': 1.2793108908226714e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008319023996591568, 'sampling/sampling_logp_difference/max': 1.0997838973999023, 'sampling/importance_sampling_ratio/min': 0.33294302225112915, 'sampling/importance_sampling_ratio/mean': 1.0013294219970703, 'sampling/importance_sampling_ratio/max': 1.3374110460281372, 'entropy': 0.07476874813437462, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.005681818351149559, 'clip_ratio/high_max': 0.005681818351149559, 'clip_ratio/region_mean': 0.009469697251915932, 'reward_total_mean': 0.9990821480751038, 'reward_meter_mean': 0.9990821480751038, 'reward_meter_std': 1.2787881132680923e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990821480751038, 'reward_total_composite_std': 1.2787881132680923e-05, 'epoch': 0.09}

 65%|██████▌   | 2158/3300 [2:56:29<2:06:23,  6.64s/it]INFO 04-12 01:33:21 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2159/3300 [2:56:34<1:54:06,  6.00s/it]2026-04-12 01:33:25,318 | INFO | train_grpo_train | metrics_logged mode=train step=2159

                                                       
{'loss': 0.0063, 'grad_norm': 3.910881280899048, 'learning_rate': 3.460606060606061e-06, 'num_tokens': 4873104.0, 'completions/mean_length': 56.125, 'completions/min_length': 56.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9947191476821899, 'rewards/meter/std': 0.0006506324862129986, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9947191476821899, 'rewards/total_composite/std': 0.0006506324862129986, 'reward': 0.9947191476821899, 'reward_std': 0.0006506353965960443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009572095237672329, 'sampling/sampling_logp_difference/max': 0.4652571678161621, 'sampling/importance_sampling_ratio/min': 0.6279736161231995, 'sampling/importance_sampling_ratio/mean': 1.002916693687439, 'sampling/importance_sampling_ratio/max': 1.261582612991333, 'entropy': 0.093318160623312, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0022321429569274187, 'clip_ratio/high_max': 0.0022321429569274187, 'clip_ratio/region_mean': 0.0022321429569274187, 'reward_total_mean': 0.9947191476821899, 'reward_meter_mean': 0.9947191476821899, 'reward_meter_std': 0.0006506324862129986, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9947191476821899, 'reward_total_composite_std': 0.0006506324862129986, 'epoch': 0.09}

 65%|██████▌   | 2159/3300 [2:56:34<1:54:06,  6.00s/it]INFO 04-12 01:33:25 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2160/3300 [2:56:38<1:46:29,  5.60s/it]2026-04-12 01:33:29,999 | INFO | train_grpo_train | metrics_logged mode=train step=2160

                                                       
{'loss': 0.0003, 'grad_norm': 1.6953576803207397, 'learning_rate': 3.4575757575757577e-06, 'num_tokens': 4875040.0, 'completions/mean_length': 87.0, 'completions/min_length': 86.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.0, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9958928227424622, 'rewards/meter/std': 0.000379539153072983, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9958928227424622, 'rewards/total_composite/std': 0.000379539153072983, 'reward': 0.9958928227424622, 'reward_std': 0.000379539153072983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018249215558171272, 'sampling/sampling_logp_difference/max': 0.9508843421936035, 'sampling/importance_sampling_ratio/min': 0.38639914989471436, 'sampling/importance_sampling_ratio/mean': 1.0018218755722046, 'sampling/importance_sampling_ratio/max': 1.3476742506027222, 'entropy': 0.11874265689402819, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.01438528229482472, 'clip_ratio/high_max': 0.01438528229482472, 'clip_ratio/region_mean': 0.01869562710635364, 'reward_total_mean': 0.9958928227424622, 'reward_meter_mean': 0.9958928227424622, 'reward_meter_std': 0.000379539153072983, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9958928227424622, 'reward_total_composite_std': 0.000379539153072983, 'epoch': 0.09}

 65%|██████▌   | 2160/3300 [2:56:38<1:46:29,  5.60s/it]INFO 04-12 01:33:30 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2161/3300 [2:56:43<1:39:24,  5.24s/it]2026-04-12 01:33:34,380 | INFO | train_grpo_train | metrics_logged mode=train step=2161

                                                       
{'loss': 0.0029, 'grad_norm': 2.359445810317993, 'learning_rate': 3.454545454545455e-06, 'num_tokens': 4876737.0, 'completions/mean_length': 56.125, 'completions/min_length': 56.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9948000907897949, 'rewards/meter/std': 0.0004447593819350004, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948000907897949, 'rewards/total_composite/std': 0.0004447593819350004, 'reward': 0.9948000907897949, 'reward_std': 0.00044475641334429383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007567924913018942, 'sampling/sampling_logp_difference/max': 0.8509788513183594, 'sampling/importance_sampling_ratio/min': 0.42699676752090454, 'sampling/importance_sampling_ratio/mean': 1.0003728866577148, 'sampling/importance_sampling_ratio/max': 1.2896134853363037, 'entropy': 0.06367568951100111, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006696428870782256, 'clip_ratio/high_max': 0.006696428870782256, 'clip_ratio/region_mean': 0.006696428870782256, 'reward_total_mean': 0.9948000907897949, 'reward_meter_mean': 0.9948000907897949, 'reward_meter_std': 0.0004447593819350004, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948000907897949, 'reward_total_composite_std': 0.0004447593819350004, 'epoch': 0.09}

 65%|██████▌   | 2161/3300 [2:56:43<1:39:24,  5.24s/it]INFO 04-12 01:33:34 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2162/3300 [2:56:48<1:36:27,  5.09s/it]2026-04-12 01:33:39,106 | INFO | train_grpo_train | metrics_logged mode=train step=2162

                                                       
{'loss': -0.0273, 'grad_norm': 4.7188615798950195, 'learning_rate': 3.451515151515152e-06, 'num_tokens': 4878413.0, 'completions/mean_length': 53.5, 'completions/min_length': 49.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7199991941452026, 'rewards/meter/std': 0.19119207561016083, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333730697632, 'rewards/repeat_penalty/std': 0.117851123213768, 'rewards/total_composite/mean': 0.709714949131012, 'rewards/total_composite/std': 0.2202804535627365, 'reward': 0.709714949131012, 'reward_std': 0.22028043866157532, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014019476249814034, 'sampling/sampling_logp_difference/max': 1.523421287536621, 'sampling/importance_sampling_ratio/min': 0.21796490252017975, 'sampling/importance_sampling_ratio/mean': 0.9992796778678894, 'sampling/importance_sampling_ratio/max': 1.4815089702606201, 'entropy': 0.08292486914433539, 'clip_ratio/low_mean': 0.0025510203558951616, 'clip_ratio/low_min': 0.0025510203558951616, 'clip_ratio/high_mean': 0.0022727272007614374, 'clip_ratio/high_max': 0.0022727272007614374, 'clip_ratio/region_mean': 0.004823747556656599, 'reward_total_mean': 0.709714949131012, 'reward_meter_mean': 0.7199991941452026, 'reward_meter_std': 0.19119207561016083, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333730697632, 'reward_repeat_penalty_std': 0.117851123213768, 'reward_total_composite_mean': 0.709714949131012, 'reward_total_composite_std': 0.2202804535627365, 'epoch': 0.09}

 66%|██████▌   | 2162/3300 [2:56:48<1:36:27,  5.09s/it]INFO 04-12 01:33:39 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2163/3300 [2:56:54<1:46:58,  5.65s/it]2026-04-12 01:33:46,062 | INFO | train_grpo_train | metrics_logged mode=train step=2163

                                                       
{'loss': 0.0136, 'grad_norm': 2.0742194652557373, 'learning_rate': 3.4484848484848486e-06, 'num_tokens': 4882602.0, 'completions/mean_length': 254.625, 'completions/min_length': 248.0, 'completions/max_length': 272.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 254.625, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 272.0, 'rewards/meter/mean': 0.9978955984115601, 'rewards/meter/std': 0.0005918731912970543, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9340659379959106, 'rewards/repeat_penalty/std': 0.062163230031728745, 'rewards/total_composite/mean': 0.8155835270881653, 'rewards/total_composite/std': 0.05420287698507309, 'reward': 0.8155835270881653, 'reward_std': 0.054202865809202194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04149610176682472, 'sampling/sampling_logp_difference/max': 16.053680419921875, 'sampling/importance_sampling_ratio/min': 1.0665352334626732e-07, 'sampling/importance_sampling_ratio/mean': 1.0046485662460327, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32685666531324387, 'clip_ratio/low_mean': 0.00980698096100241, 'clip_ratio/low_min': 0.00980698096100241, 'clip_ratio/high_mean': 0.011834641918540001, 'clip_ratio/high_max': 0.011834641918540001, 'clip_ratio/region_mean': 0.02164162287954241, 'reward_total_mean': 0.8155835270881653, 'reward_meter_mean': 0.9978955984115601, 'reward_meter_std': 0.0005918731912970543, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9340659379959106, 'reward_repeat_penalty_std': 0.062163230031728745, 'reward_total_composite_mean': 0.8155835270881653, 'reward_total_composite_std': 0.05420287698507309, 'epoch': 0.09}

 66%|██████▌   | 2163/3300 [2:56:55<1:46:58,  5.65s/it]INFO 04-12 01:33:46 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2164/3300 [2:56:59<1:40:23,  5.30s/it]2026-04-12 01:33:50,564 | INFO | train_grpo_train | metrics_logged mode=train step=2164

                                                       
{'loss': -0.007, 'grad_norm': 3.962695837020874, 'learning_rate': 3.445454545454546e-06, 'num_tokens': 4884392.0, 'completions/mean_length': 67.75, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9326764345169067, 'rewards/meter/std': 0.024432551115751266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9326764345169067, 'rewards/total_composite/std': 0.024432551115751266, 'reward': 0.9326764345169067, 'reward_std': 0.024432554841041565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021882515400648117, 'sampling/sampling_logp_difference/max': 1.0846776962280273, 'sampling/importance_sampling_ratio/min': 0.3380107283592224, 'sampling/importance_sampling_ratio/mean': 1.0046216249465942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14008523616939783, 'clip_ratio/low_mean': 0.00378874852322042, 'clip_ratio/low_min': 0.00378874852322042, 'clip_ratio/high_mean': 0.0055147059028968215, 'clip_ratio/high_max': 0.0055147059028968215, 'clip_ratio/region_mean': 0.009303454426117241, 'reward_total_mean': 0.9326764345169067, 'reward_meter_mean': 0.9326764345169067, 'reward_meter_std': 0.024432551115751266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9326764345169067, 'reward_total_composite_std': 0.024432551115751266, 'epoch': 0.09}

 66%|██████▌   | 2164/3300 [2:56:59<1:40:23,  5.30s/it]INFO 04-12 01:33:50 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2165/3300 [2:57:04<1:36:16,  5.09s/it]2026-04-12 01:33:55,158 | INFO | train_grpo_train | metrics_logged mode=train step=2165

                                                       
{'loss': -0.0061, 'grad_norm': 3.4584054946899414, 'learning_rate': 3.4424242424242427e-06, 'num_tokens': 4886154.0, 'completions/mean_length': 66.25, 'completions/min_length': 66.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9357738494873047, 'rewards/meter/std': 0.17855486273765564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9357738494873047, 'rewards/total_composite/std': 0.17855486273765564, 'reward': 0.9357738494873047, 'reward_std': 0.17855484783649445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030108362436294556, 'sampling/sampling_logp_difference/max': 7.794986724853516, 'sampling/importance_sampling_ratio/min': 0.0004117942589800805, 'sampling/importance_sampling_ratio/mean': 1.00041925907135, 'sampling/importance_sampling_ratio/max': 1.5800968408584595, 'entropy': 0.10618306044489145, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007464349502697587, 'clip_ratio/high_max': 0.007464349502697587, 'clip_ratio/region_mean': 0.007464349502697587, 'reward_total_mean': 0.9357738494873047, 'reward_meter_mean': 0.9357738494873047, 'reward_meter_std': 0.17855486273765564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9357738494873047, 'reward_total_composite_std': 0.17855486273765564, 'epoch': 0.09}

 66%|██████▌   | 2165/3300 [2:57:04<1:36:16,  5.09s/it]INFO 04-12 01:33:55 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2166/3300 [2:57:09<1:36:39,  5.11s/it]2026-04-12 01:34:00,328 | INFO | train_grpo_train | metrics_logged mode=train step=2166

                                                       
{'loss': 0.0044, 'grad_norm': 2.468413829803467, 'learning_rate': 3.4393939393939395e-06, 'num_tokens': 4888587.0, 'completions/mean_length': 116.125, 'completions/min_length': 112.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.125, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.996168851852417, 'rewards/meter/std': 0.0008722886559553444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.9249789118766785, 'rewards/total_composite/std': 0.07558359950780869, 'reward': 0.9249789118766785, 'reward_std': 0.07558359950780869, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019451651722192764, 'sampling/sampling_logp_difference/max': 0.8788247108459473, 'sampling/importance_sampling_ratio/min': 0.41527071595191956, 'sampling/importance_sampling_ratio/mean': 1.0041885375976562, 'sampling/importance_sampling_ratio/max': 1.6350321769714355, 'entropy': 0.13098172564059496, 'clip_ratio/low_mean': 0.005332982516847551, 'clip_ratio/low_min': 0.005332982516847551, 'clip_ratio/high_mean': 0.008708675275556743, 'clip_ratio/high_max': 0.008708675275556743, 'clip_ratio/region_mean': 0.014041657792404294, 'reward_total_mean': 0.9249789118766785, 'reward_meter_mean': 0.996168851852417, 'reward_meter_std': 0.0008722886559553444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.9249789118766785, 'reward_total_composite_std': 0.07558359950780869, 'epoch': 0.09}

 66%|██████▌   | 2166/3300 [2:57:09<1:36:39,  5.11s/it]INFO 04-12 01:34:00 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2167/3300 [2:57:13<1:33:35,  4.96s/it]2026-04-12 01:34:04,918 | INFO | train_grpo_train | metrics_logged mode=train step=2167

                                                       
{'loss': 0.0303, 'grad_norm': 7.104674816131592, 'learning_rate': 3.4363636363636364e-06, 'num_tokens': 4890379.0, 'completions/mean_length': 73.0, 'completions/min_length': 70.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9878561496734619, 'rewards/meter/std': 0.010053827427327633, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9878561496734619, 'rewards/total_composite/std': 0.010053827427327633, 'reward': 0.9878561496734619, 'reward_std': 0.010053832083940506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.037653032690286636, 'sampling/sampling_logp_difference/max': 1.5957527160644531, 'sampling/importance_sampling_ratio/min': 0.2027558535337448, 'sampling/importance_sampling_ratio/mean': 1.0013821125030518, 'sampling/importance_sampling_ratio/max': 1.616224765777588, 'entropy': 0.2611173652112484, 'clip_ratio/low_mean': 0.008205128135159612, 'clip_ratio/low_min': 0.008205128135159612, 'clip_ratio/high_mean': 0.024240323924459517, 'clip_ratio/high_max': 0.024240323924459517, 'clip_ratio/region_mean': 0.03244545205961913, 'reward_total_mean': 0.9878561496734619, 'reward_meter_mean': 0.9878561496734619, 'reward_meter_std': 0.010053827427327633, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9878561496734619, 'reward_total_composite_std': 0.010053827427327633, 'epoch': 0.09}

 66%|██████▌   | 2167/3300 [2:57:13<1:33:35,  4.96s/it]INFO 04-12 01:34:05 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2168/3300 [2:57:17<1:28:06,  4.67s/it]2026-04-12 01:34:08,922 | INFO | train_grpo_train | metrics_logged mode=train step=2168

                                                       
{'loss': -0.0006, 'grad_norm': 0.1801041215658188, 'learning_rate': 3.4333333333333336e-06, 'num_tokens': 4891883.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9979922771453857, 'rewards/meter/std': 1.613328822713811e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979922771453857, 'rewards/total_composite/std': 1.613328822713811e-05, 'reward': 0.9979922771453857, 'reward_std': 1.613025233382359e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005629383493214846, 'sampling/sampling_logp_difference/max': 0.34334421157836914, 'sampling/importance_sampling_ratio/min': 0.7093939781188965, 'sampling/importance_sampling_ratio/mean': 1.0019093751907349, 'sampling/importance_sampling_ratio/max': 1.0980141162872314, 'entropy': 0.051043334417045116, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0035714285913854837, 'clip_ratio/high_max': 0.0035714285913854837, 'clip_ratio/region_mean': 0.0035714285913854837, 'reward_total_mean': 0.9979922771453857, 'reward_meter_mean': 0.9979922771453857, 'reward_meter_std': 1.613328822713811e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979922771453857, 'reward_total_composite_std': 1.613328822713811e-05, 'epoch': 0.09}

 66%|██████▌   | 2168/3300 [2:57:17<1:28:06,  4.67s/it]INFO 04-12 01:34:09 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2169/3300 [2:57:25<1:47:25,  5.70s/it]2026-04-12 01:34:17,019 | INFO | train_grpo_train | metrics_logged mode=train step=2169

                                                       
{'loss': 0.0195, 'grad_norm': 1.7812119722366333, 'learning_rate': 3.4303030303030305e-06, 'num_tokens': 4896528.0, 'completions/mean_length': 346.625, 'completions/min_length': 336.0, 'completions/max_length': 369.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 346.625, 'completions/min_terminated_length': 336.0, 'completions/max_terminated_length': 369.0, 'rewards/meter/mean': 0.9944499731063843, 'rewards/meter/std': 0.0038915579207241535, 'rewards/count_adherence/mean': 0.701923131942749, 'rewards/count_adherence/std': 0.027196412906050682, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8722910284996033, 'rewards/repeat_penalty/std': 0.08643024414777756, 'rewards/total_composite/mean': 0.6071589589118958, 'rewards/total_composite/std': 0.0453052781522274, 'reward': 0.6071589589118958, 'reward_std': 0.045305285602808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03724047169089317, 'sampling/sampling_logp_difference/max': 5.4400858879089355, 'sampling/importance_sampling_ratio/min': 0.004339110571891069, 'sampling/importance_sampling_ratio/mean': 1.0060648918151855, 'sampling/importance_sampling_ratio/max': 1.909399151802063, 'entropy': 0.32041275314986706, 'clip_ratio/low_mean': 0.006560013280250132, 'clip_ratio/low_min': 0.006560013280250132, 'clip_ratio/high_mean': 0.022110585356131196, 'clip_ratio/high_max': 0.022110585356131196, 'clip_ratio/region_mean': 0.028670598636381328, 'reward_total_mean': 0.6071589589118958, 'reward_meter_mean': 0.9944499731063843, 'reward_meter_std': 0.0038915579207241535, 'reward_count_adherence_mean': 0.701923131942749, 'reward_count_adherence_std': 0.027196412906050682, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8722910284996033, 'reward_repeat_penalty_std': 0.08643024414777756, 'reward_total_composite_mean': 0.6071589589118958, 'reward_total_composite_std': 0.0453052781522274, 'epoch': 0.09}

 66%|██████▌   | 2169/3300 [2:57:26<1:47:25,  5.70s/it]INFO 04-12 01:34:17 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2170/3300 [2:57:30<1:41:15,  5.38s/it]2026-04-12 01:34:21,643 | INFO | train_grpo_train | metrics_logged mode=train step=2170

                                                       
{'loss': -0.0018, 'grad_norm': 2.5535831451416016, 'learning_rate': 3.4272727272727273e-06, 'num_tokens': 4898436.0, 'completions/mean_length': 75.5, 'completions/min_length': 73.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9986087083816528, 'rewards/meter/std': 0.0005744586233049631, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9986087083816528, 'rewards/total_composite/std': 0.0005744586233049631, 'reward': 0.9986087083816528, 'reward_std': 0.0005744565278291702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032794393599033356, 'sampling/sampling_logp_difference/max': 1.1777338981628418, 'sampling/importance_sampling_ratio/min': 0.3079758584499359, 'sampling/importance_sampling_ratio/mean': 1.0089645385742188, 'sampling/importance_sampling_ratio/max': 1.619355320930481, 'entropy': 0.34711651876568794, 'clip_ratio/low_mean': 0.008402727777138352, 'clip_ratio/low_min': 0.008402727777138352, 'clip_ratio/high_mean': 0.016413721488788724, 'clip_ratio/high_max': 0.016413721488788724, 'clip_ratio/region_mean': 0.024816449265927076, 'reward_total_mean': 0.9986087083816528, 'reward_meter_mean': 0.9986087083816528, 'reward_meter_std': 0.0005744586233049631, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9986087083816528, 'reward_total_composite_std': 0.0005744586233049631, 'epoch': 0.09}

 66%|██████▌   | 2170/3300 [2:57:30<1:41:15,  5.38s/it]INFO 04-12 01:34:21 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2171/3300 [2:57:35<1:37:28,  5.18s/it]2026-04-12 01:34:26,383 | INFO | train_grpo_train | metrics_logged mode=train step=2171

                                                       
{'loss': 0.0221, 'grad_norm': 3.1598801612854004, 'learning_rate': 3.4242424242424246e-06, 'num_tokens': 4900337.0, 'completions/mean_length': 75.625, 'completions/min_length': 72.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9978922009468079, 'rewards/meter/std': 0.00160902866628021, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978922009468079, 'rewards/total_composite/std': 0.00160902866628021, 'reward': 0.9978922009468079, 'reward_std': 0.001609029364772141, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03502906486392021, 'sampling/sampling_logp_difference/max': 1.1911673545837402, 'sampling/importance_sampling_ratio/min': 0.30386632680892944, 'sampling/importance_sampling_ratio/mean': 1.0061036348342896, 'sampling/importance_sampling_ratio/max': 1.6152563095092773, 'entropy': 0.3033796586096287, 'clip_ratio/low_mean': 0.010196133749559522, 'clip_ratio/low_min': 0.010196133749559522, 'clip_ratio/high_mean': 0.018138289218768477, 'clip_ratio/high_max': 0.018138289218768477, 'clip_ratio/region_mean': 0.028334422968328, 'reward_total_mean': 0.9978922009468079, 'reward_meter_mean': 0.9978922009468079, 'reward_meter_std': 0.00160902866628021, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978922009468079, 'reward_total_composite_std': 0.00160902866628021, 'epoch': 0.09}

 66%|██████▌   | 2171/3300 [2:57:35<1:37:28,  5.18s/it]INFO 04-12 01:34:26 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2172/3300 [2:57:41<1:44:19,  5.55s/it]2026-04-12 01:34:32,790 | INFO | train_grpo_train | metrics_logged mode=train step=2172

                                                       
{'loss': 0.0003, 'grad_norm': 2.200758695602417, 'learning_rate': 3.4212121212121214e-06, 'num_tokens': 4903772.0, 'completions/mean_length': 216.375, 'completions/min_length': 213.0, 'completions/max_length': 219.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 216.375, 'completions/min_terminated_length': 213.0, 'completions/max_terminated_length': 219.0, 'rewards/meter/mean': 0.9980002641677856, 'rewards/meter/std': 0.00031345427851192653, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9431818723678589, 'rewards/repeat_penalty/std': 0.06763852387666702, 'rewards/total_composite/mean': 0.8068255186080933, 'rewards/total_composite/std': 0.05786222964525223, 'reward': 0.8068255186080933, 'reward_std': 0.057862233370542526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028192421421408653, 'sampling/sampling_logp_difference/max': 1.0793828964233398, 'sampling/importance_sampling_ratio/min': 0.33980512619018555, 'sampling/importance_sampling_ratio/mean': 1.0067384243011475, 'sampling/importance_sampling_ratio/max': 1.921886920928955, 'entropy': 0.3047393374145031, 'clip_ratio/low_mean': 0.012229247367940843, 'clip_ratio/low_min': 0.012229247367940843, 'clip_ratio/high_mean': 0.012681030901148915, 'clip_ratio/high_max': 0.012681030901148915, 'clip_ratio/region_mean': 0.02491027826908976, 'reward_total_mean': 0.8068255186080933, 'reward_meter_mean': 0.9980002641677856, 'reward_meter_std': 0.00031345427851192653, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9431818723678589, 'reward_repeat_penalty_std': 0.06763852387666702, 'reward_total_composite_mean': 0.8068255186080933, 'reward_total_composite_std': 0.05786222964525223, 'epoch': 0.09}

 66%|██████▌   | 2172/3300 [2:57:41<1:44:19,  5.55s/it]INFO 04-12 01:34:33 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2173/3300 [2:57:46<1:38:31,  5.25s/it]2026-04-12 01:34:37,330 | INFO | train_grpo_train | metrics_logged mode=train step=2173

                                                       
{'loss': -0.0002, 'grad_norm': 4.465800762176514, 'learning_rate': 3.4181818181818182e-06, 'num_tokens': 4905495.0, 'completions/mean_length': 68.375, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9475573301315308, 'rewards/meter/std': 0.03694632276892662, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9475573301315308, 'rewards/total_composite/std': 0.03694632276892662, 'reward': 0.9475573301315308, 'reward_std': 0.036946315318346024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02550523541867733, 'sampling/sampling_logp_difference/max': 0.9391798973083496, 'sampling/importance_sampling_ratio/min': 0.3909483253955841, 'sampling/importance_sampling_ratio/mean': 1.000171184539795, 'sampling/importance_sampling_ratio/max': 1.5417495965957642, 'entropy': 0.12862383667379618, 'clip_ratio/low_mean': 0.0036498295376077294, 'clip_ratio/low_min': 0.0036498295376077294, 'clip_ratio/high_mean': 0.020063025411218405, 'clip_ratio/high_max': 0.020063025411218405, 'clip_ratio/region_mean': 0.023712854948826134, 'reward_total_mean': 0.9475573301315308, 'reward_meter_mean': 0.9475573301315308, 'reward_meter_std': 0.03694632276892662, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9475573301315308, 'reward_total_composite_std': 0.03694632276892662, 'epoch': 0.09}

 66%|██████▌   | 2173/3300 [2:57:46<1:38:31,  5.25s/it]INFO 04-12 01:34:37 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2174/3300 [2:57:51<1:36:15,  5.13s/it]2026-04-12 01:34:42,180 | INFO | train_grpo_train | metrics_logged mode=train step=2174

                                                       
{'loss': -0.0049, 'grad_norm': 1.1670396327972412, 'learning_rate': 3.415151515151515e-06, 'num_tokens': 4907490.0, 'completions/mean_length': 93.375, 'completions/min_length': 93.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.375, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9865900874137878, 'rewards/meter/std': 0.00459433114156127, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9865900874137878, 'rewards/total_composite/std': 0.00459433114156127, 'reward': 0.9865900874137878, 'reward_std': 0.0045943367294967175, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013781530782580376, 'sampling/sampling_logp_difference/max': 2.305748224258423, 'sampling/importance_sampling_ratio/min': 0.09968418627977371, 'sampling/importance_sampling_ratio/mean': 0.999477744102478, 'sampling/importance_sampling_ratio/max': 1.247177004814148, 'entropy': 0.06566301081329584, 'clip_ratio/low_mean': 0.008064516121521592, 'clip_ratio/low_min': 0.008064516121521592, 'clip_ratio/high_mean': 0.0013020833721384406, 'clip_ratio/high_max': 0.0013020833721384406, 'clip_ratio/region_mean': 0.009366599493660033, 'reward_total_mean': 0.9865900874137878, 'reward_meter_mean': 0.9865900874137878, 'reward_meter_std': 0.00459433114156127, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9865900874137878, 'reward_total_composite_std': 0.00459433114156127, 'epoch': 0.09}

 66%|██████▌   | 2174/3300 [2:57:51<1:36:15,  5.13s/it]INFO 04-12 01:34:42 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2175/3300 [2:57:55<1:33:59,  5.01s/it]2026-04-12 01:34:46,919 | INFO | train_grpo_train | metrics_logged mode=train step=2175

                                                       
{'loss': -0.0057, 'grad_norm': 2.544595956802368, 'learning_rate': 3.4121212121212123e-06, 'num_tokens': 4909423.0, 'completions/mean_length': 85.625, 'completions/min_length': 84.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.625, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9955015778541565, 'rewards/meter/std': 0.00046232930617406964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955015778541565, 'rewards/total_composite/std': 0.00046232930617406964, 'reward': 0.9955015778541565, 'reward_std': 0.00046233335160650313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009426219388842583, 'sampling/sampling_logp_difference/max': 0.2666025161743164, 'sampling/importance_sampling_ratio/min': 0.7659775018692017, 'sampling/importance_sampling_ratio/mean': 1.005947232246399, 'sampling/importance_sampling_ratio/max': 1.216043472290039, 'entropy': 0.09059183578938246, 'clip_ratio/low_mean': 0.0029411765281111, 'clip_ratio/low_min': 0.0029411765281111, 'clip_ratio/high_mean': 0.0058139534667134285, 'clip_ratio/high_max': 0.0058139534667134285, 'clip_ratio/region_mean': 0.008755129994824529, 'reward_total_mean': 0.9955015778541565, 'reward_meter_mean': 0.9955015778541565, 'reward_meter_std': 0.00046232930617406964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955015778541565, 'reward_total_composite_std': 0.00046232930617406964, 'epoch': 0.09}

 66%|██████▌   | 2175/3300 [2:57:55<1:33:59,  5.01s/it]INFO 04-12 01:34:47 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2176/3300 [2:58:00<1:31:57,  4.91s/it]2026-04-12 01:34:51,594 | INFO | train_grpo_train | metrics_logged mode=train step=2176

                                                       
{'loss': -0.001, 'grad_norm': 2.064898729324341, 'learning_rate': 3.409090909090909e-06, 'num_tokens': 4911299.0, 'completions/mean_length': 80.5, 'completions/min_length': 80.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.5, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.7326533794403076, 'rewards/meter/std': 0.055831361562013626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.69842928647995, 'rewards/total_composite/std': 0.10409945249557495, 'reward': 0.69842928647995, 'reward_std': 0.10409945994615555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008284451439976692, 'sampling/sampling_logp_difference/max': 0.7730828523635864, 'sampling/importance_sampling_ratio/min': 0.4615878760814667, 'sampling/importance_sampling_ratio/mean': 0.9991859197616577, 'sampling/importance_sampling_ratio/max': 1.3141322135925293, 'entropy': 0.035707658622413874, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006137048127129674, 'clip_ratio/high_max': 0.006137048127129674, 'clip_ratio/region_mean': 0.006137048127129674, 'reward_total_mean': 0.69842928647995, 'reward_meter_mean': 0.7326533794403076, 'reward_meter_std': 0.055831361562013626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.69842928647995, 'reward_total_composite_std': 0.10409945249557495, 'epoch': 0.09}

 66%|██████▌   | 2176/3300 [2:58:00<1:31:57,  4.91s/it]INFO 04-12 01:34:51 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2177/3300 [2:58:04<1:28:30,  4.73s/it]2026-04-12 01:34:55,896 | INFO | train_grpo_train | metrics_logged mode=train step=2177

                                                       
{'loss': -0.0033, 'grad_norm': 0.053646668791770935, 'learning_rate': 3.406060606060606e-06, 'num_tokens': 4912899.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7176748514175415, 'rewards/meter/std': 0.19788797199726105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7176748514175415, 'rewards/total_composite/std': 0.19788797199726105, 'reward': 0.7176748514175415, 'reward_std': 0.19788797199726105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00239483080804348, 'sampling/sampling_logp_difference/max': 0.5967006683349609, 'sampling/importance_sampling_ratio/min': 0.5506253242492676, 'sampling/importance_sampling_ratio/mean': 0.999964714050293, 'sampling/importance_sampling_ratio/max': 1.0113025903701782, 'entropy': 0.009889230597764254, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.7176748514175415, 'reward_meter_mean': 0.7176748514175415, 'reward_meter_std': 0.19788797199726105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7176748514175415, 'reward_total_composite_std': 0.19788797199726105, 'epoch': 0.09}

 66%|██████▌   | 2177/3300 [2:58:04<1:28:30,  4.73s/it]INFO 04-12 01:34:56 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2178/3300 [2:58:10<1:34:48,  5.07s/it]2026-04-12 01:35:01,842 | INFO | train_grpo_train | metrics_logged mode=train step=2178

                                                       
{'loss': 0.0009, 'grad_norm': 1.710945725440979, 'learning_rate': 3.4030303030303036e-06, 'num_tokens': 4915728.0, 'completions/mean_length': 170.625, 'completions/min_length': 170.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.625, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.9990792274475098, 'rewards/meter/std': 7.88669494795613e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666269302368, 'rewards/repeat_penalty/std': 0.05143444612622261, 'rewards/total_composite/mean': 0.9158225059509277, 'rewards/total_composite/std': 0.05138414725661278, 'reward': 0.9158225059509277, 'reward_std': 0.05138415843248367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01991666667163372, 'sampling/sampling_logp_difference/max': 1.5503005981445312, 'sampling/importance_sampling_ratio/min': 0.3940596580505371, 'sampling/importance_sampling_ratio/mean': 1.0056817531585693, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1335141584277153, 'clip_ratio/low_mean': 0.00733574153855443, 'clip_ratio/low_min': 0.00733574153855443, 'clip_ratio/high_mean': 0.0021929824724793434, 'clip_ratio/high_max': 0.0021929824724793434, 'clip_ratio/region_mean': 0.009528724011033773, 'reward_total_mean': 0.9158225059509277, 'reward_meter_mean': 0.9990792274475098, 'reward_meter_std': 7.88669494795613e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666269302368, 'reward_repeat_penalty_std': 0.05143444612622261, 'reward_total_composite_mean': 0.9158225059509277, 'reward_total_composite_std': 0.05138414725661278, 'epoch': 0.09}

 66%|██████▌   | 2178/3300 [2:58:10<1:34:48,  5.07s/it]INFO 04-12 01:35:02 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2179/3300 [2:58:15<1:34:21,  5.05s/it]2026-04-12 01:35:06,769 | INFO | train_grpo_train | metrics_logged mode=train step=2179

                                                       
{'loss': 0.0006, 'grad_norm': 0.20853394269943237, 'learning_rate': 3.4000000000000005e-06, 'num_tokens': 4917552.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9990785121917725, 'rewards/meter/std': 9.964550372387748e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990785121917725, 'rewards/total_composite/std': 9.964550372387748e-06, 'reward': 0.9990785121917725, 'reward_std': 9.972212865250185e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0065149967558681965, 'sampling/sampling_logp_difference/max': 0.41229283809661865, 'sampling/importance_sampling_ratio/min': 0.6621303558349609, 'sampling/importance_sampling_ratio/mean': 1.002784252166748, 'sampling/importance_sampling_ratio/max': 1.365889072418213, 'entropy': 0.05868239142000675, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.005681818351149559, 'reward_total_mean': 0.9990785121917725, 'reward_meter_mean': 0.9990785121917725, 'reward_meter_std': 9.964550372387748e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990785121917725, 'reward_total_composite_std': 9.964550372387748e-06, 'epoch': 0.09}

 66%|██████▌   | 2179/3300 [2:58:15<1:34:21,  5.05s/it]INFO 04-12 01:35:07 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2180/3300 [2:58:22<1:45:03,  5.63s/it]2026-04-12 01:35:13,743 | INFO | train_grpo_train | metrics_logged mode=train step=2180

                                                       
{'loss': -0.0137, 'grad_norm': 2.8113794326782227, 'learning_rate': 3.3969696969696973e-06, 'num_tokens': 4921059.0, 'completions/mean_length': 219.375, 'completions/min_length': 210.0, 'completions/max_length': 236.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 219.375, 'completions/min_terminated_length': 210.0, 'completions/max_terminated_length': 236.0, 'rewards/meter/mean': 0.8746868968009949, 'rewards/meter/std': 0.34104546904563904, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9318181872367859, 'rewards/repeat_penalty/std': 0.08058229833841324, 'rewards/total_composite/mean': 0.7102342844009399, 'rewards/total_composite/std': 0.2835994064807892, 'reward': 0.7102342844009399, 'reward_std': 0.2835994064807892, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033945005387067795, 'sampling/sampling_logp_difference/max': 1.3039984703063965, 'sampling/importance_sampling_ratio/min': 0.27144426107406616, 'sampling/importance_sampling_ratio/mean': 1.002653956413269, 'sampling/importance_sampling_ratio/max': 1.8379487991333008, 'entropy': 0.3587251305580139, 'clip_ratio/low_mean': 0.005902778008021414, 'clip_ratio/low_min': 0.005902778008021414, 'clip_ratio/high_mean': 0.02744697011075914, 'clip_ratio/high_max': 0.02744697011075914, 'clip_ratio/region_mean': 0.03334974811878055, 'reward_total_mean': 0.7102342844009399, 'reward_meter_mean': 0.8746868968009949, 'reward_meter_std': 0.34104546904563904, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9318181872367859, 'reward_repeat_penalty_std': 0.08058229833841324, 'reward_total_composite_mean': 0.7102342844009399, 'reward_total_composite_std': 0.2835994064807892, 'epoch': 0.09}

 66%|██████▌   | 2180/3300 [2:58:22<1:45:03,  5.63s/it]INFO 04-12 01:35:14 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2181/3300 [2:58:27<1:42:55,  5.52s/it]2026-04-12 01:35:19,020 | INFO | train_grpo_train | metrics_logged mode=train step=2181

                                                       
{'loss': 0.0015, 'grad_norm': 2.3878560066223145, 'learning_rate': 3.3939393939393946e-06, 'num_tokens': 4923237.0, 'completions/mean_length': 97.25, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.25, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.8764194250106812, 'rewards/meter/std': 0.24076080322265625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8764194250106812, 'rewards/total_composite/std': 0.24076080322265625, 'reward': 0.8764194250106812, 'reward_std': 0.24076080322265625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012246760539710522, 'sampling/sampling_logp_difference/max': 1.104780673980713, 'sampling/importance_sampling_ratio/min': 0.3312835097312927, 'sampling/importance_sampling_ratio/mean': 1.0025182962417603, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0856907800771296, 'clip_ratio/low_mean': 0.0025641699321568012, 'clip_ratio/low_min': 0.0025641699321568012, 'clip_ratio/high_mean': 0.0064301491947844625, 'clip_ratio/high_max': 0.0064301491947844625, 'clip_ratio/region_mean': 0.008994319126941264, 'reward_total_mean': 0.8764194250106812, 'reward_meter_mean': 0.8764194250106812, 'reward_meter_std': 0.24076080322265625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8764194250106812, 'reward_total_composite_std': 0.24076080322265625, 'epoch': 0.09}

 66%|██████▌   | 2181/3300 [2:58:28<1:42:55,  5.52s/it]INFO 04-12 01:35:19 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2182/3300 [2:58:33<1:43:29,  5.55s/it]2026-04-12 01:35:24,661 | INFO | train_grpo_train | metrics_logged mode=train step=2182

                                                       
{'loss': -0.0054, 'grad_norm': 2.051013946533203, 'learning_rate': 3.3909090909090914e-06, 'num_tokens': 4925782.0, 'completions/mean_length': 127.125, 'completions/min_length': 125.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.125, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9971377849578857, 'rewards/meter/std': 0.0010128796566277742, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8546895980834961, 'rewards/total_composite/std': 0.0008681902545504272, 'reward': 0.8546895980834961, 'reward_std': 0.0008681949693709612, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011722107417881489, 'sampling/sampling_logp_difference/max': 1.6086368560791016, 'sampling/importance_sampling_ratio/min': 0.20016026496887207, 'sampling/importance_sampling_ratio/mean': 1.0007635354995728, 'sampling/importance_sampling_ratio/max': 1.307826042175293, 'entropy': 0.07762610260397196, 'clip_ratio/low_mean': 0.001953125, 'clip_ratio/low_min': 0.001953125, 'clip_ratio/high_mean': 0.005906250094994903, 'clip_ratio/high_max': 0.005906250094994903, 'clip_ratio/region_mean': 0.007859375094994903, 'reward_total_mean': 0.8546895980834961, 'reward_meter_mean': 0.9971377849578857, 'reward_meter_std': 0.0010128796566277742, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8546895980834961, 'reward_total_composite_std': 0.0008681902545504272, 'epoch': 0.09}

 66%|██████▌   | 2182/3300 [2:58:33<1:43:29,  5.55s/it]INFO 04-12 01:35:24 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2183/3300 [2:58:38<1:41:26,  5.45s/it]2026-04-12 01:35:29,910 | INFO | train_grpo_train | metrics_logged mode=train step=2183

                                                       
{'loss': 0.0002, 'grad_norm': 3.130744457244873, 'learning_rate': 3.3878787878787882e-06, 'num_tokens': 4927521.0, 'completions/mean_length': 66.375, 'completions/min_length': 66.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9979628324508667, 'rewards/meter/std': 0.0002601691521704197, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979628324508667, 'rewards/total_composite/std': 0.0002601691521704197, 'reward': 0.9979628324508667, 'reward_std': 0.0002601697633508593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010011628270149231, 'sampling/sampling_logp_difference/max': 0.9755148887634277, 'sampling/importance_sampling_ratio/min': 0.3769982159137726, 'sampling/importance_sampling_ratio/mean': 1.00442373752594, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07284869812428951, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005599473137408495, 'clip_ratio/high_max': 0.005599473137408495, 'clip_ratio/region_mean': 0.005599473137408495, 'reward_total_mean': 0.9979628324508667, 'reward_meter_mean': 0.9979628324508667, 'reward_meter_std': 0.0002601691521704197, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979628324508667, 'reward_total_composite_std': 0.0002601691521704197, 'epoch': 0.09}

 66%|██████▌   | 2183/3300 [2:58:38<1:41:26,  5.45s/it]INFO 04-12 01:35:30 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2184/3300 [2:58:44<1:42:09,  5.49s/it]2026-04-12 01:35:35,454 | INFO | train_grpo_train | metrics_logged mode=train step=2184

                                                       
{'loss': 0.0252, 'grad_norm': 4.508577346801758, 'learning_rate': 3.384848484848485e-06, 'num_tokens': 4929678.0, 'completions/mean_length': 110.625, 'completions/min_length': 106.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.625, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.8917012214660645, 'rewards/meter/std': 0.3005649745464325, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.8667252063751221, 'rewards/total_composite/std': 0.29866674542427063, 'reward': 0.8667252063751221, 'reward_std': 0.29866674542427063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05617932230234146, 'sampling/sampling_logp_difference/max': 1.6177425384521484, 'sampling/importance_sampling_ratio/min': 0.19834595918655396, 'sampling/importance_sampling_ratio/mean': 1.0148777961730957, 'sampling/importance_sampling_ratio/max': 1.945900797843933, 'entropy': 0.5080476962029934, 'clip_ratio/low_mean': 0.014024864183738828, 'clip_ratio/low_min': 0.014024864183738828, 'clip_ratio/high_mean': 0.031617405358701944, 'clip_ratio/high_max': 0.031617405358701944, 'clip_ratio/region_mean': 0.04564226954244077, 'reward_total_mean': 0.8667252063751221, 'reward_meter_mean': 0.8917012214660645, 'reward_meter_std': 0.3005649745464325, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.8667252063751221, 'reward_total_composite_std': 0.29866674542427063, 'epoch': 0.09}

 66%|██████▌   | 2184/3300 [2:58:44<1:42:09,  5.49s/it]INFO 04-12 01:35:35 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2185/3300 [2:58:49<1:40:01,  5.38s/it]2026-04-12 01:35:40,563 | INFO | train_grpo_train | metrics_logged mode=train step=2185

                                                       
{'loss': 0.001, 'grad_norm': 3.390443801879883, 'learning_rate': 3.3818181818181823e-06, 'num_tokens': 4931678.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7347736954689026, 'rewards/meter/std': 0.02520820125937462, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.6960331201553345, 'rewards/total_composite/std': 0.047284774482250214, 'reward': 0.6960331201553345, 'reward_std': 0.04728476703166962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0048207431100308895, 'sampling/sampling_logp_difference/max': 0.8722677230834961, 'sampling/importance_sampling_ratio/min': 0.41800257563591003, 'sampling/importance_sampling_ratio/mean': 1.000468134880066, 'sampling/importance_sampling_ratio/max': 1.1883066892623901, 'entropy': 0.02709749317727983, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0015625000232830644, 'clip_ratio/high_max': 0.0015625000232830644, 'clip_ratio/region_mean': 0.0015625000232830644, 'reward_total_mean': 0.6960331201553345, 'reward_meter_mean': 0.7347736954689026, 'reward_meter_std': 0.02520820125937462, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.6960331201553345, 'reward_total_composite_std': 0.047284774482250214, 'epoch': 0.09}

 66%|██████▌   | 2185/3300 [2:58:49<1:40:01,  5.38s/it]INFO 04-12 01:35:40 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▌   | 2186/3300 [2:58:54<1:37:17,  5.24s/it]2026-04-12 01:35:45,476 | INFO | train_grpo_train | metrics_logged mode=train step=2186

                                                       
{'loss': -0.0159, 'grad_norm': 4.1797590255737305, 'learning_rate': 3.378787878787879e-06, 'num_tokens': 4933511.0, 'completions/mean_length': 75.125, 'completions/min_length': 70.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9933065176010132, 'rewards/meter/std': 0.01265679020434618, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9933065176010132, 'rewards/total_composite/std': 0.01265679020434618, 'reward': 0.9933065176010132, 'reward_std': 0.012656783685088158, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03881411254405975, 'sampling/sampling_logp_difference/max': 1.1373529434204102, 'sampling/importance_sampling_ratio/min': 0.32066673040390015, 'sampling/importance_sampling_ratio/mean': 1.0102565288543701, 'sampling/importance_sampling_ratio/max': 1.9696723222732544, 'entropy': 0.3504325784742832, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.036382337333634496, 'clip_ratio/high_max': 0.036382337333634496, 'clip_ratio/region_mean': 0.03816805162932724, 'reward_total_mean': 0.9933065176010132, 'reward_meter_mean': 0.9933065176010132, 'reward_meter_std': 0.01265679020434618, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9933065176010132, 'reward_total_composite_std': 0.01265679020434618, 'epoch': 0.09}

 66%|██████▌   | 2186/3300 [2:58:54<1:37:17,  5.24s/it]INFO 04-12 01:35:45 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2187/3300 [2:59:02<1:54:22,  6.17s/it]2026-04-12 01:35:53,798 | INFO | train_grpo_train | metrics_logged mode=train step=2187

                                                       
{'loss': -0.0221, 'grad_norm': 1.9298547506332397, 'learning_rate': 3.375757575757576e-06, 'num_tokens': 4938469.0, 'completions/mean_length': 314.75, 'completions/min_length': 297.0, 'completions/max_length': 354.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 314.75, 'completions/min_terminated_length': 297.0, 'completions/max_terminated_length': 354.0, 'rewards/meter/mean': 0.997092604637146, 'rewards/meter/std': 0.0013639559037983418, 'rewards/count_adherence/mean': 0.5803571939468384, 'rewards/count_adherence/std': 0.02525380253791809, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9509804248809814, 'rewards/repeat_penalty/std': 0.04682480916380882, 'rewards/total_composite/mean': 0.5502004623413086, 'rewards/total_composite/std': 0.03388208523392677, 'reward': 0.5502004623413086, 'reward_std': 0.03388208895921707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04348405450582504, 'sampling/sampling_logp_difference/max': 1.2296199798583984, 'sampling/importance_sampling_ratio/min': 0.2924036681652069, 'sampling/importance_sampling_ratio/mean': 1.0096619129180908, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45917268842458725, 'clip_ratio/low_mean': 0.015058388700708747, 'clip_ratio/low_min': 0.015058388700708747, 'clip_ratio/high_mean': 0.02572961524128914, 'clip_ratio/high_max': 0.02572961524128914, 'clip_ratio/region_mean': 0.040788003941997886, 'reward_total_mean': 0.5502004623413086, 'reward_meter_mean': 0.997092604637146, 'reward_meter_std': 0.0013639559037983418, 'reward_count_adherence_mean': 0.5803571939468384, 'reward_count_adherence_std': 0.02525380253791809, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9509804248809814, 'reward_repeat_penalty_std': 0.04682480916380882, 'reward_total_composite_mean': 0.5502004623413086, 'reward_total_composite_std': 0.03388208523392677, 'epoch': 0.09}

 66%|██████▋   | 2187/3300 [2:59:02<1:54:22,  6.17s/it]INFO 04-12 01:35:54 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2188/3300 [2:59:07<1:48:34,  5.86s/it]2026-04-12 01:35:58,939 | INFO | train_grpo_train | metrics_logged mode=train step=2188

                                                       
{'loss': 0.0088, 'grad_norm': 1.814806342124939, 'learning_rate': 3.3727272727272732e-06, 'num_tokens': 4940773.0, 'completions/mean_length': 115.0, 'completions/min_length': 113.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.0, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9959709644317627, 'rewards/meter/std': 0.00034307397436350584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9070344567298889, 'rewards/total_composite/std': 0.07348322868347168, 'reward': 0.9070344567298889, 'reward_std': 0.07348321378231049, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015008204616606236, 'sampling/sampling_logp_difference/max': 1.293868064880371, 'sampling/importance_sampling_ratio/min': 0.27420806884765625, 'sampling/importance_sampling_ratio/mean': 0.9997696280479431, 'sampling/importance_sampling_ratio/max': 1.2936325073242188, 'entropy': 0.07655041664838791, 'clip_ratio/low_mean': 0.0010775862028822303, 'clip_ratio/low_min': 0.0010775862028822303, 'clip_ratio/high_mean': 0.006579951150342822, 'clip_ratio/high_max': 0.006579951150342822, 'clip_ratio/region_mean': 0.007657537353225052, 'reward_total_mean': 0.9070344567298889, 'reward_meter_mean': 0.9959709644317627, 'reward_meter_std': 0.00034307397436350584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9070344567298889, 'reward_total_composite_std': 0.07348322868347168, 'epoch': 0.09}

 66%|██████▋   | 2188/3300 [2:59:07<1:48:34,  5.86s/it]INFO 04-12 01:35:59 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2189/3300 [2:59:12<1:41:18,  5.47s/it]2026-04-12 01:36:03,506 | INFO | train_grpo_train | metrics_logged mode=train step=2189

                                                       
{'loss': 0.0002, 'grad_norm': 0.7317746877670288, 'learning_rate': 3.36969696969697e-06, 'num_tokens': 4942493.0, 'completions/mean_length': 56.0, 'completions/min_length': 56.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9949744939804077, 'rewards/meter/std': 3.23687017953489e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949744939804077, 'rewards/total_composite/std': 3.23687017953489e-05, 'reward': 0.9949744939804077, 'reward_std': 3.236968768760562e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00435090996325016, 'sampling/sampling_logp_difference/max': 0.32142794132232666, 'sampling/importance_sampling_ratio/min': 0.8439798355102539, 'sampling/importance_sampling_ratio/mean': 1.003525972366333, 'sampling/importance_sampling_ratio/max': 1.3790956735610962, 'entropy': 0.05346089042723179, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0022321429569274187, 'clip_ratio/high_max': 0.0022321429569274187, 'clip_ratio/region_mean': 0.0022321429569274187, 'reward_total_mean': 0.9949744939804077, 'reward_meter_mean': 0.9949744939804077, 'reward_meter_std': 3.23687017953489e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949744939804077, 'reward_total_composite_std': 3.23687017953489e-05, 'epoch': 0.09}

 66%|██████▋   | 2189/3300 [2:59:12<1:41:18,  5.47s/it]INFO 04-12 01:36:03 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2190/3300 [2:59:16<1:34:38,  5.12s/it]2026-04-12 01:36:07,790 | INFO | train_grpo_train | metrics_logged mode=train step=2190

                                                       
{'loss': 0.0157, 'grad_norm': 3.3835291862487793, 'learning_rate': 3.366666666666667e-06, 'num_tokens': 4943784.0, 'completions/mean_length': 33.375, 'completions/min_length': 33.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9737765789031982, 'rewards/meter/std': 0.008460394106805325, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9737765789031982, 'rewards/total_composite/std': 0.008460394106805325, 'reward': 0.9737765789031982, 'reward_std': 0.008460381999611855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019433647394180298, 'sampling/sampling_logp_difference/max': 1.1026906967163086, 'sampling/importance_sampling_ratio/min': 0.33197662234306335, 'sampling/importance_sampling_ratio/mean': 1.0068951845169067, 'sampling/importance_sampling_ratio/max': 1.51780366897583, 'entropy': 0.14891941286623478, 'clip_ratio/low_mean': 0.011140820104628801, 'clip_ratio/low_min': 0.011140820104628801, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.014928699005395174, 'reward_total_mean': 0.9737765789031982, 'reward_meter_mean': 0.9737765789031982, 'reward_meter_std': 0.008460394106805325, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9737765789031982, 'reward_total_composite_std': 0.008460394106805325, 'epoch': 0.09}

 66%|██████▋   | 2190/3300 [2:59:16<1:34:38,  5.12s/it]INFO 04-12 01:36:08 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2191/3300 [2:59:22<1:35:50,  5.19s/it]2026-04-12 01:36:13,138 | INFO | train_grpo_train | metrics_logged mode=train step=2191

                                                       
{'loss': 0.0003, 'grad_norm': 1.943103551864624, 'learning_rate': 3.3636363636363637e-06, 'num_tokens': 4946182.0, 'completions/mean_length': 113.75, 'completions/min_length': 112.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.75, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.995752215385437, 'rewards/meter/std': 0.00029232812812551856, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9601929187774658, 'rewards/total_composite/std': 0.06590110063552856, 'reward': 0.9601929187774658, 'reward_std': 0.06590110063552856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015357456170022488, 'sampling/sampling_logp_difference/max': 1.5764732360839844, 'sampling/importance_sampling_ratio/min': 0.2067027986049652, 'sampling/importance_sampling_ratio/mean': 1.0013221502304077, 'sampling/importance_sampling_ratio/max': 1.6537041664123535, 'entropy': 0.09450818877667189, 'clip_ratio/low_mean': 0.0021739129442721605, 'clip_ratio/low_min': 0.0021739129442721605, 'clip_ratio/high_mean': 0.006637688144110143, 'clip_ratio/high_max': 0.006637688144110143, 'clip_ratio/region_mean': 0.008811601088382304, 'reward_total_mean': 0.9601929187774658, 'reward_meter_mean': 0.995752215385437, 'reward_meter_std': 0.00029232812812551856, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9601929187774658, 'reward_total_composite_std': 0.06590110063552856, 'epoch': 0.09}

 66%|██████▋   | 2191/3300 [2:59:22<1:35:50,  5.19s/it]INFO 04-12 01:36:13 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2192/3300 [2:59:27<1:36:21,  5.22s/it]2026-04-12 01:36:18,436 | INFO | train_grpo_train | metrics_logged mode=train step=2192

                                                       
{'loss': 0.0058, 'grad_norm': 3.4850220680236816, 'learning_rate': 3.360606060606061e-06, 'num_tokens': 4948311.0, 'completions/mean_length': 100.125, 'completions/min_length': 99.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.99900221824646, 'rewards/meter/std': 0.00022944888041820377, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99900221824646, 'rewards/total_composite/std': 0.00022944888041820377, 'reward': 0.99900221824646, 'reward_std': 0.00022944044030737132, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01750640757381916, 'sampling/sampling_logp_difference/max': 1.2861604690551758, 'sampling/importance_sampling_ratio/min': 0.27632972598075867, 'sampling/importance_sampling_ratio/mean': 0.9979958534240723, 'sampling/importance_sampling_ratio/max': 1.8097566366195679, 'entropy': 0.08774240221828222, 'clip_ratio/low_mean': 0.0012376237427815795, 'clip_ratio/low_min': 0.0012376237427815795, 'clip_ratio/high_mean': 0.018826007028110325, 'clip_ratio/high_max': 0.018826007028110325, 'clip_ratio/region_mean': 0.020063630770891905, 'reward_total_mean': 0.99900221824646, 'reward_meter_mean': 0.99900221824646, 'reward_meter_std': 0.00022944888041820377, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99900221824646, 'reward_total_composite_std': 0.00022944888041820377, 'epoch': 0.09}

 66%|██████▋   | 2192/3300 [2:59:27<1:36:21,  5.22s/it]INFO 04-12 01:36:18 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2193/3300 [2:59:32<1:34:39,  5.13s/it]2026-04-12 01:36:23,358 | INFO | train_grpo_train | metrics_logged mode=train step=2193

                                                       
{'loss': 0.0298, 'grad_norm': 5.769536972045898, 'learning_rate': 3.357575757575758e-06, 'num_tokens': 4950310.0, 'completions/mean_length': 75.875, 'completions/min_length': 72.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9790392518043518, 'rewards/meter/std': 0.02525605633854866, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9790392518043518, 'rewards/total_composite/std': 0.02525605633854866, 'reward': 0.9790392518043518, 'reward_std': 0.025256047025322914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06429249048233032, 'sampling/sampling_logp_difference/max': 1.042258858680725, 'sampling/importance_sampling_ratio/min': 0.35265716910362244, 'sampling/importance_sampling_ratio/mean': 1.0126549005508423, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5629463791847229, 'clip_ratio/low_mean': 0.031126924557611346, 'clip_ratio/low_min': 0.031126924557611346, 'clip_ratio/high_mean': 0.014968461007811129, 'clip_ratio/high_max': 0.014968461007811129, 'clip_ratio/region_mean': 0.046095385565422475, 'reward_total_mean': 0.9790392518043518, 'reward_meter_mean': 0.9790392518043518, 'reward_meter_std': 0.02525605633854866, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9790392518043518, 'reward_total_composite_std': 0.02525605633854866, 'epoch': 0.09}

 66%|██████▋   | 2193/3300 [2:59:32<1:34:39,  5.13s/it]INFO 04-12 01:36:23 [block_pool.py:378] Successfully reset prefix cache

 66%|██████▋   | 2194/3300 [2:59:37<1:32:39,  5.03s/it]2026-04-12 01:36:28,140 | INFO | train_grpo_train | metrics_logged mode=train step=2194

                                                       
{'loss': 0.0016, 'grad_norm': 4.066121578216553, 'learning_rate': 3.3545454545454547e-06, 'num_tokens': 4952168.0, 'completions/mean_length': 69.25, 'completions/min_length': 69.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9961166977882385, 'rewards/meter/std': 0.0030905893072485924, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961166977882385, 'rewards/total_composite/std': 0.0030905893072485924, 'reward': 0.9961166977882385, 'reward_std': 0.0030905804596841335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008842971175909042, 'sampling/sampling_logp_difference/max': 0.6939091682434082, 'sampling/importance_sampling_ratio/min': 0.49961915612220764, 'sampling/importance_sampling_ratio/mean': 0.9982797503471375, 'sampling/importance_sampling_ratio/max': 1.0906788110733032, 'entropy': 0.04704178823158145, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.003597308532334864, 'clip_ratio/high_max': 0.003597308532334864, 'clip_ratio/region_mean': 0.003597308532334864, 'reward_total_mean': 0.9961166977882385, 'reward_meter_mean': 0.9961166977882385, 'reward_meter_std': 0.0030905893072485924, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961166977882385, 'reward_total_composite_std': 0.0030905893072485924, 'epoch': 0.09}

 66%|██████▋   | 2194/3300 [2:59:37<1:32:39,  5.03s/it]INFO 04-12 01:36:28 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2195/3300 [2:59:42<1:32:11,  5.01s/it]2026-04-12 01:36:33,106 | INFO | train_grpo_train | metrics_logged mode=train step=2195

                                                       
{'loss': 0.0008, 'grad_norm': 3.4402453899383545, 'learning_rate': 3.351515151515152e-06, 'num_tokens': 4953959.0, 'completions/mean_length': 75.875, 'completions/min_length': 72.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.998624861240387, 'rewards/meter/std': 0.0006542301853187382, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998624861240387, 'rewards/total_composite/std': 0.0006542301853187382, 'reward': 0.998624861240387, 'reward_std': 0.0006542490446008742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04149501770734787, 'sampling/sampling_logp_difference/max': 1.0772240161895752, 'sampling/importance_sampling_ratio/min': 0.3405395448207855, 'sampling/importance_sampling_ratio/mean': 1.0039963722229004, 'sampling/importance_sampling_ratio/max': 1.4867199659347534, 'entropy': 0.3429943472146988, 'clip_ratio/low_mean': 0.00645240826997906, 'clip_ratio/low_min': 0.00645240826997906, 'clip_ratio/high_mean': 0.013268729439005256, 'clip_ratio/high_max': 0.013268729439005256, 'clip_ratio/region_mean': 0.019721137708984315, 'reward_total_mean': 0.998624861240387, 'reward_meter_mean': 0.998624861240387, 'reward_meter_std': 0.0006542301853187382, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998624861240387, 'reward_total_composite_std': 0.0006542301853187382, 'epoch': 0.09}

 67%|██████▋   | 2195/3300 [2:59:42<1:32:11,  5.01s/it]INFO 04-12 01:36:33 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2196/3300 [2:59:46<1:30:38,  4.93s/it]2026-04-12 01:36:37,846 | INFO | train_grpo_train | metrics_logged mode=train step=2196

                                                       
{'loss': 0.0022, 'grad_norm': 1.1731183528900146, 'learning_rate': 3.3484848484848487e-06, 'num_tokens': 4955804.0, 'completions/mean_length': 65.625, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9991217255592346, 'rewards/meter/std': 3.6965378967579454e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991217255592346, 'rewards/total_composite/std': 3.6965378967579454e-05, 'reward': 0.9991217255592346, 'reward_std': 3.6982291931053624e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007183389738202095, 'sampling/sampling_logp_difference/max': 0.5108299255371094, 'sampling/importance_sampling_ratio/min': 0.7030104994773865, 'sampling/importance_sampling_ratio/mean': 1.0017669200897217, 'sampling/importance_sampling_ratio/max': 1.6666737794876099, 'entropy': 0.04586625238880515, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.003846153849735856, 'clip_ratio/high_max': 0.003846153849735856, 'clip_ratio/region_mean': 0.003846153849735856, 'reward_total_mean': 0.9991217255592346, 'reward_meter_mean': 0.9991217255592346, 'reward_meter_std': 3.6965378967579454e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991217255592346, 'reward_total_composite_std': 3.6965378967579454e-05, 'epoch': 0.09}

 67%|██████▋   | 2196/3300 [2:59:46<1:30:38,  4.93s/it]INFO 04-12 01:36:38 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2197/3300 [2:59:51<1:31:08,  4.96s/it]2026-04-12 01:36:42,872 | INFO | train_grpo_train | metrics_logged mode=train step=2197

                                                       
{'loss': 0.0579, 'grad_norm': 9.319520950317383, 'learning_rate': 3.3454545454545456e-06, 'num_tokens': 4957640.0, 'completions/mean_length': 75.5, 'completions/min_length': 71.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9673671126365662, 'rewards/meter/std': 0.03739238530397415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9673671126365662, 'rewards/total_composite/std': 0.03739238530397415, 'reward': 0.9673671126365662, 'reward_std': 0.03739236667752266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032729994505643845, 'sampling/sampling_logp_difference/max': 1.7551674842834473, 'sampling/importance_sampling_ratio/min': 0.17287828028202057, 'sampling/importance_sampling_ratio/mean': 1.0026880502700806, 'sampling/importance_sampling_ratio/max': 1.8825844526290894, 'entropy': 0.20823358744382858, 'clip_ratio/low_mean': 0.010880606481805444, 'clip_ratio/low_min': 0.010880606481805444, 'clip_ratio/high_mean': 0.011807307717390358, 'clip_ratio/high_max': 0.011807307717390358, 'clip_ratio/region_mean': 0.022687914199195802, 'reward_total_mean': 0.9673671126365662, 'reward_meter_mean': 0.9673671126365662, 'reward_meter_std': 0.03739238530397415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9673671126365662, 'reward_total_composite_std': 0.03739238530397415, 'epoch': 0.09}

 67%|██████▋   | 2197/3300 [2:59:51<1:31:08,  4.96s/it]INFO 04-12 01:36:43 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2198/3300 [2:59:56<1:29:42,  4.88s/it]2026-04-12 01:36:47,581 | INFO | train_grpo_train | metrics_logged mode=train step=2198

                                                       
{'loss': 0.0073, 'grad_norm': 3.934696912765503, 'learning_rate': 3.3424242424242424e-06, 'num_tokens': 4959602.0, 'completions/mean_length': 75.25, 'completions/min_length': 72.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.25, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9990965127944946, 'rewards/meter/std': 0.00029607766191475093, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990965127944946, 'rewards/total_composite/std': 0.00029607766191475093, 'reward': 0.9990965127944946, 'reward_std': 0.00029608141630887985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021792318671941757, 'sampling/sampling_logp_difference/max': 0.8232808113098145, 'sampling/importance_sampling_ratio/min': 0.4389890432357788, 'sampling/importance_sampling_ratio/mean': 1.0073840618133545, 'sampling/importance_sampling_ratio/max': 1.3668806552886963, 'entropy': 0.24795558489859104, 'clip_ratio/low_mean': 0.0032051282469183207, 'clip_ratio/low_min': 0.0032051282469183207, 'clip_ratio/high_mean': 0.006831709994003177, 'clip_ratio/high_max': 0.006831709994003177, 'clip_ratio/region_mean': 0.010036838240921497, 'reward_total_mean': 0.9990965127944946, 'reward_meter_mean': 0.9990965127944946, 'reward_meter_std': 0.00029607766191475093, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990965127944946, 'reward_total_composite_std': 0.00029607766191475093, 'epoch': 0.09}

 67%|██████▋   | 2198/3300 [2:59:56<1:29:42,  4.88s/it]INFO 04-12 01:36:47 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2199/3300 [3:00:01<1:28:06,  4.80s/it]2026-04-12 01:36:52,193 | INFO | train_grpo_train | metrics_logged mode=train step=2199

                                                       
{'loss': 0.0001, 'grad_norm': 0.02073613740503788, 'learning_rate': 3.3393939393939397e-06, 'num_tokens': 4961586.0, 'completions/mean_length': 69.0, 'completions/min_length': 69.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9973071813583374, 'rewards/meter/std': 2.9068671665299917e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973071813583374, 'rewards/total_composite/std': 2.9068671665299917e-06, 'reward': 0.9973071813583374, 'reward_std': 2.895276793424273e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0027781545650213957, 'sampling/sampling_logp_difference/max': 0.15232563018798828, 'sampling/importance_sampling_ratio/min': 0.8587086200714111, 'sampling/importance_sampling_ratio/mean': 1.0010336637496948, 'sampling/importance_sampling_ratio/max': 1.067936897277832, 'entropy': 0.025351210730150342, 'clip_ratio/low_mean': 0.0036231884732842445, 'clip_ratio/low_min': 0.0036231884732842445, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0036231884732842445, 'reward_total_mean': 0.9973071813583374, 'reward_meter_mean': 0.9973071813583374, 'reward_meter_std': 2.9068671665299917e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973071813583374, 'reward_total_composite_std': 2.9068671665299917e-06, 'epoch': 0.09}

 67%|██████▋   | 2199/3300 [3:00:01<1:28:06,  4.80s/it]INFO 04-12 01:36:52 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2200/3300 [3:00:05<1:28:20,  4.82s/it]2026-04-12 01:36:57,048 | INFO | train_grpo_train | metrics_logged mode=train step=2200

                                                       
{'loss': -0.0049, 'grad_norm': 3.2771596908569336, 'learning_rate': 3.3363636363636365e-06, 'num_tokens': 4963619.0, 'completions/mean_length': 80.125, 'completions/min_length': 77.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.125, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9969666600227356, 'rewards/meter/std': 0.001407405361533165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8722573518753052, 'rewards/total_composite/std': 0.35244786739349365, 'reward': 0.8722573518753052, 'reward_std': 0.35244786739349365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038304030895233154, 'sampling/sampling_logp_difference/max': 1.204660415649414, 'sampling/importance_sampling_ratio/min': 0.2997937798500061, 'sampling/importance_sampling_ratio/mean': 1.0073878765106201, 'sampling/importance_sampling_ratio/max': 1.6984498500823975, 'entropy': 0.3972342722117901, 'clip_ratio/low_mean': 0.0031645570416003466, 'clip_ratio/low_min': 0.0031645570416003466, 'clip_ratio/high_mean': 0.03107945283409208, 'clip_ratio/high_max': 0.03107945283409208, 'clip_ratio/region_mean': 0.03424400987569243, 'reward_total_mean': 0.8722573518753052, 'reward_meter_mean': 0.9969666600227356, 'reward_meter_std': 0.001407405361533165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8722573518753052, 'reward_total_composite_std': 0.35244786739349365, 'epoch': 0.09}

 67%|██████▋   | 2200/3300 [3:00:06<1:28:20,  4.82s/it]INFO 04-12 01:36:57 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:31,  2.86s/it]

 23%|██▎       | 3/13 [00:08<00:29,  2.98s/it]

 31%|███       | 4/13 [00:13<00:32,  3.63s/it]

 38%|███▊      | 5/13 [00:18<00:32,  4.02s/it]

 46%|████▌     | 6/13 [00:24<00:32,  4.60s/it]

 54%|█████▍    | 7/13 [00:29<00:28,  4.70s/it]

 62%|██████▏   | 8/13 [00:35<00:25,  5.15s/it]

 69%|██████▉   | 9/13 [00:41<00:21,  5.42s/it]

 77%|███████▋  | 10/13 [00:45<00:15,  5.18s/it]

 85%|████████▍ | 11/13 [00:50<00:10,  5.04s/it]

 92%|█████████▏| 12/13 [00:55<00:04,  4.95s/it]

100%|██████████| 13/13 [00:59<00:00,  4.85s/it]2026-04-12 01:38:02,281 | INFO | train_grpo_train | metrics_logged mode=eval step=2200

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 65.1539, 'eval_samples_per_second': 1.596, 'eval_steps_per_second': 0.2, 'eval_num_tokens': 4963619.0, 'eval_completions/mean_length': 187.77884615384616, 'eval_completions/min_length': 62.69230769230769, 'eval_completions/max_length': 339.15384615384613, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 187.77884615384616, 'eval_completions/min_terminated_length': 62.69230769230769, 'eval_completions/max_terminated_length': 339.15384615384613, 'eval_rewards/meter/mean': 0.722443516437824, 'eval_rewards/meter/std': 0.41069405812483567, 'eval_rewards/count_adherence/mean': 0.8735798276387728, 'eval_rewards/count_adherence/std': 0.14306257034723574, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8453576381389911, 'eval_rewards/repeat_penalty/std': 0.13899515225337103, 'eval_rewards/total_composite/mean': 0.5540328117517325, 'eval_rewards/total_composite/std': 0.3618074002174231, 'eval_reward': 0.5540328117517325, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.01593467600357074, 'eval_sampling/sampling_logp_difference/max': 1.04065675001878, 'eval_sampling/importance_sampling_ratio/min': 0.3679314060853078, 'eval_sampling/importance_sampling_ratio/mean': 1.0049138986147368, 'eval_sampling/importance_sampling_ratio/max': 1.4119950441213756, 'eval_entropy': 0.1745165208211312, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5540328117517325, 'eval_reward_meter_mean': 0.722443516437824, 'eval_reward_meter_std': 0.41069405812483567, 'eval_reward_count_adherence_mean': 0.8735798276387728, 'eval_reward_count_adherence_std': 0.14306257034723574, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8453576381389911, 'eval_reward_repeat_penalty_std': 0.13899515225337103, 'eval_reward_total_composite_mean': 0.5540328117517325, 'eval_reward_total_composite_std': 0.3618074002174231, 'epoch': 0.09}

 67%|██████▋   | 2200/3300 [3:01:11<1:28:20,  4.82s/it]

100%|██████████| 13/13 [00:59<00:00,  4.85s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 67%|██████▋   | 2201/3300 [3:01:18<7:41:19, 25.19s/it]2026-04-12 01:38:09,817 | INFO | train_grpo_train | metrics_logged mode=train step=2201

                                                       
{'loss': 0.0037, 'grad_norm': 2.442410707473755, 'learning_rate': 3.3333333333333333e-06, 'num_tokens': 4965596.0, 'completions/mean_length': 69.125, 'completions/min_length': 69.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9970394968986511, 'rewards/meter/std': 0.0007004055660218, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970394968986511, 'rewards/total_composite/std': 0.0007004055660218, 'reward': 0.9970394968986511, 'reward_std': 0.0007004134822636843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005273114424198866, 'sampling/sampling_logp_difference/max': 0.7160444259643555, 'sampling/importance_sampling_ratio/min': 0.48868146538734436, 'sampling/importance_sampling_ratio/mean': 1.001259207725525, 'sampling/importance_sampling_ratio/max': 1.3967480659484863, 'entropy': 0.029302328824996948, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9970394968986511, 'reward_meter_mean': 0.9970394968986511, 'reward_meter_std': 0.0007004055660218, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970394968986511, 'reward_total_composite_std': 0.0007004055660218, 'epoch': 0.09}

 67%|██████▋   | 2201/3300 [3:01:18<7:41:19, 25.19s/it]INFO 04-12 01:38:10 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2202/3300 [3:01:26<6:04:37, 19.92s/it]2026-04-12 01:38:17,419 | INFO | train_grpo_train | metrics_logged mode=train step=2202

                                                       
{'loss': 0.0069, 'grad_norm': 2.4785351753234863, 'learning_rate': 3.3303030303030306e-06, 'num_tokens': 4969547.0, 'completions/mean_length': 277.875, 'completions/min_length': 273.0, 'completions/max_length': 285.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 277.875, 'completions/min_terminated_length': 273.0, 'completions/max_terminated_length': 285.0, 'rewards/meter/mean': 0.997477114200592, 'rewards/meter/std': 0.001005422556772828, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9519230723381042, 'rewards/repeat_penalty/std': 0.05723259970545769, 'rewards/total_composite/mean': 0.8308628797531128, 'rewards/total_composite/std': 0.05055442824959755, 'reward': 0.8308628797531128, 'reward_std': 0.05055440962314606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.046521060168743134, 'sampling/sampling_logp_difference/max': 1.3806867599487305, 'sampling/importance_sampling_ratio/min': 0.25140586495399475, 'sampling/importance_sampling_ratio/mean': 1.0075916051864624, 'sampling/importance_sampling_ratio/max': 1.8519618511199951, 'entropy': 0.4109150692820549, 'clip_ratio/low_mean': 0.014749160269275308, 'clip_ratio/low_min': 0.014749160269275308, 'clip_ratio/high_mean': 0.020743532106280327, 'clip_ratio/high_max': 0.020743532106280327, 'clip_ratio/region_mean': 0.035492692375555634, 'reward_total_mean': 0.8308628797531128, 'reward_meter_mean': 0.997477114200592, 'reward_meter_std': 0.001005422556772828, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9519230723381042, 'reward_repeat_penalty_std': 0.05723259970545769, 'reward_total_composite_mean': 0.8308628797531128, 'reward_total_composite_std': 0.05055442824959755, 'epoch': 0.09}

 67%|██████▋   | 2202/3300 [3:01:26<6:04:37, 19.92s/it]INFO 04-12 01:38:17 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2203/3300 [3:01:31<4:41:52, 15.42s/it]2026-04-12 01:38:22,322 | INFO | train_grpo_train | metrics_logged mode=train step=2203

                                                       
{'loss': 0.0081, 'grad_norm': 3.297503709793091, 'learning_rate': 3.3272727272727274e-06, 'num_tokens': 4971474.0, 'completions/mean_length': 73.875, 'completions/min_length': 71.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9979462623596191, 'rewards/meter/std': 0.002537182066589594, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979462623596191, 'rewards/total_composite/std': 0.002537182066589594, 'reward': 0.9979462623596191, 'reward_std': 0.002537175314500928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.060300085693597794, 'sampling/sampling_logp_difference/max': 1.7386038303375244, 'sampling/importance_sampling_ratio/min': 0.17576563358306885, 'sampling/importance_sampling_ratio/mean': 1.004874587059021, 'sampling/importance_sampling_ratio/max': 1.8745375871658325, 'entropy': 0.46222241409122944, 'clip_ratio/low_mean': 0.0050675676902756095, 'clip_ratio/low_min': 0.0050675676902756095, 'clip_ratio/high_mean': 0.030337184784002602, 'clip_ratio/high_max': 0.030337184784002602, 'clip_ratio/region_mean': 0.03540475247427821, 'reward_total_mean': 0.9979462623596191, 'reward_meter_mean': 0.9979462623596191, 'reward_meter_std': 0.002537182066589594, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979462623596191, 'reward_total_composite_std': 0.002537182066589594, 'epoch': 0.09}

 67%|██████▋   | 2203/3300 [3:01:31<4:41:52, 15.42s/it]INFO 04-12 01:38:22 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2204/3300 [3:01:40<4:09:10, 13.64s/it]2026-04-12 01:38:31,816 | INFO | train_grpo_train | metrics_logged mode=train step=2204

                                                       
{'loss': -0.0316, 'grad_norm': 1.867836833000183, 'learning_rate': 3.3242424242424242e-06, 'num_tokens': 4976337.0, 'completions/mean_length': 410.875, 'completions/min_length': 380.0, 'completions/max_length': 449.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 410.875, 'completions/min_terminated_length': 380.0, 'completions/max_terminated_length': 449.0, 'rewards/meter/mean': 0.9976021647453308, 'rewards/meter/std': 0.0008352479781024158, 'rewards/count_adherence/mean': 0.6029411554336548, 'rewards/count_adherence/std': 0.027230001986026764, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8183584213256836, 'rewards/repeat_penalty/std': 0.10606583207845688, 'rewards/total_composite/mean': 0.4941902458667755, 'rewards/total_composite/std': 0.08418715745210648, 'reward': 0.4941902458667755, 'reward_std': 0.08418715000152588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0467507466673851, 'sampling/sampling_logp_difference/max': 6.5194783210754395, 'sampling/importance_sampling_ratio/min': 0.0014744381187483668, 'sampling/importance_sampling_ratio/mean': 1.0103265047073364, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3659870997071266, 'clip_ratio/low_mean': 0.014909796707797796, 'clip_ratio/low_min': 0.014909796707797796, 'clip_ratio/high_mean': 0.014851224608719349, 'clip_ratio/high_max': 0.014851224608719349, 'clip_ratio/region_mean': 0.029761021316517144, 'reward_total_mean': 0.4941902458667755, 'reward_meter_mean': 0.9976021647453308, 'reward_meter_std': 0.0008352479781024158, 'reward_count_adherence_mean': 0.6029411554336548, 'reward_count_adherence_std': 0.027230001986026764, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8183584213256836, 'reward_repeat_penalty_std': 0.10606583207845688, 'reward_total_composite_mean': 0.4941902458667755, 'reward_total_composite_std': 0.08418715745210648, 'epoch': 0.09}

 67%|██████▋   | 2204/3300 [3:01:40<4:09:10, 13.64s/it]INFO 04-12 01:38:32 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2205/3300 [3:01:47<3:33:51, 11.72s/it]2026-04-12 01:38:39,048 | INFO | train_grpo_train | metrics_logged mode=train step=2205

                                                       
{'loss': 0.0016, 'grad_norm': 1.170793056488037, 'learning_rate': 3.321212121212121e-06, 'num_tokens': 4979646.0, 'completions/mean_length': 223.625, 'completions/min_length': 223.0, 'completions/max_length': 225.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 223.625, 'completions/min_terminated_length': 223.0, 'completions/max_terminated_length': 225.0, 'rewards/meter/mean': 0.9960297346115112, 'rewards/meter/std': 0.0002207064680987969, 'rewards/count_adherence/mean': 0.8888888955116272, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6666666865348816, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5902398228645325, 'rewards/total_composite/std': 0.00013078592019155622, 'reward': 0.5902398228645325, 'reward_std': 0.00013079482596367598, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005620577838271856, 'sampling/sampling_logp_difference/max': 1.2731132507324219, 'sampling/importance_sampling_ratio/min': 0.27995866537094116, 'sampling/importance_sampling_ratio/mean': 1.0011334419250488, 'sampling/importance_sampling_ratio/max': 1.5353397130966187, 'entropy': 0.02647018409334123, 'clip_ratio/low_mean': 0.0005555555690079927, 'clip_ratio/low_min': 0.0005555555690079927, 'clip_ratio/high_mean': 0.0022346453624777496, 'clip_ratio/high_max': 0.0022346453624777496, 'clip_ratio/region_mean': 0.0027902009314857423, 'reward_total_mean': 0.5902398228645325, 'reward_meter_mean': 0.9960297346115112, 'reward_meter_std': 0.0002207064680987969, 'reward_count_adherence_mean': 0.8888888955116272, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6666666865348816, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5902398228645325, 'reward_total_composite_std': 0.00013078592019155622, 'epoch': 0.09}

 67%|██████▋   | 2205/3300 [3:01:48<3:33:51, 11.72s/it]INFO 04-12 01:38:39 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2206/3300 [3:01:52<2:54:28,  9.57s/it]2026-04-12 01:38:43,601 | INFO | train_grpo_train | metrics_logged mode=train step=2206

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.3181818181818188e-06, 'num_tokens': 4981398.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0008829182479530573, 'sampling/sampling_logp_difference/max': 0.027269084006547928, 'sampling/importance_sampling_ratio/min': 0.9798296689987183, 'sampling/importance_sampling_ratio/mean': 1.0007154941558838, 'sampling/importance_sampling_ratio/max': 1.0276442766189575, 'entropy': 0.009517844882793725, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 67%|██████▋   | 2206/3300 [3:01:52<2:54:28,  9.57s/it]INFO 04-12 01:38:43 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2207/3300 [3:02:01<2:50:00,  9.33s/it]2026-04-12 01:38:52,372 | INFO | train_grpo_train | metrics_logged mode=train step=2207

                                                       
{'loss': -0.0214, 'grad_norm': 1.6692476272583008, 'learning_rate': 3.3151515151515156e-06, 'num_tokens': 4986186.0, 'completions/mean_length': 377.5, 'completions/min_length': 362.0, 'completions/max_length': 396.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 377.5, 'completions/min_terminated_length': 362.0, 'completions/max_terminated_length': 396.0, 'rewards/meter/mean': 0.9973218441009521, 'rewards/meter/std': 0.000687557680066675, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.0345032773911953, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8846749067306519, 'rewards/repeat_penalty/std': 0.11391878128051758, 'rewards/total_composite/mean': 0.5530020594596863, 'rewards/total_composite/std': 0.08831970393657684, 'reward': 0.5530020594596863, 'reward_std': 0.08831970393657684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03707325831055641, 'sampling/sampling_logp_difference/max': 2.2571840286254883, 'sampling/importance_sampling_ratio/min': 0.10464474558830261, 'sampling/importance_sampling_ratio/mean': 1.0060136318206787, 'sampling/importance_sampling_ratio/max': 1.8353766202926636, 'entropy': 0.3299342021346092, 'clip_ratio/low_mean': 0.0101922721369192, 'clip_ratio/low_min': 0.0101922721369192, 'clip_ratio/high_mean': 0.019536115461960435, 'clip_ratio/high_max': 0.019536115461960435, 'clip_ratio/region_mean': 0.029728387598879635, 'reward_total_mean': 0.5530020594596863, 'reward_meter_mean': 0.9973218441009521, 'reward_meter_std': 0.000687557680066675, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.0345032773911953, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8846749067306519, 'reward_repeat_penalty_std': 0.11391878128051758, 'reward_total_composite_mean': 0.5530020594596863, 'reward_total_composite_std': 0.08831970393657684, 'epoch': 0.09}

 67%|██████▋   | 2207/3300 [3:02:01<2:50:00,  9.33s/it]INFO 04-12 01:38:52 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2208/3300 [3:02:05<2:23:42,  7.90s/it]2026-04-12 01:38:56,921 | INFO | train_grpo_train | metrics_logged mode=train step=2208

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.3121212121212124e-06, 'num_tokens': 4988010.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0008998184930533171, 'sampling/sampling_logp_difference/max': 0.027839738875627518, 'sampling/importance_sampling_ratio/min': 0.9878785610198975, 'sampling/importance_sampling_ratio/mean': 1.0008260011672974, 'sampling/importance_sampling_ratio/max': 1.028230905532837, 'entropy': 0.00906775618204847, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 67%|██████▋   | 2208/3300 [3:02:05<2:23:42,  7.90s/it]INFO 04-12 01:38:57 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2209/3300 [3:02:10<2:05:49,  6.92s/it]2026-04-12 01:39:01,562 | INFO | train_grpo_train | metrics_logged mode=train step=2209

                                                       
{'loss': -0.0012, 'grad_norm': 2.7090210914611816, 'learning_rate': 3.3090909090909097e-06, 'num_tokens': 4989794.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7749799489974976, 'rewards/meter/std': 0.033681225031614304, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7749799489974976, 'rewards/total_composite/std': 0.033681225031614304, 'reward': 0.7749799489974976, 'reward_std': 0.033681221306324005, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00368741643615067, 'sampling/sampling_logp_difference/max': 0.8190579414367676, 'sampling/importance_sampling_ratio/min': 0.4408467411994934, 'sampling/importance_sampling_ratio/mean': 0.9990789890289307, 'sampling/importance_sampling_ratio/max': 1.0404185056686401, 'entropy': 0.011399149021599442, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7749799489974976, 'reward_meter_mean': 0.7749799489974976, 'reward_meter_std': 0.033681225031614304, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7749799489974976, 'reward_total_composite_std': 0.033681225031614304, 'epoch': 0.09}

 67%|██████▋   | 2209/3300 [3:02:10<2:05:49,  6.92s/it]INFO 04-12 01:39:01 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2210/3300 [3:02:15<1:54:52,  6.32s/it]2026-04-12 01:39:06,496 | INFO | train_grpo_train | metrics_logged mode=train step=2210

                                                       
{'loss': 0.0121, 'grad_norm': 6.713454723358154, 'learning_rate': 3.3060606060606065e-06, 'num_tokens': 4991661.0, 'completions/mean_length': 80.375, 'completions/min_length': 78.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9961856007575989, 'rewards/meter/std': 0.0025163544341921806, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961856007575989, 'rewards/total_composite/std': 0.0025163544341921806, 'reward': 0.9961856007575989, 'reward_std': 0.0025163493119180202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039697032421827316, 'sampling/sampling_logp_difference/max': 1.445500373840332, 'sampling/importance_sampling_ratio/min': 0.2356281578540802, 'sampling/importance_sampling_ratio/mean': 1.0097302198410034, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3479005917906761, 'clip_ratio/low_mean': 0.018732663709670305, 'clip_ratio/low_min': 0.018732663709670305, 'clip_ratio/high_mean': 0.01074601011350751, 'clip_ratio/high_max': 0.01074601011350751, 'clip_ratio/region_mean': 0.029478673823177814, 'reward_total_mean': 0.9961856007575989, 'reward_meter_mean': 0.9961856007575989, 'reward_meter_std': 0.0025163544341921806, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961856007575989, 'reward_total_composite_std': 0.0025163544341921806, 'epoch': 0.09}

 67%|██████▋   | 2210/3300 [3:02:15<1:54:52,  6.32s/it]INFO 04-12 01:39:06 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2211/3300 [3:02:20<1:48:13,  5.96s/it]2026-04-12 01:39:11,615 | INFO | train_grpo_train | metrics_logged mode=train step=2211

                                                       
{'loss': 0.0013, 'grad_norm': 0.7238816022872925, 'learning_rate': 3.3030303030303033e-06, 'num_tokens': 4993997.0, 'completions/mean_length': 111.0, 'completions/min_length': 111.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.0, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9956170320510864, 'rewards/meter/std': 6.146137457108125e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8533860445022583, 'rewards/total_composite/std': 5.2672676247311756e-05, 'reward': 0.8533860445022583, 'reward_std': 5.2673178288387135e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007354200817644596, 'sampling/sampling_logp_difference/max': 0.7475004196166992, 'sampling/importance_sampling_ratio/min': 0.4735487401485443, 'sampling/importance_sampling_ratio/mean': 0.9998362064361572, 'sampling/importance_sampling_ratio/max': 1.6665387153625488, 'entropy': 0.03893151739612222, 'clip_ratio/low_mean': 0.0022522523067891598, 'clip_ratio/low_min': 0.0022522523067891598, 'clip_ratio/high_mean': 0.0045045046135783195, 'clip_ratio/high_max': 0.0045045046135783195, 'clip_ratio/region_mean': 0.006756756920367479, 'reward_total_mean': 0.8533860445022583, 'reward_meter_mean': 0.9956170320510864, 'reward_meter_std': 6.146137457108125e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8533860445022583, 'reward_total_composite_std': 5.2672676247311756e-05, 'epoch': 0.09}

 67%|██████▋   | 2211/3300 [3:02:20<1:48:13,  5.96s/it]INFO 04-12 01:39:11 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2212/3300 [3:02:27<1:55:36,  6.38s/it]2026-04-12 01:39:18,955 | INFO | train_grpo_train | metrics_logged mode=train step=2212

                                                       
{'loss': 0.0064, 'grad_norm': 1.9566975831985474, 'learning_rate': 3.3000000000000006e-06, 'num_tokens': 4998173.0, 'completions/mean_length': 276.0, 'completions/min_length': 274.0, 'completions/max_length': 277.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 276.0, 'completions/min_terminated_length': 274.0, 'completions/max_terminated_length': 277.0, 'rewards/meter/mean': 0.9973146319389343, 'rewards/meter/std': 0.0006697024800814688, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8666666746139526, 'rewards/repeat_penalty/std': 0.06172133609652519, 'rewards/total_composite/mean': 0.5762275457382202, 'rewards/total_composite/std': 0.04107243940234184, 'reward': 0.5762275457382202, 'reward_std': 0.041072458028793335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015320011414587498, 'sampling/sampling_logp_difference/max': 1.8627681732177734, 'sampling/importance_sampling_ratio/min': 0.17606841027736664, 'sampling/importance_sampling_ratio/mean': 1.0007983446121216, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06808455288410187, 'clip_ratio/low_mean': 0.005421726265922189, 'clip_ratio/low_min': 0.005421726265922189, 'clip_ratio/high_mean': 0.006823194562457502, 'clip_ratio/high_max': 0.006823194562457502, 'clip_ratio/region_mean': 0.01224492082837969, 'reward_total_mean': 0.5762275457382202, 'reward_meter_mean': 0.9973146319389343, 'reward_meter_std': 0.0006697024800814688, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8666666746139526, 'reward_repeat_penalty_std': 0.06172133609652519, 'reward_total_composite_mean': 0.5762275457382202, 'reward_total_composite_std': 0.04107243940234184, 'epoch': 0.09}

 67%|██████▋   | 2212/3300 [3:02:27<1:55:36,  6.38s/it]INFO 04-12 01:39:19 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2213/3300 [3:02:32<1:43:53,  5.73s/it]2026-04-12 01:39:23,189 | INFO | train_grpo_train | metrics_logged mode=train step=2213

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.2969696969696974e-06, 'num_tokens': 4999621.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9929623007774353, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9929623007774353, 'rewards/total_composite/std': 0.0, 'reward': 0.9929623007774353, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.000746682460885495, 'sampling/sampling_logp_difference/max': 0.023203659802675247, 'sampling/importance_sampling_ratio/min': 0.9999908804893494, 'sampling/importance_sampling_ratio/mean': 1.000752329826355, 'sampling/importance_sampling_ratio/max': 1.023474931716919, 'entropy': 0.004421527351951227, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9929623007774353, 'reward_meter_mean': 0.9929623007774353, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9929623007774353, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 67%|██████▋   | 2213/3300 [3:02:32<1:43:53,  5.73s/it]INFO 04-12 01:39:23 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2214/3300 [3:02:39<1:54:39,  6.33s/it]2026-04-12 01:39:30,931 | INFO | train_grpo_train | metrics_logged mode=train step=2214

                                                       
{'loss': 0.0124, 'grad_norm': 2.7332818508148193, 'learning_rate': 3.2939393939393943e-06, 'num_tokens': 5003989.0, 'completions/mean_length': 306.0, 'completions/min_length': 291.0, 'completions/max_length': 319.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 306.0, 'completions/min_terminated_length': 291.0, 'completions/max_terminated_length': 319.0, 'rewards/meter/mean': 0.996366024017334, 'rewards/meter/std': 0.004282170441001654, 'rewards/count_adherence/mean': 0.6477272510528564, 'rewards/count_adherence/std': 0.03214123100042343, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9187728762626648, 'rewards/repeat_penalty/std': 0.04368516802787781, 'rewards/total_composite/mean': 0.5923222303390503, 'rewards/total_composite/std': 0.027357451617717743, 'reward': 0.5923222303390503, 'reward_std': 0.027357440441846848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04633094370365143, 'sampling/sampling_logp_difference/max': 1.9842910766601562, 'sampling/importance_sampling_ratio/min': 0.1374780386686325, 'sampling/importance_sampling_ratio/mean': 1.0054079294204712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3657448794692755, 'clip_ratio/low_mean': 0.025091978488489985, 'clip_ratio/low_min': 0.025091978488489985, 'clip_ratio/high_mean': 0.015862172469496727, 'clip_ratio/high_max': 0.015862172469496727, 'clip_ratio/region_mean': 0.04095415095798671, 'reward_total_mean': 0.5923222303390503, 'reward_meter_mean': 0.996366024017334, 'reward_meter_std': 0.004282170441001654, 'reward_count_adherence_mean': 0.6477272510528564, 'reward_count_adherence_std': 0.03214123100042343, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9187728762626648, 'reward_repeat_penalty_std': 0.04368516802787781, 'reward_total_composite_mean': 0.5923222303390503, 'reward_total_composite_std': 0.027357451617717743, 'epoch': 0.09}

 67%|██████▋   | 2214/3300 [3:02:39<1:54:39,  6.33s/it]INFO 04-12 01:39:31 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2215/3300 [3:02:45<1:49:01,  6.03s/it]2026-04-12 01:39:36,238 | INFO | train_grpo_train | metrics_logged mode=train step=2215

                                                       
{'loss': -0.0075, 'grad_norm': 3.505354404449463, 'learning_rate': 3.290909090909091e-06, 'num_tokens': 5006191.0, 'completions/mean_length': 122.25, 'completions/min_length': 119.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.25, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.9975504875183105, 'rewards/meter/std': 0.0010059267515316606, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975504875183105, 'rewards/total_composite/std': 0.0010059267515316606, 'reward': 0.9975504875183105, 'reward_std': 0.001005922444164753, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04226119443774223, 'sampling/sampling_logp_difference/max': 1.1195478439331055, 'sampling/importance_sampling_ratio/min': 0.3264273405075073, 'sampling/importance_sampling_ratio/mean': 1.0016958713531494, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3795941434800625, 'clip_ratio/low_mean': 0.011416577035561204, 'clip_ratio/low_min': 0.011416577035561204, 'clip_ratio/high_mean': 0.030406045261770487, 'clip_ratio/high_max': 0.030406045261770487, 'clip_ratio/region_mean': 0.04182262229733169, 'reward_total_mean': 0.9975504875183105, 'reward_meter_mean': 0.9975504875183105, 'reward_meter_std': 0.0010059267515316606, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975504875183105, 'reward_total_composite_std': 0.0010059267515316606, 'epoch': 0.09}

 67%|██████▋   | 2215/3300 [3:02:45<1:49:01,  6.03s/it]INFO 04-12 01:39:36 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2216/3300 [3:02:49<1:40:35,  5.57s/it]2026-04-12 01:39:40,731 | INFO | train_grpo_train | metrics_logged mode=train step=2216

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.2878787878787883e-06, 'num_tokens': 5007999.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9992372989654541, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992372989654541, 'rewards/total_composite/std': 0.0, 'reward': 0.9992372989654541, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0007629028987139463, 'sampling/sampling_logp_difference/max': 0.029877394437789917, 'sampling/importance_sampling_ratio/min': 0.9758114814758301, 'sampling/importance_sampling_ratio/mean': 1.0005970001220703, 'sampling/importance_sampling_ratio/max': 1.0303282737731934, 'entropy': 0.008104118285700679, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992372989654541, 'reward_meter_mean': 0.9992372989654541, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992372989654541, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 67%|██████▋   | 2216/3300 [3:02:49<1:40:35,  5.57s/it]INFO 04-12 01:39:40 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2217/3300 [3:02:54<1:38:36,  5.46s/it]2026-04-12 01:39:45,948 | INFO | train_grpo_train | metrics_logged mode=train step=2217

                                                       
{'loss': -0.0229, 'grad_norm': 3.7353017330169678, 'learning_rate': 3.284848484848485e-06, 'num_tokens': 5010225.0, 'completions/mean_length': 108.25, 'completions/min_length': 102.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.25, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9519559144973755, 'rewards/meter/std': 0.09988751262426376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8596517443656921, 'rewards/total_composite/std': 0.15026217699050903, 'reward': 0.8596517443656921, 'reward_std': 0.15026216208934784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02141641452908516, 'sampling/sampling_logp_difference/max': 1.004293441772461, 'sampling/importance_sampling_ratio/min': 0.36630338430404663, 'sampling/importance_sampling_ratio/mean': 1.003779649734497, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14850023202598095, 'clip_ratio/low_mean': 0.005999947316013277, 'clip_ratio/low_min': 0.005999947316013277, 'clip_ratio/high_mean': 0.014177569886669517, 'clip_ratio/high_max': 0.014177569886669517, 'clip_ratio/region_mean': 0.020177517202682793, 'reward_total_mean': 0.8596517443656921, 'reward_meter_mean': 0.9519559144973755, 'reward_meter_std': 0.09988751262426376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8596517443656921, 'reward_total_composite_std': 0.15026217699050903, 'epoch': 0.09}

 67%|██████▋   | 2217/3300 [3:02:54<1:38:36,  5.46s/it]INFO 04-12 01:39:46 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2218/3300 [3:03:04<2:01:27,  6.74s/it]2026-04-12 01:39:55,654 | INFO | train_grpo_train | metrics_logged mode=train step=2218

                                                       
{'loss': -0.003, 'grad_norm': 1.9956766366958618, 'learning_rate': 3.281818181818182e-06, 'num_tokens': 5015622.0, 'completions/mean_length': 474.625, 'completions/min_length': 460.0, 'completions/max_length': 487.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 474.625, 'completions/min_terminated_length': 460.0, 'completions/max_terminated_length': 487.0, 'rewards/meter/mean': 0.9962097406387329, 'rewards/meter/std': 0.003824215615168214, 'rewards/count_adherence/mean': 0.5986841917037964, 'rewards/count_adherence/std': 0.027239417657256126, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8791407942771912, 'rewards/repeat_penalty/std': 0.11821687966585159, 'rewards/total_composite/mean': 0.5244148969650269, 'rewards/total_composite/std': 0.0736735388636589, 'reward': 0.5244148969650269, 'reward_std': 0.07367353141307831, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04105058312416077, 'sampling/sampling_logp_difference/max': 1.8464699983596802, 'sampling/importance_sampling_ratio/min': 0.15779319405555725, 'sampling/importance_sampling_ratio/mean': 1.007934808731079, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28974736854434013, 'clip_ratio/low_mean': 0.011183922295458615, 'clip_ratio/low_min': 0.011183922295458615, 'clip_ratio/high_mean': 0.017237887950614095, 'clip_ratio/high_max': 0.017237887950614095, 'clip_ratio/region_mean': 0.02842181024607271, 'reward_total_mean': 0.5244148969650269, 'reward_meter_mean': 0.9962097406387329, 'reward_meter_std': 0.003824215615168214, 'reward_count_adherence_mean': 0.5986841917037964, 'reward_count_adherence_std': 0.027239417657256126, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8791407942771912, 'reward_repeat_penalty_std': 0.11821687966585159, 'reward_total_composite_mean': 0.5244148969650269, 'reward_total_composite_std': 0.0736735388636589, 'epoch': 0.09}

 67%|██████▋   | 2218/3300 [3:03:04<2:01:27,  6.74s/it]INFO 04-12 01:39:55 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2219/3300 [3:03:09<1:49:43,  6.09s/it]2026-04-12 01:40:00,238 | INFO | train_grpo_train | metrics_logged mode=train step=2219

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.2787878787878793e-06, 'num_tokens': 5017335.0, 'completions/mean_length': 64.125, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9992372989654541, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992372989654541, 'rewards/total_composite/std': 0.0, 'reward': 0.9992372989654541, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0031146432738751173, 'sampling/sampling_logp_difference/max': 0.3054380416870117, 'sampling/importance_sampling_ratio/min': 0.7368006110191345, 'sampling/importance_sampling_ratio/mean': 1.000244140625, 'sampling/importance_sampling_ratio/max': 1.1355178356170654, 'entropy': 0.01567571924533695, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992372989654541, 'reward_meter_mean': 0.9992372989654541, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992372989654541, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 67%|██████▋   | 2219/3300 [3:03:09<1:49:43,  6.09s/it]INFO 04-12 01:40:00 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2220/3300 [3:03:14<1:45:46,  5.88s/it]2026-04-12 01:40:05,612 | INFO | train_grpo_train | metrics_logged mode=train step=2220

                                                       
{'loss': 0.0018, 'grad_norm': 2.9601826667785645, 'learning_rate': 3.275757575757576e-06, 'num_tokens': 5019823.0, 'completions/mean_length': 137.0, 'completions/min_length': 137.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.0, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9971017837524414, 'rewards/meter/std': 0.0003135971783194691, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8546586632728577, 'rewards/total_composite/std': 0.0002687963715288788, 'reward': 0.8546586632728577, 'reward_std': 0.00026877920026890934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003638867987319827, 'sampling/sampling_logp_difference/max': 0.5708838701248169, 'sampling/importance_sampling_ratio/min': 0.565025806427002, 'sampling/importance_sampling_ratio/mean': 0.9999095797538757, 'sampling/importance_sampling_ratio/max': 1.3408164978027344, 'entropy': 0.018264403683133423, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.8546586632728577, 'reward_meter_mean': 0.9971017837524414, 'reward_meter_std': 0.0003135971783194691, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8546586632728577, 'reward_total_composite_std': 0.0002687963715288788, 'epoch': 0.09}

 67%|██████▋   | 2220/3300 [3:03:14<1:45:46,  5.88s/it]INFO 04-12 01:40:05 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2221/3300 [3:03:19<1:39:29,  5.53s/it]2026-04-12 01:40:10,348 | INFO | train_grpo_train | metrics_logged mode=train step=2221

                                                       
{'loss': -0.0001, 'grad_norm': 0.8054800033569336, 'learning_rate': 3.272727272727273e-06, 'num_tokens': 5021798.0, 'completions/mean_length': 82.875, 'completions/min_length': 82.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.875, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9953393936157227, 'rewards/meter/std': 4.823669223696925e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953393936157227, 'rewards/total_composite/std': 4.823669223696925e-05, 'reward': 0.9953393936157227, 'reward_std': 4.822280607186258e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009308185428380966, 'sampling/sampling_logp_difference/max': 0.9146785736083984, 'sampling/importance_sampling_ratio/min': 0.4006454050540924, 'sampling/importance_sampling_ratio/mean': 0.99955153465271, 'sampling/importance_sampling_ratio/max': 1.2980178594589233, 'entropy': 0.03914514696225524, 'clip_ratio/low_mean': 0.0015243901871144772, 'clip_ratio/low_min': 0.0015243901871144772, 'clip_ratio/high_mean': 0.0015060240402817726, 'clip_ratio/high_max': 0.0015060240402817726, 'clip_ratio/region_mean': 0.0030304142273962498, 'reward_total_mean': 0.9953393936157227, 'reward_meter_mean': 0.9953393936157227, 'reward_meter_std': 4.823669223696925e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953393936157227, 'reward_total_composite_std': 4.823669223696925e-05, 'epoch': 0.09}

 67%|██████▋   | 2221/3300 [3:03:19<1:39:29,  5.53s/it]INFO 04-12 01:40:10 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2222/3300 [3:03:26<1:49:14,  6.08s/it]2026-04-12 01:40:17,704 | INFO | train_grpo_train | metrics_logged mode=train step=2222

                                                       
{'loss': 0.0058, 'grad_norm': 1.5890640020370483, 'learning_rate': 3.2696969696969698e-06, 'num_tokens': 5025785.0, 'completions/mean_length': 290.375, 'completions/min_length': 287.0, 'completions/max_length': 293.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 290.375, 'completions/min_terminated_length': 287.0, 'completions/max_terminated_length': 293.0, 'rewards/meter/mean': 0.9988961219787598, 'rewards/meter/std': 0.0002896705409511924, 'rewards/count_adherence/mean': 0.7272727489471436, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9166666269302368, 'rewards/repeat_penalty/std': 0.0471404492855072, 'rewards/total_composite/mean': 0.6659300327301025, 'rewards/total_composite/std': 0.0342310331761837, 'reward': 0.6659300327301025, 'reward_std': 0.03423101827502251, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02183903567492962, 'sampling/sampling_logp_difference/max': 1.4761362075805664, 'sampling/importance_sampling_ratio/min': 0.22851894795894623, 'sampling/importance_sampling_ratio/mean': 1.0035452842712402, 'sampling/importance_sampling_ratio/max': 1.77455735206604, 'entropy': 0.24721034429967403, 'clip_ratio/low_mean': 0.006000505993142724, 'clip_ratio/low_min': 0.006000505993142724, 'clip_ratio/high_mean': 0.01464845088776201, 'clip_ratio/high_max': 0.01464845088776201, 'clip_ratio/region_mean': 0.020648956880904734, 'reward_total_mean': 0.6659300327301025, 'reward_meter_mean': 0.9988961219787598, 'reward_meter_std': 0.0002896705409511924, 'reward_count_adherence_mean': 0.7272727489471436, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9166666269302368, 'reward_repeat_penalty_std': 0.0471404492855072, 'reward_total_composite_mean': 0.6659300327301025, 'reward_total_composite_std': 0.0342310331761837, 'epoch': 0.09}

 67%|██████▋   | 2222/3300 [3:03:26<1:49:14,  6.08s/it]INFO 04-12 01:40:17 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2223/3300 [3:03:32<1:48:40,  6.05s/it]2026-04-12 01:40:23,696 | INFO | train_grpo_train | metrics_logged mode=train step=2223

                                                       
{'loss': 0.0074, 'grad_norm': 1.8738999366760254, 'learning_rate': 3.266666666666667e-06, 'num_tokens': 5028644.0, 'completions/mean_length': 172.375, 'completions/min_length': 171.0, 'completions/max_length': 182.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 172.375, 'completions/min_terminated_length': 171.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.9511035084724426, 'rewards/meter/std': 0.01162177324295044, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6164560317993164, 'rewards/total_composite/std': 0.00753263384103775, 'reward': 0.6164560317993164, 'reward_std': 0.0075326296500861645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012546628713607788, 'sampling/sampling_logp_difference/max': 0.6718964576721191, 'sampling/importance_sampling_ratio/min': 0.5107390880584717, 'sampling/importance_sampling_ratio/mean': 1.002518892288208, 'sampling/importance_sampling_ratio/max': 1.608762502670288, 'entropy': 0.0778789222240448, 'clip_ratio/low_mean': 0.005715410457924008, 'clip_ratio/low_min': 0.005715410457924008, 'clip_ratio/high_mean': 0.002923976629972458, 'clip_ratio/high_max': 0.002923976629972458, 'clip_ratio/region_mean': 0.008639387087896466, 'reward_total_mean': 0.6164560317993164, 'reward_meter_mean': 0.9511035084724426, 'reward_meter_std': 0.01162177324295044, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6164560317993164, 'reward_total_composite_std': 0.00753263384103775, 'epoch': 0.09}

 67%|██████▋   | 2223/3300 [3:03:32<1:48:40,  6.05s/it]INFO 04-12 01:40:23 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2224/3300 [3:03:37<1:40:00,  5.58s/it]2026-04-12 01:40:28,161 | INFO | train_grpo_train | metrics_logged mode=train step=2224

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.263636363636364e-06, 'num_tokens': 5030428.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0009967123623937368, 'sampling/sampling_logp_difference/max': 0.025670286267995834, 'sampling/importance_sampling_ratio/min': 0.9967485070228577, 'sampling/importance_sampling_ratio/mean': 1.000982403755188, 'sampling/importance_sampling_ratio/max': 1.0260026454925537, 'entropy': 0.00872120977146551, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 67%|██████▋   | 2224/3300 [3:03:37<1:40:00,  5.58s/it]INFO 04-12 01:40:28 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2225/3300 [3:03:41<1:35:26,  5.33s/it]2026-04-12 01:40:32,906 | INFO | train_grpo_train | metrics_logged mode=train step=2225

                                                       
{'loss': 0.0108, 'grad_norm': 3.8838281631469727, 'learning_rate': 3.2606060606060607e-06, 'num_tokens': 5032200.0, 'completions/mean_length': 76.5, 'completions/min_length': 73.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9987069964408875, 'rewards/meter/std': 0.0007462438079528511, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987069964408875, 'rewards/total_composite/std': 0.0007462438079528511, 'reward': 0.9987069964408875, 'reward_std': 0.0007462415378540754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03162068501114845, 'sampling/sampling_logp_difference/max': 0.9698367118835449, 'sampling/importance_sampling_ratio/min': 0.37914493680000305, 'sampling/importance_sampling_ratio/mean': 1.0066579580307007, 'sampling/importance_sampling_ratio/max': 1.6162910461425781, 'entropy': 0.27160678431391716, 'clip_ratio/low_mean': 0.0032748287776485085, 'clip_ratio/low_min': 0.0032748287776485085, 'clip_ratio/high_mean': 0.021526333526708186, 'clip_ratio/high_max': 0.021526333526708186, 'clip_ratio/region_mean': 0.024801162304356694, 'reward_total_mean': 0.9987069964408875, 'reward_meter_mean': 0.9987069964408875, 'reward_meter_std': 0.0007462438079528511, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987069964408875, 'reward_total_composite_std': 0.0007462438079528511, 'epoch': 0.09}

 67%|██████▋   | 2225/3300 [3:03:41<1:35:26,  5.33s/it]INFO 04-12 01:40:33 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2226/3300 [3:03:46<1:29:49,  5.02s/it]2026-04-12 01:40:37,206 | INFO | train_grpo_train | metrics_logged mode=train step=2226

                                                       
{'loss': 0.0153, 'grad_norm': 7.719726085662842, 'learning_rate': 3.257575757575758e-06, 'num_tokens': 5033803.0, 'completions/mean_length': 36.375, 'completions/min_length': 36.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9948956370353699, 'rewards/meter/std': 0.005730676930397749, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948956370353699, 'rewards/total_composite/std': 0.005730676930397749, 'reward': 0.9948956370353699, 'reward_std': 0.005730684380978346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028130551800131798, 'sampling/sampling_logp_difference/max': 0.9965043067932129, 'sampling/importance_sampling_ratio/min': 0.369167685508728, 'sampling/importance_sampling_ratio/mean': 1.006754994392395, 'sampling/importance_sampling_ratio/max': 1.3892041444778442, 'entropy': 0.18548811599612236, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010416666744276881, 'clip_ratio/high_max': 0.010416666744276881, 'clip_ratio/region_mean': 0.010416666744276881, 'reward_total_mean': 0.9948956370353699, 'reward_meter_mean': 0.9948956370353699, 'reward_meter_std': 0.005730676930397749, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948956370353699, 'reward_total_composite_std': 0.005730676930397749, 'epoch': 0.09}

 67%|██████▋   | 2226/3300 [3:03:46<1:29:49,  5.02s/it]INFO 04-12 01:40:37 [block_pool.py:378] Successfully reset prefix cache

 67%|██████▋   | 2227/3300 [3:03:56<1:57:05,  6.55s/it]2026-04-12 01:40:47,321 | INFO | train_grpo_train | metrics_logged mode=train step=2227

                                                       
{'loss': 0.205, 'grad_norm': 1.227808952331543, 'learning_rate': 3.2545454545454548e-06, 'num_tokens': 5038942.0, 'completions/mean_length': 477.375, 'completions/min_length': 445.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 472.4285888671875, 'completions/min_terminated_length': 445.0, 'completions/max_terminated_length': 505.0, 'rewards/meter/mean': 0.9960645437240601, 'rewards/meter/std': 0.0029486457351595163, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.0353553481400013, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8150820732116699, 'rewards/repeat_penalty/std': 0.08796744048595428, 'rewards/total_composite/mean': 0.4574677348136902, 'rewards/total_composite/std': 0.06483428180217743, 'reward': 0.4574677348136902, 'reward_std': 0.06483428180217743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03101653791964054, 'sampling/sampling_logp_difference/max': 2.6449055671691895, 'sampling/importance_sampling_ratio/min': 0.07101205736398697, 'sampling/importance_sampling_ratio/mean': 1.0072216987609863, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21459073200821877, 'clip_ratio/low_mean': 0.014457646873779595, 'clip_ratio/low_min': 0.014457646873779595, 'clip_ratio/high_mean': 0.005880667129531503, 'clip_ratio/high_max': 0.005880667129531503, 'clip_ratio/region_mean': 0.020338314003311098, 'reward_total_mean': 0.4574677348136902, 'reward_meter_mean': 0.9960645437240601, 'reward_meter_std': 0.0029486457351595163, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.0353553481400013, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8150820732116699, 'reward_repeat_penalty_std': 0.08796744048595428, 'reward_total_composite_mean': 0.4574677348136902, 'reward_total_composite_std': 0.06483428180217743, 'epoch': 0.09}

 67%|██████▋   | 2227/3300 [3:03:56<1:57:05,  6.55s/it]INFO 04-12 01:40:47 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2228/3300 [3:04:00<1:45:44,  5.92s/it]2026-04-12 01:40:51,769 | INFO | train_grpo_train | metrics_logged mode=train step=2228

                                                       
{'loss': 0.0001, 'grad_norm': 0.01318140048533678, 'learning_rate': 3.2515151515151516e-06, 'num_tokens': 5040574.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7876332998275757, 'rewards/meter/std': 1.5573279597447254e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7876332998275757, 'rewards/total_composite/std': 1.5573279597447254e-05, 'reward': 0.7876332998275757, 'reward_std': 1.558229632792063e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0009326103026978672, 'sampling/sampling_logp_difference/max': 0.06658339500427246, 'sampling/importance_sampling_ratio/min': 0.991096556186676, 'sampling/importance_sampling_ratio/mean': 1.0008612871170044, 'sampling/importance_sampling_ratio/max': 1.068850040435791, 'entropy': 0.007959658920299262, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.7876332998275757, 'reward_meter_mean': 0.7876332998275757, 'reward_meter_std': 1.5573279597447254e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7876332998275757, 'reward_total_composite_std': 1.5573279597447254e-05, 'epoch': 0.09}

 68%|██████▊   | 2228/3300 [3:04:00<1:45:44,  5.92s/it]INFO 04-12 01:40:52 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2229/3300 [3:04:06<1:45:07,  5.89s/it]2026-04-12 01:40:57,593 | INFO | train_grpo_train | metrics_logged mode=train step=2229

                                                       
{'loss': -0.0005, 'grad_norm': 2.3057451248168945, 'learning_rate': 3.2484848484848484e-06, 'num_tokens': 5043518.0, 'completions/mean_length': 166.0, 'completions/min_length': 165.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.0, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9984830617904663, 'rewards/meter/std': 6.16059624007903e-05, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.05143444612622261, 'rewards/total_composite/mean': 0.7165045738220215, 'rewards/total_composite/std': 0.042806532233953476, 'reward': 0.7165045738220215, 'reward_std': 0.042806532233953476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017393097281455994, 'sampling/sampling_logp_difference/max': 2.966561794281006, 'sampling/importance_sampling_ratio/min': 0.0514800027012825, 'sampling/importance_sampling_ratio/mean': 1.0007810592651367, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.06033759843558073, 'clip_ratio/low_mean': 0.0015151514671742916, 'clip_ratio/low_min': 0.0015151514671742916, 'clip_ratio/high_mean': 0.010519623290747404, 'clip_ratio/high_max': 0.010519623290747404, 'clip_ratio/region_mean': 0.012034774757921696, 'reward_total_mean': 0.7165045738220215, 'reward_meter_mean': 0.9984830617904663, 'reward_meter_std': 6.16059624007903e-05, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.05143444612622261, 'reward_total_composite_mean': 0.7165045738220215, 'reward_total_composite_std': 0.042806532233953476, 'epoch': 0.09}

 68%|██████▊   | 2229/3300 [3:04:06<1:45:07,  5.89s/it]INFO 04-12 01:40:57 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2230/3300 [3:04:10<1:36:44,  5.42s/it]2026-04-12 01:41:01,932 | INFO | train_grpo_train | metrics_logged mode=train step=2230

                                                       
{'loss': 0.0099, 'grad_norm': 10.35647964477539, 'learning_rate': 3.2454545454545457e-06, 'num_tokens': 5044997.0, 'completions/mean_length': 33.875, 'completions/min_length': 33.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9669080972671509, 'rewards/meter/std': 0.007700720801949501, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9669080972671509, 'rewards/total_composite/std': 0.007700720801949501, 'reward': 0.9669080972671509, 'reward_std': 0.007700727321207523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042054299265146255, 'sampling/sampling_logp_difference/max': 2.527313470840454, 'sampling/importance_sampling_ratio/min': 0.07987331598997116, 'sampling/importance_sampling_ratio/mean': 1.003213882446289, 'sampling/importance_sampling_ratio/max': 1.6818870306015015, 'entropy': 0.1314310822635889, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.02239304850809276, 'clip_ratio/high_max': 0.02239304850809276, 'clip_ratio/region_mean': 0.02239304850809276, 'reward_total_mean': 0.9669080972671509, 'reward_meter_mean': 0.9669080972671509, 'reward_meter_std': 0.007700720801949501, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9669080972671509, 'reward_total_composite_std': 0.007700720801949501, 'epoch': 0.09}

 68%|██████▊   | 2230/3300 [3:04:10<1:36:44,  5.42s/it]INFO 04-12 01:41:02 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2231/3300 [3:04:15<1:31:13,  5.12s/it]2026-04-12 01:41:06,344 | INFO | train_grpo_train | metrics_logged mode=train step=2231

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.2424242424242425e-06, 'num_tokens': 5046893.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0011728927493095398, 'sampling/sampling_logp_difference/max': 0.02333247661590576, 'sampling/importance_sampling_ratio/min': 0.9904347658157349, 'sampling/importance_sampling_ratio/mean': 1.0010844469070435, 'sampling/importance_sampling_ratio/max': 1.023606777191162, 'entropy': 0.010291360202245414, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 68%|██████▊   | 2231/3300 [3:04:15<1:31:13,  5.12s/it]INFO 04-12 01:41:06 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2232/3300 [3:04:19<1:28:30,  4.97s/it]2026-04-12 01:41:10,975 | INFO | train_grpo_train | metrics_logged mode=train step=2232

                                                       
{'loss': -0.0002, 'grad_norm': 0.10460558533668518, 'learning_rate': 3.2393939393939393e-06, 'num_tokens': 5048693.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9992489814758301, 'rewards/meter/std': 1.6926040188991465e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992489814758301, 'rewards/total_composite/std': 1.6926040188991465e-05, 'reward': 0.9992489814758301, 'reward_std': 1.6926040188991465e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006007255986332893, 'sampling/sampling_logp_difference/max': 1.3880231380462646, 'sampling/importance_sampling_ratio/min': 0.24956819415092468, 'sampling/importance_sampling_ratio/mean': 0.9996911883354187, 'sampling/importance_sampling_ratio/max': 1.2030258178710938, 'entropy': 0.019377628806978464, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0078125, 'clip_ratio/high_max': 0.0078125, 'clip_ratio/region_mean': 0.0078125, 'reward_total_mean': 0.9992489814758301, 'reward_meter_mean': 0.9992489814758301, 'reward_meter_std': 1.6926040188991465e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992489814758301, 'reward_total_composite_std': 1.6926040188991465e-05, 'epoch': 0.09}

 68%|██████▊   | 2232/3300 [3:04:19<1:28:30,  4.97s/it]INFO 04-12 01:41:11 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2233/3300 [3:04:25<1:31:52,  5.17s/it]2026-04-12 01:41:16,589 | INFO | train_grpo_train | metrics_logged mode=train step=2233

                                                       
{'loss': 0.0026, 'grad_norm': 1.9100828170776367, 'learning_rate': 3.236363636363636e-06, 'num_tokens': 5051460.0, 'completions/mean_length': 145.875, 'completions/min_length': 143.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 145.875, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.9988158941268921, 'rewards/meter/std': 0.0003107816446572542, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9107142686843872, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.7277184724807739, 'rewards/total_composite/std': 0.059221845120191574, 'reward': 0.7277184724807739, 'reward_std': 0.059221841394901276, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027437884360551834, 'sampling/sampling_logp_difference/max': 1.669896125793457, 'sampling/importance_sampling_ratio/min': 0.18826662003993988, 'sampling/importance_sampling_ratio/mean': 1.0000747442245483, 'sampling/importance_sampling_ratio/max': 1.9528251886367798, 'entropy': 0.2233992125838995, 'clip_ratio/low_mean': 0.012797378934919834, 'clip_ratio/low_min': 0.012797378934919834, 'clip_ratio/high_mean': 0.005999534041620791, 'clip_ratio/high_max': 0.005999534041620791, 'clip_ratio/region_mean': 0.018796912976540625, 'reward_total_mean': 0.7277184724807739, 'reward_meter_mean': 0.9988158941268921, 'reward_meter_std': 0.0003107816446572542, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9107142686843872, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.7277184724807739, 'reward_total_composite_std': 0.059221845120191574, 'epoch': 0.09}

 68%|██████▊   | 2233/3300 [3:04:25<1:31:52,  5.17s/it]INFO 04-12 01:41:16 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2234/3300 [3:04:30<1:29:31,  5.04s/it]2026-04-12 01:41:21,334 | INFO | train_grpo_train | metrics_logged mode=train step=2234

                                                       
{'loss': 0.0004, 'grad_norm': 3.689906120300293, 'learning_rate': 3.2333333333333334e-06, 'num_tokens': 5053322.0, 'completions/mean_length': 68.75, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9638321399688721, 'rewards/meter/std': 0.004275594372302294, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9638321399688721, 'rewards/total_composite/std': 0.004275594372302294, 'reward': 0.9638321399688721, 'reward_std': 0.004275602288544178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02401396632194519, 'sampling/sampling_logp_difference/max': 1.2271108627319336, 'sampling/importance_sampling_ratio/min': 0.2931382656097412, 'sampling/importance_sampling_ratio/mean': 0.9958019852638245, 'sampling/importance_sampling_ratio/max': 1.3294564485549927, 'entropy': 0.10225279349833727, 'clip_ratio/low_mean': 0.005488064838573337, 'clip_ratio/low_min': 0.005488064838573337, 'clip_ratio/high_mean': 0.017986543010920286, 'clip_ratio/high_max': 0.017986543010920286, 'clip_ratio/region_mean': 0.023474607849493623, 'reward_total_mean': 0.9638321399688721, 'reward_meter_mean': 0.9638321399688721, 'reward_meter_std': 0.004275594372302294, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9638321399688721, 'reward_total_composite_std': 0.004275594372302294, 'epoch': 0.09}

 68%|██████▊   | 2234/3300 [3:04:30<1:29:31,  5.04s/it]INFO 04-12 01:41:21 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2235/3300 [3:04:35<1:28:06,  4.96s/it]2026-04-12 01:41:26,118 | INFO | train_grpo_train | metrics_logged mode=train step=2235

                                                       
{'loss': 0.0039, 'grad_norm': 1.4711960554122925, 'learning_rate': 3.2303030303030307e-06, 'num_tokens': 5055211.0, 'completions/mean_length': 69.125, 'completions/min_length': 69.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9970292448997498, 'rewards/meter/std': 0.0009962875628843904, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970292448997498, 'rewards/total_composite/std': 0.0009962875628843904, 'reward': 0.9970292448997498, 'reward_std': 0.0009962901240214705, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008085141889750957, 'sampling/sampling_logp_difference/max': 0.4244270324707031, 'sampling/importance_sampling_ratio/min': 0.6541445255279541, 'sampling/importance_sampling_ratio/mean': 1.0035878419876099, 'sampling/importance_sampling_ratio/max': 1.4033783674240112, 'entropy': 0.06750913895666599, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.005434782709926367, 'clip_ratio/high_max': 0.005434782709926367, 'clip_ratio/region_mean': 0.007220497005619109, 'reward_total_mean': 0.9970292448997498, 'reward_meter_mean': 0.9970292448997498, 'reward_meter_std': 0.0009962875628843904, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970292448997498, 'reward_total_composite_std': 0.0009962875628843904, 'epoch': 0.09}

 68%|██████▊   | 2235/3300 [3:04:35<1:28:06,  4.96s/it]INFO 04-12 01:41:26 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2236/3300 [3:04:40<1:28:48,  5.01s/it]2026-04-12 01:41:31,231 | INFO | train_grpo_train | metrics_logged mode=train step=2236

                                                       
{'loss': 0.0014, 'grad_norm': 0.5810080766677856, 'learning_rate': 3.227272727272728e-06, 'num_tokens': 5057433.0, 'completions/mean_length': 102.75, 'completions/min_length': 102.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.75, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9973079562187195, 'rewards/meter/std': 4.1739614971447736e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973079562187195, 'rewards/total_composite/std': 4.1739614971447736e-05, 'reward': 0.9973079562187195, 'reward_std': 4.174207424512133e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008572143502533436, 'sampling/sampling_logp_difference/max': 0.9729032516479492, 'sampling/importance_sampling_ratio/min': 0.39697858691215515, 'sampling/importance_sampling_ratio/mean': 1.0025743246078491, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.05867503536865115, 'clip_ratio/low_mean': 0.0036407767329365015, 'clip_ratio/low_min': 0.0036407767329365015, 'clip_ratio/high_mean': 0.006067961221560836, 'clip_ratio/high_max': 0.006067961221560836, 'clip_ratio/region_mean': 0.009708737954497337, 'reward_total_mean': 0.9973079562187195, 'reward_meter_mean': 0.9973079562187195, 'reward_meter_std': 4.1739614971447736e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973079562187195, 'reward_total_composite_std': 4.1739614971447736e-05, 'epoch': 0.09}

 68%|██████▊   | 2236/3300 [3:04:40<1:28:48,  5.01s/it]INFO 04-12 01:41:31 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2237/3300 [3:04:45<1:32:50,  5.24s/it]2026-04-12 01:41:37,024 | INFO | train_grpo_train | metrics_logged mode=train step=2237

                                                       
{'loss': -0.0042, 'grad_norm': 2.2996366024017334, 'learning_rate': 3.2242424242424248e-06, 'num_tokens': 5060180.0, 'completions/mean_length': 158.375, 'completions/min_length': 155.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.375, 'completions/min_terminated_length': 155.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.9958896636962891, 'rewards/meter/std': 0.0022456336300820112, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.7824714183807373, 'rewards/total_composite/std': 0.040016695857048035, 'reward': 0.7824714183807373, 'reward_std': 0.04001671448349953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0335751473903656, 'sampling/sampling_logp_difference/max': 1.1518535614013672, 'sampling/importance_sampling_ratio/min': 0.3160504102706909, 'sampling/importance_sampling_ratio/mean': 1.003827452659607, 'sampling/importance_sampling_ratio/max': 1.8041925430297852, 'entropy': 0.28765890188515186, 'clip_ratio/low_mean': 0.004807692486792803, 'clip_ratio/low_min': 0.004807692486792803, 'clip_ratio/high_mean': 0.021151655237190425, 'clip_ratio/high_max': 0.021151655237190425, 'clip_ratio/region_mean': 0.025959347723983228, 'reward_total_mean': 0.7824714183807373, 'reward_meter_mean': 0.9958896636962891, 'reward_meter_std': 0.0022456336300820112, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.7824714183807373, 'reward_total_composite_std': 0.040016695857048035, 'epoch': 0.09}

 68%|██████▊   | 2237/3300 [3:04:46<1:32:50,  5.24s/it]INFO 04-12 01:41:37 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2238/3300 [3:04:50<1:29:58,  5.08s/it]2026-04-12 01:41:41,731 | INFO | train_grpo_train | metrics_logged mode=train step=2238

                                                       
{'loss': 0.0, 'grad_norm': 0.5047301650047302, 'learning_rate': 3.2212121212121216e-06, 'num_tokens': 5061987.0, 'completions/mean_length': 68.875, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9973945617675781, 'rewards/meter/std': 5.974572559352964e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973945617675781, 'rewards/total_composite/std': 5.974572559352964e-05, 'reward': 0.9973945617675781, 'reward_std': 5.973832230665721e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008767606690526009, 'sampling/sampling_logp_difference/max': 0.3773939609527588, 'sampling/importance_sampling_ratio/min': 0.7254200577735901, 'sampling/importance_sampling_ratio/mean': 1.0013201236724854, 'sampling/importance_sampling_ratio/max': 1.4584788084030151, 'entropy': 0.07339224359020591, 'clip_ratio/low_mean': 0.005461423774249852, 'clip_ratio/low_min': 0.005461423774249852, 'clip_ratio/high_mean': 0.010869565419852734, 'clip_ratio/high_max': 0.010869565419852734, 'clip_ratio/region_mean': 0.016330989194102585, 'reward_total_mean': 0.9973945617675781, 'reward_meter_mean': 0.9973945617675781, 'reward_meter_std': 5.974572559352964e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973945617675781, 'reward_total_composite_std': 5.974572559352964e-05, 'epoch': 0.09}

 68%|██████▊   | 2238/3300 [3:04:50<1:29:58,  5.08s/it]INFO 04-12 01:41:41 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2239/3300 [3:04:55<1:31:11,  5.16s/it]2026-04-12 01:41:47,075 | INFO | train_grpo_train | metrics_logged mode=train step=2239

                                                       
{'loss': 0.0221, 'grad_norm': 4.3548431396484375, 'learning_rate': 3.2181818181818184e-06, 'num_tokens': 5064107.0, 'completions/mean_length': 106.0, 'completions/min_length': 100.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.0, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9938052296638489, 'rewards/meter/std': 0.002556393388658762, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8500000238418579, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.8447941541671753, 'rewards/total_composite/std': 0.09272917360067368, 'reward': 0.8447941541671753, 'reward_std': 0.09272917360067368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02209414355456829, 'sampling/sampling_logp_difference/max': 1.215738296508789, 'sampling/importance_sampling_ratio/min': 0.3138589859008789, 'sampling/importance_sampling_ratio/mean': 1.0017590522766113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1653594933450222, 'clip_ratio/low_mean': 0.010472985217347741, 'clip_ratio/low_min': 0.010472985217347741, 'clip_ratio/high_mean': 0.00717289699241519, 'clip_ratio/high_max': 0.00717289699241519, 'clip_ratio/region_mean': 0.01764588220976293, 'reward_total_mean': 0.8447941541671753, 'reward_meter_mean': 0.9938052296638489, 'reward_meter_std': 0.002556393388658762, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8500000238418579, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.8447941541671753, 'reward_total_composite_std': 0.09272917360067368, 'epoch': 0.09}

 68%|██████▊   | 2239/3300 [3:04:56<1:31:11,  5.16s/it]INFO 04-12 01:41:47 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2240/3300 [3:05:01<1:31:56,  5.20s/it]2026-04-12 01:41:52,377 | INFO | train_grpo_train | metrics_logged mode=train step=2240

                                                       
{'loss': 0.0227, 'grad_norm': 6.652218341827393, 'learning_rate': 3.2151515151515157e-06, 'num_tokens': 5066446.0, 'completions/mean_length': 112.375, 'completions/min_length': 108.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.375, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.8900635242462158, 'rewards/meter/std': 0.201475590467453, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8900635242462158, 'rewards/total_composite/std': 0.201475590467453, 'reward': 0.8900635242462158, 'reward_std': 0.2014755755662918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05001511052250862, 'sampling/sampling_logp_difference/max': 1.6565048694610596, 'sampling/importance_sampling_ratio/min': 0.19080470502376556, 'sampling/importance_sampling_ratio/mean': 1.0028852224349976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3458435367792845, 'clip_ratio/low_mean': 0.008410736452788115, 'clip_ratio/low_min': 0.008410736452788115, 'clip_ratio/high_mean': 0.036260453751310706, 'clip_ratio/high_max': 0.036260453751310706, 'clip_ratio/region_mean': 0.04467119020409882, 'reward_total_mean': 0.8900635242462158, 'reward_meter_mean': 0.8900635242462158, 'reward_meter_std': 0.201475590467453, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8900635242462158, 'reward_total_composite_std': 0.201475590467453, 'epoch': 0.09}

 68%|██████▊   | 2240/3300 [3:05:01<1:31:56,  5.20s/it]INFO 04-12 01:41:52 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2241/3300 [3:05:06<1:29:29,  5.07s/it]2026-04-12 01:41:57,131 | INFO | train_grpo_train | metrics_logged mode=train step=2241

                                                       
{'loss': -0.0017, 'grad_norm': 2.789885997772217, 'learning_rate': 3.2121212121212125e-06, 'num_tokens': 5068415.0, 'completions/mean_length': 74.125, 'completions/min_length': 73.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.125, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9991567134857178, 'rewards/meter/std': 0.0002589829673524946, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991567134857178, 'rewards/total_composite/std': 0.0002589829673524946, 'reward': 0.9991567134857178, 'reward_std': 0.00025898346211761236, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029841933399438858, 'sampling/sampling_logp_difference/max': 1.069981575012207, 'sampling/importance_sampling_ratio/min': 0.34301483631134033, 'sampling/importance_sampling_ratio/mean': 1.0056432485580444, 'sampling/importance_sampling_ratio/max': 1.596889853477478, 'entropy': 0.2299621980637312, 'clip_ratio/low_mean': 0.006711711874231696, 'clip_ratio/low_min': 0.006711711874231696, 'clip_ratio/high_mean': 0.01353727001696825, 'clip_ratio/high_max': 0.01353727001696825, 'clip_ratio/region_mean': 0.020248981891199946, 'reward_total_mean': 0.9991567134857178, 'reward_meter_mean': 0.9991567134857178, 'reward_meter_std': 0.0002589829673524946, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991567134857178, 'reward_total_composite_std': 0.0002589829673524946, 'epoch': 0.09}

 68%|██████▊   | 2241/3300 [3:05:06<1:29:29,  5.07s/it]INFO 04-12 01:41:57 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2242/3300 [3:05:10<1:28:17,  5.01s/it]2026-04-12 01:42:01,996 | INFO | train_grpo_train | metrics_logged mode=train step=2242

                                                       
{'loss': -0.0007, 'grad_norm': 1.4458215236663818, 'learning_rate': 3.2090909090909094e-06, 'num_tokens': 5070277.0, 'completions/mean_length': 68.75, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9974172115325928, 'rewards/meter/std': 9.149048128165305e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974172115325928, 'rewards/total_composite/std': 9.149048128165305e-05, 'reward': 0.9974172115325928, 'reward_std': 9.149555262411013e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01175268180668354, 'sampling/sampling_logp_difference/max': 0.9791126251220703, 'sampling/importance_sampling_ratio/min': 0.3756442964076996, 'sampling/importance_sampling_ratio/mean': 1.0000346899032593, 'sampling/importance_sampling_ratio/max': 1.4807547330856323, 'entropy': 0.06746953492984176, 'clip_ratio/low_mean': 0.007326300139538944, 'clip_ratio/low_min': 0.007326300139538944, 'clip_ratio/high_mean': 0.007246376946568489, 'clip_ratio/high_max': 0.007246376946568489, 'clip_ratio/region_mean': 0.014572677086107433, 'reward_total_mean': 0.9974172115325928, 'reward_meter_mean': 0.9974172115325928, 'reward_meter_std': 9.149048128165305e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974172115325928, 'reward_total_composite_std': 9.149048128165305e-05, 'epoch': 0.09}

 68%|██████▊   | 2242/3300 [3:05:10<1:28:17,  5.01s/it]INFO 04-12 01:42:02 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2243/3300 [3:05:15<1:26:38,  4.92s/it]2026-04-12 01:42:06,705 | INFO | train_grpo_train | metrics_logged mode=train step=2243

                                                       
{'loss': -0.0004, 'grad_norm': 0.028668954968452454, 'learning_rate': 3.2060606060606066e-06, 'num_tokens': 5071893.0, 'completions/mean_length': 55.0, 'completions/min_length': 55.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9948811531066895, 'rewards/meter/std': 1.4479670653599896e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948811531066895, 'rewards/total_composite/std': 1.4479670653599896e-06, 'reward': 0.9948811531066895, 'reward_std': 1.448755483579589e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005395011510699987, 'sampling/sampling_logp_difference/max': 0.7887990474700928, 'sampling/importance_sampling_ratio/min': 0.45439019799232483, 'sampling/importance_sampling_ratio/mean': 0.9990595579147339, 'sampling/importance_sampling_ratio/max': 1.0825896263122559, 'entropy': 0.01730545365717262, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9948811531066895, 'reward_meter_mean': 0.9948811531066895, 'reward_meter_std': 1.4479670653599896e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948811531066895, 'reward_total_composite_std': 1.4479670653599896e-06, 'epoch': 0.09}

 68%|██████▊   | 2243/3300 [3:05:15<1:26:38,  4.92s/it]INFO 04-12 01:42:06 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2244/3300 [3:05:20<1:27:08,  4.95s/it]2026-04-12 01:42:11,731 | INFO | train_grpo_train | metrics_logged mode=train step=2244

                                                       
{'loss': 0.0324, 'grad_norm': 2.5679850578308105, 'learning_rate': 3.2030303030303034e-06, 'num_tokens': 5073801.0, 'completions/mean_length': 70.5, 'completions/min_length': 68.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9620155096054077, 'rewards/meter/std': 0.010956778191030025, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9620155096054077, 'rewards/total_composite/std': 0.010956778191030025, 'reward': 0.9620155096054077, 'reward_std': 0.010956759564578533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026699479669332504, 'sampling/sampling_logp_difference/max': 1.6187305450439453, 'sampling/importance_sampling_ratio/min': 0.19815008342266083, 'sampling/importance_sampling_ratio/mean': 1.0068614482879639, 'sampling/importance_sampling_ratio/max': 1.9989465475082397, 'entropy': 0.14800493232905865, 'clip_ratio/low_mean': 0.013456470100209117, 'clip_ratio/low_min': 0.013456470100209117, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.015294705401174724, 'reward_total_mean': 0.9620155096054077, 'reward_meter_mean': 0.9620155096054077, 'reward_meter_std': 0.010956778191030025, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9620155096054077, 'reward_total_composite_std': 0.010956778191030025, 'epoch': 0.09}

 68%|██████▊   | 2244/3300 [3:05:20<1:27:08,  4.95s/it]INFO 04-12 01:42:11 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2245/3300 [3:05:26<1:32:05,  5.24s/it]2026-04-12 01:42:17,663 | INFO | train_grpo_train | metrics_logged mode=train step=2245

                                                       
{'loss': 0.0006, 'grad_norm': 1.3484723567962646, 'learning_rate': 3.2000000000000003e-06, 'num_tokens': 5076689.0, 'completions/mean_length': 167.0, 'completions/min_length': 167.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 167.0, 'completions/min_terminated_length': 167.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9959465265274048, 'rewards/meter/std': 9.931746171787381e-05, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8409091234207153, 'rewards/repeat_penalty/std': 0.08058230578899384, 'rewards/total_composite/mean': 0.717853307723999, 'rewards/total_composite/std': 0.06874335557222366, 'reward': 0.717853307723999, 'reward_std': 0.06874334812164307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010420297272503376, 'sampling/sampling_logp_difference/max': 1.2074825763702393, 'sampling/importance_sampling_ratio/min': 0.29894891381263733, 'sampling/importance_sampling_ratio/mean': 0.9994251132011414, 'sampling/importance_sampling_ratio/max': 1.3406785726547241, 'entropy': 0.05933140218257904, 'clip_ratio/low_mean': 0.0022455090656876564, 'clip_ratio/low_min': 0.0022455090656876564, 'clip_ratio/high_mean': 0.005239521153271198, 'clip_ratio/high_max': 0.005239521153271198, 'clip_ratio/region_mean': 0.007485030218958855, 'reward_total_mean': 0.717853307723999, 'reward_meter_mean': 0.9959465265274048, 'reward_meter_std': 9.931746171787381e-05, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8409091234207153, 'reward_repeat_penalty_std': 0.08058230578899384, 'reward_total_composite_mean': 0.717853307723999, 'reward_total_composite_std': 0.06874335557222366, 'epoch': 0.09}

 68%|██████▊   | 2245/3300 [3:05:26<1:32:05,  5.24s/it]INFO 04-12 01:42:17 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2246/3300 [3:05:32<1:35:52,  5.46s/it]2026-04-12 01:42:23,607 | INFO | train_grpo_train | metrics_logged mode=train step=2246

                                                       
{'loss': 0.0044, 'grad_norm': 2.09184193611145, 'learning_rate': 3.196969696969697e-06, 'num_tokens': 5079477.0, 'completions/mean_length': 143.5, 'completions/min_length': 136.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.5, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.9908416867256165, 'rewards/meter/std': 0.005638771690428257, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8392857313156128, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.6652302742004395, 'rewards/total_composite/std': 0.03939948230981827, 'reward': 0.6652302742004395, 'reward_std': 0.03939947113394737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02601451985538006, 'sampling/sampling_logp_difference/max': 1.5025901794433594, 'sampling/importance_sampling_ratio/min': 0.22255297005176544, 'sampling/importance_sampling_ratio/mean': 1.004660725593567, 'sampling/importance_sampling_ratio/max': 1.8299243450164795, 'entropy': 0.19038555212318897, 'clip_ratio/low_mean': 0.0008802816737443209, 'clip_ratio/low_min': 0.0008802816737443209, 'clip_ratio/high_mean': 0.018073388375341892, 'clip_ratio/high_max': 0.018073388375341892, 'clip_ratio/region_mean': 0.018953670049086213, 'reward_total_mean': 0.6652302742004395, 'reward_meter_mean': 0.9908416867256165, 'reward_meter_std': 0.005638771690428257, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8392857313156128, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.6652302742004395, 'reward_total_composite_std': 0.03939948230981827, 'epoch': 0.09}

 68%|██████▊   | 2246/3300 [3:05:32<1:35:52,  5.46s/it]INFO 04-12 01:42:23 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2247/3300 [3:05:38<1:37:19,  5.55s/it]2026-04-12 01:42:29,356 | INFO | train_grpo_train | metrics_logged mode=train step=2247

                                                       
{'loss': 0.0006, 'grad_norm': 2.32881760597229, 'learning_rate': 3.1939393939393944e-06, 'num_tokens': 5082066.0, 'completions/mean_length': 142.625, 'completions/min_length': 141.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.625, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9939021468162537, 'rewards/meter/std': 0.0005415793275460601, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8519160747528076, 'rewards/total_composite/std': 0.0004642106650862843, 'reward': 0.8519160747528076, 'reward_std': 0.0004642065614461899, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014642242342233658, 'sampling/sampling_logp_difference/max': 1.4676084518432617, 'sampling/importance_sampling_ratio/min': 0.2304760217666626, 'sampling/importance_sampling_ratio/mean': 1.001724362373352, 'sampling/importance_sampling_ratio/max': 1.7263494729995728, 'entropy': 0.10466479323804379, 'clip_ratio/low_mean': 0.001760650658980012, 'clip_ratio/low_min': 0.001760650658980012, 'clip_ratio/high_mean': 0.0008741258643567562, 'clip_ratio/high_max': 0.0008741258643567562, 'clip_ratio/region_mean': 0.002634776523336768, 'reward_total_mean': 0.8519160747528076, 'reward_meter_mean': 0.9939021468162537, 'reward_meter_std': 0.0005415793275460601, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8519160747528076, 'reward_total_composite_std': 0.0004642106650862843, 'epoch': 0.09}

 68%|██████▊   | 2247/3300 [3:05:38<1:37:19,  5.55s/it]INFO 04-12 01:42:29 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2248/3300 [3:05:43<1:34:28,  5.39s/it]2026-04-12 01:42:34,381 | INFO | train_grpo_train | metrics_logged mode=train step=2248

                                                       
{'loss': 0.0011, 'grad_norm': 2.4443159103393555, 'learning_rate': 3.190909090909091e-06, 'num_tokens': 5084426.0, 'completions/mean_length': 111.0, 'completions/min_length': 111.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.0, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9956398010253906, 'rewards/meter/std': 9.625943494029343e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9422965049743652, 'rewards/total_composite/std': 0.07353661954402924, 'reward': 0.9422965049743652, 'reward_std': 0.07353662699460983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008148727007210255, 'sampling/sampling_logp_difference/max': 0.5144362449645996, 'sampling/importance_sampling_ratio/min': 0.5978375673294067, 'sampling/importance_sampling_ratio/mean': 1.0003148317337036, 'sampling/importance_sampling_ratio/max': 1.3495516777038574, 'entropy': 0.06026614410802722, 'clip_ratio/low_mean': 0.0022522523067891598, 'clip_ratio/low_min': 0.0022522523067891598, 'clip_ratio/high_mean': 0.0045045046135783195, 'clip_ratio/high_max': 0.0045045046135783195, 'clip_ratio/region_mean': 0.006756756920367479, 'reward_total_mean': 0.9422965049743652, 'reward_meter_mean': 0.9956398010253906, 'reward_meter_std': 9.625943494029343e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9422965049743652, 'reward_total_composite_std': 0.07353661954402924, 'epoch': 0.09}

 68%|██████▊   | 2248/3300 [3:05:43<1:34:28,  5.39s/it]INFO 04-12 01:42:34 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2249/3300 [3:05:47<1:29:53,  5.13s/it]2026-04-12 01:42:38,914 | INFO | train_grpo_train | metrics_logged mode=train step=2249

                                                       
{'loss': 0.0001, 'grad_norm': 0.3145919442176819, 'learning_rate': 3.187878787878788e-06, 'num_tokens': 5086546.0, 'completions/mean_length': 69.0, 'completions/min_length': 69.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9975405931472778, 'rewards/meter/std': 1.4270765859691892e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975405931472778, 'rewards/total_composite/std': 1.4270765859691892e-05, 'reward': 0.9975405931472778, 'reward_std': 1.4251797438191716e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006101167760789394, 'sampling/sampling_logp_difference/max': 0.688692569732666, 'sampling/importance_sampling_ratio/min': 0.5022322535514832, 'sampling/importance_sampling_ratio/mean': 1.0008234977722168, 'sampling/importance_sampling_ratio/max': 1.2142425775527954, 'entropy': 0.05507086403667927, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9975405931472778, 'reward_meter_mean': 0.9975405931472778, 'reward_meter_std': 1.4270765859691892e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975405931472778, 'reward_total_composite_std': 1.4270765859691892e-05, 'epoch': 0.09}

 68%|██████▊   | 2249/3300 [3:05:47<1:29:53,  5.13s/it]INFO 04-12 01:42:39 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2250/3300 [3:05:52<1:26:38,  4.95s/it]2026-04-12 01:42:43,442 | INFO | train_grpo_train | metrics_logged mode=train step=2250

                                                       
{'loss': -0.0011, 'grad_norm': 0.5331827402114868, 'learning_rate': 3.1848484848484853e-06, 'num_tokens': 5088505.0, 'completions/mean_length': 68.875, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9975307583808899, 'rewards/meter/std': 2.9643033485626802e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975307583808899, 'rewards/total_composite/std': 2.9643033485626802e-05, 'reward': 0.9975307583808899, 'reward_std': 2.9639279091497883e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005882658995687962, 'sampling/sampling_logp_difference/max': 0.6369953155517578, 'sampling/importance_sampling_ratio/min': 0.5288791656494141, 'sampling/importance_sampling_ratio/mean': 1.0018889904022217, 'sampling/importance_sampling_ratio/max': 1.2806227207183838, 'entropy': 0.0506328740157187, 'clip_ratio/low_mean': 0.0036498295376077294, 'clip_ratio/low_min': 0.0036498295376077294, 'clip_ratio/high_mean': 0.0036231884732842445, 'clip_ratio/high_max': 0.0036231884732842445, 'clip_ratio/region_mean': 0.007273018010891974, 'reward_total_mean': 0.9975307583808899, 'reward_meter_mean': 0.9975307583808899, 'reward_meter_std': 2.9643033485626802e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975307583808899, 'reward_total_composite_std': 2.9643033485626802e-05, 'epoch': 0.09}

 68%|██████▊   | 2250/3300 [3:05:52<1:26:38,  4.95s/it]INFO 04-12 01:42:43 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:29,  2.68s/it]

 23%|██▎       | 3/13 [00:08<00:28,  2.87s/it]

 31%|███       | 4/13 [00:12<00:30,  3.40s/it]

 38%|███▊      | 5/13 [00:17<00:30,  3.83s/it]

 46%|████▌     | 6/13 [00:22<00:30,  4.34s/it]

 54%|█████▍    | 7/13 [00:27<00:26,  4.43s/it]

 62%|██████▏   | 8/13 [00:34<00:26,  5.34s/it]

 69%|██████▉   | 9/13 [00:40<00:21,  5.42s/it]

 77%|███████▋  | 10/13 [00:45<00:15,  5.33s/it]

 85%|████████▍ | 11/13 [00:50<00:10,  5.22s/it]

 92%|█████████▏| 12/13 [00:55<00:05,  5.03s/it]

100%|██████████| 13/13 [00:59<00:00,  4.79s/it]2026-04-12 01:43:47,813 | INFO | train_grpo_train | metrics_logged mode=eval step=2250

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 64.292, 'eval_samples_per_second': 1.618, 'eval_steps_per_second': 0.202, 'eval_num_tokens': 5088505.0, 'eval_completions/mean_length': 182.43269230769232, 'eval_completions/min_length': 59.53846153846154, 'eval_completions/max_length': 330.53846153846155, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 180.01648418719952, 'eval_completions/min_terminated_length': 59.53846153846154, 'eval_completions/max_terminated_length': 319.53846153846155, 'eval_rewards/meter/mean': 0.732613939505357, 'eval_rewards/meter/std': 0.3826414684836681, 'eval_rewards/count_adherence/mean': 0.8588019150954026, 'eval_rewards/count_adherence/std': 0.1499052156622593, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.875946778517503, 'eval_rewards/repeat_penalty/std': 0.12293276706567177, 'eval_rewards/total_composite/mean': 0.5561339992743272, 'eval_rewards/total_composite/std': 0.3474527712051685, 'eval_reward': 0.5561339992743272, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0185515835451392, 'eval_sampling/sampling_logp_difference/max': 0.9445312756758469, 'eval_sampling/importance_sampling_ratio/min': 0.40484776290563435, 'eval_sampling/importance_sampling_ratio/mean': 1.0042996865052443, 'eval_sampling/importance_sampling_ratio/max': 1.410868342106159, 'eval_entropy': 0.19319924941429725, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5561339992743272, 'eval_reward_meter_mean': 0.732613939505357, 'eval_reward_meter_std': 0.3826414684836681, 'eval_reward_count_adherence_mean': 0.8588019150954026, 'eval_reward_count_adherence_std': 0.1499052156622593, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.875946778517503, 'eval_reward_repeat_penalty_std': 0.12293276706567177, 'eval_reward_total_composite_mean': 0.5561339992743272, 'eval_reward_total_composite_std': 0.3474527712051685, 'epoch': 0.09}

 68%|██████▊   | 2250/3300 [3:06:56<1:26:38,  4.95s/it]

100%|██████████| 13/13 [00:59<00:00,  4.79s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 68%|██████▊   | 2251/3300 [3:07:04<7:18:00, 25.05s/it]2026-04-12 01:43:55,401 | INFO | train_grpo_train | metrics_logged mode=train step=2251

                                                       
{'loss': 0.0069, 'grad_norm': 2.5881564617156982, 'learning_rate': 3.181818181818182e-06, 'num_tokens': 5090737.0, 'completions/mean_length': 102.0, 'completions/min_length': 101.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.0, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9657948613166809, 'rewards/meter/std': 0.005292544141411781, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9657948613166809, 'rewards/total_composite/std': 0.005292544141411781, 'reward': 0.9657948613166809, 'reward_std': 0.005292540416121483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022372471168637276, 'sampling/sampling_logp_difference/max': 1.110783576965332, 'sampling/importance_sampling_ratio/min': 0.32930082082748413, 'sampling/importance_sampling_ratio/mean': 1.007267713546753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15675346460193396, 'clip_ratio/low_mean': 0.009804157423786819, 'clip_ratio/low_min': 0.009804157423786819, 'clip_ratio/high_mean': 0.003688839729875326, 'clip_ratio/high_max': 0.003688839729875326, 'clip_ratio/region_mean': 0.013492997153662145, 'reward_total_mean': 0.9657948613166809, 'reward_meter_mean': 0.9657948613166809, 'reward_meter_std': 0.005292544141411781, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9657948613166809, 'reward_total_composite_std': 0.005292544141411781, 'epoch': 0.09}

 68%|██████▊   | 2251/3300 [3:07:04<7:18:00, 25.05s/it]INFO 04-12 01:43:55 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2252/3300 [3:07:09<5:32:16, 19.02s/it]2026-04-12 01:44:00,351 | INFO | train_grpo_train | metrics_logged mode=train step=2252

                                                       
{'loss': 0.0092, 'grad_norm': 8.668261528015137, 'learning_rate': 3.178787878787879e-06, 'num_tokens': 5092817.0, 'completions/mean_length': 71.0, 'completions/min_length': 69.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9977865219116211, 'rewards/meter/std': 0.0025578902568668127, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977865219116211, 'rewards/total_composite/std': 0.0025578902568668127, 'reward': 0.9977865219116211, 'reward_std': 0.0025578862987458706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06398905813694, 'sampling/sampling_logp_difference/max': 1.5597095489501953, 'sampling/importance_sampling_ratio/min': 0.2101971060037613, 'sampling/importance_sampling_ratio/mean': 1.0122182369232178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43995901197195053, 'clip_ratio/low_mean': 0.012399396393448114, 'clip_ratio/low_min': 0.012399396393448114, 'clip_ratio/high_mean': 0.031607592245563865, 'clip_ratio/high_max': 0.031607592245563865, 'clip_ratio/region_mean': 0.04400698863901198, 'reward_total_mean': 0.9977865219116211, 'reward_meter_mean': 0.9977865219116211, 'reward_meter_std': 0.0025578902568668127, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977865219116211, 'reward_total_composite_std': 0.0025578902568668127, 'epoch': 0.09}

 68%|██████▊   | 2252/3300 [3:07:09<5:32:16, 19.02s/it]INFO 04-12 01:44:00 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2253/3300 [3:07:13<4:15:33, 14.65s/it]2026-04-12 01:44:04,829 | INFO | train_grpo_train | metrics_logged mode=train step=2253

                                                       
{'loss': -0.0002, 'grad_norm': 0.0889202132821083, 'learning_rate': 3.1757575757575758e-06, 'num_tokens': 5094681.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7876288890838623, 'rewards/meter/std': 2.8027659936924465e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7876288890838623, 'rewards/total_composite/std': 2.8027659936924465e-05, 'reward': 0.7876288890838623, 'reward_std': 2.8021637263009325e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0013179951347410679, 'sampling/sampling_logp_difference/max': 0.07088017463684082, 'sampling/importance_sampling_ratio/min': 0.9315734505653381, 'sampling/importance_sampling_ratio/mean': 1.0006846189498901, 'sampling/importance_sampling_ratio/max': 1.052701711654663, 'entropy': 0.01149469253141433, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.7876288890838623, 'reward_meter_mean': 0.7876288890838623, 'reward_meter_std': 2.8027659936924465e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7876288890838623, 'reward_total_composite_std': 2.8027659936924465e-05, 'epoch': 0.09}

 68%|██████▊   | 2253/3300 [3:07:13<4:15:33, 14.65s/it]INFO 04-12 01:44:05 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2254/3300 [3:07:20<3:32:14, 12.17s/it]2026-04-12 01:44:11,196 | INFO | train_grpo_train | metrics_logged mode=train step=2254

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.172727272727273e-06, 'num_tokens': 5096369.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0015978036681190133, 'sampling/sampling_logp_difference/max': 0.11908148974180222, 'sampling/importance_sampling_ratio/min': 0.8877354264259338, 'sampling/importance_sampling_ratio/mean': 1.0007966756820679, 'sampling/importance_sampling_ratio/max': 1.057936429977417, 'entropy': 0.010870376718230546, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 68%|██████▊   | 2254/3300 [3:07:20<3:32:14, 12.17s/it]INFO 04-12 01:44:11 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2255/3300 [3:07:24<2:51:19,  9.84s/it]2026-04-12 01:44:15,575 | INFO | train_grpo_train | metrics_logged mode=train step=2255

                                                       
{'loss': -0.0042, 'grad_norm': 11.48292350769043, 'learning_rate': 3.16969696969697e-06, 'num_tokens': 5097900.0, 'completions/mean_length': 33.375, 'completions/min_length': 33.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9692210555076599, 'rewards/meter/std': 0.003493846394121647, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9692210555076599, 'rewards/total_composite/std': 0.003493846394121647, 'reward': 0.9692210555076599, 'reward_std': 0.0034938445314764977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03685401380062103, 'sampling/sampling_logp_difference/max': 1.65773606300354, 'sampling/importance_sampling_ratio/min': 0.1905699372291565, 'sampling/importance_sampling_ratio/mean': 1.0003447532653809, 'sampling/importance_sampling_ratio/max': 1.707613468170166, 'entropy': 0.1904709991067648, 'clip_ratio/low_mean': 0.007464349502697587, 'clip_ratio/low_min': 0.007464349502697587, 'clip_ratio/high_mean': 0.018939394503831863, 'clip_ratio/high_max': 0.018939394503831863, 'clip_ratio/region_mean': 0.02640374400652945, 'reward_total_mean': 0.9692210555076599, 'reward_meter_mean': 0.9692210555076599, 'reward_meter_std': 0.003493846394121647, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9692210555076599, 'reward_total_composite_std': 0.003493846394121647, 'epoch': 0.09}

 68%|██████▊   | 2255/3300 [3:07:24<2:51:19,  9.84s/it]INFO 04-12 01:44:15 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2256/3300 [3:07:30<2:32:12,  8.75s/it]2026-04-12 01:44:21,785 | INFO | train_grpo_train | metrics_logged mode=train step=2256

                                                       
{'loss': 0.0115, 'grad_norm': 2.849970817565918, 'learning_rate': 3.1666666666666667e-06, 'num_tokens': 5100800.0, 'completions/mean_length': 176.5, 'completions/min_length': 169.0, 'completions/max_length': 188.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 176.5, 'completions/min_terminated_length': 169.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.998549222946167, 'rewards/meter/std': 0.00034224658156745136, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9027777910232544, 'rewards/repeat_penalty/std': 0.07120776921510696, 'rewards/total_composite/mean': 0.7512168884277344, 'rewards/total_composite/std': 0.05916035175323486, 'reward': 0.7512168884277344, 'reward_std': 0.05916035920381546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04010716825723648, 'sampling/sampling_logp_difference/max': 2.1722252368927, 'sampling/importance_sampling_ratio/min': 0.1139238253235817, 'sampling/importance_sampling_ratio/mean': 1.005678415298462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.27090173214673996, 'clip_ratio/low_mean': 0.02541261224541813, 'clip_ratio/low_min': 0.02541261224541813, 'clip_ratio/high_mean': 0.011609932873398066, 'clip_ratio/high_max': 0.011609932873398066, 'clip_ratio/region_mean': 0.0370225451188162, 'reward_total_mean': 0.7512168884277344, 'reward_meter_mean': 0.998549222946167, 'reward_meter_std': 0.00034224658156745136, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9027777910232544, 'reward_repeat_penalty_std': 0.07120776921510696, 'reward_total_composite_mean': 0.7512168884277344, 'reward_total_composite_std': 0.05916035175323486, 'epoch': 0.09}

 68%|██████▊   | 2256/3300 [3:07:30<2:32:12,  8.75s/it]INFO 04-12 01:44:22 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2257/3300 [3:07:37<2:22:12,  8.18s/it]2026-04-12 01:44:28,639 | INFO | train_grpo_train | metrics_logged mode=train step=2257

                                                       
{'loss': -0.0557, 'grad_norm': 2.3163866996765137, 'learning_rate': 3.1636363636363635e-06, 'num_tokens': 5104173.0, 'completions/mean_length': 222.625, 'completions/min_length': 192.0, 'completions/max_length': 244.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 222.625, 'completions/min_terminated_length': 192.0, 'completions/max_terminated_length': 244.0, 'rewards/meter/mean': 0.9975998401641846, 'rewards/meter/std': 0.0006318566738627851, 'rewards/count_adherence/mean': 0.8035714030265808, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8901515007019043, 'rewards/repeat_penalty/std': 0.057505469769239426, 'rewards/total_composite/mean': 0.7136683464050293, 'rewards/total_composite/std': 0.08285731077194214, 'reward': 0.7136683464050293, 'reward_std': 0.08285731077194214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03854899853467941, 'sampling/sampling_logp_difference/max': 1.9127869606018066, 'sampling/importance_sampling_ratio/min': 0.14766825735569, 'sampling/importance_sampling_ratio/mean': 1.0057517290115356, 'sampling/importance_sampling_ratio/max': 1.9951218366622925, 'entropy': 0.273501418530941, 'clip_ratio/low_mean': 0.01093733258312568, 'clip_ratio/low_min': 0.01093733258312568, 'clip_ratio/high_mean': 0.01619163854047656, 'clip_ratio/high_max': 0.01619163854047656, 'clip_ratio/region_mean': 0.02712897112360224, 'reward_total_mean': 0.7136683464050293, 'reward_meter_mean': 0.9975998401641846, 'reward_meter_std': 0.0006318566738627851, 'reward_count_adherence_mean': 0.8035714030265808, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8901515007019043, 'reward_repeat_penalty_std': 0.057505469769239426, 'reward_total_composite_mean': 0.7136683464050293, 'reward_total_composite_std': 0.08285731077194214, 'epoch': 0.09}

 68%|██████▊   | 2257/3300 [3:07:37<2:22:12,  8.18s/it]INFO 04-12 01:44:28 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2258/3300 [3:07:43<2:10:13,  7.50s/it]2026-04-12 01:44:34,544 | INFO | train_grpo_train | metrics_logged mode=train step=2258

                                                       
{'loss': 0.014, 'grad_norm': 2.4006783962249756, 'learning_rate': 3.1606060606060608e-06, 'num_tokens': 5106939.0, 'completions/mean_length': 173.75, 'completions/min_length': 170.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.75, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.9978283047676086, 'rewards/meter/std': 0.0015693600289523602, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.7968448400497437, 'rewards/total_composite/std': 0.04724044352769852, 'reward': 0.7968448400497437, 'reward_std': 0.04724043607711792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03900068625807762, 'sampling/sampling_logp_difference/max': 1.3687856197357178, 'sampling/importance_sampling_ratio/min': 0.25441572070121765, 'sampling/importance_sampling_ratio/mean': 1.0058785676956177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29748900793492794, 'clip_ratio/low_mean': 0.00781669607385993, 'clip_ratio/low_min': 0.00781669607385993, 'clip_ratio/high_mean': 0.029055888997390866, 'clip_ratio/high_max': 0.029055888997390866, 'clip_ratio/region_mean': 0.036872585071250796, 'reward_total_mean': 0.7968448400497437, 'reward_meter_mean': 0.9978283047676086, 'reward_meter_std': 0.0015693600289523602, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.7968448400497437, 'reward_total_composite_std': 0.04724044352769852, 'epoch': 0.09}

 68%|██████▊   | 2258/3300 [3:07:43<2:10:13,  7.50s/it]INFO 04-12 01:44:34 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2259/3300 [3:07:48<1:58:27,  6.83s/it]2026-04-12 01:44:39,815 | INFO | train_grpo_train | metrics_logged mode=train step=2259

                                                       
{'loss': -0.0036, 'grad_norm': 5.68310022354126, 'learning_rate': 3.1575757575757576e-06, 'num_tokens': 5108781.0, 'completions/mean_length': 70.25, 'completions/min_length': 67.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9816989898681641, 'rewards/meter/std': 0.026165146380662918, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9816989898681641, 'rewards/total_composite/std': 0.026165146380662918, 'reward': 0.9816989898681641, 'reward_std': 0.02616514265537262, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031211039051413536, 'sampling/sampling_logp_difference/max': 1.0319323539733887, 'sampling/importance_sampling_ratio/min': 0.35631778836250305, 'sampling/importance_sampling_ratio/mean': 1.0080952644348145, 'sampling/importance_sampling_ratio/max': 1.8370952606201172, 'entropy': 0.27262550778687, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.024707219563424587, 'clip_ratio/high_max': 0.024707219563424587, 'clip_ratio/region_mean': 0.02827864815481007, 'reward_total_mean': 0.9816989898681641, 'reward_meter_mean': 0.9816989898681641, 'reward_meter_std': 0.026165146380662918, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9816989898681641, 'reward_total_composite_std': 0.026165146380662918, 'epoch': 0.09}

 68%|██████▊   | 2259/3300 [3:07:48<1:58:27,  6.83s/it]INFO 04-12 01:44:40 [block_pool.py:378] Successfully reset prefix cache

 68%|██████▊   | 2260/3300 [3:07:56<2:05:39,  7.25s/it]2026-04-12 01:44:48,042 | INFO | train_grpo_train | metrics_logged mode=train step=2260

                                                       
{'loss': -0.0005, 'grad_norm': 1.6112135648727417, 'learning_rate': 3.1545454545454545e-06, 'num_tokens': 5113402.0, 'completions/mean_length': 355.625, 'completions/min_length': 346.0, 'completions/max_length': 369.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 355.625, 'completions/min_terminated_length': 346.0, 'completions/max_terminated_length': 369.0, 'rewards/meter/mean': 0.9965711236000061, 'rewards/meter/std': 0.0014478074153885245, 'rewards/count_adherence/mean': 0.6428571343421936, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8088235259056091, 'rewards/repeat_penalty/std': 0.07539646327495575, 'rewards/total_composite/mean': 0.5181742906570435, 'rewards/total_composite/std': 0.04829566553235054, 'reward': 0.5181742906570435, 'reward_std': 0.048295676708221436, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034783411771059036, 'sampling/sampling_logp_difference/max': 2.516125440597534, 'sampling/importance_sampling_ratio/min': 0.08077196031808853, 'sampling/importance_sampling_ratio/mean': 1.0037370920181274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2432896252721548, 'clip_ratio/low_mean': 0.005656401859596372, 'clip_ratio/low_min': 0.005656401859596372, 'clip_ratio/high_mean': 0.0209878021851182, 'clip_ratio/high_max': 0.0209878021851182, 'clip_ratio/region_mean': 0.02664420404471457, 'reward_total_mean': 0.5181742906570435, 'reward_meter_mean': 0.9965711236000061, 'reward_meter_std': 0.0014478074153885245, 'reward_count_adherence_mean': 0.6428571343421936, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8088235259056091, 'reward_repeat_penalty_std': 0.07539646327495575, 'reward_total_composite_mean': 0.5181742906570435, 'reward_total_composite_std': 0.04829566553235054, 'epoch': 0.09}

 68%|██████▊   | 2260/3300 [3:07:57<2:05:39,  7.25s/it]INFO 04-12 01:44:48 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2261/3300 [3:08:02<1:54:24,  6.61s/it]2026-04-12 01:44:53,150 | INFO | train_grpo_train | metrics_logged mode=train step=2261

                                                       
{'loss': 0.0152, 'grad_norm': 3.7612602710723877, 'learning_rate': 3.1515151515151517e-06, 'num_tokens': 5115515.0, 'completions/mean_length': 104.125, 'completions/min_length': 100.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.125, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9984878301620483, 'rewards/meter/std': 0.000643216073513031, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9735202789306641, 'rewards/total_composite/std': 0.07053547352552414, 'reward': 0.9735202789306641, 'reward_std': 0.07053548842668533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04828030988574028, 'sampling/sampling_logp_difference/max': 2.574674606323242, 'sampling/importance_sampling_ratio/min': 0.07617860287427902, 'sampling/importance_sampling_ratio/mean': 1.0051672458648682, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34618273936212063, 'clip_ratio/low_mean': 0.00589622650295496, 'clip_ratio/low_min': 0.00589622650295496, 'clip_ratio/high_mean': 0.04133175266906619, 'clip_ratio/high_max': 0.04133175266906619, 'clip_ratio/region_mean': 0.04722797917202115, 'reward_total_mean': 0.9735202789306641, 'reward_meter_mean': 0.9984878301620483, 'reward_meter_std': 0.000643216073513031, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9735202789306641, 'reward_total_composite_std': 0.07053547352552414, 'epoch': 0.09}

 69%|██████▊   | 2261/3300 [3:08:02<1:54:24,  6.61s/it]INFO 04-12 01:44:53 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2262/3300 [3:08:06<1:43:59,  6.01s/it]2026-04-12 01:44:57,768 | INFO | train_grpo_train | metrics_logged mode=train step=2262

                                                       
{'loss': 0.0082, 'grad_norm': 3.4450576305389404, 'learning_rate': 3.1484848484848485e-06, 'num_tokens': 5117374.0, 'completions/mean_length': 68.375, 'completions/min_length': 67.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9669647812843323, 'rewards/meter/std': 0.014694334007799625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9669647812843323, 'rewards/total_composite/std': 0.014694334007799625, 'reward': 0.9669647812843323, 'reward_std': 0.014694339595735073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028783582150936127, 'sampling/sampling_logp_difference/max': 1.1841890811920166, 'sampling/importance_sampling_ratio/min': 0.3059942126274109, 'sampling/importance_sampling_ratio/mean': 1.0112721920013428, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19510209374129772, 'clip_ratio/low_mean': 0.0018115942366421223, 'clip_ratio/low_min': 0.0018115942366421223, 'clip_ratio/high_mean': 0.0202221788931638, 'clip_ratio/high_max': 0.0202221788931638, 'clip_ratio/region_mean': 0.022033773129805923, 'reward_total_mean': 0.9669647812843323, 'reward_meter_mean': 0.9669647812843323, 'reward_meter_std': 0.014694334007799625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9669647812843323, 'reward_total_composite_std': 0.014694334007799625, 'epoch': 0.09}

 69%|██████▊   | 2262/3300 [3:08:06<1:43:59,  6.01s/it]INFO 04-12 01:44:58 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2263/3300 [3:08:10<1:34:44,  5.48s/it]2026-04-12 01:45:02,016 | INFO | train_grpo_train | metrics_logged mode=train step=2263

                                                       
{'loss': 0.0034, 'grad_norm': 1.744556188583374, 'learning_rate': 3.145454545454546e-06, 'num_tokens': 5118823.0, 'completions/mean_length': 32.125, 'completions/min_length': 32.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9992284774780273, 'rewards/meter/std': 0.00014189988723956048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992284774780273, 'rewards/total_composite/std': 0.00014189988723956048, 'reward': 0.9992284774780273, 'reward_std': 0.00014191023365128785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005864746868610382, 'sampling/sampling_logp_difference/max': 0.17466145753860474, 'sampling/importance_sampling_ratio/min': 0.8397412300109863, 'sampling/importance_sampling_ratio/mean': 0.9978591203689575, 'sampling/importance_sampling_ratio/max': 1.1715004444122314, 'entropy': 0.05214253021404147, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.007694128900766373, 'reward_total_mean': 0.9992284774780273, 'reward_meter_mean': 0.9992284774780273, 'reward_meter_std': 0.00014189988723956048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992284774780273, 'reward_total_composite_std': 0.00014189988723956048, 'epoch': 0.09}

 69%|██████▊   | 2263/3300 [3:08:11<1:34:44,  5.48s/it]INFO 04-12 01:45:02 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2264/3300 [3:08:15<1:29:18,  5.17s/it]2026-04-12 01:45:06,465 | INFO | train_grpo_train | metrics_logged mode=train step=2264

                                                       
{'loss': -0.0007, 'grad_norm': 4.9907097816467285, 'learning_rate': 3.142424242424243e-06, 'num_tokens': 5120274.0, 'completions/mean_length': 33.375, 'completions/min_length': 33.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9775052070617676, 'rewards/meter/std': 0.010550287552177906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9775052070617676, 'rewards/total_composite/std': 0.010550287552177906, 'reward': 0.9775052070617676, 'reward_std': 0.010550293140113354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02235337160527706, 'sampling/sampling_logp_difference/max': 0.6383800506591797, 'sampling/importance_sampling_ratio/min': 0.528147280216217, 'sampling/importance_sampling_ratio/mean': 1.0072755813598633, 'sampling/importance_sampling_ratio/max': 1.8347361087799072, 'entropy': 0.16098480112850666, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.014823656994849443, 'clip_ratio/high_max': 0.014823656994849443, 'clip_ratio/region_mean': 0.018611535895615816, 'reward_total_mean': 0.9775052070617676, 'reward_meter_mean': 0.9775052070617676, 'reward_meter_std': 0.010550287552177906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9775052070617676, 'reward_total_composite_std': 0.010550287552177906, 'epoch': 0.09}

 69%|██████▊   | 2264/3300 [3:08:15<1:29:18,  5.17s/it]INFO 04-12 01:45:06 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2265/3300 [3:08:20<1:28:02,  5.10s/it]2026-04-12 01:45:11,410 | INFO | train_grpo_train | metrics_logged mode=train step=2265

                                                       
{'loss': 0.0095, 'grad_norm': 2.8727030754089355, 'learning_rate': 3.13939393939394e-06, 'num_tokens': 5122367.0, 'completions/mean_length': 83.625, 'completions/min_length': 83.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.625, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9946054816246033, 'rewards/meter/std': 0.0010328061180189252, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946054816246033, 'rewards/total_composite/std': 0.0010328061180189252, 'reward': 0.9946054816246033, 'reward_std': 0.0010328061180189252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018597107380628586, 'sampling/sampling_logp_difference/max': 1.4482271671295166, 'sampling/importance_sampling_ratio/min': 0.23498652875423431, 'sampling/importance_sampling_ratio/mean': 1.0014218091964722, 'sampling/importance_sampling_ratio/max': 1.7508782148361206, 'entropy': 0.08839869406074286, 'clip_ratio/low_mean': 0.01339285762514919, 'clip_ratio/low_min': 0.01339285762514919, 'clip_ratio/high_mean': 0.00745840510353446, 'clip_ratio/high_max': 0.00745840510353446, 'clip_ratio/region_mean': 0.02085126272868365, 'reward_total_mean': 0.9946054816246033, 'reward_meter_mean': 0.9946054816246033, 'reward_meter_std': 0.0010328061180189252, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946054816246033, 'reward_total_composite_std': 0.0010328061180189252, 'epoch': 0.09}

 69%|██████▊   | 2265/3300 [3:08:20<1:28:02,  5.10s/it]INFO 04-12 01:45:11 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2266/3300 [3:08:25<1:30:08,  5.23s/it]2026-04-12 01:45:16,939 | INFO | train_grpo_train | metrics_logged mode=train step=2266

                                                       
{'loss': 0.0001, 'grad_norm': 3.1527485847473145, 'learning_rate': 3.1363636363636367e-06, 'num_tokens': 5125064.0, 'completions/mean_length': 140.125, 'completions/min_length': 139.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.125, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9949551820755005, 'rewards/meter/std': 0.0007479682681150734, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.819444477558136, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.81532883644104, 'rewards/total_composite/std': 0.05752642825245857, 'reward': 0.81532883644104, 'reward_std': 0.057526424527168274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01804959587752819, 'sampling/sampling_logp_difference/max': 1.1149301528930664, 'sampling/importance_sampling_ratio/min': 0.32793816924095154, 'sampling/importance_sampling_ratio/mean': 1.0007151365280151, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09540552459657192, 'clip_ratio/low_mean': 0.006256514578126371, 'clip_ratio/low_min': 0.006256514578126371, 'clip_ratio/high_mean': 0.007080778945237398, 'clip_ratio/high_max': 0.007080778945237398, 'clip_ratio/region_mean': 0.013337293523363769, 'reward_total_mean': 0.81532883644104, 'reward_meter_mean': 0.9949551820755005, 'reward_meter_std': 0.0007479682681150734, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.819444477558136, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.81532883644104, 'reward_total_composite_std': 0.05752642825245857, 'epoch': 0.09}

 69%|██████▊   | 2266/3300 [3:08:25<1:30:08,  5.23s/it]INFO 04-12 01:45:17 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2267/3300 [3:08:31<1:30:57,  5.28s/it]2026-04-12 01:45:22,341 | INFO | train_grpo_train | metrics_logged mode=train step=2267

                                                       
{'loss': 0.0082, 'grad_norm': 2.1978232860565186, 'learning_rate': 3.133333333333334e-06, 'num_tokens': 5127489.0, 'completions/mean_length': 135.125, 'completions/min_length': 134.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.125, 'completions/min_terminated_length': 134.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9982975721359253, 'rewards/meter/std': 0.000450806604931131, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982975721359253, 'rewards/total_composite/std': 0.000450806604931131, 'reward': 0.9982975721359253, 'reward_std': 0.00045079842675477266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022603249177336693, 'sampling/sampling_logp_difference/max': 1.3583526611328125, 'sampling/importance_sampling_ratio/min': 0.2570839524269104, 'sampling/importance_sampling_ratio/mean': 0.9998656511306763, 'sampling/importance_sampling_ratio/max': 1.8831219673156738, 'entropy': 0.09932332392781973, 'clip_ratio/low_mean': 0.006448142405133694, 'clip_ratio/low_min': 0.006448142405133694, 'clip_ratio/high_mean': 0.012057766725774854, 'clip_ratio/high_max': 0.012057766725774854, 'clip_ratio/region_mean': 0.01850590913090855, 'reward_total_mean': 0.9982975721359253, 'reward_meter_mean': 0.9982975721359253, 'reward_meter_std': 0.000450806604931131, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982975721359253, 'reward_total_composite_std': 0.000450806604931131, 'epoch': 0.09}

 69%|██████▊   | 2267/3300 [3:08:31<1:30:57,  5.28s/it]INFO 04-12 01:45:22 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▊   | 2268/3300 [3:08:36<1:31:44,  5.33s/it]2026-04-12 01:45:27,797 | INFO | train_grpo_train | metrics_logged mode=train step=2268

                                                       
{'loss': 0.0066, 'grad_norm': 3.012678384780884, 'learning_rate': 3.130303030303031e-06, 'num_tokens': 5130007.0, 'completions/mean_length': 135.75, 'completions/min_length': 132.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.75, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9679410457611084, 'rewards/meter/std': 0.006429874338209629, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8470007181167603, 'rewards/total_composite/std': 0.05034321919083595, 'reward': 0.8470007181167603, 'reward_std': 0.05034322291612625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022967170923948288, 'sampling/sampling_logp_difference/max': 1.6587629318237305, 'sampling/importance_sampling_ratio/min': 0.3170107305049896, 'sampling/importance_sampling_ratio/mean': 1.0043870210647583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1438667206093669, 'clip_ratio/low_mean': 0.013712125248275697, 'clip_ratio/low_min': 0.013712125248275697, 'clip_ratio/high_mean': 0.005474452394992113, 'clip_ratio/high_max': 0.005474452394992113, 'clip_ratio/region_mean': 0.01918657764326781, 'reward_total_mean': 0.8470007181167603, 'reward_meter_mean': 0.9679410457611084, 'reward_meter_std': 0.006429874338209629, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8470007181167603, 'reward_total_composite_std': 0.05034321919083595, 'epoch': 0.09}

 69%|██████▊   | 2268/3300 [3:08:36<1:31:44,  5.33s/it]INFO 04-12 01:45:28 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2269/3300 [3:08:42<1:35:05,  5.53s/it]2026-04-12 01:45:33,798 | INFO | train_grpo_train | metrics_logged mode=train step=2269

                                                       
{'loss': 0.0004, 'grad_norm': 2.066251754760742, 'learning_rate': 3.1272727272727276e-06, 'num_tokens': 5133013.0, 'completions/mean_length': 178.75, 'completions/min_length': 176.0, 'completions/max_length': 181.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.75, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.9990208148956299, 'rewards/meter/std': 0.00016956882609520108, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8888888955116272, 'rewards/repeat_penalty/std': 0.10286889225244522, 'rewards/total_composite/mean': 0.7400156855583191, 'rewards/total_composite/std': 0.08564041554927826, 'reward': 0.7400156855583191, 'reward_std': 0.08564040064811707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023759307339787483, 'sampling/sampling_logp_difference/max': 1.466665267944336, 'sampling/importance_sampling_ratio/min': 0.2306935042142868, 'sampling/importance_sampling_ratio/mean': 1.0039939880371094, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19854111224412918, 'clip_ratio/low_mean': 0.0028210257878527045, 'clip_ratio/low_min': 0.0028210257878527045, 'clip_ratio/high_mean': 0.012555458233691752, 'clip_ratio/high_max': 0.012555458233691752, 'clip_ratio/region_mean': 0.015376484021544456, 'reward_total_mean': 0.7400156855583191, 'reward_meter_mean': 0.9990208148956299, 'reward_meter_std': 0.00016956882609520108, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8888888955116272, 'reward_repeat_penalty_std': 0.10286889225244522, 'reward_total_composite_mean': 0.7400156855583191, 'reward_total_composite_std': 0.08564041554927826, 'epoch': 0.09}

 69%|██████▉   | 2269/3300 [3:08:42<1:35:05,  5.53s/it]INFO 04-12 01:45:34 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2270/3300 [3:08:48<1:37:42,  5.69s/it]2026-04-12 01:45:39,854 | INFO | train_grpo_train | metrics_logged mode=train step=2270

                                                       
{'loss': 0.0089, 'grad_norm': 3.698033094406128, 'learning_rate': 3.1242424242424245e-06, 'num_tokens': 5136127.0, 'completions/mean_length': 170.25, 'completions/min_length': 162.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.25, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.9971339702606201, 'rewards/meter/std': 0.0020925395656377077, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.7962738275527954, 'rewards/total_composite/std': 0.04691069945693016, 'reward': 0.7962738275527954, 'reward_std': 0.04691069573163986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0504683218896389, 'sampling/sampling_logp_difference/max': 2.827153444290161, 'sampling/importance_sampling_ratio/min': 0.05918107181787491, 'sampling/importance_sampling_ratio/mean': 1.011202096939087, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.38951336964964867, 'clip_ratio/low_mean': 0.01330830343067646, 'clip_ratio/low_min': 0.01330830343067646, 'clip_ratio/high_mean': 0.02643707417882979, 'clip_ratio/high_max': 0.02643707417882979, 'clip_ratio/region_mean': 0.03974537760950625, 'reward_total_mean': 0.7962738275527954, 'reward_meter_mean': 0.9971339702606201, 'reward_meter_std': 0.0020925395656377077, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.7962738275527954, 'reward_total_composite_std': 0.04691069945693016, 'epoch': 0.09}

 69%|██████▉   | 2270/3300 [3:08:48<1:37:42,  5.69s/it]INFO 04-12 01:45:40 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2271/3300 [3:08:53<1:31:54,  5.36s/it]2026-04-12 01:45:44,437 | INFO | train_grpo_train | metrics_logged mode=train step=2271

                                                       
{'loss': 0.0327, 'grad_norm': 6.407037258148193, 'learning_rate': 3.1212121212121217e-06, 'num_tokens': 5137976.0, 'completions/mean_length': 62.125, 'completions/min_length': 57.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.125, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.09545363485813141, 'rewards/meter/std': 0.17621159553527832, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.09545363485813141, 'rewards/total_composite/std': 0.17621159553527832, 'reward': 0.09545363485813141, 'reward_std': 0.17621161043643951, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09075994789600372, 'sampling/sampling_logp_difference/max': 2.8358354568481445, 'sampling/importance_sampling_ratio/min': 0.05866949260234833, 'sampling/importance_sampling_ratio/mean': 0.9893724918365479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3145434930920601, 'clip_ratio/low_mean': 0.0624148678034544, 'clip_ratio/low_min': 0.0624148678034544, 'clip_ratio/high_mean': 0.026904894039034843, 'clip_ratio/high_max': 0.026904894039034843, 'clip_ratio/region_mean': 0.08931976184248924, 'reward_total_mean': 0.09545363485813141, 'reward_meter_mean': 0.09545363485813141, 'reward_meter_std': 0.17621159553527832, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.09545363485813141, 'reward_total_composite_std': 0.17621159553527832, 'epoch': 0.09}

 69%|██████▉   | 2271/3300 [3:08:53<1:31:54,  5.36s/it]INFO 04-12 01:45:44 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2272/3300 [3:08:57<1:27:50,  5.13s/it]2026-04-12 01:45:49,021 | INFO | train_grpo_train | metrics_logged mode=train step=2272

                                                       
{'loss': -0.0059, 'grad_norm': 5.84548282623291, 'learning_rate': 3.1181818181818186e-06, 'num_tokens': 5139832.0, 'completions/mean_length': 69.0, 'completions/min_length': 67.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9690738916397095, 'rewards/meter/std': 0.02259281650185585, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9690738916397095, 'rewards/total_composite/std': 0.02259281650185585, 'reward': 0.9690738916397095, 'reward_std': 0.022592833265662193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031493693590164185, 'sampling/sampling_logp_difference/max': 1.107264757156372, 'sampling/importance_sampling_ratio/min': 0.33046165108680725, 'sampling/importance_sampling_ratio/mean': 1.0030878782272339, 'sampling/importance_sampling_ratio/max': 1.7162621021270752, 'entropy': 0.21663649939000607, 'clip_ratio/low_mean': 0.00911509501747787, 'clip_ratio/low_min': 0.00911509501747787, 'clip_ratio/high_mean': 0.01436411531176418, 'clip_ratio/high_max': 0.01436411531176418, 'clip_ratio/region_mean': 0.02347921032924205, 'reward_total_mean': 0.9690738916397095, 'reward_meter_mean': 0.9690738916397095, 'reward_meter_std': 0.02259281650185585, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9690738916397095, 'reward_total_composite_std': 0.02259281650185585, 'epoch': 0.09}

 69%|██████▉   | 2272/3300 [3:08:58<1:27:50,  5.13s/it]INFO 04-12 01:45:49 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2273/3300 [3:09:05<1:40:28,  5.87s/it]2026-04-12 01:45:56,629 | INFO | train_grpo_train | metrics_logged mode=train step=2273

                                                       
{'loss': 0.0065, 'grad_norm': 2.4946935176849365, 'learning_rate': 3.1151515151515154e-06, 'num_tokens': 5144049.0, 'completions/mean_length': 308.125, 'completions/min_length': 295.0, 'completions/max_length': 318.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 308.125, 'completions/min_terminated_length': 295.0, 'completions/max_terminated_length': 318.0, 'rewards/meter/mean': 0.985474705696106, 'rewards/meter/std': 0.02723194845020771, 'rewards/count_adherence/mean': 0.6634615659713745, 'rewards/count_adherence/std': 0.039811473339796066, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7977941036224365, 'rewards/repeat_penalty/std': 0.08265344798564911, 'rewards/total_composite/mean': 0.5205361843109131, 'rewards/total_composite/std': 0.053465962409973145, 'reward': 0.5205361843109131, 'reward_std': 0.05346594378352165, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02342665195465088, 'sampling/sampling_logp_difference/max': 1.336355209350586, 'sampling/importance_sampling_ratio/min': 0.26280176639556885, 'sampling/importance_sampling_ratio/mean': 1.0016851425170898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16073014307767153, 'clip_ratio/low_mean': 0.011750375153496861, 'clip_ratio/low_min': 0.011750375153496861, 'clip_ratio/high_mean': 0.008914577425457537, 'clip_ratio/high_max': 0.008914577425457537, 'clip_ratio/region_mean': 0.0206649525789544, 'reward_total_mean': 0.5205361843109131, 'reward_meter_mean': 0.985474705696106, 'reward_meter_std': 0.02723194845020771, 'reward_count_adherence_mean': 0.6634615659713745, 'reward_count_adherence_std': 0.039811473339796066, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7977941036224365, 'reward_repeat_penalty_std': 0.08265344798564911, 'reward_total_composite_mean': 0.5205361843109131, 'reward_total_composite_std': 0.053465962409973145, 'epoch': 0.09}

 69%|██████▉   | 2273/3300 [3:09:05<1:40:28,  5.87s/it]INFO 04-12 01:45:56 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2274/3300 [3:09:11<1:39:17,  5.81s/it]2026-04-12 01:46:02,287 | INFO | train_grpo_train | metrics_logged mode=train step=2274

                                                       
{'loss': -0.006, 'grad_norm': 2.2683939933776855, 'learning_rate': 3.1121212121212126e-06, 'num_tokens': 5146825.0, 'completions/mean_length': 154.0, 'completions/min_length': 150.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.0, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.9939357042312622, 'rewards/meter/std': 0.003742816159501672, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.976262092590332, 'rewards/total_composite/std': 0.051798708736896515, 'reward': 0.976262092590332, 'reward_std': 0.05179871618747711, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03507037088274956, 'sampling/sampling_logp_difference/max': 1.2760963439941406, 'sampling/importance_sampling_ratio/min': 0.27912476658821106, 'sampling/importance_sampling_ratio/mean': 1.006531834602356, 'sampling/importance_sampling_ratio/max': 1.880401372909546, 'entropy': 0.29750954918563366, 'clip_ratio/low_mean': 0.0016666667070239782, 'clip_ratio/low_min': 0.0016666667070239782, 'clip_ratio/high_mean': 0.02496329549467191, 'clip_ratio/high_max': 0.02496329549467191, 'clip_ratio/region_mean': 0.02662996220169589, 'reward_total_mean': 0.976262092590332, 'reward_meter_mean': 0.9939357042312622, 'reward_meter_std': 0.003742816159501672, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.976262092590332, 'reward_total_composite_std': 0.051798708736896515, 'epoch': 0.09}

 69%|██████▉   | 2274/3300 [3:09:11<1:39:17,  5.81s/it]INFO 04-12 01:46:02 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2275/3300 [3:09:15<1:31:47,  5.37s/it]2026-04-12 01:46:06,646 | INFO | train_grpo_train | metrics_logged mode=train step=2275

                                                       
{'loss': 0.0025, 'grad_norm': 2.3654937744140625, 'learning_rate': 3.1090909090909095e-06, 'num_tokens': 5148537.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7876189947128296, 'rewards/meter/std': 3.669682701001875e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7876189947128296, 'rewards/total_composite/std': 3.669682701001875e-05, 'reward': 0.7876189947128296, 'reward_std': 3.669681245810352e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0021758992224931717, 'sampling/sampling_logp_difference/max': 0.25645923614501953, 'sampling/importance_sampling_ratio/min': 0.9631562829017639, 'sampling/importance_sampling_ratio/mean': 1.0020716190338135, 'sampling/importance_sampling_ratio/max': 1.2923461198806763, 'entropy': 0.012424270971678197, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.7876189947128296, 'reward_meter_mean': 0.7876189947128296, 'reward_meter_std': 3.669682701001875e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7876189947128296, 'reward_total_composite_std': 3.669682701001875e-05, 'epoch': 0.09}

 69%|██████▉   | 2275/3300 [3:09:15<1:31:47,  5.37s/it]INFO 04-12 01:46:06 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2276/3300 [3:09:21<1:32:48,  5.44s/it]2026-04-12 01:46:12,260 | INFO | train_grpo_train | metrics_logged mode=train step=2276

                                                       
{'loss': -0.0014, 'grad_norm': 1.912965178489685, 'learning_rate': 3.1060606060606063e-06, 'num_tokens': 5151308.0, 'completions/mean_length': 142.375, 'completions/min_length': 141.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.375, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9956492185592651, 'rewards/meter/std': 0.00017072130867745727, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.05143444612622261, 'rewards/total_composite/mean': 0.8573694229125977, 'rewards/total_composite/std': 0.05129777640104294, 'reward': 0.8573694229125977, 'reward_std': 0.05129777267575264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01284986175596714, 'sampling/sampling_logp_difference/max': 1.0029168128967285, 'sampling/importance_sampling_ratio/min': 0.3668079972267151, 'sampling/importance_sampling_ratio/mean': 1.000401258468628, 'sampling/importance_sampling_ratio/max': 1.4935818910598755, 'entropy': 0.07997249066829681, 'clip_ratio/low_mean': 0.0008865247946232557, 'clip_ratio/low_min': 0.0008865247946232557, 'clip_ratio/high_mean': 0.009610342094674706, 'clip_ratio/high_max': 0.009610342094674706, 'clip_ratio/region_mean': 0.010496866889297962, 'reward_total_mean': 0.8573694229125977, 'reward_meter_mean': 0.9956492185592651, 'reward_meter_std': 0.00017072130867745727, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.05143444612622261, 'reward_total_composite_mean': 0.8573694229125977, 'reward_total_composite_std': 0.05129777640104294, 'epoch': 0.09}

 69%|██████▉   | 2276/3300 [3:09:21<1:32:48,  5.44s/it]INFO 04-12 01:46:12 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2277/3300 [3:09:27<1:37:03,  5.69s/it]2026-04-12 01:46:18,521 | INFO | train_grpo_train | metrics_logged mode=train step=2277

                                                       
{'loss': -0.0004, 'grad_norm': 2.30692195892334, 'learning_rate': 3.103030303030303e-06, 'num_tokens': 5154060.0, 'completions/mean_length': 170.0, 'completions/min_length': 157.0, 'completions/max_length': 205.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.0, 'completions/min_terminated_length': 157.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.9647563695907593, 'rewards/meter/std': 0.03250608593225479, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7992424368858337, 'rewards/repeat_penalty/std': 0.11809822916984558, 'rewards/total_composite/mean': 0.658399760723114, 'rewards/total_composite/std': 0.10351461172103882, 'reward': 0.658399760723114, 'reward_std': 0.10351462662220001, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02872909978032112, 'sampling/sampling_logp_difference/max': 1.5842926502227783, 'sampling/importance_sampling_ratio/min': 0.20509281754493713, 'sampling/importance_sampling_ratio/mean': 0.9995459318161011, 'sampling/importance_sampling_ratio/max': 1.7825084924697876, 'entropy': 0.18850434944033623, 'clip_ratio/low_mean': 0.009245004854165018, 'clip_ratio/low_min': 0.009245004854165018, 'clip_ratio/high_mean': 0.01534117921255529, 'clip_ratio/high_max': 0.01534117921255529, 'clip_ratio/region_mean': 0.024586184066720307, 'reward_total_mean': 0.658399760723114, 'reward_meter_mean': 0.9647563695907593, 'reward_meter_std': 0.03250608593225479, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7992424368858337, 'reward_repeat_penalty_std': 0.11809822916984558, 'reward_total_composite_mean': 0.658399760723114, 'reward_total_composite_std': 0.10351461172103882, 'epoch': 0.09}

 69%|██████▉   | 2277/3300 [3:09:27<1:37:03,  5.69s/it]INFO 04-12 01:46:18 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2278/3300 [3:09:32<1:31:50,  5.39s/it]2026-04-12 01:46:23,237 | INFO | train_grpo_train | metrics_logged mode=train step=2278

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.1000000000000004e-06, 'num_tokens': 5155988.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.000632537470664829, 'sampling/sampling_logp_difference/max': 0.022958219051361084, 'sampling/importance_sampling_ratio/min': 0.9780369400978088, 'sampling/importance_sampling_ratio/mean': 1.0005130767822266, 'sampling/importance_sampling_ratio/max': 1.023223876953125, 'entropy': 0.006176002731081098, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 69%|██████▉   | 2278/3300 [3:09:32<1:31:50,  5.39s/it]INFO 04-12 01:46:23 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2279/3300 [3:09:36<1:26:12,  5.07s/it]2026-04-12 01:46:27,684 | INFO | train_grpo_train | metrics_logged mode=train step=2279

                                                       
{'loss': 0.0082, 'grad_norm': 13.398720741271973, 'learning_rate': 3.0969696969696972e-06, 'num_tokens': 5157395.0, 'completions/mean_length': 35.875, 'completions/min_length': 35.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9567666053771973, 'rewards/meter/std': 0.09672017395496368, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9567666053771973, 'rewards/total_composite/std': 0.09672017395496368, 'reward': 0.9567666053771973, 'reward_std': 0.09672017395496368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.055583853274583817, 'sampling/sampling_logp_difference/max': 1.244422435760498, 'sampling/importance_sampling_ratio/min': 0.2881072461605072, 'sampling/importance_sampling_ratio/mean': 0.9919366836547852, 'sampling/importance_sampling_ratio/max': 1.522891640663147, 'entropy': 0.3067398630082607, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.04533730214461684, 'clip_ratio/high_max': 0.04533730214461684, 'clip_ratio/region_mean': 0.048809524392709136, 'reward_total_mean': 0.9567666053771973, 'reward_meter_mean': 0.9567666053771973, 'reward_meter_std': 0.09672017395496368, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9567666053771973, 'reward_total_composite_std': 0.09672017395496368, 'epoch': 0.09}

 69%|██████▉   | 2279/3300 [3:09:36<1:26:12,  5.07s/it]INFO 04-12 01:46:27 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2280/3300 [3:09:42<1:30:53,  5.35s/it]2026-04-12 01:46:33,697 | INFO | train_grpo_train | metrics_logged mode=train step=2280

                                                       
{'loss': 0.0149, 'grad_norm': 5.9251813888549805, 'learning_rate': 3.093939393939394e-06, 'num_tokens': 5159888.0, 'completions/mean_length': 137.625, 'completions/min_length': 131.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.625, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9712404012680054, 'rewards/meter/std': 0.020027272403240204, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8324918150901794, 'rewards/total_composite/std': 0.017166240140795708, 'reward': 0.8324918150901794, 'reward_std': 0.017166240140795708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03175117075443268, 'sampling/sampling_logp_difference/max': 1.4596366882324219, 'sampling/importance_sampling_ratio/min': 0.2323206663131714, 'sampling/importance_sampling_ratio/mean': 1.007045865058899, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18117596302181482, 'clip_ratio/low_mean': 0.009806309011764824, 'clip_ratio/low_min': 0.009806309011764824, 'clip_ratio/high_mean': 0.008134586038067937, 'clip_ratio/high_max': 0.008134586038067937, 'clip_ratio/region_mean': 0.01794089504983276, 'reward_total_mean': 0.8324918150901794, 'reward_meter_mean': 0.9712404012680054, 'reward_meter_std': 0.020027272403240204, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8324918150901794, 'reward_total_composite_std': 0.017166240140795708, 'epoch': 0.09}

 69%|██████▉   | 2280/3300 [3:09:42<1:30:53,  5.35s/it]INFO 04-12 01:46:33 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2281/3300 [3:09:47<1:26:57,  5.12s/it]2026-04-12 01:46:38,256 | INFO | train_grpo_train | metrics_logged mode=train step=2281

                                                       
{'loss': 0.0073, 'grad_norm': 2.3620288372039795, 'learning_rate': 3.090909090909091e-06, 'num_tokens': 5161346.0, 'completions/mean_length': 32.25, 'completions/min_length': 32.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9992298483848572, 'rewards/meter/std': 7.94332881923765e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992298483848572, 'rewards/total_composite/std': 7.94332881923765e-05, 'reward': 0.9992298483848572, 'reward_std': 7.940443902043626e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016870969906449318, 'sampling/sampling_logp_difference/max': 0.7187650203704834, 'sampling/importance_sampling_ratio/min': 0.5924263596534729, 'sampling/importance_sampling_ratio/mean': 1.0008995532989502, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.057134561240673065, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.015625, 'clip_ratio/high_max': 0.015625, 'clip_ratio/region_mean': 0.023200757801532745, 'reward_total_mean': 0.9992298483848572, 'reward_meter_mean': 0.9992298483848572, 'reward_meter_std': 7.94332881923765e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992298483848572, 'reward_total_composite_std': 7.94332881923765e-05, 'epoch': 0.09}

 69%|██████▉   | 2281/3300 [3:09:47<1:26:57,  5.12s/it]INFO 04-12 01:46:38 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2282/3300 [3:09:52<1:29:18,  5.26s/it]2026-04-12 01:46:43,710 | INFO | train_grpo_train | metrics_logged mode=train step=2282

                                                       
{'loss': 0.0173, 'grad_norm': 2.3963494300842285, 'learning_rate': 3.087878787878788e-06, 'num_tokens': 5163820.0, 'completions/mean_length': 132.25, 'completions/min_length': 127.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.25, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9742553234100342, 'rewards/meter/std': 0.023107483983039856, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.1322600096464157, 'rewards/total_composite/mean': 0.8361626863479614, 'rewards/total_composite/std': 0.13729579746723175, 'reward': 0.8361626863479614, 'reward_std': 0.13729579746723175, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03228021413087845, 'sampling/sampling_logp_difference/max': 1.481900691986084, 'sampling/importance_sampling_ratio/min': 0.22720544040203094, 'sampling/importance_sampling_ratio/mean': 1.0067442655563354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23136323876678944, 'clip_ratio/low_mean': 0.005548747314605862, 'clip_ratio/low_min': 0.005548747314605862, 'clip_ratio/high_mean': 0.01910717412829399, 'clip_ratio/high_max': 0.01910717412829399, 'clip_ratio/region_mean': 0.024655921442899853, 'reward_total_mean': 0.8361626863479614, 'reward_meter_mean': 0.9742553234100342, 'reward_meter_std': 0.023107483983039856, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.1322600096464157, 'reward_total_composite_mean': 0.8361626863479614, 'reward_total_composite_std': 0.13729579746723175, 'epoch': 0.09}

 69%|██████▉   | 2282/3300 [3:09:52<1:29:18,  5.26s/it]INFO 04-12 01:46:43 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2283/3300 [3:09:58<1:32:50,  5.48s/it]2026-04-12 01:46:49,691 | INFO | train_grpo_train | metrics_logged mode=train step=2283

                                                       
{'loss': 0.0041, 'grad_norm': 2.038731098175049, 'learning_rate': 3.084848484848485e-06, 'num_tokens': 5166635.0, 'completions/mean_length': 168.875, 'completions/min_length': 167.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.875, 'completions/min_terminated_length': 167.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.9957842230796814, 'rewards/meter/std': 0.0002837553038261831, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8636363744735718, 'rewards/repeat_penalty/std': 0.08416546881198883, 'rewards/total_composite/mean': 0.8599779009819031, 'rewards/total_composite/std': 0.08360190689563751, 'reward': 0.8599779009819031, 'reward_std': 0.08360189944505692, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021027982234954834, 'sampling/sampling_logp_difference/max': 1.4786853790283203, 'sampling/importance_sampling_ratio/min': 0.22793714702129364, 'sampling/importance_sampling_ratio/mean': 1.0022042989730835, 'sampling/importance_sampling_ratio/max': 1.8878965377807617, 'entropy': 0.11252321023494005, 'clip_ratio/low_mean': 0.008858592598699033, 'clip_ratio/low_min': 0.008858592598699033, 'clip_ratio/high_mean': 0.012570029124617577, 'clip_ratio/high_max': 0.012570029124617577, 'clip_ratio/region_mean': 0.02142862172331661, 'reward_total_mean': 0.8599779009819031, 'reward_meter_mean': 0.9957842230796814, 'reward_meter_std': 0.0002837553038261831, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8636363744735718, 'reward_repeat_penalty_std': 0.08416546881198883, 'reward_total_composite_mean': 0.8599779009819031, 'reward_total_composite_std': 0.08360190689563751, 'epoch': 0.09}

 69%|██████▉   | 2283/3300 [3:09:58<1:32:50,  5.48s/it]INFO 04-12 01:46:49 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2284/3300 [3:10:02<1:25:38,  5.06s/it]2026-04-12 01:46:53,770 | INFO | train_grpo_train | metrics_logged mode=train step=2284

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.081818181818182e-06, 'num_tokens': 5168059.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9929623007774353, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9929623007774353, 'rewards/total_composite/std': 0.0, 'reward': 0.9929623007774353, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002442288678139448, 'sampling/sampling_logp_difference/max': 0.0076376767829060555, 'sampling/importance_sampling_ratio/min': 0.9986435174942017, 'sampling/importance_sampling_ratio/mean': 1.0002309083938599, 'sampling/importance_sampling_ratio/max': 1.0076669454574585, 'entropy': 0.0025252392806578428, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9929623007774353, 'reward_meter_mean': 0.9929623007774353, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9929623007774353, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 69%|██████▉   | 2284/3300 [3:10:02<1:25:38,  5.06s/it]INFO 04-12 01:46:54 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2285/3300 [3:10:08<1:30:01,  5.32s/it]2026-04-12 01:46:59,711 | INFO | train_grpo_train | metrics_logged mode=train step=2285

                                                       
{'loss': -0.0656, 'grad_norm': 1.7548184394836426, 'learning_rate': 3.078787878787879e-06, 'num_tokens': 5170897.0, 'completions/mean_length': 171.75, 'completions/min_length': 143.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.75, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.9989954233169556, 'rewards/meter/std': 0.0003574988222680986, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.8935445547103882, 'rewards/total_composite/std': 0.07573916018009186, 'reward': 0.8935445547103882, 'reward_std': 0.07573915272951126, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022320233285427094, 'sampling/sampling_logp_difference/max': 1.9766596555709839, 'sampling/importance_sampling_ratio/min': 0.13853120803833008, 'sampling/importance_sampling_ratio/mean': 1.0065274238586426, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19338966719806194, 'clip_ratio/low_mean': 0.014906731550581753, 'clip_ratio/low_min': 0.014906731550581753, 'clip_ratio/high_mean': 0.006868339143693447, 'clip_ratio/high_max': 0.006868339143693447, 'clip_ratio/region_mean': 0.0217750706942752, 'reward_total_mean': 0.8935445547103882, 'reward_meter_mean': 0.9989954233169556, 'reward_meter_std': 0.0003574988222680986, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.8935445547103882, 'reward_total_composite_std': 0.07573916018009186, 'epoch': 0.09}

 69%|██████▉   | 2285/3300 [3:10:08<1:30:01,  5.32s/it]INFO 04-12 01:46:59 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2286/3300 [3:10:13<1:28:02,  5.21s/it]2026-04-12 01:47:04,653 | INFO | train_grpo_train | metrics_logged mode=train step=2286

                                                       
{'loss': -0.0008, 'grad_norm': 1.9542547464370728, 'learning_rate': 3.075757575757576e-06, 'num_tokens': 5173297.0, 'completions/mean_length': 106.0, 'completions/min_length': 106.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.0, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.680180549621582, 'rewards/meter/std': 0.013479826971888542, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5830118656158447, 'rewards/total_composite/std': 0.011554129421710968, 'reward': 0.5830118656158447, 'reward_std': 0.011554128490388393, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0026654843240976334, 'sampling/sampling_logp_difference/max': 0.7275528907775879, 'sampling/importance_sampling_ratio/min': 0.48308974504470825, 'sampling/importance_sampling_ratio/mean': 1.0005828142166138, 'sampling/importance_sampling_ratio/max': 1.85325026512146, 'entropy': 0.009184476220980287, 'clip_ratio/low_mean': 0.002358490601181984, 'clip_ratio/low_min': 0.002358490601181984, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002358490601181984, 'reward_total_mean': 0.5830118656158447, 'reward_meter_mean': 0.680180549621582, 'reward_meter_std': 0.013479826971888542, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5830118656158447, 'reward_total_composite_std': 0.011554129421710968, 'epoch': 0.09}

 69%|██████▉   | 2286/3300 [3:10:13<1:28:02,  5.21s/it]INFO 04-12 01:47:04 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2287/3300 [3:10:18<1:25:37,  5.07s/it]2026-04-12 01:47:09,404 | INFO | train_grpo_train | metrics_logged mode=train step=2287

                                                       
{'loss': -0.0098, 'grad_norm': 5.288448810577393, 'learning_rate': 3.0727272727272727e-06, 'num_tokens': 5175104.0, 'completions/mean_length': 73.875, 'completions/min_length': 70.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8793976306915283, 'rewards/meter/std': 0.33859992027282715, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8793976306915283, 'rewards/total_composite/std': 0.33859992027282715, 'reward': 0.8793976306915283, 'reward_std': 0.33859992027282715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027358917519450188, 'sampling/sampling_logp_difference/max': 0.9376602172851562, 'sampling/importance_sampling_ratio/min': 0.39154288172721863, 'sampling/importance_sampling_ratio/mean': 1.005669355392456, 'sampling/importance_sampling_ratio/max': 1.7006129026412964, 'entropy': 0.24043426476418972, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.016635622712783515, 'clip_ratio/high_max': 0.016635622712783515, 'clip_ratio/region_mean': 0.018421337008476257, 'reward_total_mean': 0.8793976306915283, 'reward_meter_mean': 0.8793976306915283, 'reward_meter_std': 0.33859992027282715, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8793976306915283, 'reward_total_composite_std': 0.33859992027282715, 'epoch': 0.09}

 69%|██████▉   | 2287/3300 [3:10:18<1:25:37,  5.07s/it]INFO 04-12 01:47:09 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2288/3300 [3:10:24<1:29:02,  5.28s/it]2026-04-12 01:47:15,166 | INFO | train_grpo_train | metrics_logged mode=train step=2288

                                                       
{'loss': 0.013, 'grad_norm': 7.310388088226318, 'learning_rate': 3.0696969696969696e-06, 'num_tokens': 5178040.0, 'completions/mean_length': 168.0, 'completions/min_length': 166.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.0, 'completions/min_terminated_length': 166.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.7916620969772339, 'rewards/meter/std': 0.292292982339859, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.930555522441864, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.621074378490448, 'rewards/total_composite/std': 0.2475537359714508, 'reward': 0.621074378490448, 'reward_std': 0.2475537210702896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03382183238863945, 'sampling/sampling_logp_difference/max': 1.4603016376495361, 'sampling/importance_sampling_ratio/min': 0.23216624557971954, 'sampling/importance_sampling_ratio/mean': 1.000993251800537, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17776653543114662, 'clip_ratio/low_mean': 0.008171421475708485, 'clip_ratio/low_min': 0.008171421475708485, 'clip_ratio/high_mean': 0.017823605565354228, 'clip_ratio/high_max': 0.017823605565354228, 'clip_ratio/region_mean': 0.025995027041062713, 'reward_total_mean': 0.621074378490448, 'reward_meter_mean': 0.7916620969772339, 'reward_meter_std': 0.292292982339859, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.930555522441864, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.621074378490448, 'reward_total_composite_std': 0.2475537359714508, 'epoch': 0.09}

 69%|██████▉   | 2288/3300 [3:10:24<1:29:02,  5.28s/it]INFO 04-12 01:47:15 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2289/3300 [3:10:28<1:24:58,  5.04s/it]2026-04-12 01:47:19,658 | INFO | train_grpo_train | metrics_logged mode=train step=2289

                                                       
{'loss': 0.012, 'grad_norm': 1.6638261079788208, 'learning_rate': 3.066666666666667e-06, 'num_tokens': 5179986.0, 'completions/mean_length': 68.25, 'completions/min_length': 66.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9976264834403992, 'rewards/meter/std': 0.00037816373514942825, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976264834403992, 'rewards/total_composite/std': 0.00037816373514942825, 'reward': 0.9976264834403992, 'reward_std': 0.00037817005068063736, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01430835947394371, 'sampling/sampling_logp_difference/max': 2.2254679203033447, 'sampling/importance_sampling_ratio/min': 0.10801687091588974, 'sampling/importance_sampling_ratio/mean': 0.996768593788147, 'sampling/importance_sampling_ratio/max': 1.2755680084228516, 'entropy': 0.04147487785667181, 'clip_ratio/low_mean': 0.005434782709926367, 'clip_ratio/low_min': 0.005434782709926367, 'clip_ratio/high_mean': 0.005681818351149559, 'clip_ratio/high_max': 0.005681818351149559, 'clip_ratio/region_mean': 0.011116601061075926, 'reward_total_mean': 0.9976264834403992, 'reward_meter_mean': 0.9976264834403992, 'reward_meter_std': 0.00037816373514942825, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976264834403992, 'reward_total_composite_std': 0.00037816373514942825, 'epoch': 0.09}

 69%|██████▉   | 2289/3300 [3:10:28<1:24:58,  5.04s/it]INFO 04-12 01:47:19 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2290/3300 [3:10:33<1:22:09,  4.88s/it]2026-04-12 01:47:24,156 | INFO | train_grpo_train | metrics_logged mode=train step=2290

                                                       
{'loss': 0.0052, 'grad_norm': 1.542847752571106, 'learning_rate': 3.0636363636363636e-06, 'num_tokens': 5181793.0, 'completions/mean_length': 64.875, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9991232752799988, 'rewards/meter/std': 0.00012234247697051615, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991232752799988, 'rewards/total_composite/std': 0.00012234247697051615, 'reward': 0.9991232752799988, 'reward_std': 0.00012233950837980956, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01243247464299202, 'sampling/sampling_logp_difference/max': 1.3001289367675781, 'sampling/importance_sampling_ratio/min': 0.27249664068222046, 'sampling/importance_sampling_ratio/mean': 0.9971635937690735, 'sampling/importance_sampling_ratio/max': 1.3870518207550049, 'entropy': 0.04461120581254363, 'clip_ratio/low_mean': 0.001923076924867928, 'clip_ratio/low_min': 0.001923076924867928, 'clip_ratio/high_mean': 0.005859375, 'clip_ratio/high_max': 0.005859375, 'clip_ratio/region_mean': 0.007782451924867928, 'reward_total_mean': 0.9991232752799988, 'reward_meter_mean': 0.9991232752799988, 'reward_meter_std': 0.00012234247697051615, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991232752799988, 'reward_total_composite_std': 0.00012234247697051615, 'epoch': 0.09}

 69%|██████▉   | 2290/3300 [3:10:33<1:22:09,  4.88s/it]INFO 04-12 01:47:24 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2291/3300 [3:10:37<1:19:43,  4.74s/it]2026-04-12 01:47:28,573 | INFO | train_grpo_train | metrics_logged mode=train step=2291

                                                       
{'loss': -0.0007, 'grad_norm': 0.26676464080810547, 'learning_rate': 3.0606060606060605e-06, 'num_tokens': 5183553.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.998096764087677, 'rewards/meter/std': 3.926327553926967e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998096764087677, 'rewards/total_composite/std': 3.926327553926967e-05, 'reward': 0.998096764087677, 'reward_std': 3.925973942386918e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005647474434226751, 'sampling/sampling_logp_difference/max': 0.8950796127319336, 'sampling/importance_sampling_ratio/min': 0.4085750877857208, 'sampling/importance_sampling_ratio/mean': 0.999199628829956, 'sampling/importance_sampling_ratio/max': 1.0820250511169434, 'entropy': 0.036194659769535065, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.998096764087677, 'reward_meter_mean': 0.998096764087677, 'reward_meter_std': 3.926327553926967e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998096764087677, 'reward_total_composite_std': 3.926327553926967e-05, 'epoch': 0.09}

 69%|██████▉   | 2291/3300 [3:10:37<1:19:43,  4.74s/it]INFO 04-12 01:47:28 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2292/3300 [3:10:43<1:25:48,  5.11s/it]2026-04-12 01:47:34,539 | INFO | train_grpo_train | metrics_logged mode=train step=2292

                                                       
{'loss': -0.0461, 'grad_norm': 3.2664194107055664, 'learning_rate': 3.057575757575758e-06, 'num_tokens': 5186427.0, 'completions/mean_length': 159.25, 'completions/min_length': 148.0, 'completions/max_length': 201.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.25, 'completions/min_terminated_length': 148.0, 'completions/max_terminated_length': 201.0, 'rewards/meter/mean': 0.9959465861320496, 'rewards/meter/std': 0.003204776206985116, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9682539701461792, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.793583869934082, 'rewards/total_composite/std': 0.05434796214103699, 'reward': 0.793583869934082, 'reward_std': 0.05434795841574669, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03989114239811897, 'sampling/sampling_logp_difference/max': 2.4048454761505127, 'sampling/importance_sampling_ratio/min': 0.09027944505214691, 'sampling/importance_sampling_ratio/mean': 1.0047212839126587, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30428963899612427, 'clip_ratio/low_mean': 0.005605590064078569, 'clip_ratio/low_min': 0.005605590064078569, 'clip_ratio/high_mean': 0.024386633071117103, 'clip_ratio/high_max': 0.024386633071117103, 'clip_ratio/region_mean': 0.029992223135195673, 'reward_total_mean': 0.793583869934082, 'reward_meter_mean': 0.9959465861320496, 'reward_meter_std': 0.003204776206985116, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9682539701461792, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.793583869934082, 'reward_total_composite_std': 0.05434796214103699, 'epoch': 0.09}

 69%|██████▉   | 2292/3300 [3:10:43<1:25:48,  5.11s/it]INFO 04-12 01:47:34 [block_pool.py:378] Successfully reset prefix cache

 69%|██████▉   | 2293/3300 [3:10:51<1:42:58,  6.14s/it]2026-04-12 01:47:43,075 | INFO | train_grpo_train | metrics_logged mode=train step=2293

                                                       
{'loss': 0.0327, 'grad_norm': 1.504547119140625, 'learning_rate': 3.054545454545455e-06, 'num_tokens': 5191210.0, 'completions/mean_length': 363.875, 'completions/min_length': 343.0, 'completions/max_length': 392.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 363.875, 'completions/min_terminated_length': 343.0, 'completions/max_terminated_length': 392.0, 'rewards/meter/mean': 0.9957020282745361, 'rewards/meter/std': 0.001336081069894135, 'rewards/count_adherence/mean': 0.6083333492279053, 'rewards/count_adherence/std': 0.0235702246427536, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8192294836044312, 'rewards/repeat_penalty/std': 0.07807844877243042, 'rewards/total_composite/mean': 0.4955565929412842, 'rewards/total_composite/std': 0.04254494607448578, 'reward': 0.4955565929412842, 'reward_std': 0.04254494979977608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032130178064107895, 'sampling/sampling_logp_difference/max': 2.1043951511383057, 'sampling/importance_sampling_ratio/min': 0.12191939353942871, 'sampling/importance_sampling_ratio/mean': 1.0060222148895264, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26295676827430725, 'clip_ratio/low_mean': 0.010282430681400001, 'clip_ratio/low_min': 0.010282430681400001, 'clip_ratio/high_mean': 0.017152261221781373, 'clip_ratio/high_max': 0.017152261221781373, 'clip_ratio/region_mean': 0.027434691903181374, 'reward_total_mean': 0.4955565929412842, 'reward_meter_mean': 0.9957020282745361, 'reward_meter_std': 0.001336081069894135, 'reward_count_adherence_mean': 0.6083333492279053, 'reward_count_adherence_std': 0.0235702246427536, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8192294836044312, 'reward_repeat_penalty_std': 0.07807844877243042, 'reward_total_composite_mean': 0.4955565929412842, 'reward_total_composite_std': 0.04254494607448578, 'epoch': 0.09}

 69%|██████▉   | 2293/3300 [3:10:52<1:42:58,  6.14s/it]INFO 04-12 01:47:43 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2294/3300 [3:11:00<1:52:18,  6.70s/it]2026-04-12 01:47:51,080 | INFO | train_grpo_train | metrics_logged mode=train step=2294

                                                       
{'loss': -0.0027, 'grad_norm': 1.8439388275146484, 'learning_rate': 3.051515151515152e-06, 'num_tokens': 5195522.0, 'completions/mean_length': 309.0, 'completions/min_length': 287.0, 'completions/max_length': 326.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 309.0, 'completions/min_terminated_length': 287.0, 'completions/max_terminated_length': 326.0, 'rewards/meter/mean': 0.99519944190979, 'rewards/meter/std': 0.003783464664593339, 'rewards/count_adherence/mean': 0.7875000238418579, 'rewards/count_adherence/std': 0.0353553481400013, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8807692527770996, 'rewards/repeat_penalty/std': 0.07008695602416992, 'rewards/total_composite/mean': 0.690649151802063, 'rewards/total_composite/std': 0.06761835515499115, 'reward': 0.690649151802063, 'reward_std': 0.06761834770441055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03343411162495613, 'sampling/sampling_logp_difference/max': 1.6177313327789307, 'sampling/importance_sampling_ratio/min': 0.19834816455841064, 'sampling/importance_sampling_ratio/mean': 1.0094666481018066, 'sampling/importance_sampling_ratio/max': 1.937671184539795, 'entropy': 0.25325608998537064, 'clip_ratio/low_mean': 0.00620853912550956, 'clip_ratio/low_min': 0.00620853912550956, 'clip_ratio/high_mean': 0.014976657228544354, 'clip_ratio/high_max': 0.014976657228544354, 'clip_ratio/region_mean': 0.021185196354053915, 'reward_total_mean': 0.690649151802063, 'reward_meter_mean': 0.99519944190979, 'reward_meter_std': 0.003783464664593339, 'reward_count_adherence_mean': 0.7875000238418579, 'reward_count_adherence_std': 0.0353553481400013, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8807692527770996, 'reward_repeat_penalty_std': 0.07008695602416992, 'reward_total_composite_mean': 0.690649151802063, 'reward_total_composite_std': 0.06761835515499115, 'epoch': 0.09}

 70%|██████▉   | 2294/3300 [3:11:00<1:52:18,  6.70s/it]INFO 04-12 01:47:51 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2295/3300 [3:11:04<1:43:11,  6.16s/it]2026-04-12 01:47:55,989 | INFO | train_grpo_train | metrics_logged mode=train step=2295

                                                       
{'loss': 0.0008, 'grad_norm': 3.1970479488372803, 'learning_rate': 3.048484848484849e-06, 'num_tokens': 5197722.0, 'completions/mean_length': 106.0, 'completions/min_length': 106.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.0, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.6495559215545654, 'rewards/meter/std': 0.1058802381157875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5567622184753418, 'rewards/total_composite/std': 0.09075448662042618, 'reward': 0.5567622184753418, 'reward_std': 0.09075447171926498, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003019496565684676, 'sampling/sampling_logp_difference/max': 0.5486016273498535, 'sampling/importance_sampling_ratio/min': 0.5777571797370911, 'sampling/importance_sampling_ratio/mean': 0.9996184706687927, 'sampling/importance_sampling_ratio/max': 1.2651506662368774, 'entropy': 0.02729532145895064, 'clip_ratio/low_mean': 0.001179245300590992, 'clip_ratio/low_min': 0.001179245300590992, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.001179245300590992, 'reward_total_mean': 0.5567622184753418, 'reward_meter_mean': 0.6495559215545654, 'reward_meter_std': 0.1058802381157875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5567622184753418, 'reward_total_composite_std': 0.09075448662042618, 'epoch': 0.09}

 70%|██████▉   | 2295/3300 [3:11:04<1:43:11,  6.16s/it]INFO 04-12 01:47:56 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2296/3300 [3:11:09<1:34:20,  5.64s/it]2026-04-12 01:48:00,411 | INFO | train_grpo_train | metrics_logged mode=train step=2296

                                                       
{'loss': 0.005, 'grad_norm': 2.241980791091919, 'learning_rate': 3.045454545454546e-06, 'num_tokens': 5199420.0, 'completions/mean_length': 66.25, 'completions/min_length': 66.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.998041033744812, 'rewards/meter/std': 8.582558075431734e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998041033744812, 'rewards/total_composite/std': 8.582558075431734e-05, 'reward': 0.998041033744812, 'reward_std': 8.582806185586378e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014799090102314949, 'sampling/sampling_logp_difference/max': 1.3824272155761719, 'sampling/importance_sampling_ratio/min': 0.25096866488456726, 'sampling/importance_sampling_ratio/mean': 0.99764084815979, 'sampling/importance_sampling_ratio/max': 1.2699003219604492, 'entropy': 0.06576015800237656, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.0037878789007663727, 'clip_ratio/high_max': 0.0037878789007663727, 'clip_ratio/region_mean': 0.00562611420173198, 'reward_total_mean': 0.998041033744812, 'reward_meter_mean': 0.998041033744812, 'reward_meter_std': 8.582558075431734e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998041033744812, 'reward_total_composite_std': 8.582558075431734e-05, 'epoch': 0.09}

 70%|██████▉   | 2296/3300 [3:11:09<1:34:20,  5.64s/it]INFO 04-12 01:48:00 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2297/3300 [3:11:13<1:26:51,  5.20s/it]2026-04-12 01:48:04,576 | INFO | train_grpo_train | metrics_logged mode=train step=2297

                                                       
{'loss': 0.0003, 'grad_norm': 0.5168666839599609, 'learning_rate': 3.0424242424242427e-06, 'num_tokens': 5201164.0, 'completions/mean_length': 55.0, 'completions/min_length': 55.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9948955774307251, 'rewards/meter/std': 1.7123466022894718e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948955774307251, 'rewards/total_composite/std': 1.7123466022894718e-05, 'reward': 0.9948955774307251, 'reward_std': 1.7113467038143426e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0070167421363294125, 'sampling/sampling_logp_difference/max': 0.6335644721984863, 'sampling/importance_sampling_ratio/min': 0.5306968092918396, 'sampling/importance_sampling_ratio/mean': 0.9979526400566101, 'sampling/importance_sampling_ratio/max': 1.3264281749725342, 'entropy': 0.037125169299542904, 'clip_ratio/low_mean': 0.0022727272007614374, 'clip_ratio/low_min': 0.0022727272007614374, 'clip_ratio/high_mean': 0.0022727272007614374, 'clip_ratio/high_max': 0.0022727272007614374, 'clip_ratio/region_mean': 0.004545454401522875, 'reward_total_mean': 0.9948955774307251, 'reward_meter_mean': 0.9948955774307251, 'reward_meter_std': 1.7123466022894718e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948955774307251, 'reward_total_composite_std': 1.7123466022894718e-05, 'epoch': 0.09}

 70%|██████▉   | 2297/3300 [3:11:13<1:26:51,  5.20s/it]INFO 04-12 01:48:04 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2298/3300 [3:11:17<1:23:09,  4.98s/it]2026-04-12 01:48:09,047 | INFO | train_grpo_train | metrics_logged mode=train step=2298

                                                       
{'loss': -0.0044, 'grad_norm': 2.7533559799194336, 'learning_rate': 3.03939393939394e-06, 'num_tokens': 5203008.0, 'completions/mean_length': 76.5, 'completions/min_length': 75.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9956209659576416, 'rewards/meter/std': 0.0032750838436186314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9956209659576416, 'rewards/total_composite/std': 0.0032750838436186314, 'reward': 0.9956209659576416, 'reward_std': 0.00327507546171546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0241053719073534, 'sampling/sampling_logp_difference/max': 1.008596420288086, 'sampling/importance_sampling_ratio/min': 0.3647305369377136, 'sampling/importance_sampling_ratio/mean': 1.0014814138412476, 'sampling/importance_sampling_ratio/max': 1.4816969633102417, 'entropy': 0.21845361031591892, 'clip_ratio/low_mean': 0.004934210330247879, 'clip_ratio/low_min': 0.004934210330247879, 'clip_ratio/high_mean': 0.014611484948545694, 'clip_ratio/high_max': 0.014611484948545694, 'clip_ratio/region_mean': 0.019545695278793573, 'reward_total_mean': 0.9956209659576416, 'reward_meter_mean': 0.9956209659576416, 'reward_meter_std': 0.0032750838436186314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9956209659576416, 'reward_total_composite_std': 0.0032750838436186314, 'epoch': 0.09}

 70%|██████▉   | 2298/3300 [3:11:18<1:23:09,  4.98s/it]INFO 04-12 01:48:09 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2299/3300 [3:11:23<1:24:33,  5.07s/it]2026-04-12 01:48:14,320 | INFO | train_grpo_train | metrics_logged mode=train step=2299

                                                       
{'loss': 0.013, 'grad_norm': 4.413851737976074, 'learning_rate': 3.036363636363637e-06, 'num_tokens': 5205419.0, 'completions/mean_length': 136.375, 'completions/min_length': 131.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.375, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.8257510662078857, 'rewards/meter/std': 0.3110801577568054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.1079898476600647, 'rewards/total_composite/mean': 0.754673421382904, 'rewards/total_composite/std': 0.2842678129673004, 'reward': 0.754673421382904, 'reward_std': 0.284267783164978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02921801060438156, 'sampling/sampling_logp_difference/max': 1.611825942993164, 'sampling/importance_sampling_ratio/min': 0.1995229572057724, 'sampling/importance_sampling_ratio/mean': 1.0080208778381348, 'sampling/importance_sampling_ratio/max': 1.8313679695129395, 'entropy': 0.24156404845416546, 'clip_ratio/low_mean': 0.01099004433490336, 'clip_ratio/low_min': 0.01099004433490336, 'clip_ratio/high_mean': 0.02088787977118045, 'clip_ratio/high_max': 0.02088787977118045, 'clip_ratio/region_mean': 0.03187792410608381, 'reward_total_mean': 0.754673421382904, 'reward_meter_mean': 0.8257510662078857, 'reward_meter_std': 0.3110801577568054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.1079898476600647, 'reward_total_composite_mean': 0.754673421382904, 'reward_total_composite_std': 0.2842678129673004, 'epoch': 0.09}

 70%|██████▉   | 2299/3300 [3:11:23<1:24:33,  5.07s/it]INFO 04-12 01:48:14 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2300/3300 [3:11:27<1:21:20,  4.88s/it]2026-04-12 01:48:18,765 | INFO | train_grpo_train | metrics_logged mode=train step=2300

                                                       
{'loss': -0.0163, 'grad_norm': 2.978400230407715, 'learning_rate': 3.0333333333333337e-06, 'num_tokens': 5207160.0, 'completions/mean_length': 65.625, 'completions/min_length': 62.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.996369481086731, 'rewards/meter/std': 0.00487491674721241, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.996369481086731, 'rewards/total_composite/std': 0.00487491674721241, 'reward': 0.996369481086731, 'reward_std': 0.004874910227954388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007210928946733475, 'sampling/sampling_logp_difference/max': 0.5058443546295166, 'sampling/importance_sampling_ratio/min': 0.7272458076477051, 'sampling/importance_sampling_ratio/mean': 1.004469871520996, 'sampling/importance_sampling_ratio/max': 1.6583852767944336, 'entropy': 0.04499326320365071, 'clip_ratio/low_mean': 0.002016128972172737, 'clip_ratio/low_min': 0.002016128972172737, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.0039100684225559235, 'reward_total_mean': 0.996369481086731, 'reward_meter_mean': 0.996369481086731, 'reward_meter_std': 0.00487491674721241, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.996369481086731, 'reward_total_composite_std': 0.00487491674721241, 'epoch': 0.09}

 70%|██████▉   | 2300/3300 [3:11:27<1:21:20,  4.88s/it]INFO 04-12 01:48:18 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:04<00:26,  2.42s/it]

 23%|██▎       | 3/13 [00:07<00:26,  2.64s/it]

 31%|███       | 4/13 [00:12<00:30,  3.40s/it]

 38%|███▊      | 5/13 [00:17<00:31,  3.91s/it]

 46%|████▌     | 6/13 [00:22<00:30,  4.37s/it]

 54%|█████▍    | 7/13 [00:27<00:26,  4.40s/it]

 62%|██████▏   | 8/13 [00:32<00:23,  4.79s/it]

 69%|██████▉   | 9/13 [00:38<00:20,  5.16s/it]

 77%|███████▋  | 10/13 [00:42<00:14,  4.82s/it]

 85%|████████▍ | 11/13 [00:47<00:09,  4.75s/it]

 92%|█████████▏| 12/13 [00:52<00:04,  4.76s/it]

100%|██████████| 13/13 [00:56<00:00,  4.66s/it]2026-04-12 01:49:20,733 | INFO | train_grpo_train | metrics_logged mode=eval step=2300

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 61.8872, 'eval_samples_per_second': 1.68, 'eval_steps_per_second': 0.21, 'eval_num_tokens': 5207160.0, 'eval_completions/mean_length': 179.41346153846155, 'eval_completions/min_length': 60.46153846153846, 'eval_completions/max_length': 324.3076923076923, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 179.41346153846155, 'eval_completions/min_terminated_length': 60.46153846153846, 'eval_completions/max_terminated_length': 324.3076923076923, 'eval_rewards/meter/mean': 0.7395852987582867, 'eval_rewards/meter/std': 0.3918028657252972, 'eval_rewards/count_adherence/mean': 0.8744415503281814, 'eval_rewards/count_adherence/std': 0.1353673808849775, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.9028620398961581, 'eval_rewards/repeat_penalty/std': 0.10280066442031127, 'eval_rewards/total_composite/mean': 0.5927730821646177, 'eval_rewards/total_composite/std': 0.35657222683613116, 'eval_reward': 0.5927730821646177, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.02075868207388199, 'eval_sampling/sampling_logp_difference/max': 0.9916172761183518, 'eval_sampling/importance_sampling_ratio/min': 0.38055819043746364, 'eval_sampling/importance_sampling_ratio/mean': 1.004779577255249, 'eval_sampling/importance_sampling_ratio/max': 1.405327586027292, 'eval_entropy': 0.21972922522288102, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.5927730821646177, 'eval_reward_meter_mean': 0.7395852987582867, 'eval_reward_meter_std': 0.3918028657252972, 'eval_reward_count_adherence_mean': 0.8744415503281814, 'eval_reward_count_adherence_std': 0.1353673808849775, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.9028620398961581, 'eval_reward_repeat_penalty_std': 0.10280066442031127, 'eval_reward_total_composite_mean': 0.5927730821646177, 'eval_reward_total_composite_std': 0.35657222683613116, 'epoch': 0.09}

 70%|██████▉   | 2300/3300 [3:12:29<1:21:20,  4.88s/it]

100%|██████████| 13/13 [00:56<00:00,  4.66s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 70%|██████▉   | 2301/3300 [3:12:36<6:42:22, 24.17s/it]2026-04-12 01:49:27,932 | INFO | train_grpo_train | metrics_logged mode=train step=2301

                                                       
{'loss': -0.0022, 'grad_norm': 1.7726956605911255, 'learning_rate': 3.0303030303030305e-06, 'num_tokens': 5209084.0, 'completions/mean_length': 73.5, 'completions/min_length': 73.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9988920092582703, 'rewards/meter/std': 0.0004005288355983794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988920092582703, 'rewards/total_composite/std': 0.0004005288355983794, 'reward': 0.9988920092582703, 'reward_std': 0.0004005288355983794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023180123418569565, 'sampling/sampling_logp_difference/max': 0.5834887027740479, 'sampling/importance_sampling_ratio/min': 0.5579484701156616, 'sampling/importance_sampling_ratio/mean': 1.0048654079437256, 'sampling/importance_sampling_ratio/max': 1.46310555934906, 'entropy': 0.21760859712958336, 'clip_ratio/low_mean': 0.008561643771827221, 'clip_ratio/low_min': 0.008561643771827221, 'clip_ratio/high_mean': 0.01874305820092559, 'clip_ratio/high_max': 0.01874305820092559, 'clip_ratio/region_mean': 0.02730470197275281, 'reward_total_mean': 0.9988920092582703, 'reward_meter_mean': 0.9988920092582703, 'reward_meter_std': 0.0004005288355983794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988920092582703, 'reward_total_composite_std': 0.0004005288355983794, 'epoch': 0.09}

 70%|██████▉   | 2301/3300 [3:12:36<6:42:22, 24.17s/it]INFO 04-12 01:49:28 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2302/3300 [3:12:41<5:04:55, 18.33s/it]2026-04-12 01:49:32,653 | INFO | train_grpo_train | metrics_logged mode=train step=2302

                                                       
{'loss': 0.0094, 'grad_norm': 4.9054951667785645, 'learning_rate': 3.0272727272727277e-06, 'num_tokens': 5210937.0, 'completions/mean_length': 68.625, 'completions/min_length': 67.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.8321632742881775, 'rewards/meter/std': 0.27767401933670044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8321632742881775, 'rewards/total_composite/std': 0.27767401933670044, 'reward': 0.8321632742881775, 'reward_std': 0.27767398953437805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03458629176020622, 'sampling/sampling_logp_difference/max': 1.4874401092529297, 'sampling/importance_sampling_ratio/min': 0.22595034539699554, 'sampling/importance_sampling_ratio/mean': 1.0079941749572754, 'sampling/importance_sampling_ratio/max': 1.5462456941604614, 'entropy': 0.26816077157855034, 'clip_ratio/low_mean': 0.0018115942366421223, 'clip_ratio/low_min': 0.0018115942366421223, 'clip_ratio/high_mean': 0.025409106630831957, 'clip_ratio/high_max': 0.025409106630831957, 'clip_ratio/region_mean': 0.02722070086747408, 'reward_total_mean': 0.8321632742881775, 'reward_meter_mean': 0.8321632742881775, 'reward_meter_std': 0.27767401933670044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8321632742881775, 'reward_total_composite_std': 0.27767401933670044, 'epoch': 0.09}

 70%|██████▉   | 2302/3300 [3:12:41<5:04:55, 18.33s/it]INFO 04-12 01:49:32 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2303/3300 [3:12:50<4:17:13, 15.48s/it]2026-04-12 01:49:41,481 | INFO | train_grpo_train | metrics_logged mode=train step=2303

                                                       
{'loss': -0.0043, 'grad_norm': 1.882811188697815, 'learning_rate': 3.0242424242424246e-06, 'num_tokens': 5215896.0, 'completions/mean_length': 387.875, 'completions/min_length': 366.0, 'completions/max_length': 422.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 387.875, 'completions/min_terminated_length': 366.0, 'completions/max_terminated_length': 422.0, 'rewards/meter/mean': 0.9971157908439636, 'rewards/meter/std': 0.001221620594151318, 'rewards/count_adherence/mean': 0.5882353186607361, 'rewards/count_adherence/std': 0.03144249692559242, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9170373678207397, 'rewards/repeat_penalty/std': 0.06482076644897461, 'rewards/total_composite/mean': 0.5379288196563721, 'rewards/total_composite/std': 0.04898425191640854, 'reward': 0.5379288196563721, 'reward_std': 0.04898424819111824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035993542522192, 'sampling/sampling_logp_difference/max': 2.7829220294952393, 'sampling/importance_sampling_ratio/min': 0.25277209281921387, 'sampling/importance_sampling_ratio/mean': 1.0075607299804688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2818215675652027, 'clip_ratio/low_mean': 0.010549622820690274, 'clip_ratio/low_min': 0.010549622820690274, 'clip_ratio/high_mean': 0.011514384881593287, 'clip_ratio/high_max': 0.011514384881593287, 'clip_ratio/region_mean': 0.02206400770228356, 'reward_total_mean': 0.5379288196563721, 'reward_meter_mean': 0.9971157908439636, 'reward_meter_std': 0.001221620594151318, 'reward_count_adherence_mean': 0.5882353186607361, 'reward_count_adherence_std': 0.03144249692559242, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9170373678207397, 'reward_repeat_penalty_std': 0.06482076644897461, 'reward_total_composite_mean': 0.5379288196563721, 'reward_total_composite_std': 0.04898425191640854, 'epoch': 0.09}

 70%|██████▉   | 2303/3300 [3:12:50<4:17:13, 15.48s/it]INFO 04-12 01:49:41 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2304/3300 [3:12:54<3:21:24, 12.13s/it]2026-04-12 01:49:45,805 | INFO | train_grpo_train | metrics_logged mode=train step=2304

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.0212121212121214e-06, 'num_tokens': 5217688.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00046610343270003796, 'sampling/sampling_logp_difference/max': 0.01851455122232437, 'sampling/importance_sampling_ratio/min': 0.9936742186546326, 'sampling/importance_sampling_ratio/mean': 1.000432014465332, 'sampling/importance_sampling_ratio/max': 1.0186870098114014, 'entropy': 0.004575719474814832, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 70%|██████▉   | 2304/3300 [3:12:54<3:21:24, 12.13s/it]INFO 04-12 01:49:46 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2305/3300 [3:13:00<2:51:19, 10.33s/it]2026-04-12 01:49:51,930 | INFO | train_grpo_train | metrics_logged mode=train step=2305

                                                       
{'loss': -0.0071, 'grad_norm': 2.1037492752075195, 'learning_rate': 3.0181818181818182e-06, 'num_tokens': 5220865.0, 'completions/mean_length': 188.125, 'completions/min_length': 184.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 188.125, 'completions/min_terminated_length': 184.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.9963282942771912, 'rewards/meter/std': 0.0020138672553002834, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9305555820465088, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.7725695371627808, 'rewards/total_composite/std': 0.06809280812740326, 'reward': 0.7725695371627808, 'reward_std': 0.06809281557798386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033207789063453674, 'sampling/sampling_logp_difference/max': 1.2795171737670898, 'sampling/importance_sampling_ratio/min': 0.278171569108963, 'sampling/importance_sampling_ratio/mean': 1.0070801973342896, 'sampling/importance_sampling_ratio/max': 1.9226046800613403, 'entropy': 0.2865954786539078, 'clip_ratio/low_mean': 0.014036171836778522, 'clip_ratio/low_min': 0.014036171836778522, 'clip_ratio/high_mean': 0.014565107179805636, 'clip_ratio/high_max': 0.014565107179805636, 'clip_ratio/region_mean': 0.028601279016584158, 'reward_total_mean': 0.7725695371627808, 'reward_meter_mean': 0.9963282942771912, 'reward_meter_std': 0.0020138672553002834, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9305555820465088, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.7725695371627808, 'reward_total_composite_std': 0.06809280812740326, 'epoch': 0.09}

 70%|██████▉   | 2305/3300 [3:13:00<2:51:19, 10.33s/it]INFO 04-12 01:49:52 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2306/3300 [3:13:04<2:19:36,  8.43s/it]2026-04-12 01:49:55,911 | INFO | train_grpo_train | metrics_logged mode=train step=2306

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.0151515151515155e-06, 'num_tokens': 5222265.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.997984766960144, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997984766960144, 'rewards/total_composite/std': 0.0, 'reward': 0.997984766960144, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00116938806604594, 'sampling/sampling_logp_difference/max': 0.035293105989694595, 'sampling/importance_sampling_ratio/min': 0.9653224349021912, 'sampling/importance_sampling_ratio/mean': 1.0008283853530884, 'sampling/importance_sampling_ratio/max': 1.025589942932129, 'entropy': 0.009295969794038683, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.997984766960144, 'reward_meter_mean': 0.997984766960144, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997984766960144, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 70%|██████▉   | 2306/3300 [3:13:04<2:19:36,  8.43s/it]INFO 04-12 01:49:56 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2307/3300 [3:13:09<1:58:46,  7.18s/it]2026-04-12 01:50:00,175 | INFO | train_grpo_train | metrics_logged mode=train step=2307

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.0121212121212123e-06, 'num_tokens': 5223977.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005960848648101091, 'sampling/sampling_logp_difference/max': 0.03551316633820534, 'sampling/importance_sampling_ratio/min': 0.9723794460296631, 'sampling/importance_sampling_ratio/mean': 1.0003576278686523, 'sampling/importance_sampling_ratio/max': 1.0361512899398804, 'entropy': 0.00621713837608695, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 70%|██████▉   | 2307/3300 [3:13:09<1:58:46,  7.18s/it]INFO 04-12 01:50:00 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2308/3300 [3:13:13<1:43:07,  6.24s/it]2026-04-12 01:50:04,215 | INFO | train_grpo_train | metrics_logged mode=train step=2308

                                                       
{'loss': -0.0001, 'grad_norm': 1.3243682384490967, 'learning_rate': 3.009090909090909e-06, 'num_tokens': 5225489.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992917776107788, 'rewards/meter/std': 1.824959326768294e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992917776107788, 'rewards/total_composite/std': 1.824959326768294e-05, 'reward': 0.9992917776107788, 'reward_std': 1.82435705937678e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0022324356250464916, 'sampling/sampling_logp_difference/max': 0.30353307723999023, 'sampling/importance_sampling_ratio/min': 0.7382054924964905, 'sampling/importance_sampling_ratio/mean': 0.9991614818572998, 'sampling/importance_sampling_ratio/max': 1.0328819751739502, 'entropy': 0.007502294494770467, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992917776107788, 'reward_meter_mean': 0.9992917776107788, 'reward_meter_std': 1.824959326768294e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992917776107788, 'reward_total_composite_std': 1.824959326768294e-05, 'epoch': 0.09}

 70%|██████▉   | 2308/3300 [3:13:13<1:43:07,  6.24s/it]INFO 04-12 01:50:04 [block_pool.py:378] Successfully reset prefix cache

 70%|██████▉   | 2309/3300 [3:13:17<1:34:22,  5.71s/it]2026-04-12 01:50:08,709 | INFO | train_grpo_train | metrics_logged mode=train step=2309

                                                       
{'loss': 0.0018, 'grad_norm': 1.3253436088562012, 'learning_rate': 3.0060606060606064e-06, 'num_tokens': 5227402.0, 'completions/mean_length': 64.125, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9992837905883789, 'rewards/meter/std': 7.721914880676195e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992837905883789, 'rewards/total_composite/std': 7.721914880676195e-05, 'reward': 0.9992837905883789, 'reward_std': 7.722593727521598e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004359652288258076, 'sampling/sampling_logp_difference/max': 1.3222827911376953, 'sampling/importance_sampling_ratio/min': 0.7730492949485779, 'sampling/importance_sampling_ratio/mean': 1.002130150794983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.013516030623577535, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992837905883789, 'reward_meter_mean': 0.9992837905883789, 'reward_meter_std': 7.721914880676195e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992837905883789, 'reward_total_composite_std': 7.721914880676195e-05, 'epoch': 0.09}

 70%|██████▉   | 2309/3300 [3:13:17<1:34:22,  5.71s/it]INFO 04-12 01:50:08 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2310/3300 [3:13:21<1:26:37,  5.25s/it]2026-04-12 01:50:12,880 | INFO | train_grpo_train | metrics_logged mode=train step=2310

                                                       
{'loss': -0.001, 'grad_norm': 0.4977571368217468, 'learning_rate': 3.0030303030303032e-06, 'num_tokens': 5229178.0, 'completions/mean_length': 55.0, 'completions/min_length': 55.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9948618412017822, 'rewards/meter/std': 9.404453885508701e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948618412017822, 'rewards/total_composite/std': 9.404453885508701e-05, 'reward': 0.9948618412017822, 'reward_std': 9.405182208865881e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008051915094256401, 'sampling/sampling_logp_difference/max': 0.535893440246582, 'sampling/importance_sampling_ratio/min': 0.5851462483406067, 'sampling/importance_sampling_ratio/mean': 1.000832200050354, 'sampling/importance_sampling_ratio/max': 1.473840355873108, 'entropy': 0.03565134946256876, 'clip_ratio/low_mean': 0.004545454401522875, 'clip_ratio/low_min': 0.004545454401522875, 'clip_ratio/high_mean': 0.006818181602284312, 'clip_ratio/high_max': 0.006818181602284312, 'clip_ratio/region_mean': 0.011363636003807187, 'reward_total_mean': 0.9948618412017822, 'reward_meter_mean': 0.9948618412017822, 'reward_meter_std': 9.404453885508701e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948618412017822, 'reward_total_composite_std': 9.404453885508701e-05, 'epoch': 0.09}

 70%|███████   | 2310/3300 [3:13:21<1:26:37,  5.25s/it]INFO 04-12 01:50:13 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2311/3300 [3:13:25<1:21:19,  4.93s/it]2026-04-12 01:50:17,072 | INFO | train_grpo_train | metrics_logged mode=train step=2311

                                                       
{'loss': -0.0002, 'grad_norm': 0.3652920722961426, 'learning_rate': 3e-06, 'num_tokens': 5230866.0, 'completions/mean_length': 55.0, 'completions/min_length': 55.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9948963522911072, 'rewards/meter/std': 1.816852636693511e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948963522911072, 'rewards/total_composite/std': 1.816852636693511e-05, 'reward': 0.9948963522911072, 'reward_std': 1.817071097320877e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00472200708463788, 'sampling/sampling_logp_difference/max': 0.3054179549217224, 'sampling/importance_sampling_ratio/min': 0.7490553259849548, 'sampling/importance_sampling_ratio/mean': 0.999363124370575, 'sampling/importance_sampling_ratio/max': 1.3571921586990356, 'entropy': 0.032598731108009815, 'clip_ratio/low_mean': 0.0022727272007614374, 'clip_ratio/low_min': 0.0022727272007614374, 'clip_ratio/high_mean': 0.004545454401522875, 'clip_ratio/high_max': 0.004545454401522875, 'clip_ratio/region_mean': 0.006818181602284312, 'reward_total_mean': 0.9948963522911072, 'reward_meter_mean': 0.9948963522911072, 'reward_meter_std': 1.816852636693511e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948963522911072, 'reward_total_composite_std': 1.816852636693511e-05, 'epoch': 0.09}

 70%|███████   | 2311/3300 [3:13:26<1:21:19,  4.93s/it]INFO 04-12 01:50:17 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2312/3300 [3:13:30<1:20:32,  4.89s/it]2026-04-12 01:50:21,867 | INFO | train_grpo_train | metrics_logged mode=train step=2312

                                                       
{'loss': -0.0133, 'grad_norm': 2.4858875274658203, 'learning_rate': 2.996969696969697e-06, 'num_tokens': 5232710.0, 'completions/mean_length': 76.5, 'completions/min_length': 75.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9963726997375488, 'rewards/meter/std': 0.000908972229808569, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9963726997375488, 'rewards/total_composite/std': 0.000908972229808569, 'reward': 0.9963726997375488, 'reward_std': 0.0009089733357541263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024638459086418152, 'sampling/sampling_logp_difference/max': 1.1124024391174316, 'sampling/importance_sampling_ratio/min': 0.32876816391944885, 'sampling/importance_sampling_ratio/mean': 1.0097090005874634, 'sampling/importance_sampling_ratio/max': 1.9459387063980103, 'entropy': 0.18640629388391972, 'clip_ratio/low_mean': 0.011666666949167848, 'clip_ratio/low_min': 0.011666666949167848, 'clip_ratio/high_mean': 0.006349399336613715, 'clip_ratio/high_max': 0.006349399336613715, 'clip_ratio/region_mean': 0.018016066285781562, 'reward_total_mean': 0.9963726997375488, 'reward_meter_mean': 0.9963726997375488, 'reward_meter_std': 0.000908972229808569, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9963726997375488, 'reward_total_composite_std': 0.000908972229808569, 'epoch': 0.09}

 70%|███████   | 2312/3300 [3:13:30<1:20:32,  4.89s/it]INFO 04-12 01:50:22 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2313/3300 [3:13:35<1:17:11,  4.69s/it]2026-04-12 01:50:26,094 | INFO | train_grpo_train | metrics_logged mode=train step=2313

                                                       
{'loss': 0.001, 'grad_norm': 1.909967303276062, 'learning_rate': 2.993939393939394e-06, 'num_tokens': 5234365.0, 'completions/mean_length': 54.875, 'completions/min_length': 54.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9950602054595947, 'rewards/meter/std': 0.0005074667278677225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9950602054595947, 'rewards/total_composite/std': 0.0005074667278677225, 'reward': 0.9950602054595947, 'reward_std': 0.0005074667278677225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007333268411457539, 'sampling/sampling_logp_difference/max': 0.7493400573730469, 'sampling/importance_sampling_ratio/min': 0.47267839312553406, 'sampling/importance_sampling_ratio/mean': 0.9969737529754639, 'sampling/importance_sampling_ratio/max': 1.1668245792388916, 'entropy': 0.02833111328072846, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.9950602054595947, 'reward_meter_mean': 0.9950602054595947, 'reward_meter_std': 0.0005074667278677225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9950602054595947, 'reward_total_composite_std': 0.0005074667278677225, 'epoch': 0.09}

 70%|███████   | 2313/3300 [3:13:35<1:17:11,  4.69s/it]INFO 04-12 01:50:26 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2314/3300 [3:13:40<1:20:53,  4.92s/it]2026-04-12 01:50:31,552 | INFO | train_grpo_train | metrics_logged mode=train step=2314

                                                       
{'loss': -0.005, 'grad_norm': 1.915865182876587, 'learning_rate': 2.990909090909091e-06, 'num_tokens': 5237243.0, 'completions/mean_length': 151.75, 'completions/min_length': 149.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.75, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.9968797564506531, 'rewards/meter/std': 0.0010569265577942133, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.10101525485515594, 'rewards/total_composite/mean': 0.9612998962402344, 'rewards/total_composite/std': 0.10089994966983795, 'reward': 0.9612998962402344, 'reward_std': 0.10089994221925735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027893707156181335, 'sampling/sampling_logp_difference/max': 1.2294130325317383, 'sampling/importance_sampling_ratio/min': 0.29246416687965393, 'sampling/importance_sampling_ratio/mean': 1.0044018030166626, 'sampling/importance_sampling_ratio/max': 1.6189171075820923, 'entropy': 0.21811402402818203, 'clip_ratio/low_mean': 0.0008389261784031987, 'clip_ratio/low_min': 0.0008389261784031987, 'clip_ratio/high_mean': 0.0221836578566581, 'clip_ratio/high_max': 0.0221836578566581, 'clip_ratio/region_mean': 0.0230225840350613, 'reward_total_mean': 0.9612998962402344, 'reward_meter_mean': 0.9968797564506531, 'reward_meter_std': 0.0010569265577942133, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.10101525485515594, 'reward_total_composite_mean': 0.9612998962402344, 'reward_total_composite_std': 0.10089994966983795, 'epoch': 0.09}

 70%|███████   | 2314/3300 [3:13:40<1:20:53,  4.92s/it]INFO 04-12 01:50:31 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2315/3300 [3:13:45<1:18:56,  4.81s/it]2026-04-12 01:50:36,095 | INFO | train_grpo_train | metrics_logged mode=train step=2315

                                                       
{'loss': -0.001, 'grad_norm': 2.175835132598877, 'learning_rate': 2.987878787878788e-06, 'num_tokens': 5239130.0, 'completions/mean_length': 74.875, 'completions/min_length': 73.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.875, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9990985989570618, 'rewards/meter/std': 0.00011346396786393598, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990985989570618, 'rewards/total_composite/std': 0.00011346396786393598, 'reward': 0.9990985989570618, 'reward_std': 0.00011346933752065524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02791663445532322, 'sampling/sampling_logp_difference/max': 1.1362063884735107, 'sampling/importance_sampling_ratio/min': 0.32103458046913147, 'sampling/importance_sampling_ratio/mean': 1.0061297416687012, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21749908849596977, 'clip_ratio/low_mean': 0.005022522644139826, 'clip_ratio/low_min': 0.005022522644139826, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.005022522644139826, 'reward_total_mean': 0.9990985989570618, 'reward_meter_mean': 0.9990985989570618, 'reward_meter_std': 0.00011346396786393598, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990985989570618, 'reward_total_composite_std': 0.00011346396786393598, 'epoch': 0.09}

 70%|███████   | 2315/3300 [3:13:45<1:18:56,  4.81s/it]INFO 04-12 01:50:36 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2316/3300 [3:13:50<1:20:33,  4.91s/it]2026-04-12 01:50:41,249 | INFO | train_grpo_train | metrics_logged mode=train step=2316

                                                       
{'loss': 0.0011, 'grad_norm': 4.3104939460754395, 'learning_rate': 2.984848484848485e-06, 'num_tokens': 5241562.0, 'completions/mean_length': 128.0, 'completions/min_length': 127.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.0, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9043974876403809, 'rewards/meter/std': 0.2645852863788605, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7751978635787964, 'rewards/total_composite/std': 0.22678740322589874, 'reward': 0.7751978635787964, 'reward_std': 0.22678740322589874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01084654126316309, 'sampling/sampling_logp_difference/max': 0.9293702840805054, 'sampling/importance_sampling_ratio/min': 0.3948022723197937, 'sampling/importance_sampling_ratio/mean': 1.003368854522705, 'sampling/importance_sampling_ratio/max': 1.5058414936065674, 'entropy': 0.0738038350827992, 'clip_ratio/low_mean': 0.0049212598241865635, 'clip_ratio/low_min': 0.0049212598241865635, 'clip_ratio/high_mean': 0.0029296875, 'clip_ratio/high_max': 0.0029296875, 'clip_ratio/region_mean': 0.007850947324186563, 'reward_total_mean': 0.7751978635787964, 'reward_meter_mean': 0.9043974876403809, 'reward_meter_std': 0.2645852863788605, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7751978635787964, 'reward_total_composite_std': 0.22678740322589874, 'epoch': 0.09}

 70%|███████   | 2316/3300 [3:13:50<1:20:33,  4.91s/it]INFO 04-12 01:50:41 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2317/3300 [3:13:59<1:41:38,  6.20s/it]2026-04-12 01:50:50,466 | INFO | train_grpo_train | metrics_logged mode=train step=2317

                                                       
{'loss': -0.0076, 'grad_norm': 1.892166256904602, 'learning_rate': 2.981818181818182e-06, 'num_tokens': 5246963.0, 'completions/mean_length': 424.125, 'completions/min_length': 401.0, 'completions/max_length': 450.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 424.125, 'completions/min_terminated_length': 401.0, 'completions/max_terminated_length': 450.0, 'rewards/meter/mean': 0.992563009262085, 'rewards/meter/std': 0.0028387887869030237, 'rewards/count_adherence/mean': 0.6447368264198303, 'rewards/count_adherence/std': 0.024363704025745392, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9372192025184631, 'rewards/repeat_penalty/std': 0.03141416236758232, 'rewards/total_composite/mean': 0.5997676849365234, 'rewards/total_composite/std': 0.030326971784234047, 'reward': 0.5997676849365234, 'reward_std': 0.030326958745718002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0303457360714674, 'sampling/sampling_logp_difference/max': 1.870690107345581, 'sampling/importance_sampling_ratio/min': 0.1540173441171646, 'sampling/importance_sampling_ratio/mean': 1.0031834840774536, 'sampling/importance_sampling_ratio/max': 1.956535816192627, 'entropy': 0.2362643126398325, 'clip_ratio/low_mean': 0.012263738899491727, 'clip_ratio/low_min': 0.012263738899491727, 'clip_ratio/high_mean': 0.009331883396953344, 'clip_ratio/high_max': 0.009331883396953344, 'clip_ratio/region_mean': 0.02159562229644507, 'reward_total_mean': 0.5997676849365234, 'reward_meter_mean': 0.992563009262085, 'reward_meter_std': 0.0028387887869030237, 'reward_count_adherence_mean': 0.6447368264198303, 'reward_count_adherence_std': 0.024363704025745392, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9372192025184631, 'reward_repeat_penalty_std': 0.03141416236758232, 'reward_total_composite_mean': 0.5997676849365234, 'reward_total_composite_std': 0.030326971784234047, 'epoch': 0.09}

 70%|███████   | 2317/3300 [3:13:59<1:41:38,  6.20s/it]INFO 04-12 01:50:50 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2318/3300 [3:14:04<1:34:03,  5.75s/it]2026-04-12 01:50:55,149 | INFO | train_grpo_train | metrics_logged mode=train step=2318

                                                       
{'loss': -0.0195, 'grad_norm': 1.0140128135681152, 'learning_rate': 2.9787878787878787e-06, 'num_tokens': 5248971.0, 'completions/mean_length': 66.0, 'completions/min_length': 63.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9957113265991211, 'rewards/meter/std': 0.006800828501582146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957113265991211, 'rewards/total_composite/std': 0.006800828501582146, 'reward': 0.9957113265991211, 'reward_std': 0.0068008191883563995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006797808688133955, 'sampling/sampling_logp_difference/max': 1.166569709777832, 'sampling/importance_sampling_ratio/min': 0.3114334046840668, 'sampling/importance_sampling_ratio/mean': 0.9992537498474121, 'sampling/importance_sampling_ratio/max': 1.1265363693237305, 'entropy': 0.037344205658882856, 'clip_ratio/low_mean': 0.003968254197388887, 'clip_ratio/low_min': 0.003968254197388887, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.007699597394093871, 'reward_total_mean': 0.9957113265991211, 'reward_meter_mean': 0.9957113265991211, 'reward_meter_std': 0.006800828501582146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957113265991211, 'reward_total_composite_std': 0.006800828501582146, 'epoch': 0.09}

 70%|███████   | 2318/3300 [3:14:04<1:34:03,  5.75s/it]INFO 04-12 01:50:55 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2319/3300 [3:14:08<1:27:34,  5.36s/it]2026-04-12 01:50:59,593 | INFO | train_grpo_train | metrics_logged mode=train step=2319

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.9757575757575756e-06, 'num_tokens': 5250619.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00042203269549645483, 'sampling/sampling_logp_difference/max': 0.014573629945516586, 'sampling/importance_sampling_ratio/min': 0.9997897744178772, 'sampling/importance_sampling_ratio/mean': 1.0004198551177979, 'sampling/importance_sampling_ratio/max': 1.0146803855895996, 'entropy': 0.0041495013574603945, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 70%|███████   | 2319/3300 [3:14:08<1:27:34,  5.36s/it]INFO 04-12 01:50:59 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2320/3300 [3:14:14<1:32:28,  5.66s/it]2026-04-12 01:51:05,968 | INFO | train_grpo_train | metrics_logged mode=train step=2320

                                                       
{'loss': 0.0058, 'grad_norm': 1.6090915203094482, 'learning_rate': 2.9727272727272733e-06, 'num_tokens': 5253749.0, 'completions/mean_length': 202.25, 'completions/min_length': 200.0, 'completions/max_length': 204.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 202.25, 'completions/min_terminated_length': 200.0, 'completions/max_terminated_length': 204.0, 'rewards/meter/mean': 0.9957519769668579, 'rewards/meter/std': 0.004873792175203562, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9204546213150024, 'rewards/repeat_penalty/std': 0.03214120864868164, 'rewards/total_composite/mean': 0.7856208682060242, 'rewards/total_composite/std': 0.02808825671672821, 'reward': 0.7856208682060242, 'reward_std': 0.028088262304663658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02219260297715664, 'sampling/sampling_logp_difference/max': 3.1726443767547607, 'sampling/importance_sampling_ratio/min': 0.04189267382025719, 'sampling/importance_sampling_ratio/mean': 0.9995094537734985, 'sampling/importance_sampling_ratio/max': 1.7268435955047607, 'entropy': 0.12941910233348608, 'clip_ratio/low_mean': 0.01485577883431688, 'clip_ratio/low_min': 0.01485577883431688, 'clip_ratio/high_mean': 0.0024999999441206455, 'clip_ratio/high_max': 0.0024999999441206455, 'clip_ratio/region_mean': 0.017355778778437525, 'reward_total_mean': 0.7856208682060242, 'reward_meter_mean': 0.9957519769668579, 'reward_meter_std': 0.004873792175203562, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9204546213150024, 'reward_repeat_penalty_std': 0.03214120864868164, 'reward_total_composite_mean': 0.7856208682060242, 'reward_total_composite_std': 0.02808825671672821, 'epoch': 0.09}

 70%|███████   | 2320/3300 [3:14:14<1:32:28,  5.66s/it]INFO 04-12 01:51:06 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2321/3300 [3:14:19<1:26:21,  5.29s/it]2026-04-12 01:51:10,402 | INFO | train_grpo_train | metrics_logged mode=train step=2321

                                                       
{'loss': -0.0002, 'grad_norm': 0.006876571103930473, 'learning_rate': 2.96969696969697e-06, 'num_tokens': 5255469.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7876332998275757, 'rewards/meter/std': 1.5573279597447254e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7876332998275757, 'rewards/total_composite/std': 1.5573279597447254e-05, 'reward': 0.7876332998275757, 'reward_std': 1.558229632792063e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0014006540877744555, 'sampling/sampling_logp_difference/max': 0.3940424919128418, 'sampling/importance_sampling_ratio/min': 0.6743254065513611, 'sampling/importance_sampling_ratio/mean': 0.9997325539588928, 'sampling/importance_sampling_ratio/max': 1.0165002346038818, 'entropy': 0.004189703788142651, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.7876332998275757, 'reward_meter_mean': 0.7876332998275757, 'reward_meter_std': 1.5573279597447254e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7876332998275757, 'reward_total_composite_std': 1.5573279597447254e-05, 'epoch': 0.09}

 70%|███████   | 2321/3300 [3:14:19<1:26:21,  5.29s/it]INFO 04-12 01:51:10 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2322/3300 [3:14:23<1:23:08,  5.10s/it]2026-04-12 01:51:15,053 | INFO | train_grpo_train | metrics_logged mode=train step=2322

                                                       
{'loss': 0.0375, 'grad_norm': 7.9188666343688965, 'learning_rate': 2.9666666666666673e-06, 'num_tokens': 5257324.0, 'completions/mean_length': 67.875, 'completions/min_length': 61.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.875, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.7128692269325256, 'rewards/meter/std': 0.2854415774345398, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.14880475401878357, 'rewards/total_composite/mean': 0.5824298858642578, 'rewards/total_composite/std': 0.28861746191978455, 'reward': 0.5824298858642578, 'reward_std': 0.28861746191978455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07370011508464813, 'sampling/sampling_logp_difference/max': 1.0788640975952148, 'sampling/importance_sampling_ratio/min': 0.3399814963340759, 'sampling/importance_sampling_ratio/mean': 1.0028222799301147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3680881727486849, 'clip_ratio/low_mean': 0.032118565402925014, 'clip_ratio/low_min': 0.032118565402925014, 'clip_ratio/high_mean': 0.04014339251443744, 'clip_ratio/high_max': 0.04014339251443744, 'clip_ratio/region_mean': 0.07226195791736245, 'reward_total_mean': 0.5824298858642578, 'reward_meter_mean': 0.7128692269325256, 'reward_meter_std': 0.2854415774345398, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.14880475401878357, 'reward_total_composite_mean': 0.5824298858642578, 'reward_total_composite_std': 0.28861746191978455, 'epoch': 0.09}

 70%|███████   | 2322/3300 [3:14:24<1:23:08,  5.10s/it]INFO 04-12 01:51:15 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2323/3300 [3:14:28<1:20:57,  4.97s/it]2026-04-12 01:51:19,724 | INFO | train_grpo_train | metrics_logged mode=train step=2323

                                                       
{'loss': 0.0003, 'grad_norm': 0.16411831974983215, 'learning_rate': 2.963636363636364e-06, 'num_tokens': 5259053.0, 'completions/mean_length': 64.125, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9993196725845337, 'rewards/meter/std': 1.4266728612710722e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993196725845337, 'rewards/total_composite/std': 1.4266728612710722e-05, 'reward': 0.9993196725845337, 'reward_std': 1.4269724488258362e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0027244146913290024, 'sampling/sampling_logp_difference/max': 0.8163547515869141, 'sampling/importance_sampling_ratio/min': 0.4420400857925415, 'sampling/importance_sampling_ratio/mean': 0.9990805387496948, 'sampling/importance_sampling_ratio/max': 1.0394550561904907, 'entropy': 0.009595987561624497, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.001923076924867928, 'clip_ratio/high_max': 0.001923076924867928, 'clip_ratio/region_mean': 0.001923076924867928, 'reward_total_mean': 0.9993196725845337, 'reward_meter_mean': 0.9993196725845337, 'reward_meter_std': 1.4266728612710722e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993196725845337, 'reward_total_composite_std': 1.4266728612710722e-05, 'epoch': 0.09}

 70%|███████   | 2323/3300 [3:14:28<1:20:57,  4.97s/it]INFO 04-12 01:51:20 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2324/3300 [3:14:33<1:22:40,  5.08s/it]2026-04-12 01:51:25,066 | INFO | train_grpo_train | metrics_logged mode=train step=2324

                                                       
{'loss': 0.0001, 'grad_norm': 3.402935028076172, 'learning_rate': 2.960606060606061e-06, 'num_tokens': 5261376.0, 'completions/mean_length': 115.375, 'completions/min_length': 112.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.375, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9961607456207275, 'rewards/meter/std': 0.003092888044193387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961607456207275, 'rewards/total_composite/std': 0.003092888044193387, 'reward': 0.9961607456207275, 'reward_std': 0.0030928829219192266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027808500453829765, 'sampling/sampling_logp_difference/max': 1.3722867965698242, 'sampling/importance_sampling_ratio/min': 0.2535265386104584, 'sampling/importance_sampling_ratio/mean': 1.0031054019927979, 'sampling/importance_sampling_ratio/max': 1.6382875442504883, 'entropy': 0.2400868460536003, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.021577789448201656, 'clip_ratio/high_max': 0.021577789448201656, 'clip_ratio/region_mean': 0.025888134259730577, 'reward_total_mean': 0.9961607456207275, 'reward_meter_mean': 0.9961607456207275, 'reward_meter_std': 0.003092888044193387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961607456207275, 'reward_total_composite_std': 0.003092888044193387, 'epoch': 0.09}

 70%|███████   | 2324/3300 [3:14:34<1:22:40,  5.08s/it]INFO 04-12 01:51:25 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2325/3300 [3:14:38<1:21:15,  5.00s/it]2026-04-12 01:51:29,876 | INFO | train_grpo_train | metrics_logged mode=train step=2325

                                                       
{'loss': 0.0074, 'grad_norm': 4.780300140380859, 'learning_rate': 2.957575757575758e-06, 'num_tokens': 5263331.0, 'completions/mean_length': 67.375, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9977126121520996, 'rewards/meter/std': 0.0028336539398878813, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977126121520996, 'rewards/total_composite/std': 0.0028336539398878813, 'reward': 0.9977126121520996, 'reward_std': 0.0028336546383798122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049458228051662445, 'sampling/sampling_logp_difference/max': 1.1433992385864258, 'sampling/importance_sampling_ratio/min': 0.31873372197151184, 'sampling/importance_sampling_ratio/mean': 1.0002453327178955, 'sampling/importance_sampling_ratio/max': 1.6862565279006958, 'entropy': 0.3584112077951431, 'clip_ratio/low_mean': 0.009615384973585606, 'clip_ratio/low_min': 0.009615384973585606, 'clip_ratio/high_mean': 0.02963537711184472, 'clip_ratio/high_max': 0.02963537711184472, 'clip_ratio/region_mean': 0.039250762085430324, 'reward_total_mean': 0.9977126121520996, 'reward_meter_mean': 0.9977126121520996, 'reward_meter_std': 0.0028336539398878813, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977126121520996, 'reward_total_composite_std': 0.0028336539398878813, 'epoch': 0.09}

 70%|███████   | 2325/3300 [3:14:38<1:21:15,  5.00s/it]INFO 04-12 01:51:30 [block_pool.py:378] Successfully reset prefix cache

 70%|███████   | 2326/3300 [3:14:43<1:20:25,  4.95s/it]2026-04-12 01:51:34,721 | INFO | train_grpo_train | metrics_logged mode=train step=2326

                                                       
{'loss': -0.0182, 'grad_norm': 3.8735227584838867, 'learning_rate': 2.954545454545455e-06, 'num_tokens': 5265267.0, 'completions/mean_length': 78.0, 'completions/min_length': 75.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.0, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9950919151306152, 'rewards/meter/std': 0.007245616987347603, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9950919151306152, 'rewards/total_composite/std': 0.007245616987347603, 'reward': 0.9950919151306152, 'reward_std': 0.007245623506605625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01848999224603176, 'sampling/sampling_logp_difference/max': 0.8240165710449219, 'sampling/importance_sampling_ratio/min': 0.4386661946773529, 'sampling/importance_sampling_ratio/mean': 1.0039349794387817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16736672818660736, 'clip_ratio/low_mean': 0.0033333334140479565, 'clip_ratio/low_min': 0.0033333334140479565, 'clip_ratio/high_mean': 0.011138874455355108, 'clip_ratio/high_max': 0.011138874455355108, 'clip_ratio/region_mean': 0.014472207869403064, 'reward_total_mean': 0.9950919151306152, 'reward_meter_mean': 0.9950919151306152, 'reward_meter_std': 0.007245616987347603, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9950919151306152, 'reward_total_composite_std': 0.007245616987347603, 'epoch': 0.09}

 70%|███████   | 2326/3300 [3:14:43<1:20:25,  4.95s/it]INFO 04-12 01:51:34 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2327/3300 [3:14:48<1:17:30,  4.78s/it]2026-04-12 01:51:39,092 | INFO | train_grpo_train | metrics_logged mode=train step=2327

                                                       
{'loss': -0.0017, 'grad_norm': 7.240959167480469, 'learning_rate': 2.951515151515152e-06, 'num_tokens': 5266851.0, 'completions/mean_length': 39.0, 'completions/min_length': 37.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9987367391586304, 'rewards/meter/std': 0.0013684039004147053, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987367391586304, 'rewards/total_composite/std': 0.0013684039004147053, 'reward': 0.9987367391586304, 'reward_std': 0.0013684096047654748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032071422785520554, 'sampling/sampling_logp_difference/max': 0.8518095016479492, 'sampling/importance_sampling_ratio/min': 0.42664220929145813, 'sampling/importance_sampling_ratio/mean': 1.003873348236084, 'sampling/importance_sampling_ratio/max': 1.3673170804977417, 'entropy': 0.19926494918763638, 'clip_ratio/low_mean': 0.00937500037252903, 'clip_ratio/low_min': 0.00937500037252903, 'clip_ratio/high_mean': 0.009627727791666985, 'clip_ratio/high_max': 0.009627727791666985, 'clip_ratio/region_mean': 0.019002728164196014, 'reward_total_mean': 0.9987367391586304, 'reward_meter_mean': 0.9987367391586304, 'reward_meter_std': 0.0013684039004147053, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987367391586304, 'reward_total_composite_std': 0.0013684039004147053, 'epoch': 0.09}

 71%|███████   | 2327/3300 [3:14:48<1:17:30,  4.78s/it]INFO 04-12 01:51:39 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2328/3300 [3:14:53<1:21:09,  5.01s/it]2026-04-12 01:51:44,641 | INFO | train_grpo_train | metrics_logged mode=train step=2328

                                                       
{'loss': 0.0007, 'grad_norm': 0.9024956822395325, 'learning_rate': 2.9484848484848488e-06, 'num_tokens': 5269464.0, 'completions/mean_length': 159.625, 'completions/min_length': 158.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.625, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9978625774383545, 'rewards/meter/std': 3.399623528821394e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.819444477558136, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.8176930546760559, 'rewards/total_composite/std': 0.057384345680475235, 'reward': 0.8176930546760559, 'reward_std': 0.05738433822989464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010679478757083416, 'sampling/sampling_logp_difference/max': 1.0905389785766602, 'sampling/importance_sampling_ratio/min': 0.3360353410243988, 'sampling/importance_sampling_ratio/mean': 0.9999021291732788, 'sampling/importance_sampling_ratio/max': 1.274688720703125, 'entropy': 0.060957128182053566, 'clip_ratio/low_mean': 0.0007812500116415322, 'clip_ratio/low_min': 0.0007812500116415322, 'clip_ratio/high_mean': 0.0031250000465661287, 'clip_ratio/high_max': 0.0031250000465661287, 'clip_ratio/region_mean': 0.003906250058207661, 'reward_total_mean': 0.8176930546760559, 'reward_meter_mean': 0.9978625774383545, 'reward_meter_std': 3.399623528821394e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.819444477558136, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.8176930546760559, 'reward_total_composite_std': 0.057384345680475235, 'epoch': 0.09}

 71%|███████   | 2328/3300 [3:14:53<1:21:09,  5.01s/it]INFO 04-12 01:51:44 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2329/3300 [3:14:58<1:18:59,  4.88s/it]2026-04-12 01:51:49,221 | INFO | train_grpo_train | metrics_logged mode=train step=2329

                                                       
{'loss': 0.0043, 'grad_norm': 3.704277992248535, 'learning_rate': 2.9454545454545456e-06, 'num_tokens': 5271316.0, 'completions/mean_length': 66.5, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9816258549690247, 'rewards/meter/std': 0.004315865226089954, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9816258549690247, 'rewards/total_composite/std': 0.004315865226089954, 'reward': 0.9816258549690247, 'reward_std': 0.004315854981541634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022112296894192696, 'sampling/sampling_logp_difference/max': 1.1111278533935547, 'sampling/importance_sampling_ratio/min': 0.3291874825954437, 'sampling/importance_sampling_ratio/mean': 1.0040234327316284, 'sampling/importance_sampling_ratio/max': 1.3998769521713257, 'entropy': 0.1675149966031313, 'clip_ratio/low_mean': 0.013287583249621093, 'clip_ratio/low_min': 0.013287583249621093, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.017018926446326077, 'reward_total_mean': 0.9816258549690247, 'reward_meter_mean': 0.9816258549690247, 'reward_meter_std': 0.004315865226089954, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9816258549690247, 'reward_total_composite_std': 0.004315865226089954, 'epoch': 0.09}

 71%|███████   | 2329/3300 [3:14:58<1:18:59,  4.88s/it]INFO 04-12 01:51:49 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2330/3300 [3:15:02<1:17:02,  4.77s/it]2026-04-12 01:51:53,718 | INFO | train_grpo_train | metrics_logged mode=train step=2330

                                                       
{'loss': 0.001, 'grad_norm': 4.507577419281006, 'learning_rate': 2.942424242424243e-06, 'num_tokens': 5272995.0, 'completions/mean_length': 67.875, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9815051555633545, 'rewards/meter/std': 0.004705195315182209, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9815051555633545, 'rewards/total_composite/std': 0.004705195315182209, 'reward': 0.9815051555633545, 'reward_std': 0.004705206956714392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015049697831273079, 'sampling/sampling_logp_difference/max': 0.6336269378662109, 'sampling/importance_sampling_ratio/min': 0.5306636691093445, 'sampling/importance_sampling_ratio/mean': 1.0061925649642944, 'sampling/importance_sampling_ratio/max': 1.3674873113632202, 'entropy': 0.0950354440137744, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0036764706019312143, 'clip_ratio/high_max': 0.0036764706019312143, 'clip_ratio/region_mean': 0.005542142200283706, 'reward_total_mean': 0.9815051555633545, 'reward_meter_mean': 0.9815051555633545, 'reward_meter_std': 0.004705195315182209, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9815051555633545, 'reward_total_composite_std': 0.004705195315182209, 'epoch': 0.09}

 71%|███████   | 2330/3300 [3:15:02<1:17:02,  4.77s/it]INFO 04-12 01:51:53 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2331/3300 [3:15:07<1:16:00,  4.71s/it]2026-04-12 01:51:58,289 | INFO | train_grpo_train | metrics_logged mode=train step=2331

                                                       
{'loss': 0.0002, 'grad_norm': 3.0374279022216797, 'learning_rate': 2.9393939393939397e-06, 'num_tokens': 5274705.0, 'completions/mean_length': 65.75, 'completions/min_length': 64.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9988436698913574, 'rewards/meter/std': 0.00024908926570788026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988436698913574, 'rewards/total_composite/std': 0.00024908926570788026, 'reward': 0.9988436698913574, 'reward_std': 0.000249094155151397, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04332128167152405, 'sampling/sampling_logp_difference/max': 1.0077180862426758, 'sampling/importance_sampling_ratio/min': 0.3650510609149933, 'sampling/importance_sampling_ratio/mean': 1.0106713771820068, 'sampling/importance_sampling_ratio/max': 1.7954013347625732, 'entropy': 0.3256809823215008, 'clip_ratio/low_mean': 0.020951705053448677, 'clip_ratio/low_min': 0.020951705053448677, 'clip_ratio/high_mean': 0.018828062573447824, 'clip_ratio/high_max': 0.018828062573447824, 'clip_ratio/region_mean': 0.0397797676268965, 'reward_total_mean': 0.9988436698913574, 'reward_meter_mean': 0.9988436698913574, 'reward_meter_std': 0.00024908926570788026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988436698913574, 'reward_total_composite_std': 0.00024908926570788026, 'epoch': 0.09}

 71%|███████   | 2331/3300 [3:15:07<1:16:00,  4.71s/it]INFO 04-12 01:51:58 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2332/3300 [3:15:11<1:14:42,  4.63s/it]2026-04-12 01:52:02,739 | INFO | train_grpo_train | metrics_logged mode=train step=2332

                                                       
{'loss': -0.0003, 'grad_norm': 0.5694947242736816, 'learning_rate': 2.9363636363636365e-06, 'num_tokens': 5276497.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.998113214969635, 'rewards/meter/std': 3.099795139860362e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998113214969635, 'rewards/total_composite/std': 3.099795139860362e-05, 'reward': 0.998113214969635, 'reward_std': 3.100236426689662e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007214725017547607, 'sampling/sampling_logp_difference/max': 0.7471715211868286, 'sampling/importance_sampling_ratio/min': 0.4737045168876648, 'sampling/importance_sampling_ratio/mean': 1.002618670463562, 'sampling/importance_sampling_ratio/max': 1.6667267084121704, 'entropy': 0.05256783217191696, 'clip_ratio/low_mean': 0.005597014795057476, 'clip_ratio/low_min': 0.005597014795057476, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.998113214969635, 'reward_meter_mean': 0.998113214969635, 'reward_meter_std': 3.099795139860362e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998113214969635, 'reward_total_composite_std': 3.099795139860362e-05, 'epoch': 0.09}

 71%|███████   | 2332/3300 [3:15:11<1:14:42,  4.63s/it]INFO 04-12 01:52:02 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2333/3300 [3:15:16<1:14:50,  4.64s/it]2026-04-12 01:52:07,416 | INFO | train_grpo_train | metrics_logged mode=train step=2333

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.9333333333333338e-06, 'num_tokens': 5278425.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0004051509313285351, 'sampling/sampling_logp_difference/max': 0.01761876977980137, 'sampling/importance_sampling_ratio/min': 0.9928526878356934, 'sampling/importance_sampling_ratio/mean': 1.0003622770309448, 'sampling/importance_sampling_ratio/max': 1.0177748203277588, 'entropy': 0.0035651897196657956, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 71%|███████   | 2333/3300 [3:15:16<1:14:50,  4.64s/it]INFO 04-12 01:52:07 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2334/3300 [3:15:22<1:23:12,  5.17s/it]2026-04-12 01:52:13,805 | INFO | train_grpo_train | metrics_logged mode=train step=2334

                                                       
{'loss': -0.0067, 'grad_norm': 8.575316429138184, 'learning_rate': 2.9303030303030306e-06, 'num_tokens': 5280271.0, 'completions/mean_length': 66.75, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9969208836555481, 'rewards/meter/std': 0.0032654430251568556, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969208836555481, 'rewards/total_composite/std': 0.0032654430251568556, 'reward': 0.9969208836555481, 'reward_std': 0.003265430685132742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012270020321011543, 'sampling/sampling_logp_difference/max': 1.2344985008239746, 'sampling/importance_sampling_ratio/min': 0.29098066687583923, 'sampling/importance_sampling_ratio/mean': 0.9968401193618774, 'sampling/importance_sampling_ratio/max': 1.4511247873306274, 'entropy': 0.0710354926995933, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.011166593292728066, 'clip_ratio/high_max': 0.011166593292728066, 'clip_ratio/region_mean': 0.011166593292728066, 'reward_total_mean': 0.9969208836555481, 'reward_meter_mean': 0.9969208836555481, 'reward_meter_std': 0.0032654430251568556, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969208836555481, 'reward_total_composite_std': 0.0032654430251568556, 'epoch': 0.09}

 71%|███████   | 2334/3300 [3:15:22<1:23:12,  5.17s/it]INFO 04-12 01:52:14 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2335/3300 [3:15:27<1:19:09,  4.92s/it]2026-04-12 01:52:18,153 | INFO | train_grpo_train | metrics_logged mode=train step=2335

                                                       
{'loss': 0.0011, 'grad_norm': 2.4751315116882324, 'learning_rate': 2.9272727272727274e-06, 'num_tokens': 5281911.0, 'completions/mean_length': 55.0, 'completions/min_length': 55.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9950971603393555, 'rewards/meter/std': 0.0004900125786662102, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9950971603393555, 'rewards/total_composite/std': 0.0004900125786662102, 'reward': 0.9950971603393555, 'reward_std': 0.0004900065250694752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005754906218498945, 'sampling/sampling_logp_difference/max': 0.3944000005722046, 'sampling/importance_sampling_ratio/min': 0.6740844249725342, 'sampling/importance_sampling_ratio/mean': 0.9981791973114014, 'sampling/importance_sampling_ratio/max': 1.2244776487350464, 'entropy': 0.022867268649861217, 'clip_ratio/low_mean': 0.0022727272007614374, 'clip_ratio/low_min': 0.0022727272007614374, 'clip_ratio/high_mean': 0.0022727272007614374, 'clip_ratio/high_max': 0.0022727272007614374, 'clip_ratio/region_mean': 0.004545454401522875, 'reward_total_mean': 0.9950971603393555, 'reward_meter_mean': 0.9950971603393555, 'reward_meter_std': 0.0004900125786662102, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9950971603393555, 'reward_total_composite_std': 0.0004900125786662102, 'epoch': 0.09}

 71%|███████   | 2335/3300 [3:15:27<1:19:09,  4.92s/it]INFO 04-12 01:52:18 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2336/3300 [3:15:31<1:17:29,  4.82s/it]2026-04-12 01:52:22,744 | INFO | train_grpo_train | metrics_logged mode=train step=2336

                                                       
{'loss': 0.0022, 'grad_norm': 2.559536933898926, 'learning_rate': 2.9242424242424243e-06, 'num_tokens': 5283616.0, 'completions/mean_length': 67.125, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9981128573417664, 'rewards/meter/std': 5.023560152039863e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981128573417664, 'rewards/total_composite/std': 5.023560152039863e-05, 'reward': 0.9981128573417664, 'reward_std': 5.024260462960228e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004402663093060255, 'sampling/sampling_logp_difference/max': 0.3112151622772217, 'sampling/importance_sampling_ratio/min': 0.8633265495300293, 'sampling/importance_sampling_ratio/mean': 1.00238835811615, 'sampling/importance_sampling_ratio/max': 1.3650829792022705, 'entropy': 0.04221090069040656, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981128573417664, 'reward_meter_mean': 0.9981128573417664, 'reward_meter_std': 5.023560152039863e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981128573417664, 'reward_total_composite_std': 5.023560152039863e-05, 'epoch': 0.09}

 71%|███████   | 2336/3300 [3:15:31<1:17:29,  4.82s/it]INFO 04-12 01:52:22 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2337/3300 [3:15:37<1:24:40,  5.28s/it]2026-04-12 01:52:29,077 | INFO | train_grpo_train | metrics_logged mode=train step=2337

                                                       
{'loss': 0.0008, 'grad_norm': 1.5094997882843018, 'learning_rate': 2.9212121212121215e-06, 'num_tokens': 5287016.0, 'completions/mean_length': 207.0, 'completions/min_length': 206.0, 'completions/max_length': 208.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 207.0, 'completions/min_terminated_length': 206.0, 'completions/max_terminated_length': 208.0, 'rewards/meter/mean': 0.9978703260421753, 'rewards/meter/std': 3.329759420012124e-05, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8020833730697632, 'rewards/repeat_penalty/std': 0.10853918641805649, 'rewards/total_composite/mean': 0.686035692691803, 'rewards/total_composite/std': 0.09283558279275894, 'reward': 0.686035692691803, 'reward_std': 0.09283558279275894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013423108495771885, 'sampling/sampling_logp_difference/max': 1.0543880462646484, 'sampling/importance_sampling_ratio/min': 0.34840553998947144, 'sampling/importance_sampling_ratio/mean': 1.0003283023834229, 'sampling/importance_sampling_ratio/max': 1.4435890913009644, 'entropy': 0.09781584423035383, 'clip_ratio/low_mean': 0.007231889001559466, 'clip_ratio/low_min': 0.007231889001559466, 'clip_ratio/high_mean': 0.0036231884150765836, 'clip_ratio/high_max': 0.0036231884150765836, 'clip_ratio/region_mean': 0.01085507741663605, 'reward_total_mean': 0.686035692691803, 'reward_meter_mean': 0.9978703260421753, 'reward_meter_std': 3.329759420012124e-05, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8020833730697632, 'reward_repeat_penalty_std': 0.10853918641805649, 'reward_total_composite_mean': 0.686035692691803, 'reward_total_composite_std': 0.09283558279275894, 'epoch': 0.09}

 71%|███████   | 2337/3300 [3:15:38<1:24:40,  5.28s/it]INFO 04-12 01:52:29 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2338/3300 [3:15:42<1:21:15,  5.07s/it]2026-04-12 01:52:33,663 | INFO | train_grpo_train | metrics_logged mode=train step=2338

                                                       
{'loss': -0.0001, 'grad_norm': 0.4725949764251709, 'learning_rate': 2.9181818181818183e-06, 'num_tokens': 5288928.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981162548065186, 'rewards/meter/std': 2.446455619065091e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981162548065186, 'rewards/total_composite/std': 2.446455619065091e-05, 'reward': 0.9981162548065186, 'reward_std': 2.446935832267627e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006431423127651215, 'sampling/sampling_logp_difference/max': 0.8494582176208496, 'sampling/importance_sampling_ratio/min': 0.42764657735824585, 'sampling/importance_sampling_ratio/mean': 0.9997783899307251, 'sampling/importance_sampling_ratio/max': 1.1632411479949951, 'entropy': 0.04713014606386423, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981162548065186, 'reward_meter_mean': 0.9981162548065186, 'reward_meter_std': 2.446455619065091e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981162548065186, 'reward_total_composite_std': 2.446455619065091e-05, 'epoch': 0.09}

 71%|███████   | 2338/3300 [3:15:42<1:21:15,  5.07s/it]INFO 04-12 01:52:33 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2339/3300 [3:15:47<1:22:07,  5.13s/it]2026-04-12 01:52:39,022 | INFO | train_grpo_train | metrics_logged mode=train step=2339

                                                       
{'loss': 0.0017, 'grad_norm': 0.770327091217041, 'learning_rate': 2.915151515151515e-06, 'num_tokens': 5291330.0, 'completions/mean_length': 129.25, 'completions/min_length': 129.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.25, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.9979046583175659, 'rewards/meter/std': 4.928474663756788e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8553467988967896, 'rewards/total_composite/std': 4.225334123475477e-05, 'reward': 0.8553467988967896, 'reward_std': 4.226414966979064e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010605924762785435, 'sampling/sampling_logp_difference/max': 1.4333992004394531, 'sampling/importance_sampling_ratio/min': 0.23849685490131378, 'sampling/importance_sampling_ratio/mean': 0.9997851848602295, 'sampling/importance_sampling_ratio/max': 1.9889717102050781, 'entropy': 0.040207551792263985, 'clip_ratio/low_mean': 0.001923076924867928, 'clip_ratio/low_min': 0.001923076924867928, 'clip_ratio/high_mean': 0.00484496122226119, 'clip_ratio/high_max': 0.00484496122226119, 'clip_ratio/region_mean': 0.0067680381471291184, 'reward_total_mean': 0.8553467988967896, 'reward_meter_mean': 0.9979046583175659, 'reward_meter_std': 4.928474663756788e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8553467988967896, 'reward_total_composite_std': 4.225334123475477e-05, 'epoch': 0.09}

 71%|███████   | 2339/3300 [3:15:48<1:22:07,  5.13s/it]INFO 04-12 01:52:39 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2340/3300 [3:15:55<1:32:47,  5.80s/it]2026-04-12 01:52:46,295 | INFO | train_grpo_train | metrics_logged mode=train step=2340

                                                       
{'loss': -0.0245, 'grad_norm': 1.771281361579895, 'learning_rate': 2.9121212121212124e-06, 'num_tokens': 5295161.0, 'completions/mean_length': 262.875, 'completions/min_length': 252.0, 'completions/max_length': 289.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 262.875, 'completions/min_terminated_length': 252.0, 'completions/max_terminated_length': 289.0, 'rewards/meter/mean': 0.9991946220397949, 'rewards/meter/std': 0.00014320346235763282, 'rewards/count_adherence/mean': 0.8055555820465088, 'rewards/count_adherence/std': 0.05143444985151291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8057692050933838, 'rewards/repeat_penalty/std': 0.08968339115381241, 'rewards/total_composite/mean': 0.6484043002128601, 'rewards/total_composite/std': 0.07946042716503143, 'reward': 0.6484043002128601, 'reward_std': 0.07946040481328964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017933662980794907, 'sampling/sampling_logp_difference/max': 1.0136165618896484, 'sampling/importance_sampling_ratio/min': 0.3629041314125061, 'sampling/importance_sampling_ratio/mean': 1.00568687915802, 'sampling/importance_sampling_ratio/max': 1.6187031269073486, 'entropy': 0.18291304260492325, 'clip_ratio/low_mean': 0.004446686594747007, 'clip_ratio/low_min': 0.004446686594747007, 'clip_ratio/high_mean': 0.011448833451140672, 'clip_ratio/high_max': 0.011448833451140672, 'clip_ratio/region_mean': 0.01589552004588768, 'reward_total_mean': 0.6484043002128601, 'reward_meter_mean': 0.9991946220397949, 'reward_meter_std': 0.00014320346235763282, 'reward_count_adherence_mean': 0.8055555820465088, 'reward_count_adherence_std': 0.05143444985151291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8057692050933838, 'reward_repeat_penalty_std': 0.08968339115381241, 'reward_total_composite_mean': 0.6484043002128601, 'reward_total_composite_std': 0.07946042716503143, 'epoch': 0.09}

 71%|███████   | 2340/3300 [3:15:55<1:32:47,  5.80s/it]INFO 04-12 01:52:46 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2341/3300 [3:16:02<1:39:55,  6.25s/it]2026-04-12 01:52:53,609 | INFO | train_grpo_train | metrics_logged mode=train step=2341

                                                       
{'loss': -0.0113, 'grad_norm': 1.6412862539291382, 'learning_rate': 2.9090909090909093e-06, 'num_tokens': 5299382.0, 'completions/mean_length': 265.625, 'completions/min_length': 250.0, 'completions/max_length': 290.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 265.625, 'completions/min_terminated_length': 250.0, 'completions/max_terminated_length': 290.0, 'rewards/meter/mean': 0.9990073442459106, 'rewards/meter/std': 0.0001355017302557826, 'rewards/count_adherence/mean': 0.737500011920929, 'rewards/count_adherence/std': 0.05175492912530899, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8301282525062561, 'rewards/repeat_penalty/std': 0.06518138945102692, 'rewards/total_composite/mean': 0.6096493005752563, 'rewards/total_composite/std': 0.03883464261889458, 'reward': 0.6096493005752563, 'reward_std': 0.03883464261889458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022991513833403587, 'sampling/sampling_logp_difference/max': 2.072540044784546, 'sampling/importance_sampling_ratio/min': 0.30119454860687256, 'sampling/importance_sampling_ratio/mean': 1.0052927732467651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20100159756839275, 'clip_ratio/low_mean': 0.011894350522197783, 'clip_ratio/low_min': 0.011894350522197783, 'clip_ratio/high_mean': 0.00551724131219089, 'clip_ratio/high_max': 0.00551724131219089, 'clip_ratio/region_mean': 0.017411591834388673, 'reward_total_mean': 0.6096493005752563, 'reward_meter_mean': 0.9990073442459106, 'reward_meter_std': 0.0001355017302557826, 'reward_count_adherence_mean': 0.737500011920929, 'reward_count_adherence_std': 0.05175492912530899, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8301282525062561, 'reward_repeat_penalty_std': 0.06518138945102692, 'reward_total_composite_mean': 0.6096493005752563, 'reward_total_composite_std': 0.03883464261889458, 'epoch': 0.09}

 71%|███████   | 2341/3300 [3:16:02<1:39:55,  6.25s/it]INFO 04-12 01:52:53 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2342/3300 [3:16:06<1:31:01,  5.70s/it]2026-04-12 01:52:58,022 | INFO | train_grpo_train | metrics_logged mode=train step=2342

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.906060606060606e-06, 'num_tokens': 5301206.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00041229455382563174, 'sampling/sampling_logp_difference/max': 0.0067067258059978485, 'sampling/importance_sampling_ratio/min': 0.9993113875389099, 'sampling/importance_sampling_ratio/mean': 1.000400185585022, 'sampling/importance_sampling_ratio/max': 1.0067293643951416, 'entropy': 0.003808505309280008, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 71%|███████   | 2342/3300 [3:16:07<1:31:01,  5.70s/it]INFO 04-12 01:52:58 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2343/3300 [3:16:11<1:25:26,  5.36s/it]2026-04-12 01:53:02,574 | INFO | train_grpo_train | metrics_logged mode=train step=2343

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.903030303030303e-06, 'num_tokens': 5303262.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00038630847120657563, 'sampling/sampling_logp_difference/max': 0.01495426706969738, 'sampling/importance_sampling_ratio/min': 0.9851569533348083, 'sampling/importance_sampling_ratio/mean': 1.0003089904785156, 'sampling/importance_sampling_ratio/max': 1.0098844766616821, 'entropy': 0.0037699798413086683, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 71%|███████   | 2343/3300 [3:16:11<1:25:26,  5.36s/it]INFO 04-12 01:53:02 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2344/3300 [3:16:16<1:21:52,  5.14s/it]2026-04-12 01:53:07,199 | INFO | train_grpo_train | metrics_logged mode=train step=2344

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.9e-06, 'num_tokens': 5305230.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00038092854083515704, 'sampling/sampling_logp_difference/max': 0.01068283524364233, 'sampling/importance_sampling_ratio/min': 0.989374041557312, 'sampling/importance_sampling_ratio/mean': 1.000275731086731, 'sampling/importance_sampling_ratio/max': 1.0104482173919678, 'entropy': 0.0038317264115903527, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 71%|███████   | 2344/3300 [3:16:16<1:21:52,  5.14s/it]INFO 04-12 01:53:07 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2345/3300 [3:16:20<1:19:44,  5.01s/it]2026-04-12 01:53:11,920 | INFO | train_grpo_train | metrics_logged mode=train step=2345

                                                       
{'loss': 0.0324, 'grad_norm': 9.732110023498535, 'learning_rate': 2.896969696969697e-06, 'num_tokens': 5306953.0, 'completions/mean_length': 64.375, 'completions/min_length': 56.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7194586992263794, 'rewards/meter/std': 0.26375213265419006, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.7001276016235352, 'rewards/total_composite/std': 0.2649116516113281, 'reward': 0.7001276016235352, 'reward_std': 0.26491162180900574, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06900904327630997, 'sampling/sampling_logp_difference/max': 2.044987678527832, 'sampling/importance_sampling_ratio/min': 0.12938177585601807, 'sampling/importance_sampling_ratio/mean': 1.0110507011413574, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3568100146949291, 'clip_ratio/low_mean': 0.023096519173122942, 'clip_ratio/low_min': 0.023096519173122942, 'clip_ratio/high_mean': 0.029499810189008713, 'clip_ratio/high_max': 0.029499810189008713, 'clip_ratio/region_mean': 0.052596329362131655, 'reward_total_mean': 0.7001276016235352, 'reward_meter_mean': 0.7194586992263794, 'reward_meter_std': 0.26375213265419006, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.7001276016235352, 'reward_total_composite_std': 0.2649116516113281, 'epoch': 0.09}

 71%|███████   | 2345/3300 [3:16:20<1:19:44,  5.01s/it]INFO 04-12 01:53:12 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2346/3300 [3:16:27<1:28:53,  5.59s/it]2026-04-12 01:53:18,868 | INFO | train_grpo_train | metrics_logged mode=train step=2346

                                                       
{'loss': -0.0015, 'grad_norm': 3.087636947631836, 'learning_rate': 2.893939393939394e-06, 'num_tokens': 5310264.0, 'completions/mean_length': 237.875, 'completions/min_length': 228.0, 'completions/max_length': 248.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 237.875, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 248.0, 'rewards/meter/mean': 0.9981396794319153, 'rewards/meter/std': 0.0003048716171178967, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.942307710647583, 'rewards/repeat_penalty/std': 0.07962294667959213, 'rewards/total_composite/mean': 0.9405449032783508, 'rewards/total_composite/std': 0.07935542613267899, 'reward': 0.9405449032783508, 'reward_std': 0.07935541868209839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0508662573993206, 'sampling/sampling_logp_difference/max': 5.533112049102783, 'sampling/importance_sampling_ratio/min': 0.003953665494918823, 'sampling/importance_sampling_ratio/mean': 1.0058072805404663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4049236960709095, 'clip_ratio/low_mean': 0.014375669648870826, 'clip_ratio/low_min': 0.014375669648870826, 'clip_ratio/high_mean': 0.02505582571029663, 'clip_ratio/high_max': 0.02505582571029663, 'clip_ratio/region_mean': 0.03943149535916746, 'reward_total_mean': 0.9405449032783508, 'reward_meter_mean': 0.9981396794319153, 'reward_meter_std': 0.0003048716171178967, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.942307710647583, 'reward_repeat_penalty_std': 0.07962294667959213, 'reward_total_composite_mean': 0.9405449032783508, 'reward_total_composite_std': 0.07935542613267899, 'epoch': 0.09}

 71%|███████   | 2346/3300 [3:16:27<1:28:53,  5.59s/it]INFO 04-12 01:53:19 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2347/3300 [3:16:32<1:26:54,  5.47s/it]2026-04-12 01:53:24,060 | INFO | train_grpo_train | metrics_logged mode=train step=2347

                                                       
{'loss': -0.0006, 'grad_norm': 3.2611429691314697, 'learning_rate': 2.8909090909090907e-06, 'num_tokens': 5312388.0, 'completions/mean_length': 100.5, 'completions/min_length': 97.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.5, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9978408217430115, 'rewards/meter/std': 0.0007257388206198812, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978408217430115, 'rewards/total_composite/std': 0.0007257388206198812, 'reward': 0.9978408217430115, 'reward_std': 0.0007257444667629898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.056275609880685806, 'sampling/sampling_logp_difference/max': 2.3127644062042236, 'sampling/importance_sampling_ratio/min': 0.09898723661899567, 'sampling/importance_sampling_ratio/mean': 1.0071477890014648, 'sampling/importance_sampling_ratio/max': 1.796935796737671, 'entropy': 0.4347502291202545, 'clip_ratio/low_mean': 0.017315812641754746, 'clip_ratio/low_min': 0.017315812641754746, 'clip_ratio/high_mean': 0.031063538044691086, 'clip_ratio/high_max': 0.031063538044691086, 'clip_ratio/region_mean': 0.04837935068644583, 'reward_total_mean': 0.9978408217430115, 'reward_meter_mean': 0.9978408217430115, 'reward_meter_std': 0.0007257388206198812, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978408217430115, 'reward_total_composite_std': 0.0007257388206198812, 'epoch': 0.09}

 71%|███████   | 2347/3300 [3:16:33<1:26:54,  5.47s/it]INFO 04-12 01:53:24 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2348/3300 [3:16:38<1:29:07,  5.62s/it]2026-04-12 01:53:30,021 | INFO | train_grpo_train | metrics_logged mode=train step=2348

                                                       
{'loss': -0.0016, 'grad_norm': 2.218722105026245, 'learning_rate': 2.8878787878787884e-06, 'num_tokens': 5315212.0, 'completions/mean_length': 165.0, 'completions/min_length': 163.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.0, 'completions/min_terminated_length': 163.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9987882375717163, 'rewards/meter/std': 0.0007174185593612492, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.930555522441864, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.9294023513793945, 'rewards/total_composite/std': 0.05695149675011635, 'reward': 0.9294023513793945, 'reward_std': 0.05695149302482605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018433764576911926, 'sampling/sampling_logp_difference/max': 1.2507028579711914, 'sampling/importance_sampling_ratio/min': 0.28630349040031433, 'sampling/importance_sampling_ratio/mean': 1.0003210306167603, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10382581874728203, 'clip_ratio/low_mean': 0.00988566328305751, 'clip_ratio/low_min': 0.00988566328305751, 'clip_ratio/high_mean': 0.003007704159244895, 'clip_ratio/high_max': 0.003007704159244895, 'clip_ratio/region_mean': 0.012893367442302406, 'reward_total_mean': 0.9294023513793945, 'reward_meter_mean': 0.9987882375717163, 'reward_meter_std': 0.0007174185593612492, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.930555522441864, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.9294023513793945, 'reward_total_composite_std': 0.05695149675011635, 'epoch': 0.09}

 71%|███████   | 2348/3300 [3:16:39<1:29:07,  5.62s/it]INFO 04-12 01:53:30 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2349/3300 [3:16:43<1:25:49,  5.41s/it]2026-04-12 01:53:34,953 | INFO | train_grpo_train | metrics_logged mode=train step=2349

                                                       
{'loss': -0.0002, 'grad_norm': 2.1028835773468018, 'learning_rate': 2.884848484848485e-06, 'num_tokens': 5317189.0, 'completions/mean_length': 83.125, 'completions/min_length': 83.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.125, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9964075684547424, 'rewards/meter/std': 0.0002765149693004787, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9964075684547424, 'rewards/total_composite/std': 0.0002765149693004787, 'reward': 0.9964075684547424, 'reward_std': 0.0002765233220998198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014425117522478104, 'sampling/sampling_logp_difference/max': 0.9964022636413574, 'sampling/importance_sampling_ratio/min': 0.3692053556442261, 'sampling/importance_sampling_ratio/mean': 1.00690495967865, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04944844264537096, 'clip_ratio/low_mean': 0.013554216362535954, 'clip_ratio/low_min': 0.013554216362535954, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.013554216362535954, 'reward_total_mean': 0.9964075684547424, 'reward_meter_mean': 0.9964075684547424, 'reward_meter_std': 0.0002765149693004787, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9964075684547424, 'reward_total_composite_std': 0.0002765149693004787, 'epoch': 0.09}

 71%|███████   | 2349/3300 [3:16:43<1:25:49,  5.41s/it]INFO 04-12 01:53:35 [block_pool.py:378] Successfully reset prefix cache

 71%|███████   | 2350/3300 [3:16:50<1:30:47,  5.73s/it]2026-04-12 01:53:41,434 | INFO | train_grpo_train | metrics_logged mode=train step=2350

                                                       
{'loss': -0.0123, 'grad_norm': 2.5682480335235596, 'learning_rate': 2.8818181818181824e-06, 'num_tokens': 5320279.0, 'completions/mean_length': 193.25, 'completions/min_length': 185.0, 'completions/max_length': 203.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 193.25, 'completions/min_terminated_length': 185.0, 'completions/max_terminated_length': 203.0, 'rewards/meter/mean': 0.9980719089508057, 'rewards/meter/std': 0.00038988571031950414, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9659091234207153, 'rewards/repeat_penalty/std': 0.06763853132724762, 'rewards/total_composite/mean': 0.9640417098999023, 'rewards/total_composite/std': 0.06744071841239929, 'reward': 0.9640417098999023, 'reward_std': 0.06744074076414108, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05063340812921524, 'sampling/sampling_logp_difference/max': 1.2536392211914062, 'sampling/importance_sampling_ratio/min': 0.2854640483856201, 'sampling/importance_sampling_ratio/mean': 1.0081247091293335, 'sampling/importance_sampling_ratio/max': 1.9953590631484985, 'entropy': 0.40191996842622757, 'clip_ratio/low_mean': 0.004700823919847608, 'clip_ratio/low_min': 0.004700823919847608, 'clip_ratio/high_mean': 0.03450171835720539, 'clip_ratio/high_max': 0.03450171835720539, 'clip_ratio/region_mean': 0.039202542277053, 'reward_total_mean': 0.9640417098999023, 'reward_meter_mean': 0.9980719089508057, 'reward_meter_std': 0.00038988571031950414, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9659091234207153, 'reward_repeat_penalty_std': 0.06763853132724762, 'reward_total_composite_mean': 0.9640417098999023, 'reward_total_composite_std': 0.06744071841239929, 'epoch': 0.09}

 71%|███████   | 2350/3300 [3:16:50<1:30:47,  5.73s/it]INFO 04-12 01:53:41 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:05<00:31,  2.89s/it]

 23%|██▎       | 3/13 [00:08<00:30,  3.02s/it]

 31%|███       | 4/13 [00:14<00:34,  3.79s/it]

 38%|███▊      | 5/13 [00:18<00:32,  4.04s/it]

 46%|████▌     | 6/13 [00:23<00:31,  4.43s/it]

 54%|█████▍    | 7/13 [00:28<00:27,  4.61s/it]

 62%|██████▏   | 8/13 [00:34<00:25,  5.07s/it]

 69%|██████▉   | 9/13 [00:41<00:22,  5.52s/it]

 77%|███████▋  | 10/13 [00:45<00:15,  5.19s/it]

 85%|████████▍ | 11/13 [00:51<00:10,  5.20s/it]

 92%|█████████▏| 12/13 [00:57<00:05,  5.52s/it]

100%|██████████| 13/13 [01:02<00:00,  5.37s/it]2026-04-12 01:54:49,201 | INFO | train_grpo_train | metrics_logged mode=eval step=2350

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 67.6666, 'eval_samples_per_second': 1.537, 'eval_steps_per_second': 0.192, 'eval_num_tokens': 5320279.0, 'eval_completions/mean_length': 194.6346153846154, 'eval_completions/min_length': 61.30769230769231, 'eval_completions/max_length': 349.3076923076923, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 194.6346153846154, 'eval_completions/min_terminated_length': 61.30769230769231, 'eval_completions/max_terminated_length': 349.3076923076923, 'eval_rewards/meter/mean': 0.7545687923064599, 'eval_rewards/meter/std': 0.3586408014480884, 'eval_rewards/count_adherence/mean': 0.9230444156206571, 'eval_rewards/count_adherence/std': 0.10953088907095102, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.885237611257113, 'eval_rewards/repeat_penalty/std': 0.1111061366704794, 'eval_rewards/total_composite/mean': 0.6291824304140531, 'eval_rewards/total_composite/std': 0.33865179121494293, 'eval_reward': 0.6291824304140531, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.021095395159835998, 'eval_sampling/sampling_logp_difference/max': 1.0696188119741588, 'eval_sampling/importance_sampling_ratio/min': 0.3466451099285713, 'eval_sampling/importance_sampling_ratio/mean': 1.0047486470295832, 'eval_sampling/importance_sampling_ratio/max': 1.4322827320832472, 'eval_entropy': 0.21742234894862542, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.6291824304140531, 'eval_reward_meter_mean': 0.7545687923064599, 'eval_reward_meter_std': 0.3586408014480884, 'eval_reward_count_adherence_mean': 0.9230444156206571, 'eval_reward_count_adherence_std': 0.10953088907095102, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.885237611257113, 'eval_reward_repeat_penalty_std': 0.1111061366704794, 'eval_reward_total_composite_mean': 0.6291824304140531, 'eval_reward_total_composite_std': 0.33865179121494293, 'epoch': 0.09}

 71%|███████   | 2350/3300 [3:17:58<1:30:47,  5.73s/it]

100%|██████████| 13/13 [01:02<00:00,  5.37s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 71%|███████   | 2351/3300 [3:18:05<6:58:39, 26.47s/it]2026-04-12 01:54:56,287 | INFO | train_grpo_train | metrics_logged mode=train step=2351

                                                       
{'loss': 0.0052, 'grad_norm': 4.469809532165527, 'learning_rate': 2.8787878787878793e-06, 'num_tokens': 5322067.0, 'completions/mean_length': 75.5, 'completions/min_length': 73.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.998935878276825, 'rewards/meter/std': 0.0006324206478893757, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998935878276825, 'rewards/total_composite/std': 0.0006324206478893757, 'reward': 0.998935878276825, 'reward_std': 0.0006324192509055138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030197838321328163, 'sampling/sampling_logp_difference/max': 1.0953636169433594, 'sampling/importance_sampling_ratio/min': 0.33441799879074097, 'sampling/importance_sampling_ratio/mean': 1.0036641359329224, 'sampling/importance_sampling_ratio/max': 1.4723999500274658, 'entropy': 0.20778764225542545, 'clip_ratio/low_mean': 0.008248508209362626, 'clip_ratio/low_min': 0.008248508209362626, 'clip_ratio/high_mean': 0.01327340246643871, 'clip_ratio/high_max': 0.01327340246643871, 'clip_ratio/region_mean': 0.021521910675801337, 'reward_total_mean': 0.998935878276825, 'reward_meter_mean': 0.998935878276825, 'reward_meter_std': 0.0006324206478893757, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998935878276825, 'reward_total_composite_std': 0.0006324206478893757, 'epoch': 0.09}

 71%|███████   | 2351/3300 [3:18:05<6:58:39, 26.47s/it]INFO 04-12 01:54:56 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2352/3300 [3:18:09<5:15:00, 19.94s/it]2026-04-12 01:55:00,982 | INFO | train_grpo_train | metrics_logged mode=train step=2352

                                                       
{'loss': -0.0028, 'grad_norm': 2.229356288909912, 'learning_rate': 2.875757575757576e-06, 'num_tokens': 5323867.0, 'completions/mean_length': 55.0, 'completions/min_length': 55.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9961405992507935, 'rewards/meter/std': 0.0004942560917697847, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9961405992507935, 'rewards/total_composite/std': 0.0004942560917697847, 'reward': 0.9961405992507935, 'reward_std': 0.0004942500381730497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006679588928818703, 'sampling/sampling_logp_difference/max': 1.3542317152023315, 'sampling/importance_sampling_ratio/min': 0.2581455409526825, 'sampling/importance_sampling_ratio/mean': 0.9985120892524719, 'sampling/importance_sampling_ratio/max': 1.1960804462432861, 'entropy': 0.018420133623294532, 'clip_ratio/low_mean': 0.0022727272007614374, 'clip_ratio/low_min': 0.0022727272007614374, 'clip_ratio/high_mean': 0.004545454401522875, 'clip_ratio/high_max': 0.004545454401522875, 'clip_ratio/region_mean': 0.006818181602284312, 'reward_total_mean': 0.9961405992507935, 'reward_meter_mean': 0.9961405992507935, 'reward_meter_std': 0.0004942560917697847, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9961405992507935, 'reward_total_composite_std': 0.0004942560917697847, 'epoch': 0.09}

 71%|███████▏  | 2352/3300 [3:18:09<5:15:00, 19.94s/it]INFO 04-12 01:55:01 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2353/3300 [3:18:19<4:26:43, 16.90s/it]2026-04-12 01:55:10,789 | INFO | train_grpo_train | metrics_logged mode=train step=2353

                                                       
{'loss': -0.0155, 'grad_norm': 1.3235185146331787, 'learning_rate': 2.872727272727273e-06, 'num_tokens': 5329206.0, 'completions/mean_length': 421.375, 'completions/min_length': 398.0, 'completions/max_length': 455.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 421.375, 'completions/min_terminated_length': 398.0, 'completions/max_terminated_length': 455.0, 'rewards/meter/mean': 0.9977288246154785, 'rewards/meter/std': 0.0008687536465004086, 'rewards/count_adherence/mean': 0.6470588445663452, 'rewards/count_adherence/std': 0.03144249692559242, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8467603325843811, 'rewards/repeat_penalty/std': 0.08341042697429657, 'rewards/total_composite/mean': 0.5467392206192017, 'rewards/total_composite/std': 0.06185723468661308, 'reward': 0.5467392206192017, 'reward_std': 0.061857227236032486, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02322297915816307, 'sampling/sampling_logp_difference/max': 1.9221699237823486, 'sampling/importance_sampling_ratio/min': 0.1462891846895218, 'sampling/importance_sampling_ratio/mean': 1.0037055015563965, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19907046295702457, 'clip_ratio/low_mean': 0.01009113050531596, 'clip_ratio/low_min': 0.01009113050531596, 'clip_ratio/high_mean': 0.007577877026051283, 'clip_ratio/high_max': 0.007577877026051283, 'clip_ratio/region_mean': 0.017669007531367242, 'reward_total_mean': 0.5467392206192017, 'reward_meter_mean': 0.9977288246154785, 'reward_meter_std': 0.0008687536465004086, 'reward_count_adherence_mean': 0.6470588445663452, 'reward_count_adherence_std': 0.03144249692559242, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8467603325843811, 'reward_repeat_penalty_std': 0.08341042697429657, 'reward_total_composite_mean': 0.5467392206192017, 'reward_total_composite_std': 0.06185723468661308, 'epoch': 0.09}

 71%|███████▏  | 2353/3300 [3:18:19<4:26:43, 16.90s/it]INFO 04-12 01:55:11 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2354/3300 [3:18:24<3:28:39, 13.23s/it]2026-04-12 01:55:15,477 | INFO | train_grpo_train | metrics_logged mode=train step=2354

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.86969696969697e-06, 'num_tokens': 5330886.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003920203307643533, 'sampling/sampling_logp_difference/max': 0.0068152910098433495, 'sampling/importance_sampling_ratio/min': 1.0, 'sampling/importance_sampling_ratio/mean': 1.0003925561904907, 'sampling/importance_sampling_ratio/max': 1.0068386793136597, 'entropy': 0.0035217964032199234, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.09}

 71%|███████▏  | 2354/3300 [3:18:24<3:28:39, 13.23s/it]INFO 04-12 01:55:15 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2355/3300 [3:18:32<3:02:08, 11.56s/it]2026-04-12 01:55:23,150 | INFO | train_grpo_train | metrics_logged mode=train step=2355

                                                       
{'loss': 0.0071, 'grad_norm': 2.712887763977051, 'learning_rate': 2.866666666666667e-06, 'num_tokens': 5333484.0, 'completions/mean_length': 168.75, 'completions/min_length': 165.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.75, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.9781259298324585, 'rewards/meter/std': 0.03459608182311058, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611111640930176, 'rewards/repeat_penalty/std': 0.07856741547584534, 'rewards/total_composite/mean': 0.8432746529579163, 'rewards/total_composite/std': 0.09240440279245377, 'reward': 0.8432746529579163, 'reward_std': 0.09240438789129257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027000876143574715, 'sampling/sampling_logp_difference/max': 1.2758288383483887, 'sampling/importance_sampling_ratio/min': 0.2791994512081146, 'sampling/importance_sampling_ratio/mean': 1.0031359195709229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1752516869455576, 'clip_ratio/low_mean': 0.008841085364110768, 'clip_ratio/low_min': 0.008841085364110768, 'clip_ratio/high_mean': 0.017101168166846037, 'clip_ratio/high_max': 0.017101168166846037, 'clip_ratio/region_mean': 0.025942253530956805, 'reward_total_mean': 0.8432746529579163, 'reward_meter_mean': 0.9781259298324585, 'reward_meter_std': 0.03459608182311058, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8611111640930176, 'reward_repeat_penalty_std': 0.07856741547584534, 'reward_total_composite_mean': 0.8432746529579163, 'reward_total_composite_std': 0.09240440279245377, 'epoch': 0.09}

 71%|███████▏  | 2355/3300 [3:18:32<3:02:08, 11.56s/it]INFO 04-12 01:55:23 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2356/3300 [3:18:37<2:30:49,  9.59s/it]2026-04-12 01:55:28,125 | INFO | train_grpo_train | metrics_logged mode=train step=2356

                                                       
{'loss': 0.034, 'grad_norm': 12.347071647644043, 'learning_rate': 2.863636363636364e-06, 'num_tokens': 5335199.0, 'completions/mean_length': 44.375, 'completions/min_length': 38.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9444069266319275, 'rewards/meter/std': 0.008035099133849144, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9444069266319275, 'rewards/total_composite/std': 0.008035099133849144, 'reward': 0.9444069266319275, 'reward_std': 0.00803509633988142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07534459978342056, 'sampling/sampling_logp_difference/max': 2.617283821105957, 'sampling/importance_sampling_ratio/min': 0.3516010642051697, 'sampling/importance_sampling_ratio/mean': 1.0038548707962036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32474897243082523, 'clip_ratio/low_mean': 0.030935976887121797, 'clip_ratio/low_min': 0.030935976887121797, 'clip_ratio/high_mean': 0.01955222897231579, 'clip_ratio/high_max': 0.01955222897231579, 'clip_ratio/region_mean': 0.050488205859437585, 'reward_total_mean': 0.9444069266319275, 'reward_meter_mean': 0.9444069266319275, 'reward_meter_std': 0.008035099133849144, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9444069266319275, 'reward_total_composite_std': 0.008035099133849144, 'epoch': 0.09}

 71%|███████▏  | 2356/3300 [3:18:37<2:30:49,  9.59s/it]INFO 04-12 01:55:28 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2357/3300 [3:18:42<2:09:57,  8.27s/it]2026-04-12 01:55:33,314 | INFO | train_grpo_train | metrics_logged mode=train step=2357

                                                       
{'loss': -0.0003, 'grad_norm': 2.1621651649475098, 'learning_rate': 2.860606060606061e-06, 'num_tokens': 5337204.0, 'completions/mean_length': 78.625, 'completions/min_length': 77.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9971969127655029, 'rewards/meter/std': 0.0005990342469885945, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971969127655029, 'rewards/total_composite/std': 0.0005990342469885945, 'reward': 0.9971969127655029, 'reward_std': 0.0005990189965814352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021038906648755074, 'sampling/sampling_logp_difference/max': 1.367281436920166, 'sampling/importance_sampling_ratio/min': 0.2547987103462219, 'sampling/importance_sampling_ratio/mean': 0.9987841844558716, 'sampling/importance_sampling_ratio/max': 1.4412025213241577, 'entropy': 0.12107675615698099, 'clip_ratio/low_mean': 0.00634939968585968, 'clip_ratio/low_min': 0.00634939968585968, 'clip_ratio/high_mean': 0.009516005055047572, 'clip_ratio/high_max': 0.009516005055047572, 'clip_ratio/region_mean': 0.015865404740907252, 'reward_total_mean': 0.9971969127655029, 'reward_meter_mean': 0.9971969127655029, 'reward_meter_std': 0.0005990342469885945, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971969127655029, 'reward_total_composite_std': 0.0005990342469885945, 'epoch': 0.09}

 71%|███████▏  | 2357/3300 [3:18:42<2:09:57,  8.27s/it]INFO 04-12 01:55:33 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2358/3300 [3:18:49<2:06:55,  8.08s/it]2026-04-12 01:55:40,957 | INFO | train_grpo_train | metrics_logged mode=train step=2358

                                                       
{'loss': -0.0096, 'grad_norm': 3.5978167057037354, 'learning_rate': 2.857575757575758e-06, 'num_tokens': 5339049.0, 'completions/mean_length': 73.625, 'completions/min_length': 71.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9990395307540894, 'rewards/meter/std': 0.00045778730418533087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990395307540894, 'rewards/total_composite/std': 0.00045778730418533087, 'reward': 0.9990395307540894, 'reward_std': 0.0004578051157295704, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0385035015642643, 'sampling/sampling_logp_difference/max': 1.5631723403930664, 'sampling/importance_sampling_ratio/min': 0.2094704955816269, 'sampling/importance_sampling_ratio/mean': 0.999204158782959, 'sampling/importance_sampling_ratio/max': 1.805698275566101, 'entropy': 0.25467516854405403, 'clip_ratio/low_mean': 0.02065259451046586, 'clip_ratio/low_min': 0.02065259451046586, 'clip_ratio/high_mean': 0.008355856058187783, 'clip_ratio/high_max': 0.008355856058187783, 'clip_ratio/region_mean': 0.029008450568653643, 'reward_total_mean': 0.9990395307540894, 'reward_meter_mean': 0.9990395307540894, 'reward_meter_std': 0.00045778730418533087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990395307540894, 'reward_total_composite_std': 0.00045778730418533087, 'epoch': 0.09}

 71%|███████▏  | 2358/3300 [3:18:49<2:06:55,  8.08s/it]INFO 04-12 01:55:41 [block_pool.py:378] Successfully reset prefix cache

 71%|███████▏  | 2359/3300 [3:18:54<1:52:34,  7.18s/it]2026-04-12 01:55:46,028 | INFO | train_grpo_train | metrics_logged mode=train step=2359

                                                       
{'loss': 0.015, 'grad_norm': 6.299208164215088, 'learning_rate': 2.8545454545454548e-06, 'num_tokens': 5340832.0, 'completions/mean_length': 66.875, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9785959720611572, 'rewards/meter/std': 0.018975911661982536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9785959720611572, 'rewards/total_composite/std': 0.018975911661982536, 'reward': 0.9785959720611572, 'reward_std': 0.018975909799337387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0360257662832737, 'sampling/sampling_logp_difference/max': 1.296480655670166, 'sampling/importance_sampling_ratio/min': 0.27349260449409485, 'sampling/importance_sampling_ratio/mean': 1.0064629316329956, 'sampling/importance_sampling_ratio/max': 1.6261426210403442, 'entropy': 0.2618895582854748, 'clip_ratio/low_mean': 0.007326300255954266, 'clip_ratio/low_min': 0.007326300255954266, 'clip_ratio/high_mean': 0.016904115793295205, 'clip_ratio/high_max': 0.016904115793295205, 'clip_ratio/region_mean': 0.02423041604924947, 'reward_total_mean': 0.9785959720611572, 'reward_meter_mean': 0.9785959720611572, 'reward_meter_std': 0.018975911661982536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9785959720611572, 'reward_total_composite_std': 0.018975911661982536, 'epoch': 0.09}

 71%|███████▏  | 2359/3300 [3:18:55<1:52:34,  7.18s/it]INFO 04-12 01:55:46 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2360/3300 [3:18:59<1:41:43,  6.49s/it]2026-04-12 01:55:50,916 | INFO | train_grpo_train | metrics_logged mode=train step=2360

                                                       
{'loss': 0.0207, 'grad_norm': 5.956533432006836, 'learning_rate': 2.8515151515151516e-06, 'num_tokens': 5342577.0, 'completions/mean_length': 67.125, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9699335098266602, 'rewards/meter/std': 0.04688005894422531, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9699335098266602, 'rewards/total_composite/std': 0.04688005894422531, 'reward': 0.9699335098266602, 'reward_std': 0.04688006266951561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03932439908385277, 'sampling/sampling_logp_difference/max': 1.0124170780181885, 'sampling/importance_sampling_ratio/min': 0.36333972215652466, 'sampling/importance_sampling_ratio/mean': 1.0127477645874023, 'sampling/importance_sampling_ratio/max': 1.829169750213623, 'entropy': 0.30276480689644814, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.03938204434234649, 'clip_ratio/high_max': 0.03938204434234649, 'clip_ratio/region_mean': 0.04116775863803923, 'reward_total_mean': 0.9699335098266602, 'reward_meter_mean': 0.9699335098266602, 'reward_meter_std': 0.04688005894422531, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9699335098266602, 'reward_total_composite_std': 0.04688005894422531, 'epoch': 0.09}

 72%|███████▏  | 2360/3300 [3:18:59<1:41:43,  6.49s/it]INFO 04-12 01:55:51 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2361/3300 [3:19:06<1:42:45,  6.57s/it]2026-04-12 01:55:57,680 | INFO | train_grpo_train | metrics_logged mode=train step=2361

                                                       
{'loss': 0.0078, 'grad_norm': 1.9891157150268555, 'learning_rate': 2.848484848484849e-06, 'num_tokens': 5345741.0, 'completions/mean_length': 214.5, 'completions/min_length': 210.0, 'completions/max_length': 219.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 214.5, 'completions/min_terminated_length': 210.0, 'completions/max_terminated_length': 219.0, 'rewards/meter/mean': 0.999156653881073, 'rewards/meter/std': 0.00014262554759625345, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8409091234207153, 'rewards/repeat_penalty/std': 0.10590588301420212, 'rewards/total_composite/mean': 0.8401910066604614, 'rewards/total_composite/std': 0.10572723299264908, 'reward': 0.8401910066604614, 'reward_std': 0.10572723299264908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021222621202468872, 'sampling/sampling_logp_difference/max': 1.204646110534668, 'sampling/importance_sampling_ratio/min': 0.29979807138442993, 'sampling/importance_sampling_ratio/mean': 1.002423882484436, 'sampling/importance_sampling_ratio/max': 1.6807578802108765, 'entropy': 0.19607393071055412, 'clip_ratio/low_mean': 0.009289931389503181, 'clip_ratio/low_min': 0.009289931389503181, 'clip_ratio/high_mean': 0.00470314035192132, 'clip_ratio/high_max': 0.00470314035192132, 'clip_ratio/region_mean': 0.013993071741424501, 'reward_total_mean': 0.8401910066604614, 'reward_meter_mean': 0.999156653881073, 'reward_meter_std': 0.00014262554759625345, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8409091234207153, 'reward_repeat_penalty_std': 0.10590588301420212, 'reward_total_composite_mean': 0.8401910066604614, 'reward_total_composite_std': 0.10572723299264908, 'epoch': 0.09}

 72%|███████▏  | 2361/3300 [3:19:06<1:42:45,  6.57s/it]INFO 04-12 01:55:57 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2362/3300 [3:19:11<1:33:04,  5.95s/it]2026-04-12 01:56:02,181 | INFO | train_grpo_train | metrics_logged mode=train step=2362

                                                       
{'loss': -0.0082, 'grad_norm': 3.921265125274658, 'learning_rate': 2.8454545454545457e-06, 'num_tokens': 5347168.0, 'completions/mean_length': 40.375, 'completions/min_length': 39.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9963768124580383, 'rewards/meter/std': 0.0008168795611709356, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9963768124580383, 'rewards/total_composite/std': 0.0008168795611709356, 'reward': 0.9963768124580383, 'reward_std': 0.0008168675121851265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018412932753562927, 'sampling/sampling_logp_difference/max': 0.573250412940979, 'sampling/importance_sampling_ratio/min': 0.5636902451515198, 'sampling/importance_sampling_ratio/mean': 0.9972823858261108, 'sampling/importance_sampling_ratio/max': 1.4037843942642212, 'entropy': 0.08906339993700385, 'clip_ratio/low_mean': 0.0032051282469183207, 'clip_ratio/low_min': 0.0032051282469183207, 'clip_ratio/high_mean': 0.006097560748457909, 'clip_ratio/high_max': 0.006097560748457909, 'clip_ratio/region_mean': 0.00930268899537623, 'reward_total_mean': 0.9963768124580383, 'reward_meter_mean': 0.9963768124580383, 'reward_meter_std': 0.0008168795611709356, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9963768124580383, 'reward_total_composite_std': 0.0008168795611709356, 'epoch': 0.09}

 72%|███████▏  | 2362/3300 [3:19:11<1:33:04,  5.95s/it]INFO 04-12 01:56:02 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2363/3300 [3:19:16<1:31:16,  5.85s/it]2026-04-12 01:56:07,800 | INFO | train_grpo_train | metrics_logged mode=train step=2363

                                                       
{'loss': -0.0237, 'grad_norm': 4.108429908752441, 'learning_rate': 2.8424242424242425e-06, 'num_tokens': 5348939.0, 'completions/mean_length': 74.375, 'completions/min_length': 70.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9990479946136475, 'rewards/meter/std': 0.0008222111500799656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990479946136475, 'rewards/total_composite/std': 0.0008222111500799656, 'reward': 0.9990479946136475, 'reward_std': 0.0008222072501666844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03708229959011078, 'sampling/sampling_logp_difference/max': 0.9699416160583496, 'sampling/importance_sampling_ratio/min': 0.3791051506996155, 'sampling/importance_sampling_ratio/mean': 0.9991539120674133, 'sampling/importance_sampling_ratio/max': 1.825773000717163, 'entropy': 0.2678355220705271, 'clip_ratio/low_mean': 0.00890342053025961, 'clip_ratio/low_min': 0.00890342053025961, 'clip_ratio/high_mean': 0.021532694692723453, 'clip_ratio/high_max': 0.021532694692723453, 'clip_ratio/region_mean': 0.030436115222983062, 'reward_total_mean': 0.9990479946136475, 'reward_meter_mean': 0.9990479946136475, 'reward_meter_std': 0.0008222111500799656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990479946136475, 'reward_total_composite_std': 0.0008222111500799656, 'epoch': 0.09}

 72%|███████▏  | 2363/3300 [3:19:16<1:31:16,  5.85s/it]INFO 04-12 01:56:08 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2364/3300 [3:19:22<1:33:05,  5.97s/it]2026-04-12 01:56:14,065 | INFO | train_grpo_train | metrics_logged mode=train step=2364

                                                       
{'loss': 0.0027, 'grad_norm': 1.7912240028381348, 'learning_rate': 2.83939393939394e-06, 'num_tokens': 5351730.0, 'completions/mean_length': 138.875, 'completions/min_length': 138.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.875, 'completions/min_terminated_length': 138.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.9963522553443909, 'rewards/meter/std': 0.0003084685595240444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8333333730697632, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.8302946090698242, 'rewards/total_composite/std': 0.0591915100812912, 'reward': 0.8302946090698242, 'reward_std': 0.05919152498245239, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015195811167359352, 'sampling/sampling_logp_difference/max': 1.2972681522369385, 'sampling/importance_sampling_ratio/min': 0.2732773423194885, 'sampling/importance_sampling_ratio/mean': 0.9980447292327881, 'sampling/importance_sampling_ratio/max': 1.877577304840088, 'entropy': 0.07086402643471956, 'clip_ratio/low_mean': 0.0054152331431396306, 'clip_ratio/low_min': 0.0054152331431396306, 'clip_ratio/high_mean': 0.01258992834482342, 'clip_ratio/high_max': 0.01258992834482342, 'clip_ratio/region_mean': 0.01800516148796305, 'reward_total_mean': 0.8302946090698242, 'reward_meter_mean': 0.9963522553443909, 'reward_meter_std': 0.0003084685595240444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8333333730697632, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.8302946090698242, 'reward_total_composite_std': 0.0591915100812912, 'epoch': 0.09}

 72%|███████▏  | 2364/3300 [3:19:23<1:33:05,  5.97s/it]INFO 04-12 01:56:14 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2365/3300 [3:19:27<1:27:55,  5.64s/it]2026-04-12 01:56:18,907 | INFO | train_grpo_train | metrics_logged mode=train step=2365

                                                       
{'loss': -0.0049, 'grad_norm': 0.9692533612251282, 'learning_rate': 2.8363636363636366e-06, 'num_tokens': 5353652.0, 'completions/mean_length': 73.25, 'completions/min_length': 72.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.25, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9992170333862305, 'rewards/meter/std': 0.00020559101540129632, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992170333862305, 'rewards/total_composite/std': 0.00020559101540129632, 'reward': 0.9992170333862305, 'reward_std': 0.00020559433323796839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023841191083192825, 'sampling/sampling_logp_difference/max': 1.589078426361084, 'sampling/importance_sampling_ratio/min': 0.20411363244056702, 'sampling/importance_sampling_ratio/mean': 1.0054048299789429, 'sampling/importance_sampling_ratio/max': 1.8762344121932983, 'entropy': 0.20148388668894768, 'clip_ratio/low_mean': 0.010369744850322604, 'clip_ratio/low_min': 0.010369744850322604, 'clip_ratio/high_mean': 0.01530504459515214, 'clip_ratio/high_max': 0.01530504459515214, 'clip_ratio/region_mean': 0.025674789445474744, 'reward_total_mean': 0.9992170333862305, 'reward_meter_mean': 0.9992170333862305, 'reward_meter_std': 0.00020559101540129632, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992170333862305, 'reward_total_composite_std': 0.00020559101540129632, 'epoch': 0.09}

 72%|███████▏  | 2365/3300 [3:19:27<1:27:55,  5.64s/it]INFO 04-12 01:56:19 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2366/3300 [3:19:34<1:31:51,  5.90s/it]2026-04-12 01:56:25,406 | INFO | train_grpo_train | metrics_logged mode=train step=2366

                                                       
{'loss': 0.0089, 'grad_norm': 4.017276763916016, 'learning_rate': 2.8333333333333335e-06, 'num_tokens': 5356221.0, 'completions/mean_length': 132.125, 'completions/min_length': 129.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.125, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.9979519844055176, 'rewards/meter/std': 0.0010567301651462913, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9623122215270996, 'rewards/total_composite/std': 0.06602418422698975, 'reward': 0.9623122215270996, 'reward_std': 0.06602419167757034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05144790932536125, 'sampling/sampling_logp_difference/max': 10.138971328735352, 'sampling/importance_sampling_ratio/min': 3.95094248233363e-05, 'sampling/importance_sampling_ratio/mean': 1.0044370889663696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3485673386603594, 'clip_ratio/low_mean': 0.01210981048643589, 'clip_ratio/low_min': 0.01210981048643589, 'clip_ratio/high_mean': 0.03128110943362117, 'clip_ratio/high_max': 0.03128110943362117, 'clip_ratio/region_mean': 0.04339091992005706, 'reward_total_mean': 0.9623122215270996, 'reward_meter_mean': 0.9979519844055176, 'reward_meter_std': 0.0010567301651462913, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9623122215270996, 'reward_total_composite_std': 0.06602418422698975, 'epoch': 0.1}

 72%|███████▏  | 2366/3300 [3:19:34<1:31:51,  5.90s/it]INFO 04-12 01:56:25 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2367/3300 [3:19:40<1:32:49,  5.97s/it]2026-04-12 01:56:31,598 | INFO | train_grpo_train | metrics_logged mode=train step=2367

                                                       
{'loss': 0.0164, 'grad_norm': 1.9662439823150635, 'learning_rate': 2.8303030303030303e-06, 'num_tokens': 5359095.0, 'completions/mean_length': 154.25, 'completions/min_length': 147.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.25, 'completions/min_terminated_length': 147.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9776645302772522, 'rewards/meter/std': 0.028479592874646187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.930555522441864, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.9088918566703796, 'rewards/total_composite/std': 0.04527009651064873, 'reward': 0.9088918566703796, 'reward_std': 0.04527008906006813, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023603463545441628, 'sampling/sampling_logp_difference/max': 1.6123571395874023, 'sampling/importance_sampling_ratio/min': 0.19941700994968414, 'sampling/importance_sampling_ratio/mean': 1.0010682344436646, 'sampling/importance_sampling_ratio/max': 1.7584105730056763, 'entropy': 0.16689980775117874, 'clip_ratio/low_mean': 0.007995150808710605, 'clip_ratio/low_min': 0.007995150808710605, 'clip_ratio/high_mean': 0.009202249813824892, 'clip_ratio/high_max': 0.009202249813824892, 'clip_ratio/region_mean': 0.017197400622535497, 'reward_total_mean': 0.9088918566703796, 'reward_meter_mean': 0.9776645302772522, 'reward_meter_std': 0.028479592874646187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.930555522441864, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.9088918566703796, 'reward_total_composite_std': 0.04527009651064873, 'epoch': 0.1}

 72%|███████▏  | 2367/3300 [3:19:40<1:32:49,  5.97s/it]INFO 04-12 01:56:31 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2368/3300 [3:19:46<1:31:54,  5.92s/it]2026-04-12 01:56:37,335 | INFO | train_grpo_train | metrics_logged mode=train step=2368

                                                       
{'loss': -0.005, 'grad_norm': 3.341148853302002, 'learning_rate': 2.8272727272727275e-06, 'num_tokens': 5360933.0, 'completions/mean_length': 67.75, 'completions/min_length': 66.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.998802125453949, 'rewards/meter/std': 0.00023434955801349133, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998802125453949, 'rewards/total_composite/std': 0.00023434955801349133, 'reward': 0.998802125453949, 'reward_std': 0.00023435504408553243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036128342151641846, 'sampling/sampling_logp_difference/max': 0.9030179977416992, 'sampling/importance_sampling_ratio/min': 0.40534451603889465, 'sampling/importance_sampling_ratio/mean': 1.0024096965789795, 'sampling/importance_sampling_ratio/max': 1.526517629623413, 'entropy': 0.2899752501398325, 'clip_ratio/low_mean': 0.01125139684882015, 'clip_ratio/low_min': 0.01125139684882015, 'clip_ratio/high_mean': 0.02565775695256889, 'clip_ratio/high_max': 0.02565775695256889, 'clip_ratio/region_mean': 0.03690915380138904, 'reward_total_mean': 0.998802125453949, 'reward_meter_mean': 0.998802125453949, 'reward_meter_std': 0.00023434955801349133, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998802125453949, 'reward_total_composite_std': 0.00023434955801349133, 'epoch': 0.1}

 72%|███████▏  | 2368/3300 [3:19:46<1:31:54,  5.92s/it]INFO 04-12 01:56:37 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2369/3300 [3:19:51<1:26:36,  5.58s/it]2026-04-12 01:56:42,131 | INFO | train_grpo_train | metrics_logged mode=train step=2369

                                                       
{'loss': 0.0001, 'grad_norm': 1.1244981288909912, 'learning_rate': 2.8242424242424244e-06, 'num_tokens': 5362693.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9980802536010742, 'rewards/meter/std': 5.821814920636825e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980802536010742, 'rewards/total_composite/std': 5.821814920636825e-05, 'reward': 0.9980802536010742, 'reward_std': 5.822332968818955e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011381209827959538, 'sampling/sampling_logp_difference/max': 0.39493346214294434, 'sampling/importance_sampling_ratio/min': 0.673724889755249, 'sampling/importance_sampling_ratio/mean': 1.003217101097107, 'sampling/importance_sampling_ratio/max': 1.3787553310394287, 'entropy': 0.09162188414484262, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9980802536010742, 'reward_meter_mean': 0.9980802536010742, 'reward_meter_std': 5.821814920636825e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980802536010742, 'reward_total_composite_std': 5.821814920636825e-05, 'epoch': 0.1}

 72%|███████▏  | 2369/3300 [3:19:51<1:26:36,  5.58s/it]INFO 04-12 01:56:42 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2370/3300 [3:19:55<1:20:22,  5.19s/it]2026-04-12 01:56:46,390 | INFO | train_grpo_train | metrics_logged mode=train step=2370

                                                       
{'loss': 0.0382, 'grad_norm': 12.456171035766602, 'learning_rate': 2.821212121212121e-06, 'num_tokens': 5364201.0, 'completions/mean_length': 34.5, 'completions/min_length': 34.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9742438793182373, 'rewards/meter/std': 0.03374186530709267, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9742438793182373, 'rewards/total_composite/std': 0.03374186530709267, 'reward': 0.9742438793182373, 'reward_std': 0.03374185785651207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02670416608452797, 'sampling/sampling_logp_difference/max': 1.062800407409668, 'sampling/importance_sampling_ratio/min': 0.34548693895339966, 'sampling/importance_sampling_ratio/mean': 1.0038917064666748, 'sampling/importance_sampling_ratio/max': 1.4533801078796387, 'entropy': 0.21421233005821705, 'clip_ratio/low_mean': 0.013544891262426972, 'clip_ratio/low_min': 0.013544891262426972, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.013544891262426972, 'reward_total_mean': 0.9742438793182373, 'reward_meter_mean': 0.9742438793182373, 'reward_meter_std': 0.03374186530709267, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9742438793182373, 'reward_total_composite_std': 0.03374186530709267, 'epoch': 0.1}

 72%|███████▏  | 2370/3300 [3:19:55<1:20:22,  5.19s/it]INFO 04-12 01:56:46 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2371/3300 [3:20:00<1:18:03,  5.04s/it]2026-04-12 01:56:51,096 | INFO | train_grpo_train | metrics_logged mode=train step=2371

                                                       
{'loss': 0.0014, 'grad_norm': 1.3944817781448364, 'learning_rate': 2.818181818181818e-06, 'num_tokens': 5366225.0, 'completions/mean_length': 83.0, 'completions/min_length': 83.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.0, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9963405728340149, 'rewards/meter/std': 0.00012567358498927206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9963405728340149, 'rewards/total_composite/std': 0.00012567358498927206, 'reward': 0.9963405728340149, 'reward_std': 0.00012567655357997864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006072253920137882, 'sampling/sampling_logp_difference/max': 1.5704057216644287, 'sampling/importance_sampling_ratio/min': 0.20796078443527222, 'sampling/importance_sampling_ratio/mean': 1.0000357627868652, 'sampling/importance_sampling_ratio/max': 1.3552405834197998, 'entropy': 0.030938314041122794, 'clip_ratio/low_mean': 0.0015060240402817726, 'clip_ratio/low_min': 0.0015060240402817726, 'clip_ratio/high_mean': 0.0030120480805635452, 'clip_ratio/high_max': 0.0030120480805635452, 'clip_ratio/region_mean': 0.004518072120845318, 'reward_total_mean': 0.9963405728340149, 'reward_meter_mean': 0.9963405728340149, 'reward_meter_std': 0.00012567358498927206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9963405728340149, 'reward_total_composite_std': 0.00012567358498927206, 'epoch': 0.1}

 72%|███████▏  | 2371/3300 [3:20:00<1:18:03,  5.04s/it]INFO 04-12 01:56:51 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2372/3300 [3:20:05<1:19:51,  5.16s/it]2026-04-12 01:56:56,548 | INFO | train_grpo_train | metrics_logged mode=train step=2372

                                                       
{'loss': 0.0096, 'grad_norm': 3.56820011138916, 'learning_rate': 2.8151515151515153e-06, 'num_tokens': 5368757.0, 'completions/mean_length': 132.5, 'completions/min_length': 128.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.5, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.9986804723739624, 'rewards/meter/std': 0.0002552097721491009, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9808491468429565, 'rewards/total_composite/std': 0.0504879355430603, 'reward': 0.9808491468429565, 'reward_std': 0.050487954169511795, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04589386284351349, 'sampling/sampling_logp_difference/max': 1.5482797622680664, 'sampling/importance_sampling_ratio/min': 0.21261338889598846, 'sampling/importance_sampling_ratio/mean': 1.0080647468566895, 'sampling/importance_sampling_ratio/max': 1.8720965385437012, 'entropy': 0.38280507549643517, 'clip_ratio/low_mean': 0.0055555556900799274, 'clip_ratio/low_min': 0.0055555556900799274, 'clip_ratio/high_mean': 0.03797508799470961, 'clip_ratio/high_max': 0.03797508799470961, 'clip_ratio/region_mean': 0.04353064368478954, 'reward_total_mean': 0.9808491468429565, 'reward_meter_mean': 0.9986804723739624, 'reward_meter_std': 0.0002552097721491009, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9808491468429565, 'reward_total_composite_std': 0.0504879355430603, 'epoch': 0.1}

 72%|███████▏  | 2372/3300 [3:20:05<1:19:51,  5.16s/it]INFO 04-12 01:56:56 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2373/3300 [3:20:10<1:21:18,  5.26s/it]2026-04-12 01:57:02,040 | INFO | train_grpo_train | metrics_logged mode=train step=2373

                                                       
{'loss': 0.0088, 'grad_norm': 2.409925699234009, 'learning_rate': 2.812121212121212e-06, 'num_tokens': 5371394.0, 'completions/mean_length': 143.625, 'completions/min_length': 141.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.625, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9991071224212646, 'rewards/meter/std': 0.00023018992214929312, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9634268879890442, 'rewards/total_composite/std': 0.06610530614852905, 'reward': 0.9634268879890442, 'reward_std': 0.06610530614852905, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022872699424624443, 'sampling/sampling_logp_difference/max': 1.4115514755249023, 'sampling/importance_sampling_ratio/min': 0.24376478791236877, 'sampling/importance_sampling_ratio/mean': 1.0038185119628906, 'sampling/importance_sampling_ratio/max': 1.691011667251587, 'entropy': 0.2039720769971609, 'clip_ratio/low_mean': 0.00603448273614049, 'clip_ratio/low_min': 0.00603448273614049, 'clip_ratio/high_mean': 0.02439746167510748, 'clip_ratio/high_max': 0.02439746167510748, 'clip_ratio/region_mean': 0.03043194441124797, 'reward_total_mean': 0.9634268879890442, 'reward_meter_mean': 0.9991071224212646, 'reward_meter_std': 0.00023018992214929312, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9634268879890442, 'reward_total_composite_std': 0.06610530614852905, 'epoch': 0.1}

 72%|███████▏  | 2373/3300 [3:20:11<1:21:18,  5.26s/it]INFO 04-12 01:57:02 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2374/3300 [3:20:15<1:18:38,  5.10s/it]2026-04-12 01:57:06,744 | INFO | train_grpo_train | metrics_logged mode=train step=2374

                                                       
{'loss': 0.0065, 'grad_norm': 1.9035519361495972, 'learning_rate': 2.809090909090909e-06, 'num_tokens': 5373190.0, 'completions/mean_length': 73.5, 'completions/min_length': 71.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9992323517799377, 'rewards/meter/std': 0.00020660254813265055, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992323517799377, 'rewards/total_composite/std': 0.00020660254813265055, 'reward': 0.9992323517799377, 'reward_std': 0.00020661571761593223, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031100856140255928, 'sampling/sampling_logp_difference/max': 1.3491382598876953, 'sampling/importance_sampling_ratio/min': 0.25946375727653503, 'sampling/importance_sampling_ratio/mean': 1.0032356977462769, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20485389418900013, 'clip_ratio/low_mean': 0.011739226174540818, 'clip_ratio/low_min': 0.011739226174540818, 'clip_ratio/high_mean': 0.013676133356057107, 'clip_ratio/high_max': 0.013676133356057107, 'clip_ratio/region_mean': 0.025415359530597925, 'reward_total_mean': 0.9992323517799377, 'reward_meter_mean': 0.9992323517799377, 'reward_meter_std': 0.00020660254813265055, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992323517799377, 'reward_total_composite_std': 0.00020660254813265055, 'epoch': 0.1}

 72%|███████▏  | 2374/3300 [3:20:15<1:18:38,  5.10s/it]INFO 04-12 01:57:06 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2375/3300 [3:20:23<1:32:43,  6.01s/it]2026-04-12 01:57:14,911 | INFO | train_grpo_train | metrics_logged mode=train step=2375

                                                       
{'loss': 0.0053, 'grad_norm': 1.6295384168624878, 'learning_rate': 2.806060606060606e-06, 'num_tokens': 5377395.0, 'completions/mean_length': 319.625, 'completions/min_length': 317.0, 'completions/max_length': 322.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 319.625, 'completions/min_terminated_length': 317.0, 'completions/max_terminated_length': 322.0, 'rewards/meter/mean': 0.9988389015197754, 'rewards/meter/std': 0.0002653080737218261, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8529411554336548, 'rewards/repeat_penalty/std': 0.03144249692559242, 'rewards/total_composite/mean': 0.7667578458786011, 'rewards/total_composite/std': 0.02833147719502449, 'reward': 0.7667578458786011, 'reward_std': 0.02833147533237934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02020336128771305, 'sampling/sampling_logp_difference/max': 1.3237390518188477, 'sampling/importance_sampling_ratio/min': 0.26613834500312805, 'sampling/importance_sampling_ratio/mean': 1.0060769319534302, 'sampling/importance_sampling_ratio/max': 1.622670292854309, 'entropy': 0.20801045559346676, 'clip_ratio/low_mean': 0.00859655940439552, 'clip_ratio/low_min': 0.00859655940439552, 'clip_ratio/high_mean': 0.006286898045800626, 'clip_ratio/high_max': 0.006286898045800626, 'clip_ratio/region_mean': 0.014883457450196147, 'reward_total_mean': 0.7667578458786011, 'reward_meter_mean': 0.9988389015197754, 'reward_meter_std': 0.0002653080737218261, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8529411554336548, 'reward_repeat_penalty_std': 0.03144249692559242, 'reward_total_composite_mean': 0.7667578458786011, 'reward_total_composite_std': 0.02833147719502449, 'epoch': 0.1}

 72%|███████▏  | 2375/3300 [3:20:23<1:32:43,  6.01s/it]INFO 04-12 01:57:15 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2376/3300 [3:20:28<1:26:41,  5.63s/it]2026-04-12 01:57:19,632 | INFO | train_grpo_train | metrics_logged mode=train step=2376

                                                       
{'loss': 0.0009, 'grad_norm': 0.18841330707073212, 'learning_rate': 2.803030303030303e-06, 'num_tokens': 5379314.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981135129928589, 'rewards/meter/std': 1.4713088603457436e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981135129928589, 'rewards/total_composite/std': 1.4713088603457436e-05, 'reward': 0.9981135129928589, 'reward_std': 1.4704202840221114e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010610595345497131, 'sampling/sampling_logp_difference/max': 0.893486499786377, 'sampling/importance_sampling_ratio/min': 0.40922650694847107, 'sampling/importance_sampling_ratio/mean': 0.9993737936019897, 'sampling/importance_sampling_ratio/max': 1.2641924619674683, 'entropy': 0.0682340869680047, 'clip_ratio/low_mean': 0.005597014795057476, 'clip_ratio/low_min': 0.005597014795057476, 'clip_ratio/high_mean': 0.007490954245440662, 'clip_ratio/high_max': 0.007490954245440662, 'clip_ratio/region_mean': 0.013087969040498137, 'reward_total_mean': 0.9981135129928589, 'reward_meter_mean': 0.9981135129928589, 'reward_meter_std': 1.4713088603457436e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981135129928589, 'reward_total_composite_std': 1.4713088603457436e-05, 'epoch': 0.1}

 72%|███████▏  | 2376/3300 [3:20:28<1:26:41,  5.63s/it]INFO 04-12 01:57:19 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2377/3300 [3:20:33<1:21:53,  5.32s/it]2026-04-12 01:57:24,257 | INFO | train_grpo_train | metrics_logged mode=train step=2377

                                                       
{'loss': -0.001, 'grad_norm': 0.5763253569602966, 'learning_rate': 2.8000000000000003e-06, 'num_tokens': 5381154.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981157779693604, 'rewards/meter/std': 2.6794468794832937e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981157779693604, 'rewards/total_composite/std': 2.6794468794832937e-05, 'reward': 0.9981157779693604, 'reward_std': 2.6794450604938902e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008252451196312904, 'sampling/sampling_logp_difference/max': 0.4723522663116455, 'sampling/importance_sampling_ratio/min': 0.6235338449478149, 'sampling/importance_sampling_ratio/mean': 1.001448631286621, 'sampling/importance_sampling_ratio/max': 1.3120979070663452, 'entropy': 0.06305565731599927, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005597014795057476, 'clip_ratio/high_max': 0.005597014795057476, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9981157779693604, 'reward_meter_mean': 0.9981157779693604, 'reward_meter_std': 2.6794468794832937e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981157779693604, 'reward_total_composite_std': 2.6794468794832937e-05, 'epoch': 0.1}

 72%|███████▏  | 2377/3300 [3:20:33<1:21:53,  5.32s/it]INFO 04-12 01:57:24 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2378/3300 [3:20:39<1:25:19,  5.55s/it]2026-04-12 01:57:30,330 | INFO | train_grpo_train | metrics_logged mode=train step=2378

                                                       
{'loss': 0.0034, 'grad_norm': 1.9172272682189941, 'learning_rate': 2.7969696969696976e-06, 'num_tokens': 5384219.0, 'completions/mean_length': 186.125, 'completions/min_length': 184.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.125, 'completions/min_terminated_length': 184.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9751391410827637, 'rewards/meter/std': 0.04846389219164848, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8181818723678589, 'rewards/repeat_penalty/std': 0.08416548371315002, 'rewards/total_composite/mean': 0.7978767156600952, 'rewards/total_composite/std': 0.09277582913637161, 'reward': 0.7978767156600952, 'reward_std': 0.09277583658695221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02777726948261261, 'sampling/sampling_logp_difference/max': 0.9550704956054688, 'sampling/importance_sampling_ratio/min': 0.3847850263118744, 'sampling/importance_sampling_ratio/mean': 1.005282998085022, 'sampling/importance_sampling_ratio/max': 1.715010404586792, 'entropy': 0.2459548767656088, 'clip_ratio/low_mean': 0.004003660404123366, 'clip_ratio/low_min': 0.004003660404123366, 'clip_ratio/high_mean': 0.015454705455340445, 'clip_ratio/high_max': 0.015454705455340445, 'clip_ratio/region_mean': 0.01945836585946381, 'reward_total_mean': 0.7978767156600952, 'reward_meter_mean': 0.9751391410827637, 'reward_meter_std': 0.04846389219164848, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8181818723678589, 'reward_repeat_penalty_std': 0.08416548371315002, 'reward_total_composite_mean': 0.7978767156600952, 'reward_total_composite_std': 0.09277582913637161, 'epoch': 0.1}

 72%|███████▏  | 2378/3300 [3:20:39<1:25:19,  5.55s/it]INFO 04-12 01:57:30 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2379/3300 [3:20:44<1:24:20,  5.49s/it]2026-04-12 01:57:35,693 | INFO | train_grpo_train | metrics_logged mode=train step=2379

                                                       
{'loss': 0.0033, 'grad_norm': 2.5564944744110107, 'learning_rate': 2.7939393939393944e-06, 'num_tokens': 5385952.0, 'completions/mean_length': 67.625, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9991129636764526, 'rewards/meter/std': 0.00021267922420520335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991129636764526, 'rewards/total_composite/std': 0.00021267922420520335, 'reward': 0.9991129636764526, 'reward_std': 0.00021267762349452823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03177187591791153, 'sampling/sampling_logp_difference/max': 0.6790390014648438, 'sampling/importance_sampling_ratio/min': 0.5071040987968445, 'sampling/importance_sampling_ratio/mean': 1.011904001235962, 'sampling/importance_sampling_ratio/max': 1.5362749099731445, 'entropy': 0.32505345717072487, 'clip_ratio/low_mean': 0.01117487601004541, 'clip_ratio/low_min': 0.01117487601004541, 'clip_ratio/high_mean': 0.00916612590663135, 'clip_ratio/high_max': 0.00916612590663135, 'clip_ratio/region_mean': 0.02034100191667676, 'reward_total_mean': 0.9991129636764526, 'reward_meter_mean': 0.9991129636764526, 'reward_meter_std': 0.00021267922420520335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991129636764526, 'reward_total_composite_std': 0.00021267922420520335, 'epoch': 0.1}

 72%|███████▏  | 2379/3300 [3:20:44<1:24:20,  5.49s/it]INFO 04-12 01:57:35 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2380/3300 [3:20:49<1:20:56,  5.28s/it]2026-04-12 01:57:40,472 | INFO | train_grpo_train | metrics_logged mode=train step=2380

                                                       
{'loss': -0.0004, 'grad_norm': 0.42718541622161865, 'learning_rate': 2.7909090909090912e-06, 'num_tokens': 5387928.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981181621551514, 'rewards/meter/std': 2.501497874618508e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981181621551514, 'rewards/total_composite/std': 2.501497874618508e-05, 'reward': 0.9981181621551514, 'reward_std': 2.5023065973073244e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006671510171145201, 'sampling/sampling_logp_difference/max': 1.0379719734191895, 'sampling/importance_sampling_ratio/min': 0.3541722297668457, 'sampling/importance_sampling_ratio/mean': 0.9986084699630737, 'sampling/importance_sampling_ratio/max': 1.1040713787078857, 'entropy': 0.024645814672112465, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.0037313431967049837, 'reward_total_mean': 0.9981181621551514, 'reward_meter_mean': 0.9981181621551514, 'reward_meter_std': 2.501497874618508e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981181621551514, 'reward_total_composite_std': 2.501497874618508e-05, 'epoch': 0.1}

 72%|███████▏  | 2380/3300 [3:20:49<1:20:56,  5.28s/it]INFO 04-12 01:57:40 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2381/3300 [3:20:54<1:22:20,  5.38s/it]2026-04-12 01:57:46,071 | INFO | train_grpo_train | metrics_logged mode=train step=2381

                                                       
{'loss': 0.0134, 'grad_norm': 3.057466745376587, 'learning_rate': 2.7878787878787885e-06, 'num_tokens': 5390258.0, 'completions/mean_length': 107.25, 'completions/min_length': 105.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9992028474807739, 'rewards/meter/std': 7.249469490488991e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9742231369018555, 'rewards/total_composite/std': 0.07065970450639725, 'reward': 0.9742231369018555, 'reward_std': 0.07065972685813904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020719556137919426, 'sampling/sampling_logp_difference/max': 0.7676267623901367, 'sampling/importance_sampling_ratio/min': 0.4641132354736328, 'sampling/importance_sampling_ratio/mean': 1.0024255514144897, 'sampling/importance_sampling_ratio/max': 1.8616714477539062, 'entropy': 0.1429230272769928, 'clip_ratio/low_mean': 0.0011261261533945799, 'clip_ratio/low_min': 0.0011261261533945799, 'clip_ratio/high_mean': 0.02113948343321681, 'clip_ratio/high_max': 0.02113948343321681, 'clip_ratio/region_mean': 0.02226560958661139, 'reward_total_mean': 0.9742231369018555, 'reward_meter_mean': 0.9992028474807739, 'reward_meter_std': 7.249469490488991e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9742231369018555, 'reward_total_composite_std': 0.07065970450639725, 'epoch': 0.1}

 72%|███████▏  | 2381/3300 [3:20:55<1:22:20,  5.38s/it]INFO 04-12 01:57:46 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2382/3300 [3:20:59<1:19:31,  5.20s/it]2026-04-12 01:57:50,930 | INFO | train_grpo_train | metrics_logged mode=train step=2382

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.7848484848484853e-06, 'num_tokens': 5392402.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0010491692228242755, 'sampling/sampling_logp_difference/max': 0.07737492024898529, 'sampling/importance_sampling_ratio/min': 0.9255428314208984, 'sampling/importance_sampling_ratio/mean': 1.0004812479019165, 'sampling/importance_sampling_ratio/max': 1.0541869401931763, 'entropy': 0.008781549287959933, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 72%|███████▏  | 2382/3300 [3:20:59<1:19:31,  5.20s/it]INFO 04-12 01:57:51 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2383/3300 [3:21:05<1:20:46,  5.29s/it]2026-04-12 01:57:56,338 | INFO | train_grpo_train | metrics_logged mode=train step=2383

                                                       
{'loss': 0.006, 'grad_norm': 1.7132611274719238, 'learning_rate': 2.781818181818182e-06, 'num_tokens': 5394915.0, 'completions/mean_length': 130.125, 'completions/min_length': 130.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.125, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.999283492565155, 'rewards/meter/std': 2.7988780857413076e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9457492828369141, 'rewards/total_composite/std': 0.07386652380228043, 'reward': 0.9457492828369141, 'reward_std': 0.07386651635169983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016906656324863434, 'sampling/sampling_logp_difference/max': 1.372336745262146, 'sampling/importance_sampling_ratio/min': 0.2535138428211212, 'sampling/importance_sampling_ratio/mean': 1.003693699836731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09555556159466505, 'clip_ratio/low_mean': 0.0076776278438046575, 'clip_ratio/low_min': 0.0076776278438046575, 'clip_ratio/high_mean': 0.014423077227547765, 'clip_ratio/high_max': 0.014423077227547765, 'clip_ratio/region_mean': 0.022100705071352422, 'reward_total_mean': 0.9457492828369141, 'reward_meter_mean': 0.999283492565155, 'reward_meter_std': 2.7988780857413076e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9457492828369141, 'reward_total_composite_std': 0.07386652380228043, 'epoch': 0.1}

 72%|███████▏  | 2383/3300 [3:21:05<1:20:46,  5.29s/it]INFO 04-12 01:57:56 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2384/3300 [3:21:10<1:19:24,  5.20s/it]2026-04-12 01:58:01,347 | INFO | train_grpo_train | metrics_logged mode=train step=2384

                                                       
{'loss': 0.0197, 'grad_norm': 4.2915167808532715, 'learning_rate': 2.778787878787879e-06, 'num_tokens': 5397220.0, 'completions/mean_length': 101.125, 'completions/min_length': 99.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.125, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9984184503555298, 'rewards/meter/std': 0.0006285551935434341, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984184503555298, 'rewards/total_composite/std': 0.0006285551935434341, 'reward': 0.9984184503555298, 'reward_std': 0.0006285490817390382, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05373314768075943, 'sampling/sampling_logp_difference/max': 1.02024507522583, 'sampling/importance_sampling_ratio/min': 0.3605065941810608, 'sampling/importance_sampling_ratio/mean': 1.0017614364624023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41279827430844307, 'clip_ratio/low_mean': 0.01336255669593811, 'clip_ratio/low_min': 0.01336255669593811, 'clip_ratio/high_mean': 0.03354567987844348, 'clip_ratio/high_max': 0.03354567987844348, 'clip_ratio/region_mean': 0.04690823657438159, 'reward_total_mean': 0.9984184503555298, 'reward_meter_mean': 0.9984184503555298, 'reward_meter_std': 0.0006285551935434341, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984184503555298, 'reward_total_composite_std': 0.0006285551935434341, 'epoch': 0.1}

 72%|███████▏  | 2384/3300 [3:21:10<1:19:24,  5.20s/it]INFO 04-12 01:58:01 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2385/3300 [3:21:14<1:14:35,  4.89s/it]2026-04-12 01:58:05,510 | INFO | train_grpo_train | metrics_logged mode=train step=2385

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.7757575757575762e-06, 'num_tokens': 5398716.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0007748155039735138, 'sampling/sampling_logp_difference/max': 0.023845985531806946, 'sampling/importance_sampling_ratio/min': 0.9885799884796143, 'sampling/importance_sampling_ratio/mean': 1.0006725788116455, 'sampling/importance_sampling_ratio/max': 1.024132490158081, 'entropy': 0.009345798986032605, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 72%|███████▏  | 2385/3300 [3:21:14<1:14:35,  4.89s/it]INFO 04-12 01:58:05 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2386/3300 [3:21:18<1:12:12,  4.74s/it]2026-04-12 01:58:09,899 | INFO | train_grpo_train | metrics_logged mode=train step=2386

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.772727272727273e-06, 'num_tokens': 5400332.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00041769127710722387, 'sampling/sampling_logp_difference/max': 0.010340280830860138, 'sampling/importance_sampling_ratio/min': 0.9897130131721497, 'sampling/importance_sampling_ratio/mean': 1.0003135204315186, 'sampling/importance_sampling_ratio/max': 1.0083444118499756, 'entropy': 0.004995487950509414, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 72%|███████▏  | 2386/3300 [3:21:18<1:12:12,  4.74s/it]INFO 04-12 01:58:10 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2387/3300 [3:21:23<1:12:35,  4.77s/it]2026-04-12 01:58:14,746 | INFO | train_grpo_train | metrics_logged mode=train step=2387

                                                       
{'loss': 0.0021, 'grad_norm': 3.3865015506744385, 'learning_rate': 2.76969696969697e-06, 'num_tokens': 5402149.0, 'completions/mean_length': 68.125, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9987808465957642, 'rewards/meter/std': 0.0002124947786796838, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987808465957642, 'rewards/total_composite/std': 0.0002124947786796838, 'reward': 0.9987808465957642, 'reward_std': 0.00021250976715236902, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05237334594130516, 'sampling/sampling_logp_difference/max': 1.7108659744262695, 'sampling/importance_sampling_ratio/min': 0.18070924282073975, 'sampling/importance_sampling_ratio/mean': 1.0035470724105835, 'sampling/importance_sampling_ratio/max': 1.5863909721374512, 'entropy': 0.371359147131443, 'clip_ratio/low_mean': 0.016549831489101052, 'clip_ratio/low_min': 0.016549831489101052, 'clip_ratio/high_mean': 0.019927536603063345, 'clip_ratio/high_max': 0.019927536603063345, 'clip_ratio/region_mean': 0.0364773680921644, 'reward_total_mean': 0.9987808465957642, 'reward_meter_mean': 0.9987808465957642, 'reward_meter_std': 0.0002124947786796838, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987808465957642, 'reward_total_composite_std': 0.0002124947786796838, 'epoch': 0.1}

 72%|███████▏  | 2387/3300 [3:21:23<1:12:35,  4.77s/it]INFO 04-12 01:58:14 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2388/3300 [3:21:27<1:10:01,  4.61s/it]2026-04-12 01:58:18,983 | INFO | train_grpo_train | metrics_logged mode=train step=2388

                                                       
{'loss': 0.0005, 'grad_norm': 2.705258846282959, 'learning_rate': 2.766666666666667e-06, 'num_tokens': 5403597.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992798566818237, 'rewards/meter/std': 3.3992837416008115e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992798566818237, 'rewards/total_composite/std': 3.3992837416008115e-05, 'reward': 0.9992798566818237, 'reward_std': 3.3992837416008115e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00878862477838993, 'sampling/sampling_logp_difference/max': 1.0551605224609375, 'sampling/importance_sampling_ratio/min': 0.3481365442276001, 'sampling/importance_sampling_ratio/mean': 0.9955343008041382, 'sampling/importance_sampling_ratio/max': 1.0776525735855103, 'entropy': 0.015011629671789706, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.9992798566818237, 'reward_meter_mean': 0.9992798566818237, 'reward_meter_std': 3.3992837416008115e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992798566818237, 'reward_total_composite_std': 3.3992837416008115e-05, 'epoch': 0.1}

 72%|███████▏  | 2388/3300 [3:21:27<1:10:01,  4.61s/it]INFO 04-12 01:58:19 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2389/3300 [3:21:32<1:10:54,  4.67s/it]2026-04-12 01:58:23,787 | INFO | train_grpo_train | metrics_logged mode=train step=2389

                                                       
{'loss': 0.005, 'grad_norm': 3.947572946548462, 'learning_rate': 2.763636363636364e-06, 'num_tokens': 5405494.0, 'completions/mean_length': 68.125, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9979345798492432, 'rewards/meter/std': 0.0023531855549663305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979345798492432, 'rewards/total_composite/std': 0.0023531855549663305, 'reward': 0.9979345798492432, 'reward_std': 0.0023531648330390453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03587198257446289, 'sampling/sampling_logp_difference/max': 0.9495134353637695, 'sampling/importance_sampling_ratio/min': 0.3869292438030243, 'sampling/importance_sampling_ratio/mean': 1.0087910890579224, 'sampling/importance_sampling_ratio/max': 1.5146595239639282, 'entropy': 0.3192524388432503, 'clip_ratio/low_mean': 0.0071428571827709675, 'clip_ratio/low_min': 0.0071428571827709675, 'clip_ratio/high_mean': 0.027370785945095122, 'clip_ratio/high_max': 0.027370785945095122, 'clip_ratio/region_mean': 0.03451364312786609, 'reward_total_mean': 0.9979345798492432, 'reward_meter_mean': 0.9979345798492432, 'reward_meter_std': 0.0023531855549663305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979345798492432, 'reward_total_composite_std': 0.0023531855549663305, 'epoch': 0.1}

 72%|███████▏  | 2389/3300 [3:21:32<1:10:54,  4.67s/it]INFO 04-12 01:58:24 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2390/3300 [3:21:37<1:12:01,  4.75s/it]2026-04-12 01:58:28,737 | INFO | train_grpo_train | metrics_logged mode=train step=2390

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.760606060606061e-06, 'num_tokens': 5407279.0, 'completions/mean_length': 54.125, 'completions/min_length': 54.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0010611445177346468, 'sampling/sampling_logp_difference/max': 0.17297124862670898, 'sampling/importance_sampling_ratio/min': 0.841161847114563, 'sampling/importance_sampling_ratio/mean': 0.9997290968894958, 'sampling/importance_sampling_ratio/max': 1.0198215246200562, 'entropy': 0.004519450158113614, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 72%|███████▏  | 2390/3300 [3:21:37<1:12:01,  4.75s/it]INFO 04-12 01:58:29 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2391/3300 [3:21:42<1:14:29,  4.92s/it]2026-04-12 01:58:34,028 | INFO | train_grpo_train | metrics_logged mode=train step=2391

                                                       
{'loss': -0.0026, 'grad_norm': 1.7748810052871704, 'learning_rate': 2.7575757575757576e-06, 'num_tokens': 5409667.0, 'completions/mean_length': 121.5, 'completions/min_length': 121.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.5, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9796151518821716, 'rewards/meter/std': 0.002762306947261095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9271951913833618, 'rewards/total_composite/std': 0.0733107179403305, 'reward': 0.9271951913833618, 'reward_std': 0.07331071048974991, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008518277667462826, 'sampling/sampling_logp_difference/max': 1.3747358322143555, 'sampling/importance_sampling_ratio/min': 0.2529063820838928, 'sampling/importance_sampling_ratio/mean': 1.0016850233078003, 'sampling/importance_sampling_ratio/max': 1.2646301984786987, 'entropy': 0.05755401449277997, 'clip_ratio/low_mean': 0.001033057807944715, 'clip_ratio/low_min': 0.001033057807944715, 'clip_ratio/high_mean': 0.005099173518829048, 'clip_ratio/high_max': 0.005099173518829048, 'clip_ratio/region_mean': 0.006132231326773763, 'reward_total_mean': 0.9271951913833618, 'reward_meter_mean': 0.9796151518821716, 'reward_meter_std': 0.002762306947261095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9271951913833618, 'reward_total_composite_std': 0.0733107179403305, 'epoch': 0.1}

 72%|███████▏  | 2391/3300 [3:21:43<1:14:29,  4.92s/it]INFO 04-12 01:58:34 [block_pool.py:378] Successfully reset prefix cache

 72%|███████▏  | 2392/3300 [3:21:48<1:17:47,  5.14s/it]2026-04-12 01:58:39,687 | INFO | train_grpo_train | metrics_logged mode=train step=2392

                                                       
{'loss': -0.002, 'grad_norm': 2.4178664684295654, 'learning_rate': 2.754545454545455e-06, 'num_tokens': 5412340.0, 'completions/mean_length': 154.125, 'completions/min_length': 151.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.125, 'completions/min_terminated_length': 151.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.9878800511360168, 'rewards/meter/std': 0.02659641019999981, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9002017974853516, 'rewards/total_composite/std': 0.08359287679195404, 'reward': 0.9002017974853516, 'reward_std': 0.08359287679195404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03138583526015282, 'sampling/sampling_logp_difference/max': 1.6040925979614258, 'sampling/importance_sampling_ratio/min': 0.20107191801071167, 'sampling/importance_sampling_ratio/mean': 1.0050873756408691, 'sampling/importance_sampling_ratio/max': 1.3937476873397827, 'entropy': 0.23401772044599056, 'clip_ratio/low_mean': 0.011445078765973449, 'clip_ratio/low_min': 0.011445078765973449, 'clip_ratio/high_mean': 0.007223089050967246, 'clip_ratio/high_max': 0.007223089050967246, 'clip_ratio/region_mean': 0.018668167816940695, 'reward_total_mean': 0.9002017974853516, 'reward_meter_mean': 0.9878800511360168, 'reward_meter_std': 0.02659641019999981, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9002017974853516, 'reward_total_composite_std': 0.08359287679195404, 'epoch': 0.1}

 72%|███████▏  | 2392/3300 [3:21:48<1:17:47,  5.14s/it]INFO 04-12 01:58:39 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2393/3300 [3:21:53<1:15:47,  5.01s/it]2026-04-12 01:58:44,422 | INFO | train_grpo_train | metrics_logged mode=train step=2393

                                                       
{'loss': 0.0047, 'grad_norm': 1.324416995048523, 'learning_rate': 2.7515151515151517e-06, 'num_tokens': 5414319.0, 'completions/mean_length': 72.375, 'completions/min_length': 72.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9992770552635193, 'rewards/meter/std': 0.00018226148677058518, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992770552635193, 'rewards/total_composite/std': 0.00018226148677058518, 'reward': 0.9992770552635193, 'reward_std': 0.00018225403618998826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020829232409596443, 'sampling/sampling_logp_difference/max': 1.3297691345214844, 'sampling/importance_sampling_ratio/min': 0.26453831791877747, 'sampling/importance_sampling_ratio/mean': 1.0025677680969238, 'sampling/importance_sampling_ratio/max': 1.7886258363723755, 'entropy': 0.11769631877541542, 'clip_ratio/low_mean': 0.006779896095395088, 'clip_ratio/low_min': 0.006779896095395088, 'clip_ratio/high_mean': 0.005208333372138441, 'clip_ratio/high_max': 0.005208333372138441, 'clip_ratio/region_mean': 0.011988229467533529, 'reward_total_mean': 0.9992770552635193, 'reward_meter_mean': 0.9992770552635193, 'reward_meter_std': 0.00018226148677058518, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992770552635193, 'reward_total_composite_std': 0.00018226148677058518, 'epoch': 0.1}

 73%|███████▎  | 2393/3300 [3:21:53<1:15:47,  5.01s/it]INFO 04-12 01:58:44 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2394/3300 [3:21:57<1:13:27,  4.86s/it]2026-04-12 01:58:48,931 | INFO | train_grpo_train | metrics_logged mode=train step=2394

                                                       
{'loss': 0.0004, 'grad_norm': 1.3910056352615356, 'learning_rate': 2.7484848484848486e-06, 'num_tokens': 5416127.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993087649345398, 'rewards/meter/std': 5.955179949523881e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993087649345398, 'rewards/total_composite/std': 5.955179949523881e-05, 'reward': 0.9993087649345398, 'reward_std': 5.955949382041581e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012217044830322266, 'sampling/sampling_logp_difference/max': 0.748956024646759, 'sampling/importance_sampling_ratio/min': 0.4728599488735199, 'sampling/importance_sampling_ratio/mean': 0.9990496635437012, 'sampling/importance_sampling_ratio/max': 1.3544596433639526, 'entropy': 0.047464726492762566, 'clip_ratio/low_mean': 0.01171875, 'clip_ratio/low_min': 0.01171875, 'clip_ratio/high_mean': 0.0078125, 'clip_ratio/high_max': 0.0078125, 'clip_ratio/region_mean': 0.01953125, 'reward_total_mean': 0.9993087649345398, 'reward_meter_mean': 0.9993087649345398, 'reward_meter_std': 5.955179949523881e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993087649345398, 'reward_total_composite_std': 5.955179949523881e-05, 'epoch': 0.1}

 73%|███████▎  | 2394/3300 [3:21:57<1:13:27,  4.86s/it]INFO 04-12 01:58:49 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2395/3300 [3:22:02<1:11:47,  4.76s/it]2026-04-12 01:58:53,439 | INFO | train_grpo_train | metrics_logged mode=train step=2395

                                                       
{'loss': 0.0, 'grad_norm': 0.06802152097225189, 'learning_rate': 2.7454545454545454e-06, 'num_tokens': 5417895.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993706345558167, 'rewards/meter/std': 5.743691417592345e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993706345558167, 'rewards/total_composite/std': 5.743691417592345e-06, 'reward': 0.9993706345558167, 'reward_std': 5.761006832472049e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008546598255634308, 'sampling/sampling_logp_difference/max': 0.5136269330978394, 'sampling/importance_sampling_ratio/min': 0.5983215570449829, 'sampling/importance_sampling_ratio/mean': 1.002130389213562, 'sampling/importance_sampling_ratio/max': 1.6340101957321167, 'entropy': 0.038509635254740715, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0078125, 'clip_ratio/high_max': 0.0078125, 'clip_ratio/region_mean': 0.0078125, 'reward_total_mean': 0.9993706345558167, 'reward_meter_mean': 0.9993706345558167, 'reward_meter_std': 5.743691417592345e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993706345558167, 'reward_total_composite_std': 5.743691417592345e-06, 'epoch': 0.1}

 73%|███████▎  | 2395/3300 [3:22:02<1:11:47,  4.76s/it]INFO 04-12 01:58:53 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2396/3300 [3:22:07<1:15:00,  4.98s/it]2026-04-12 01:58:58,930 | INFO | train_grpo_train | metrics_logged mode=train step=2396

                                                       
{'loss': 0.0113, 'grad_norm': 1.6420077085494995, 'learning_rate': 2.7424242424242426e-06, 'num_tokens': 5420082.0, 'completions/mean_length': 102.375, 'completions/min_length': 100.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.375, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9934667944908142, 'rewards/meter/std': 0.006176711525768042, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.968574583530426, 'rewards/total_composite/std': 0.06977622210979462, 'reward': 0.968574583530426, 'reward_std': 0.06977621465921402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022047532722353935, 'sampling/sampling_logp_difference/max': 0.9480593204498291, 'sampling/importance_sampling_ratio/min': 0.38789549469947815, 'sampling/importance_sampling_ratio/mean': 1.0029125213623047, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.13659420423209667, 'clip_ratio/low_mean': 0.0012019231216982007, 'clip_ratio/low_min': 0.0012019231216982007, 'clip_ratio/high_mean': 0.015861412626691163, 'clip_ratio/high_max': 0.015861412626691163, 'clip_ratio/region_mean': 0.017063335748389363, 'reward_total_mean': 0.968574583530426, 'reward_meter_mean': 0.9934667944908142, 'reward_meter_std': 0.006176711525768042, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.968574583530426, 'reward_total_composite_std': 0.06977622210979462, 'epoch': 0.1}

 73%|███████▎  | 2396/3300 [3:22:07<1:15:00,  4.98s/it]INFO 04-12 01:58:59 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2397/3300 [3:22:13<1:16:44,  5.10s/it]2026-04-12 01:59:04,315 | INFO | train_grpo_train | metrics_logged mode=train step=2397

                                                       
{'loss': -0.0055, 'grad_norm': 11.099533081054688, 'learning_rate': 2.7393939393939395e-06, 'num_tokens': 5422174.0, 'completions/mean_length': 84.5, 'completions/min_length': 84.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.5, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9455694556236267, 'rewards/meter/std': 0.006192359142005444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9455694556236267, 'rewards/total_composite/std': 0.006192359142005444, 'reward': 0.9455694556236267, 'reward_std': 0.00619234936311841, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014308757148683071, 'sampling/sampling_logp_difference/max': 0.9953222274780273, 'sampling/importance_sampling_ratio/min': 0.36960431933403015, 'sampling/importance_sampling_ratio/mean': 0.9979245662689209, 'sampling/importance_sampling_ratio/max': 1.6855740547180176, 'entropy': 0.06057103397324681, 'clip_ratio/low_mean': 0.008893557591363788, 'clip_ratio/low_min': 0.008893557591363788, 'clip_ratio/high_mean': 0.0028735632076859474, 'clip_ratio/high_max': 0.0028735632076859474, 'clip_ratio/region_mean': 0.011767120799049735, 'reward_total_mean': 0.9455694556236267, 'reward_meter_mean': 0.9455694556236267, 'reward_meter_std': 0.006192359142005444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9455694556236267, 'reward_total_composite_std': 0.006192359142005444, 'epoch': 0.1}

 73%|███████▎  | 2397/3300 [3:22:13<1:16:44,  5.10s/it]INFO 04-12 01:59:04 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2398/3300 [3:22:19<1:22:25,  5.48s/it]2026-04-12 01:59:10,734 | INFO | train_grpo_train | metrics_logged mode=train step=2398

                                                       
{'loss': -0.0017, 'grad_norm': 1.1203114986419678, 'learning_rate': 2.7363636363636363e-06, 'num_tokens': 5424845.0, 'completions/mean_length': 147.875, 'completions/min_length': 144.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.875, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9743971824645996, 'rewards/meter/std': 0.0041354927234351635, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.819444477558136, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.7985101938247681, 'rewards/total_composite/std': 0.05690953880548477, 'reward': 0.7985101938247681, 'reward_std': 0.05690954998135567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011587115004658699, 'sampling/sampling_logp_difference/max': 1.6029415130615234, 'sampling/importance_sampling_ratio/min': 0.20130351185798645, 'sampling/importance_sampling_ratio/mean': 0.9987753629684448, 'sampling/importance_sampling_ratio/max': 1.5559734106063843, 'entropy': 0.04592153197154403, 'clip_ratio/low_mean': 0.003401839407160878, 'clip_ratio/low_min': 0.003401839407160878, 'clip_ratio/high_mean': 0.004206492216326296, 'clip_ratio/high_max': 0.004206492216326296, 'clip_ratio/region_mean': 0.0076083316234871745, 'reward_total_mean': 0.7985101938247681, 'reward_meter_mean': 0.9743971824645996, 'reward_meter_std': 0.0041354927234351635, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.819444477558136, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.7985101938247681, 'reward_total_composite_std': 0.05690953880548477, 'epoch': 0.1}

 73%|███████▎  | 2398/3300 [3:22:19<1:22:25,  5.48s/it]INFO 04-12 01:59:11 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2399/3300 [3:22:24<1:20:28,  5.36s/it]2026-04-12 01:59:15,753 | INFO | train_grpo_train | metrics_logged mode=train step=2399

                                                       
{'loss': 0.0016, 'grad_norm': 1.154660701751709, 'learning_rate': 2.7333333333333336e-06, 'num_tokens': 5427116.0, 'completions/mean_length': 96.875, 'completions/min_length': 96.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.875, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.9992877244949341, 'rewards/meter/std': 5.900250835111365e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992877244949341, 'rewards/total_composite/std': 5.900250835111365e-05, 'reward': 0.9992877244949341, 'reward_std': 5.900421820115298e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012875314801931381, 'sampling/sampling_logp_difference/max': 1.2882099151611328, 'sampling/importance_sampling_ratio/min': 0.27576398849487305, 'sampling/importance_sampling_ratio/mean': 0.9967014789581299, 'sampling/importance_sampling_ratio/max': 1.3556400537490845, 'entropy': 0.05716710351407528, 'clip_ratio/low_mean': 0.0013020833721384406, 'clip_ratio/low_min': 0.0013020833721384406, 'clip_ratio/high_mean': 0.015463917283341289, 'clip_ratio/high_max': 0.015463917283341289, 'clip_ratio/region_mean': 0.01676600065547973, 'reward_total_mean': 0.9992877244949341, 'reward_meter_mean': 0.9992877244949341, 'reward_meter_std': 5.900250835111365e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992877244949341, 'reward_total_composite_std': 5.900250835111365e-05, 'epoch': 0.1}

 73%|███████▎  | 2399/3300 [3:22:24<1:20:28,  5.36s/it]INFO 04-12 01:59:15 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2400/3300 [3:22:29<1:18:54,  5.26s/it]2026-04-12 01:59:20,788 | INFO | train_grpo_train | metrics_logged mode=train step=2400

                                                       
{'loss': 0.0041, 'grad_norm': 1.3548955917358398, 'learning_rate': 2.7303030303030304e-06, 'num_tokens': 5428869.0, 'completions/mean_length': 72.125, 'completions/min_length': 71.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9993604421615601, 'rewards/meter/std': 9.307587606599554e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993604421615601, 'rewards/total_composite/std': 9.307587606599554e-05, 'reward': 0.9993604421615601, 'reward_std': 9.308403969043866e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01761528290808201, 'sampling/sampling_logp_difference/max': 1.1383390426635742, 'sampling/importance_sampling_ratio/min': 0.32035067677497864, 'sampling/importance_sampling_ratio/mean': 1.0040256977081299, 'sampling/importance_sampling_ratio/max': 1.8390227556228638, 'entropy': 0.07446162821725011, 'clip_ratio/low_mean': 0.004999999888241291, 'clip_ratio/low_min': 0.004999999888241291, 'clip_ratio/high_mean': 0.0034966744715347886, 'clip_ratio/high_max': 0.0034966744715347886, 'clip_ratio/region_mean': 0.00849667435977608, 'reward_total_mean': 0.9993604421615601, 'reward_meter_mean': 0.9993604421615601, 'reward_meter_std': 9.307587606599554e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993604421615601, 'reward_total_composite_std': 9.307587606599554e-05, 'epoch': 0.1}

 73%|███████▎  | 2400/3300 [3:22:29<1:18:54,  5.26s/it]INFO 04-12 01:59:21 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:34,  3.10s/it]

 23%|██▎       | 3/13 [00:09<00:31,  3.20s/it]

 31%|███       | 4/13 [00:15<00:36,  4.08s/it]

 38%|███▊      | 5/13 [00:20<00:35,  4.46s/it]

 46%|████▌     | 6/13 [00:25<00:33,  4.83s/it]

 54%|█████▍    | 7/13 [00:30<00:29,  4.88s/it]

 62%|██████▏   | 8/13 [00:36<00:26,  5.26s/it]

 69%|██████▉   | 9/13 [00:44<00:23,  6.00s/it]

 77%|███████▋  | 10/13 [00:49<00:16,  5.60s/it]

 85%|████████▍ | 11/13 [00:54<00:10,  5.47s/it]

 92%|█████████▏| 12/13 [00:59<00:05,  5.49s/it]

100%|██████████| 13/13 [01:04<00:00,  5.34s/it]2026-04-12 02:00:31,987 | INFO | train_grpo_train | metrics_logged mode=eval step=2400

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 71.1062, 'eval_samples_per_second': 1.463, 'eval_steps_per_second': 0.183, 'eval_num_tokens': 5428869.0, 'eval_completions/mean_length': 196.8653846153846, 'eval_completions/min_length': 59.46153846153846, 'eval_completions/max_length': 372.46153846153845, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 196.8653846153846, 'eval_completions/min_terminated_length': 59.46153846153846, 'eval_completions/max_terminated_length': 372.46153846153845, 'eval_rewards/meter/mean': 0.7441063798390902, 'eval_rewards/meter/std': 0.3844332993030548, 'eval_rewards/count_adherence/mean': 0.9244015262677119, 'eval_rewards/count_adherence/std': 0.09978143326365031, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8776137416179364, 'eval_rewards/repeat_penalty/std': 0.1198913437815813, 'eval_rewards/total_composite/mean': 0.6122724092923678, 'eval_rewards/total_composite/std': 0.35310536279128146, 'eval_reward': 0.6122724092923678, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.01914287731051445, 'eval_sampling/sampling_logp_difference/max': 1.0589995751014123, 'eval_sampling/importance_sampling_ratio/min': 0.36076818865079147, 'eval_sampling/importance_sampling_ratio/mean': 1.004656663307777, 'eval_sampling/importance_sampling_ratio/max': 1.4638345608344445, 'eval_entropy': 0.19041176713429964, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.6122724092923678, 'eval_reward_meter_mean': 0.7441063798390902, 'eval_reward_meter_std': 0.3844332993030548, 'eval_reward_count_adherence_mean': 0.9244015262677119, 'eval_reward_count_adherence_std': 0.09978143326365031, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8776137416179364, 'eval_reward_repeat_penalty_std': 0.1198913437815813, 'eval_reward_total_composite_mean': 0.6122724092923678, 'eval_reward_total_composite_std': 0.35310536279128146, 'epoch': 0.1}

 73%|███████▎  | 2400/3300 [3:23:40<1:18:54,  5.26s/it]

100%|██████████| 13/13 [01:05<00:00,  5.34s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 73%|███████▎  | 2401/3300 [3:23:48<6:48:02, 27.23s/it]2026-04-12 02:00:39,299 | INFO | train_grpo_train | metrics_logged mode=train step=2401

                                                       
{'loss': 0.0072, 'grad_norm': 2.6642069816589355, 'learning_rate': 2.7272727272727272e-06, 'num_tokens': 5430680.0, 'completions/mean_length': 65.375, 'completions/min_length': 65.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9385327100753784, 'rewards/meter/std': 0.10577936470508575, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9385327100753784, 'rewards/total_composite/std': 0.10577936470508575, 'reward': 0.9385327100753784, 'reward_std': 0.10577936470508575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02753767929971218, 'sampling/sampling_logp_difference/max': 0.5753829479217529, 'sampling/importance_sampling_ratio/min': 0.56248939037323, 'sampling/importance_sampling_ratio/mean': 1.0091955661773682, 'sampling/importance_sampling_ratio/max': 1.7399368286132812, 'entropy': 0.1730100642889738, 'clip_ratio/low_mean': 0.009360859869048, 'clip_ratio/low_min': 0.009360859869048, 'clip_ratio/high_mean': 0.01538461574818939, 'clip_ratio/high_max': 0.01538461574818939, 'clip_ratio/region_mean': 0.02474547561723739, 'reward_total_mean': 0.9385327100753784, 'reward_meter_mean': 0.9385327100753784, 'reward_meter_std': 0.10577936470508575, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9385327100753784, 'reward_total_composite_std': 0.10577936470508575, 'epoch': 0.1}

 73%|███████▎  | 2401/3300 [3:23:48<6:48:02, 27.23s/it]INFO 04-12 02:00:39 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2402/3300 [3:23:53<5:08:15, 20.60s/it]2026-04-12 02:00:44,409 | INFO | train_grpo_train | metrics_logged mode=train step=2402

                                                       
{'loss': 0.0097, 'grad_norm': 5.85192346572876, 'learning_rate': 2.724242424242424e-06, 'num_tokens': 5432706.0, 'completions/mean_length': 87.25, 'completions/min_length': 84.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.25, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9501332640647888, 'rewards/meter/std': 0.0800556018948555, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9501332640647888, 'rewards/total_composite/std': 0.0800556018948555, 'reward': 0.9501332640647888, 'reward_std': 0.0800556018948555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03878026455640793, 'sampling/sampling_logp_difference/max': 1.6748466491699219, 'sampling/importance_sampling_ratio/min': 0.18733689188957214, 'sampling/importance_sampling_ratio/mean': 0.9965555667877197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17476310580968857, 'clip_ratio/low_mean': 0.007134926971048117, 'clip_ratio/low_min': 0.007134926971048117, 'clip_ratio/high_mean': 0.028704919386655092, 'clip_ratio/high_max': 0.028704919386655092, 'clip_ratio/region_mean': 0.03583984635770321, 'reward_total_mean': 0.9501332640647888, 'reward_meter_mean': 0.9501332640647888, 'reward_meter_std': 0.0800556018948555, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9501332640647888, 'reward_total_composite_std': 0.0800556018948555, 'epoch': 0.1}

 73%|███████▎  | 2402/3300 [3:23:53<5:08:15, 20.60s/it]INFO 04-12 02:00:44 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2403/3300 [3:23:58<3:57:32, 15.89s/it]2026-04-12 02:00:49,313 | INFO | train_grpo_train | metrics_logged mode=train step=2403

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.7212121212121213e-06, 'num_tokens': 5434442.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00035840104101225734, 'sampling/sampling_logp_difference/max': 0.004777892027050257, 'sampling/importance_sampling_ratio/min': 0.9962913393974304, 'sampling/importance_sampling_ratio/mean': 1.0003162622451782, 'sampling/importance_sampling_ratio/max': 1.0047893524169922, 'entropy': 0.0028671720647253096, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 73%|███████▎  | 2403/3300 [3:23:58<3:57:32, 15.89s/it]INFO 04-12 02:00:49 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2404/3300 [3:24:08<3:32:25, 14.22s/it]2026-04-12 02:00:59,665 | INFO | train_grpo_train | metrics_logged mode=train step=2404

                                                       
{'loss': -0.1192, 'grad_norm': 0.281558632850647, 'learning_rate': 2.718181818181818e-06, 'num_tokens': 5437697.0, 'completions/mean_length': 506.875, 'completions/min_length': 489.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 498.3333435058594, 'completions/min_terminated_length': 489.0, 'completions/max_terminated_length': 512.0, 'rewards/meter/mean': 0.9645444750785828, 'rewards/meter/std': 0.09066132456064224, 'rewards/count_adherence/mean': 0.6907894611358643, 'rewards/count_adherence/std': 0.01860806532204151, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9167022705078125, 'rewards/repeat_penalty/std': 0.039733175188302994, 'rewards/total_composite/mean': 0.6099640130996704, 'rewards/total_composite/std': 0.05774744227528572, 'reward': 0.6099640130996704, 'reward_std': 0.057747457176446915, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03504529967904091, 'sampling/sampling_logp_difference/max': 2.621081829071045, 'sampling/importance_sampling_ratio/min': 0.07272414118051529, 'sampling/importance_sampling_ratio/mean': 1.0061771869659424, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08459514752030373, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010291029699146748, 'clip_ratio/high_max': 0.010291029699146748, 'clip_ratio/region_mean': 0.010291029699146748, 'reward_total_mean': 0.6099640130996704, 'reward_meter_mean': 0.9645444750785828, 'reward_meter_std': 0.09066132456064224, 'reward_count_adherence_mean': 0.6907894611358643, 'reward_count_adherence_std': 0.01860806532204151, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9167022705078125, 'reward_repeat_penalty_std': 0.039733175188302994, 'reward_total_composite_mean': 0.6099640130996704, 'reward_total_composite_std': 0.05774744227528572, 'epoch': 0.1}

 73%|███████▎  | 2404/3300 [3:24:08<3:32:25, 14.22s/it]INFO 04-12 02:00:59 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2405/3300 [3:24:13<2:48:53, 11.32s/it]2026-04-12 02:01:04,204 | INFO | train_grpo_train | metrics_logged mode=train step=2405

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.715151515151516e-06, 'num_tokens': 5439465.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00042045500595122576, 'sampling/sampling_logp_difference/max': 0.008271539583802223, 'sampling/importance_sampling_ratio/min': 0.9947773814201355, 'sampling/importance_sampling_ratio/mean': 1.0003942251205444, 'sampling/importance_sampling_ratio/max': 1.0083059072494507, 'entropy': 0.003762025124160573, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 73%|███████▎  | 2405/3300 [3:24:13<2:48:53, 11.32s/it]INFO 04-12 02:01:04 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2406/3300 [3:24:17<2:19:08,  9.34s/it]2026-04-12 02:01:08,923 | INFO | train_grpo_train | metrics_logged mode=train step=2406

                                                       
{'loss': 0.0059, 'grad_norm': 4.301211357116699, 'learning_rate': 2.7121212121212127e-06, 'num_tokens': 5441252.0, 'completions/mean_length': 67.375, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9776328802108765, 'rewards/meter/std': 0.05134515464305878, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9776328802108765, 'rewards/total_composite/std': 0.05134515464305878, 'reward': 0.9776328802108765, 'reward_std': 0.051345136016607285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02041078545153141, 'sampling/sampling_logp_difference/max': 1.1301698684692383, 'sampling/importance_sampling_ratio/min': 0.3229783773422241, 'sampling/importance_sampling_ratio/mean': 1.0020570755004883, 'sampling/importance_sampling_ratio/max': 1.4865719079971313, 'entropy': 0.13846461195498705, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005542142200283706, 'clip_ratio/high_max': 0.005542142200283706, 'clip_ratio/region_mean': 0.005542142200283706, 'reward_total_mean': 0.9776328802108765, 'reward_meter_mean': 0.9776328802108765, 'reward_meter_std': 0.05134515464305878, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9776328802108765, 'reward_total_composite_std': 0.05134515464305878, 'epoch': 0.1}

 73%|███████▎  | 2406/3300 [3:24:17<2:19:08,  9.34s/it]INFO 04-12 02:01:09 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2407/3300 [3:24:22<1:58:33,  7.97s/it]2026-04-12 02:01:13,672 | INFO | train_grpo_train | metrics_logged mode=train step=2407

                                                       
{'loss': 0.0436, 'grad_norm': 13.151045799255371, 'learning_rate': 2.7090909090909095e-06, 'num_tokens': 5443171.0, 'completions/mean_length': 65.875, 'completions/min_length': 61.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9381490349769592, 'rewards/meter/std': 0.009724577888846397, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9381490349769592, 'rewards/total_composite/std': 0.009724577888846397, 'reward': 0.9381490349769592, 'reward_std': 0.009724575094878674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08292688429355621, 'sampling/sampling_logp_difference/max': 2.107654094696045, 'sampling/importance_sampling_ratio/min': 0.12152271717786789, 'sampling/importance_sampling_ratio/mean': 0.9993854761123657, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.38669581711292267, 'clip_ratio/low_mean': 0.020202020648866892, 'clip_ratio/low_min': 0.020202020648866892, 'clip_ratio/high_mean': 0.06410424876958132, 'clip_ratio/high_max': 0.06410424876958132, 'clip_ratio/region_mean': 0.08430626941844821, 'reward_total_mean': 0.9381490349769592, 'reward_meter_mean': 0.9381490349769592, 'reward_meter_std': 0.009724577888846397, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9381490349769592, 'reward_total_composite_std': 0.009724577888846397, 'epoch': 0.1}

 73%|███████▎  | 2407/3300 [3:24:22<1:58:33,  7.97s/it]INFO 04-12 02:01:13 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2408/3300 [3:24:27<1:45:12,  7.08s/it]2026-04-12 02:01:18,678 | INFO | train_grpo_train | metrics_logged mode=train step=2408

                                                       
{'loss': -0.0002, 'grad_norm': 0.9304388165473938, 'learning_rate': 2.7060606060606063e-06, 'num_tokens': 5445313.0, 'completions/mean_length': 97.75, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.75, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9989974498748779, 'rewards/meter/std': 0.0009041029843501747, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989974498748779, 'rewards/total_composite/std': 0.0009041029843501747, 'reward': 0.9989974498748779, 'reward_std': 0.0009041179437190294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012694788165390491, 'sampling/sampling_logp_difference/max': 1.2907949686050415, 'sampling/importance_sampling_ratio/min': 0.2750520408153534, 'sampling/importance_sampling_ratio/mean': 0.9973183274269104, 'sampling/importance_sampling_ratio/max': 1.6353120803833008, 'entropy': 0.04637799598276615, 'clip_ratio/low_mean': 0.002577319508418441, 'clip_ratio/low_min': 0.002577319508418441, 'clip_ratio/high_mean': 0.01663423073478043, 'clip_ratio/high_max': 0.01663423073478043, 'clip_ratio/region_mean': 0.01921155024319887, 'reward_total_mean': 0.9989974498748779, 'reward_meter_mean': 0.9989974498748779, 'reward_meter_std': 0.0009041029843501747, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989974498748779, 'reward_total_composite_std': 0.0009041029843501747, 'epoch': 0.1}

 73%|███████▎  | 2408/3300 [3:24:27<1:45:12,  7.08s/it]INFO 04-12 02:01:18 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2409/3300 [3:24:33<1:39:29,  6.70s/it]2026-04-12 02:01:24,506 | INFO | train_grpo_train | metrics_logged mode=train step=2409

                                                       
{'loss': 0.0069, 'grad_norm': 2.528715133666992, 'learning_rate': 2.7030303030303036e-06, 'num_tokens': 5448370.0, 'completions/mean_length': 154.125, 'completions/min_length': 150.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.125, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.997596800327301, 'rewards/meter/std': 0.0010310488287359476, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9619561433792114, 'rewards/total_composite/std': 0.06579122692346573, 'reward': 0.9619561433792114, 'reward_std': 0.06579122692346573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028948156163096428, 'sampling/sampling_logp_difference/max': 1.3726117610931396, 'sampling/importance_sampling_ratio/min': 0.2534441351890564, 'sampling/importance_sampling_ratio/mean': 1.004657506942749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21369013004004955, 'clip_ratio/low_mean': 0.001623376621864736, 'clip_ratio/low_min': 0.001623376621864736, 'clip_ratio/high_mean': 0.02023474802263081, 'clip_ratio/high_max': 0.02023474802263081, 'clip_ratio/region_mean': 0.021858124644495547, 'reward_total_mean': 0.9619561433792114, 'reward_meter_mean': 0.997596800327301, 'reward_meter_std': 0.0010310488287359476, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9619561433792114, 'reward_total_composite_std': 0.06579122692346573, 'epoch': 0.1}

 73%|███████▎  | 2409/3300 [3:24:33<1:39:29,  6.70s/it]INFO 04-12 02:01:24 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2410/3300 [3:24:38<1:30:39,  6.11s/it]2026-04-12 02:01:29,237 | INFO | train_grpo_train | metrics_logged mode=train step=2410

                                                       
{'loss': -0.0112, 'grad_norm': 2.699066400527954, 'learning_rate': 2.7000000000000004e-06, 'num_tokens': 5450115.0, 'completions/mean_length': 67.125, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9942950010299683, 'rewards/meter/std': 0.0035084497649222612, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9942950010299683, 'rewards/total_composite/std': 0.0035084497649222612, 'reward': 0.9942950010299683, 'reward_std': 0.003508440451696515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02465406060218811, 'sampling/sampling_logp_difference/max': 1.0274019241333008, 'sampling/importance_sampling_ratio/min': 0.35793569684028625, 'sampling/importance_sampling_ratio/mean': 1.0015225410461426, 'sampling/importance_sampling_ratio/max': 1.9520822763442993, 'entropy': 0.1753720510751009, 'clip_ratio/low_mean': 0.007634902489371598, 'clip_ratio/low_min': 0.007634902489371598, 'clip_ratio/high_mean': 0.016363060451112688, 'clip_ratio/high_max': 0.016363060451112688, 'clip_ratio/region_mean': 0.023997962940484285, 'reward_total_mean': 0.9942950010299683, 'reward_meter_mean': 0.9942950010299683, 'reward_meter_std': 0.0035084497649222612, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9942950010299683, 'reward_total_composite_std': 0.0035084497649222612, 'epoch': 0.1}

 73%|███████▎  | 2410/3300 [3:24:38<1:30:39,  6.11s/it]INFO 04-12 02:01:29 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2411/3300 [3:24:43<1:29:09,  6.02s/it]2026-04-12 02:01:35,067 | INFO | train_grpo_train | metrics_logged mode=train step=2411

                                                       
{'loss': 0.0472, 'grad_norm': 7.062148571014404, 'learning_rate': 2.6969696969696972e-06, 'num_tokens': 5452646.0, 'completions/mean_length': 135.375, 'completions/min_length': 130.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.375, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9934861660003662, 'rewards/meter/std': 0.012767082080245018, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9934861660003662, 'rewards/total_composite/std': 0.012767082080245018, 'reward': 0.9934861660003662, 'reward_std': 0.012767070904374123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05575111508369446, 'sampling/sampling_logp_difference/max': 1.5884981155395508, 'sampling/importance_sampling_ratio/min': 0.2042321115732193, 'sampling/importance_sampling_ratio/mean': 1.0023579597473145, 'sampling/importance_sampling_ratio/max': 1.7682102918624878, 'entropy': 0.38587315008044243, 'clip_ratio/low_mean': 0.007401315961033106, 'clip_ratio/low_min': 0.007401315961033106, 'clip_ratio/high_mean': 0.043241268722340465, 'clip_ratio/high_max': 0.043241268722340465, 'clip_ratio/region_mean': 0.05064258468337357, 'reward_total_mean': 0.9934861660003662, 'reward_meter_mean': 0.9934861660003662, 'reward_meter_std': 0.012767082080245018, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9934861660003662, 'reward_total_composite_std': 0.012767082080245018, 'epoch': 0.1}

 73%|███████▎  | 2411/3300 [3:24:44<1:29:09,  6.02s/it]INFO 04-12 02:01:35 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2412/3300 [3:24:49<1:25:21,  5.77s/it]2026-04-12 02:01:40,214 | INFO | train_grpo_train | metrics_logged mode=train step=2412

                                                       
{'loss': 0.0204, 'grad_norm': 3.8710875511169434, 'learning_rate': 2.6939393939393945e-06, 'num_tokens': 5454382.0, 'completions/mean_length': 58.0, 'completions/min_length': 55.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9953839778900146, 'rewards/meter/std': 0.0009903458412736654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953839778900146, 'rewards/total_composite/std': 0.0009903458412736654, 'reward': 0.9953839778900146, 'reward_std': 0.0009903304744511843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02391960285604, 'sampling/sampling_logp_difference/max': 1.1877875328063965, 'sampling/importance_sampling_ratio/min': 0.44029319286346436, 'sampling/importance_sampling_ratio/mean': 1.004859447479248, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11028503353009, 'clip_ratio/low_mean': 0.015032077208161354, 'clip_ratio/low_min': 0.015032077208161354, 'clip_ratio/high_mean': 0.017077806871384382, 'clip_ratio/high_max': 0.017077806871384382, 'clip_ratio/region_mean': 0.032109884079545736, 'reward_total_mean': 0.9953839778900146, 'reward_meter_mean': 0.9953839778900146, 'reward_meter_std': 0.0009903458412736654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953839778900146, 'reward_total_composite_std': 0.0009903458412736654, 'epoch': 0.1}

 73%|███████▎  | 2412/3300 [3:24:49<1:25:21,  5.77s/it]INFO 04-12 02:01:40 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2413/3300 [3:24:57<1:37:56,  6.63s/it]2026-04-12 02:01:48,873 | INFO | train_grpo_train | metrics_logged mode=train step=2413

                                                       
{'loss': -0.0163, 'grad_norm': 1.422156810760498, 'learning_rate': 2.6909090909090913e-06, 'num_tokens': 5459457.0, 'completions/mean_length': 345.375, 'completions/min_length': 318.0, 'completions/max_length': 355.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 345.375, 'completions/min_terminated_length': 318.0, 'completions/max_terminated_length': 355.0, 'rewards/meter/mean': 0.9982728958129883, 'rewards/meter/std': 0.0010707140900194645, 'rewards/count_adherence/mean': 0.887499988079071, 'rewards/count_adherence/std': 0.035355325788259506, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9245098233222961, 'rewards/repeat_penalty/std': 0.06273634731769562, 'rewards/total_composite/mean': 0.8197740316390991, 'rewards/total_composite/std': 0.07304169237613678, 'reward': 0.8197740316390991, 'reward_std': 0.07304168492555618, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031854599714279175, 'sampling/sampling_logp_difference/max': 0.9705367088317871, 'sampling/importance_sampling_ratio/min': 0.3788796067237854, 'sampling/importance_sampling_ratio/mean': 1.0035998821258545, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2573400232940912, 'clip_ratio/low_mean': 0.006614301120862365, 'clip_ratio/low_min': 0.006614301120862365, 'clip_ratio/high_mean': 0.02032158919610083, 'clip_ratio/high_max': 0.02032158919610083, 'clip_ratio/region_mean': 0.026935890316963196, 'reward_total_mean': 0.8197740316390991, 'reward_meter_mean': 0.9982728958129883, 'reward_meter_std': 0.0010707140900194645, 'reward_count_adherence_mean': 0.887499988079071, 'reward_count_adherence_std': 0.035355325788259506, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9245098233222961, 'reward_repeat_penalty_std': 0.06273634731769562, 'reward_total_composite_mean': 0.8197740316390991, 'reward_total_composite_std': 0.07304169237613678, 'epoch': 0.1}

 73%|███████▎  | 2413/3300 [3:24:57<1:37:56,  6.63s/it]INFO 04-12 02:01:49 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2414/3300 [3:25:02<1:29:36,  6.07s/it]2026-04-12 02:01:53,649 | INFO | train_grpo_train | metrics_logged mode=train step=2414

                                                       
{'loss': -0.0, 'grad_norm': 0.009874632582068443, 'learning_rate': 2.687878787878788e-06, 'num_tokens': 5461241.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981490969657898, 'rewards/meter/std': 8.935132427723147e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981490969657898, 'rewards/total_composite/std': 8.935132427723147e-06, 'reward': 0.9981490969657898, 'reward_std': 8.935131518228445e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0008906561997719109, 'sampling/sampling_logp_difference/max': 0.16530299186706543, 'sampling/importance_sampling_ratio/min': 0.980025053024292, 'sampling/importance_sampling_ratio/mean': 1.0007779598236084, 'sampling/importance_sampling_ratio/max': 1.1797505617141724, 'entropy': 0.006271830061450601, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.9981490969657898, 'reward_meter_mean': 0.9981490969657898, 'reward_meter_std': 8.935132427723147e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981490969657898, 'reward_total_composite_std': 8.935132427723147e-06, 'epoch': 0.1}

 73%|███████▎  | 2414/3300 [3:25:02<1:29:36,  6.07s/it]INFO 04-12 02:01:53 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2415/3300 [3:25:07<1:22:30,  5.59s/it]2026-04-12 02:01:58,106 | INFO | train_grpo_train | metrics_logged mode=train step=2415

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.684848484848485e-06, 'num_tokens': 5462745.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00018749816808849573, 'sampling/sampling_logp_difference/max': 0.003093225881457329, 'sampling/importance_sampling_ratio/min': 0.9994993209838867, 'sampling/importance_sampling_ratio/mean': 1.000183343887329, 'sampling/importance_sampling_ratio/max': 1.0030980110168457, 'entropy': 0.001466380010242574, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 73%|███████▎  | 2415/3300 [3:25:07<1:22:30,  5.59s/it]INFO 04-12 02:01:58 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2416/3300 [3:25:12<1:20:29,  5.46s/it]2026-04-12 02:02:03,260 | INFO | train_grpo_train | metrics_logged mode=train step=2416

                                                       
{'loss': 0.0002, 'grad_norm': 0.24654512107372284, 'learning_rate': 2.6818181818181822e-06, 'num_tokens': 5464841.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9980032444000244, 'rewards/meter/std': 8.327968316734768e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980032444000244, 'rewards/total_composite/std': 8.327968316734768e-06, 'reward': 0.9980032444000244, 'reward_std': 8.322024768858682e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0037318698596209288, 'sampling/sampling_logp_difference/max': 0.787595272064209, 'sampling/importance_sampling_ratio/min': 0.45493748784065247, 'sampling/importance_sampling_ratio/mean': 1.0002275705337524, 'sampling/importance_sampling_ratio/max': 1.1473532915115356, 'entropy': 0.023096754681319, 'clip_ratio/low_mean': 0.0012755101779475808, 'clip_ratio/low_min': 0.0012755101779475808, 'clip_ratio/high_mean': 0.0012755101779475808, 'clip_ratio/high_max': 0.0012755101779475808, 'clip_ratio/region_mean': 0.0025510203558951616, 'reward_total_mean': 0.9980032444000244, 'reward_meter_mean': 0.9980032444000244, 'reward_meter_std': 8.327968316734768e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980032444000244, 'reward_total_composite_std': 8.327968316734768e-06, 'epoch': 0.1}

 73%|███████▎  | 2416/3300 [3:25:12<1:20:29,  5.46s/it]INFO 04-12 02:02:03 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2417/3300 [3:25:16<1:15:13,  5.11s/it]2026-04-12 02:02:07,544 | INFO | train_grpo_train | metrics_logged mode=train step=2417

                                                       
{'loss': 0.0088, 'grad_norm': 3.3394386768341064, 'learning_rate': 2.678787878787879e-06, 'num_tokens': 5466420.0, 'completions/mean_length': 35.375, 'completions/min_length': 34.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.999369204044342, 'rewards/meter/std': 0.0002697483287192881, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999369204044342, 'rewards/total_composite/std': 0.0002697483287192881, 'reward': 0.999369204044342, 'reward_std': 0.00026974434149451554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02549952082335949, 'sampling/sampling_logp_difference/max': 0.7178728580474854, 'sampling/importance_sampling_ratio/min': 0.48778876662254333, 'sampling/importance_sampling_ratio/mean': 1.0062538385391235, 'sampling/importance_sampling_ratio/max': 1.802677035331726, 'entropy': 0.18990459106862545, 'clip_ratio/low_mean': 0.006756756920367479, 'clip_ratio/low_min': 0.006756756920367479, 'clip_ratio/high_mean': 0.02460317499935627, 'clip_ratio/high_max': 0.02460317499935627, 'clip_ratio/region_mean': 0.03135993191972375, 'reward_total_mean': 0.999369204044342, 'reward_meter_mean': 0.999369204044342, 'reward_meter_std': 0.0002697483287192881, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999369204044342, 'reward_total_composite_std': 0.0002697483287192881, 'epoch': 0.1}

 73%|███████▎  | 2417/3300 [3:25:16<1:15:13,  5.11s/it]INFO 04-12 02:02:07 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2418/3300 [3:25:21<1:13:47,  5.02s/it]2026-04-12 02:02:12,353 | INFO | train_grpo_train | metrics_logged mode=train step=2418

                                                       
{'loss': 0.0167, 'grad_norm': 11.102910041809082, 'learning_rate': 2.675757575757576e-06, 'num_tokens': 5468444.0, 'completions/mean_length': 82.0, 'completions/min_length': 79.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.0, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.8613059520721436, 'rewards/meter/std': 0.17161336541175842, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.8016860485076904, 'rewards/total_composite/std': 0.17636200785636902, 'reward': 0.8016860485076904, 'reward_std': 0.1763620227575302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08092934638261795, 'sampling/sampling_logp_difference/max': 2.628443717956543, 'sampling/importance_sampling_ratio/min': 0.07219072431325912, 'sampling/importance_sampling_ratio/mean': 0.9934130311012268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3536304421722889, 'clip_ratio/low_mean': 0.012442129664123058, 'clip_ratio/low_min': 0.012442129664123058, 'clip_ratio/high_mean': 0.05160611355677247, 'clip_ratio/high_max': 0.05160611355677247, 'clip_ratio/region_mean': 0.06404824322089553, 'reward_total_mean': 0.8016860485076904, 'reward_meter_mean': 0.8613059520721436, 'reward_meter_std': 0.17161336541175842, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.8016860485076904, 'reward_total_composite_std': 0.17636200785636902, 'epoch': 0.1}

 73%|███████▎  | 2418/3300 [3:25:21<1:13:47,  5.02s/it]INFO 04-12 02:02:12 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2419/3300 [3:25:26<1:14:18,  5.06s/it]2026-04-12 02:02:17,563 | INFO | train_grpo_train | metrics_logged mode=train step=2419

                                                       
{'loss': -0.0001, 'grad_norm': 0.25810882449150085, 'learning_rate': 2.6727272727272727e-06, 'num_tokens': 5470156.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7876332998275757, 'rewards/meter/std': 1.5573279597447254e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7876332998275757, 'rewards/total_composite/std': 1.5573279597447254e-05, 'reward': 0.7876332998275757, 'reward_std': 1.558229632792063e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002996515017002821, 'sampling/sampling_logp_difference/max': 1.1518197059631348, 'sampling/importance_sampling_ratio/min': 0.3160611093044281, 'sampling/importance_sampling_ratio/mean': 0.9987244606018066, 'sampling/importance_sampling_ratio/max': 1.006128191947937, 'entropy': 0.00269945221953094, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7876332998275757, 'reward_meter_mean': 0.7876332998275757, 'reward_meter_std': 1.5573279597447254e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7876332998275757, 'reward_total_composite_std': 1.5573279597447254e-05, 'epoch': 0.1}

 73%|███████▎  | 2419/3300 [3:25:26<1:14:18,  5.06s/it]INFO 04-12 02:02:17 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2420/3300 [3:25:32<1:18:47,  5.37s/it]2026-04-12 02:02:23,606 | INFO | train_grpo_train | metrics_logged mode=train step=2420

                                                       
{'loss': 0.005, 'grad_norm': 3.2308475971221924, 'learning_rate': 2.66969696969697e-06, 'num_tokens': 5472925.0, 'completions/mean_length': 169.125, 'completions/min_length': 163.0, 'completions/max_length': 176.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 169.125, 'completions/min_terminated_length': 163.0, 'completions/max_terminated_length': 176.0, 'rewards/meter/mean': 0.9988895654678345, 'rewards/meter/std': 0.00021211641433183104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.9572727680206299, 'rewards/total_composite/std': 0.057507775723934174, 'reward': 0.9572727680206299, 'reward_std': 0.05750780552625656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04539603367447853, 'sampling/sampling_logp_difference/max': 1.4913952350616455, 'sampling/importance_sampling_ratio/min': 0.2250584363937378, 'sampling/importance_sampling_ratio/mean': 1.0061159133911133, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3278270773589611, 'clip_ratio/low_mean': 0.011130106868222356, 'clip_ratio/low_min': 0.011130106868222356, 'clip_ratio/high_mean': 0.028763922629877925, 'clip_ratio/high_max': 0.028763922629877925, 'clip_ratio/region_mean': 0.03989402949810028, 'reward_total_mean': 0.9572727680206299, 'reward_meter_mean': 0.9988895654678345, 'reward_meter_std': 0.00021211641433183104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.9572727680206299, 'reward_total_composite_std': 0.057507775723934174, 'epoch': 0.1}

 73%|███████▎  | 2420/3300 [3:25:32<1:18:47,  5.37s/it]INFO 04-12 02:02:23 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2421/3300 [3:25:37<1:17:56,  5.32s/it]2026-04-12 02:02:28,805 | INFO | train_grpo_train | metrics_logged mode=train step=2421

                                                       
{'loss': -0.0028, 'grad_norm': 0.6984437704086304, 'learning_rate': 2.666666666666667e-06, 'num_tokens': 5475014.0, 'completions/mean_length': 106.125, 'completions/min_length': 105.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.125, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9992530345916748, 'rewards/meter/std': 0.00013918116746935993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992530345916748, 'rewards/total_composite/std': 0.00013918116746935993, 'reward': 0.9992530345916748, 'reward_std': 0.0001391761179547757, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008111930452287197, 'sampling/sampling_logp_difference/max': 0.6127934455871582, 'sampling/importance_sampling_ratio/min': 0.5418351888656616, 'sampling/importance_sampling_ratio/mean': 1.000658392906189, 'sampling/importance_sampling_ratio/max': 1.1507465839385986, 'entropy': 0.056452156975865364, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.004716981202363968, 'clip_ratio/high_max': 0.004716981202363968, 'clip_ratio/region_mean': 0.004716981202363968, 'reward_total_mean': 0.9992530345916748, 'reward_meter_mean': 0.9992530345916748, 'reward_meter_std': 0.00013918116746935993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992530345916748, 'reward_total_composite_std': 0.00013918116746935993, 'epoch': 0.1}

 73%|███████▎  | 2421/3300 [3:25:37<1:17:56,  5.32s/it]INFO 04-12 02:02:29 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2422/3300 [3:25:43<1:18:28,  5.36s/it]2026-04-12 02:02:34,269 | INFO | train_grpo_train | metrics_logged mode=train step=2422

                                                       
{'loss': 0.0142, 'grad_norm': 2.5043861865997314, 'learning_rate': 2.6636363636363637e-06, 'num_tokens': 5477412.0, 'completions/mean_length': 134.75, 'completions/min_length': 132.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.75, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9812095761299133, 'rewards/meter/std': 0.03474520519375801, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.1079898476600647, 'rewards/total_composite/mean': 0.841211199760437, 'rewards/total_composite/std': 0.1127474382519722, 'reward': 0.841211199760437, 'reward_std': 0.1127474457025528, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026558201760053635, 'sampling/sampling_logp_difference/max': 1.1920127868652344, 'sampling/importance_sampling_ratio/min': 0.36081838607788086, 'sampling/importance_sampling_ratio/mean': 1.0124142169952393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23221887275576591, 'clip_ratio/low_mean': 0.010076848906464875, 'clip_ratio/low_min': 0.010076848906464875, 'clip_ratio/high_mean': 0.012170514499302953, 'clip_ratio/high_max': 0.012170514499302953, 'clip_ratio/region_mean': 0.022247363405767828, 'reward_total_mean': 0.841211199760437, 'reward_meter_mean': 0.9812095761299133, 'reward_meter_std': 0.03474520519375801, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.1079898476600647, 'reward_total_composite_mean': 0.841211199760437, 'reward_total_composite_std': 0.1127474382519722, 'epoch': 0.1}

 73%|███████▎  | 2422/3300 [3:25:43<1:18:28,  5.36s/it]INFO 04-12 02:02:34 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2423/3300 [3:25:47<1:15:21,  5.16s/it]2026-04-12 02:02:38,938 | INFO | train_grpo_train | metrics_logged mode=train step=2423

                                                       
{'loss': 0.0077, 'grad_norm': 5.193559169769287, 'learning_rate': 2.660606060606061e-06, 'num_tokens': 5479245.0, 'completions/mean_length': 64.125, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9993478059768677, 'rewards/meter/std': 0.0001315748959314078, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993478059768677, 'rewards/total_composite/std': 0.0001315748959314078, 'reward': 0.9993478059768677, 'reward_std': 0.00013158157526049763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006412571296095848, 'sampling/sampling_logp_difference/max': 1.5588197708129883, 'sampling/importance_sampling_ratio/min': 0.21038421988487244, 'sampling/importance_sampling_ratio/mean': 0.9980660080909729, 'sampling/importance_sampling_ratio/max': 1.234817385673523, 'entropy': 0.021533598424866796, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.009765625, 'clip_ratio/high_max': 0.009765625, 'clip_ratio/region_mean': 0.009765625, 'reward_total_mean': 0.9993478059768677, 'reward_meter_mean': 0.9993478059768677, 'reward_meter_std': 0.0001315748959314078, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993478059768677, 'reward_total_composite_std': 0.0001315748959314078, 'epoch': 0.1}

 73%|███████▎  | 2423/3300 [3:25:47<1:15:21,  5.16s/it]INFO 04-12 02:02:39 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2424/3300 [3:25:57<1:34:41,  6.49s/it]2026-04-12 02:02:48,529 | INFO | train_grpo_train | metrics_logged mode=train step=2424

                                                       
{'loss': 0.1719, 'grad_norm': 1.59493887424469, 'learning_rate': 2.6575757575757577e-06, 'num_tokens': 5481641.0, 'completions/mean_length': 187.5, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 79.33333587646484, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9973663687705994, 'rewards/meter/std': 0.0009385403245687485, 'rewards/count_adherence/mean': 0.23749999701976776, 'rewards/count_adherence/std': 0.25460052490234375, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9709615707397461, 'rewards/repeat_penalty/std': 0.06743928045034409, 'rewards/total_composite/mean': 0.21816782653331757, 'rewards/total_composite/std': 0.22094322741031647, 'reward': 0.21816782653331757, 'reward_std': 0.22094321250915527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041650205850601196, 'sampling/sampling_logp_difference/max': 1.7469613552093506, 'sampling/importance_sampling_ratio/min': 0.17430277168750763, 'sampling/importance_sampling_ratio/mean': 0.99797523021698, 'sampling/importance_sampling_ratio/max': 1.734250783920288, 'entropy': 0.1750364489853382, 'clip_ratio/low_mean': 0.022042680298909545, 'clip_ratio/low_min': 0.022042680298909545, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.022042680298909545, 'reward_total_mean': 0.21816782653331757, 'reward_meter_mean': 0.9973663687705994, 'reward_meter_std': 0.0009385403245687485, 'reward_count_adherence_mean': 0.23749999701976776, 'reward_count_adherence_std': 0.25460052490234375, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9709615707397461, 'reward_repeat_penalty_std': 0.06743928045034409, 'reward_total_composite_mean': 0.21816782653331757, 'reward_total_composite_std': 0.22094322741031647, 'epoch': 0.1}

 73%|███████▎  | 2424/3300 [3:25:57<1:34:41,  6.49s/it]INFO 04-12 02:02:48 [block_pool.py:378] Successfully reset prefix cache

 73%|███████▎  | 2425/3300 [3:26:02<1:28:23,  6.06s/it]2026-04-12 02:02:53,600 | INFO | train_grpo_train | metrics_logged mode=train step=2425

                                                       
{'loss': 0.0094, 'grad_norm': 3.781071186065674, 'learning_rate': 2.6545454545454546e-06, 'num_tokens': 5483726.0, 'completions/mean_length': 101.625, 'completions/min_length': 98.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.625, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.8708481192588806, 'rewards/meter/std': 0.34927064180374146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8708481192588806, 'rewards/total_composite/std': 0.34927064180374146, 'reward': 0.8708481192588806, 'reward_std': 0.34927064180374146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02785804308950901, 'sampling/sampling_logp_difference/max': 0.8949475288391113, 'sampling/importance_sampling_ratio/min': 0.40862902998924255, 'sampling/importance_sampling_ratio/mean': 1.0064235925674438, 'sampling/importance_sampling_ratio/max': 1.9378726482391357, 'entropy': 0.20133807510137558, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.017109742388129234, 'clip_ratio/high_max': 0.017109742388129234, 'clip_ratio/region_mean': 0.02078621299006045, 'reward_total_mean': 0.8708481192588806, 'reward_meter_mean': 0.8708481192588806, 'reward_meter_std': 0.34927064180374146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8708481192588806, 'reward_total_composite_std': 0.34927064180374146, 'epoch': 0.1}

 73%|███████▎  | 2425/3300 [3:26:02<1:28:23,  6.06s/it]INFO 04-12 02:02:53 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2426/3300 [3:26:07<1:22:02,  5.63s/it]2026-04-12 02:02:58,232 | INFO | train_grpo_train | metrics_logged mode=train step=2426

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.6515151515151514e-06, 'num_tokens': 5485422.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00044331722892820835, 'sampling/sampling_logp_difference/max': 0.03401622548699379, 'sampling/importance_sampling_ratio/min': 0.9665557742118835, 'sampling/importance_sampling_ratio/mean': 1.0000972747802734, 'sampling/importance_sampling_ratio/max': 1.0289011001586914, 'entropy': 0.004820480738999322, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▎  | 2426/3300 [3:26:07<1:22:02,  5.63s/it]INFO 04-12 02:02:58 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2427/3300 [3:26:11<1:17:40,  5.34s/it]2026-04-12 02:03:02,881 | INFO | train_grpo_train | metrics_logged mode=train step=2427

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.6484848484848487e-06, 'num_tokens': 5487246.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0009743176633492112, 'sampling/sampling_logp_difference/max': 0.0836307555437088, 'sampling/importance_sampling_ratio/min': 0.9197707772254944, 'sampling/importance_sampling_ratio/mean': 1.0001752376556396, 'sampling/importance_sampling_ratio/max': 1.0430419445037842, 'entropy': 0.011770866345614195, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▎  | 2427/3300 [3:26:11<1:17:40,  5.34s/it]INFO 04-12 02:03:03 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2428/3300 [3:26:15<1:12:23,  4.98s/it]2026-04-12 02:03:07,032 | INFO | train_grpo_train | metrics_logged mode=train step=2428

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.6454545454545455e-06, 'num_tokens': 5488686.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.007659213151782751, 'sampling/sampling_logp_difference/max': 0.8227400779724121, 'sampling/importance_sampling_ratio/min': 0.4392264783382416, 'sampling/importance_sampling_ratio/mean': 0.9964240789413452, 'sampling/importance_sampling_ratio/max': 1.0674165487289429, 'entropy': 0.01067307055927813, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▎  | 2428/3300 [3:26:16<1:12:23,  4.98s/it]INFO 04-12 02:03:07 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2429/3300 [3:26:20<1:10:28,  4.85s/it]2026-04-12 02:03:11,592 | INFO | train_grpo_train | metrics_logged mode=train step=2429

                                                       
{'loss': 0.0, 'grad_norm': 0.09379827976226807, 'learning_rate': 2.6424242424242423e-06, 'num_tokens': 5490494.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993983507156372, 'rewards/meter/std': 3.034573182958411e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993983507156372, 'rewards/total_composite/std': 3.034573182958411e-06, 'reward': 0.9993983507156372, 'reward_std': 3.034573182958411e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0012968340888619423, 'sampling/sampling_logp_difference/max': 0.16098451614379883, 'sampling/importance_sampling_ratio/min': 0.9115949869155884, 'sampling/importance_sampling_ratio/mean': 1.000643253326416, 'sampling/importance_sampling_ratio/max': 1.1746667623519897, 'entropy': 0.011805565096437931, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993983507156372, 'reward_meter_mean': 0.9993983507156372, 'reward_meter_std': 3.034573182958411e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993983507156372, 'reward_total_composite_std': 3.034573182958411e-06, 'epoch': 0.1}

 74%|███████▎  | 2429/3300 [3:26:20<1:10:28,  4.85s/it]INFO 04-12 02:03:11 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2430/3300 [3:26:25<1:09:33,  4.80s/it]2026-04-12 02:03:16,255 | INFO | train_grpo_train | metrics_logged mode=train step=2430

                                                       
{'loss': 0.0007, 'grad_norm': 5.21270751953125, 'learning_rate': 2.6393939393939396e-06, 'num_tokens': 5492246.0, 'completions/mean_length': 67.0, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9988259077072144, 'rewards/meter/std': 0.0006359845283441246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988259077072144, 'rewards/total_composite/std': 0.0006359845283441246, 'reward': 0.9988259077072144, 'reward_std': 0.0006359879043884575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05176051706075668, 'sampling/sampling_logp_difference/max': 1.786757230758667, 'sampling/importance_sampling_ratio/min': 0.16750246286392212, 'sampling/importance_sampling_ratio/mean': 1.002232551574707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2827076967805624, 'clip_ratio/low_mean': 0.009443168761208653, 'clip_ratio/low_min': 0.009443168761208653, 'clip_ratio/high_mean': 0.02971602266188711, 'clip_ratio/high_max': 0.02971602266188711, 'clip_ratio/region_mean': 0.03915919142309576, 'reward_total_mean': 0.9988259077072144, 'reward_meter_mean': 0.9988259077072144, 'reward_meter_std': 0.0006359845283441246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988259077072144, 'reward_total_composite_std': 0.0006359845283441246, 'epoch': 0.1}

 74%|███████▎  | 2430/3300 [3:26:25<1:09:33,  4.80s/it]INFO 04-12 02:03:16 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2431/3300 [3:26:30<1:11:42,  4.95s/it]2026-04-12 02:03:21,567 | INFO | train_grpo_train | metrics_logged mode=train step=2431

                                                       
{'loss': 0.0117, 'grad_norm': 3.1840193271636963, 'learning_rate': 2.6363636363636364e-06, 'num_tokens': 5494352.0, 'completions/mean_length': 102.25, 'completions/min_length': 98.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.25, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.998919665813446, 'rewards/meter/std': 0.00028825365006923676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998919665813446, 'rewards/total_composite/std': 0.00028825365006923676, 'reward': 0.998919665813446, 'reward_std': 0.0002882632543332875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03998712822794914, 'sampling/sampling_logp_difference/max': 1.0248355865478516, 'sampling/importance_sampling_ratio/min': 0.3685198724269867, 'sampling/importance_sampling_ratio/mean': 1.0071525573730469, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3069701548665762, 'clip_ratio/low_mean': 0.019234189530834556, 'clip_ratio/low_min': 0.019234189530834556, 'clip_ratio/high_mean': 0.023538942215964198, 'clip_ratio/high_max': 0.023538942215964198, 'clip_ratio/region_mean': 0.042773131746798754, 'reward_total_mean': 0.998919665813446, 'reward_meter_mean': 0.998919665813446, 'reward_meter_std': 0.00028825365006923676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998919665813446, 'reward_total_composite_std': 0.00028825365006923676, 'epoch': 0.1}

 74%|███████▎  | 2431/3300 [3:26:30<1:11:42,  4.95s/it]INFO 04-12 02:03:21 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2432/3300 [3:26:37<1:20:36,  5.57s/it]2026-04-12 02:03:28,590 | INFO | train_grpo_train | metrics_logged mode=train step=2432

                                                       
{'loss': 0.0053, 'grad_norm': 0.9189408421516418, 'learning_rate': 2.6333333333333332e-06, 'num_tokens': 5497991.0, 'completions/mean_length': 246.875, 'completions/min_length': 245.0, 'completions/max_length': 250.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 246.875, 'completions/min_terminated_length': 245.0, 'completions/max_terminated_length': 250.0, 'rewards/meter/mean': 0.999138593673706, 'rewards/meter/std': 0.00011935058864764869, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7307692766189575, 'rewards/repeat_penalty/std': 0.041117113083601, 'rewards/total_composite/mean': 0.7301368713378906, 'rewards/total_composite/std': 0.041023507714271545, 'reward': 0.7301368713378906, 'reward_std': 0.041023530066013336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013350401073694229, 'sampling/sampling_logp_difference/max': 1.0539226531982422, 'sampling/importance_sampling_ratio/min': 0.3485677242279053, 'sampling/importance_sampling_ratio/mean': 1.0021779537200928, 'sampling/importance_sampling_ratio/max': 1.3559163808822632, 'entropy': 0.08605087455362082, 'clip_ratio/low_mean': 0.0015080645098350942, 'clip_ratio/low_min': 0.0015080645098350942, 'clip_ratio/high_mean': 0.0035590012557804585, 'clip_ratio/high_max': 0.0035590012557804585, 'clip_ratio/region_mean': 0.005067065765615553, 'reward_total_mean': 0.7301368713378906, 'reward_meter_mean': 0.999138593673706, 'reward_meter_std': 0.00011935058864764869, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7307692766189575, 'reward_repeat_penalty_std': 0.041117113083601, 'reward_total_composite_mean': 0.7301368713378906, 'reward_total_composite_std': 0.041023507714271545, 'epoch': 0.1}

 74%|███████▎  | 2432/3300 [3:26:37<1:20:36,  5.57s/it]INFO 04-12 02:03:28 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▎  | 2433/3300 [3:26:43<1:23:02,  5.75s/it]2026-04-12 02:03:34,744 | INFO | train_grpo_train | metrics_logged mode=train step=2433

                                                       
{'loss': -0.0002, 'grad_norm': 0.08106043934822083, 'learning_rate': 2.63030303030303e-06, 'num_tokens': 5500582.0, 'completions/mean_length': 176.875, 'completions/min_length': 176.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 176.875, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.9992530345916748, 'rewards/meter/std': 1.491289003752172e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7771967649459839, 'rewards/total_composite/std': 1.1599345270951744e-05, 'reward': 0.7771967649459839, 'reward_std': 1.1595264368224889e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007231563795357943, 'sampling/sampling_logp_difference/max': 0.7993268966674805, 'sampling/importance_sampling_ratio/min': 0.4496315121650696, 'sampling/importance_sampling_ratio/mean': 1.002611517906189, 'sampling/importance_sampling_ratio/max': 1.2289493083953857, 'entropy': 0.05232963711023331, 'clip_ratio/low_mean': 0.0021186440717428923, 'clip_ratio/low_min': 0.0021186440717428923, 'clip_ratio/high_mean': 0.0014204545877873898, 'clip_ratio/high_max': 0.0014204545877873898, 'clip_ratio/region_mean': 0.003539098659530282, 'reward_total_mean': 0.7771967649459839, 'reward_meter_mean': 0.9992530345916748, 'reward_meter_std': 1.491289003752172e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7771967649459839, 'reward_total_composite_std': 1.1599345270951744e-05, 'epoch': 0.1}

 74%|███████▎  | 2433/3300 [3:26:43<1:23:02,  5.75s/it]INFO 04-12 02:03:34 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2434/3300 [3:26:48<1:18:17,  5.42s/it]2026-04-12 02:03:39,422 | INFO | train_grpo_train | metrics_logged mode=train step=2434

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.6272727272727278e-06, 'num_tokens': 5502654.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005770151037722826, 'sampling/sampling_logp_difference/max': 0.009044456295669079, 'sampling/importance_sampling_ratio/min': 0.9965153932571411, 'sampling/importance_sampling_ratio/mean': 1.0005393028259277, 'sampling/importance_sampling_ratio/max': 1.0090855360031128, 'entropy': 0.005529700662009418, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▍  | 2434/3300 [3:26:48<1:18:17,  5.42s/it]INFO 04-12 02:03:39 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2435/3300 [3:26:53<1:16:49,  5.33s/it]2026-04-12 02:03:44,518 | INFO | train_grpo_train | metrics_logged mode=train step=2435

                                                       
{'loss': 0.003, 'grad_norm': 5.022347927093506, 'learning_rate': 2.6242424242424246e-06, 'num_tokens': 5504823.0, 'completions/mean_length': 90.125, 'completions/min_length': 84.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.125, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9843810200691223, 'rewards/meter/std': 0.03011130914092064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9594758749008179, 'rewards/total_composite/std': 0.07208064943552017, 'reward': 0.9594758749008179, 'reward_std': 0.07208065688610077, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03877629339694977, 'sampling/sampling_logp_difference/max': 1.5140687227249146, 'sampling/importance_sampling_ratio/min': 0.22001299262046814, 'sampling/importance_sampling_ratio/mean': 1.0028375387191772, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2525391634553671, 'clip_ratio/low_mean': 0.00696005008649081, 'clip_ratio/low_min': 0.00696005008649081, 'clip_ratio/high_mean': 0.03053560631815344, 'clip_ratio/high_max': 0.03053560631815344, 'clip_ratio/region_mean': 0.03749565640464425, 'reward_total_mean': 0.9594758749008179, 'reward_meter_mean': 0.9843810200691223, 'reward_meter_std': 0.03011130914092064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9594758749008179, 'reward_total_composite_std': 0.07208064943552017, 'epoch': 0.1}

 74%|███████▍  | 2435/3300 [3:26:53<1:16:49,  5.33s/it]INFO 04-12 02:03:44 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2436/3300 [3:26:58<1:14:07,  5.15s/it]2026-04-12 02:03:49,243 | INFO | train_grpo_train | metrics_logged mode=train step=2436

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.621212121212122e-06, 'num_tokens': 5506655.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0004700902500189841, 'sampling/sampling_logp_difference/max': 0.009367566555738449, 'sampling/importance_sampling_ratio/min': 0.9982850551605225, 'sampling/importance_sampling_ratio/mean': 1.0004562139511108, 'sampling/importance_sampling_ratio/max': 1.0094116926193237, 'entropy': 0.004693651368143037, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▍  | 2436/3300 [3:26:58<1:14:07,  5.15s/it]INFO 04-12 02:03:49 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2437/3300 [3:27:03<1:13:55,  5.14s/it]2026-04-12 02:03:54,361 | INFO | train_grpo_train | metrics_logged mode=train step=2437

                                                       
{'loss': 0.006, 'grad_norm': 0.9422140717506409, 'learning_rate': 2.6181818181818187e-06, 'num_tokens': 5508881.0, 'completions/mean_length': 100.25, 'completions/min_length': 98.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.25, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9990655183792114, 'rewards/meter/std': 0.00015276693738996983, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990655183792114, 'rewards/total_composite/std': 0.00015276693738996983, 'reward': 0.9990655183792114, 'reward_std': 0.00015277891361620277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010102971456944942, 'sampling/sampling_logp_difference/max': 1.0181710720062256, 'sampling/importance_sampling_ratio/min': 0.36125504970550537, 'sampling/importance_sampling_ratio/mean': 0.9993890523910522, 'sampling/importance_sampling_ratio/max': 1.4387589693069458, 'entropy': 0.04010937362909317, 'clip_ratio/low_mean': 0.00376287626568228, 'clip_ratio/low_min': 0.00376287626568228, 'clip_ratio/high_mean': 0.003763133892789483, 'clip_ratio/high_max': 0.003763133892789483, 'clip_ratio/region_mean': 0.007526010158471763, 'reward_total_mean': 0.9990655183792114, 'reward_meter_mean': 0.9990655183792114, 'reward_meter_std': 0.00015276693738996983, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990655183792114, 'reward_total_composite_std': 0.00015276693738996983, 'epoch': 0.1}

 74%|███████▍  | 2437/3300 [3:27:03<1:13:55,  5.14s/it]INFO 04-12 02:03:54 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2438/3300 [3:27:07<1:11:15,  4.96s/it]2026-04-12 02:03:58,902 | INFO | train_grpo_train | metrics_logged mode=train step=2438

                                                       
{'loss': 0.0011, 'grad_norm': 1.3743865489959717, 'learning_rate': 2.6151515151515155e-06, 'num_tokens': 5510611.0, 'completions/mean_length': 64.25, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9993854761123657, 'rewards/meter/std': 2.579814099590294e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993854761123657, 'rewards/total_composite/std': 2.579814099590294e-05, 'reward': 0.9993854761123657, 'reward_std': 2.5807335987337865e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0030143873300403357, 'sampling/sampling_logp_difference/max': 0.5796234607696533, 'sampling/importance_sampling_ratio/min': 0.5601092576980591, 'sampling/importance_sampling_ratio/mean': 1.0008904933929443, 'sampling/importance_sampling_ratio/max': 1.3340281248092651, 'entropy': 0.010751945606898516, 'clip_ratio/low_mean': 0.001923076924867928, 'clip_ratio/low_min': 0.001923076924867928, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.001923076924867928, 'reward_total_mean': 0.9993854761123657, 'reward_meter_mean': 0.9993854761123657, 'reward_meter_std': 2.579814099590294e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993854761123657, 'reward_total_composite_std': 2.579814099590294e-05, 'epoch': 0.1}

 74%|███████▍  | 2438/3300 [3:27:07<1:11:15,  4.96s/it]INFO 04-12 02:03:59 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2439/3300 [3:27:13<1:13:35,  5.13s/it]2026-04-12 02:04:04,429 | INFO | train_grpo_train | metrics_logged mode=train step=2439

                                                       
{'loss': 0.0054, 'grad_norm': 3.004152774810791, 'learning_rate': 2.6121212121212123e-06, 'num_tokens': 5512728.0, 'completions/mean_length': 98.625, 'completions/min_length': 94.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.625, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9990460872650146, 'rewards/meter/std': 0.00018491577066015452, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990460872650146, 'rewards/total_composite/std': 0.00018491577066015452, 'reward': 0.9990460872650146, 'reward_std': 0.00018490130605641752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039266087114810944, 'sampling/sampling_logp_difference/max': 1.522506594657898, 'sampling/importance_sampling_ratio/min': 0.2181643694639206, 'sampling/importance_sampling_ratio/mean': 1.0018714666366577, 'sampling/importance_sampling_ratio/max': 1.664224624633789, 'entropy': 0.282865421846509, 'clip_ratio/low_mean': 0.01263557211495936, 'clip_ratio/low_min': 0.01263557211495936, 'clip_ratio/high_mean': 0.02017652615904808, 'clip_ratio/high_max': 0.02017652615904808, 'clip_ratio/region_mean': 0.03281209827400744, 'reward_total_mean': 0.9990460872650146, 'reward_meter_mean': 0.9990460872650146, 'reward_meter_std': 0.00018491577066015452, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990460872650146, 'reward_total_composite_std': 0.00018491577066015452, 'epoch': 0.1}

 74%|███████▍  | 2439/3300 [3:27:13<1:13:35,  5.13s/it]INFO 04-12 02:04:04 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2440/3300 [3:27:21<1:26:21,  6.02s/it]2026-04-12 02:04:12,542 | INFO | train_grpo_train | metrics_logged mode=train step=2440

                                                       
{'loss': -0.0037, 'grad_norm': 1.5796713829040527, 'learning_rate': 2.6090909090909096e-06, 'num_tokens': 5517263.0, 'completions/mean_length': 344.875, 'completions/min_length': 329.0, 'completions/max_length': 358.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 344.875, 'completions/min_terminated_length': 329.0, 'completions/max_terminated_length': 358.0, 'rewards/meter/mean': 0.9985554218292236, 'rewards/meter/std': 0.00046954461140558124, 'rewards/count_adherence/mean': 0.8181818127632141, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9117647409439087, 'rewards/repeat_penalty/std': 0.04446640983223915, 'rewards/total_composite/mean': 0.7449136972427368, 'rewards/total_composite/std': 0.036383431404829025, 'reward': 0.7449136972427368, 'reward_std': 0.036383435130119324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030231278389692307, 'sampling/sampling_logp_difference/max': 1.6795151233673096, 'sampling/importance_sampling_ratio/min': 0.1864643543958664, 'sampling/importance_sampling_ratio/mean': 1.0063951015472412, 'sampling/importance_sampling_ratio/max': 1.8043537139892578, 'entropy': 0.2484235055744648, 'clip_ratio/low_mean': 0.00623403606005013, 'clip_ratio/low_min': 0.00623403606005013, 'clip_ratio/high_mean': 0.01647980441339314, 'clip_ratio/high_max': 0.01647980441339314, 'clip_ratio/region_mean': 0.02271384047344327, 'reward_total_mean': 0.7449136972427368, 'reward_meter_mean': 0.9985554218292236, 'reward_meter_std': 0.00046954461140558124, 'reward_count_adherence_mean': 0.8181818127632141, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9117647409439087, 'reward_repeat_penalty_std': 0.04446640983223915, 'reward_total_composite_mean': 0.7449136972427368, 'reward_total_composite_std': 0.036383431404829025, 'epoch': 0.1}

 74%|███████▍  | 2440/3300 [3:27:21<1:26:21,  6.02s/it]INFO 04-12 02:04:12 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2441/3300 [3:27:26<1:20:41,  5.64s/it]2026-04-12 02:04:17,274 | INFO | train_grpo_train | metrics_logged mode=train step=2441

                                                       
{'loss': -0.0108, 'grad_norm': 5.441035270690918, 'learning_rate': 2.6060606060606064e-06, 'num_tokens': 5519192.0, 'completions/mean_length': 79.125, 'completions/min_length': 76.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9938977360725403, 'rewards/meter/std': 0.009339890442788601, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8725272417068481, 'rewards/total_composite/std': 0.35255616903305054, 'reward': 0.8725272417068481, 'reward_std': 0.35255616903305054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03632466122508049, 'sampling/sampling_logp_difference/max': 2.1114463806152344, 'sampling/importance_sampling_ratio/min': 0.1210627406835556, 'sampling/importance_sampling_ratio/mean': 0.9990738034248352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.312020493671298, 'clip_ratio/low_mean': 0.003289473708719015, 'clip_ratio/low_min': 0.003289473708719015, 'clip_ratio/high_mean': 0.025135348900221288, 'clip_ratio/high_max': 0.025135348900221288, 'clip_ratio/region_mean': 0.028424822608940303, 'reward_total_mean': 0.8725272417068481, 'reward_meter_mean': 0.9938977360725403, 'reward_meter_std': 0.009339890442788601, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8725272417068481, 'reward_total_composite_std': 0.35255616903305054, 'epoch': 0.1}

 74%|███████▍  | 2441/3300 [3:27:26<1:20:41,  5.64s/it]INFO 04-12 02:04:17 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2442/3300 [3:27:30<1:16:38,  5.36s/it]2026-04-12 02:04:21,985 | INFO | train_grpo_train | metrics_logged mode=train step=2442

                                                       
{'loss': -0.0071, 'grad_norm': 1.337915062904358, 'learning_rate': 2.6030303030303033e-06, 'num_tokens': 5521213.0, 'completions/mean_length': 71.625, 'completions/min_length': 69.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9993535876274109, 'rewards/meter/std': 0.00013377754657994956, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993535876274109, 'rewards/total_composite/std': 0.00013377754657994956, 'reward': 0.9993535876274109, 'reward_std': 0.00013379388838075101, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013958643190562725, 'sampling/sampling_logp_difference/max': 0.9804167747497559, 'sampling/importance_sampling_ratio/min': 0.3751547038555145, 'sampling/importance_sampling_ratio/mean': 0.9988890290260315, 'sampling/importance_sampling_ratio/max': 1.1496003866195679, 'entropy': 0.06646110257133842, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0034722222480922937, 'clip_ratio/high_max': 0.0034722222480922937, 'clip_ratio/region_mean': 0.0034722222480922937, 'reward_total_mean': 0.9993535876274109, 'reward_meter_mean': 0.9993535876274109, 'reward_meter_std': 0.00013377754657994956, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993535876274109, 'reward_total_composite_std': 0.00013377754657994956, 'epoch': 0.1}

 74%|███████▍  | 2442/3300 [3:27:30<1:16:38,  5.36s/it]INFO 04-12 02:04:22 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2443/3300 [3:27:35<1:14:57,  5.25s/it]2026-04-12 02:04:26,972 | INFO | train_grpo_train | metrics_logged mode=train step=2443

                                                       
{'loss': 0.1046, 'grad_norm': 7.693568229675293, 'learning_rate': 2.6e-06, 'num_tokens': 5523328.0, 'completions/mean_length': 97.375, 'completions/min_length': 84.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.375, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9511677026748657, 'rewards/meter/std': 0.009777599945664406, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.908730149269104, 'rewards/repeat_penalty/std': 0.08302231132984161, 'rewards/total_composite/mean': 0.729360044002533, 'rewards/total_composite/std': 0.13022364675998688, 'reward': 0.729360044002533, 'reward_std': 0.13022364675998688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06830313801765442, 'sampling/sampling_logp_difference/max': 2.807176351547241, 'sampling/importance_sampling_ratio/min': 0.06037522852420807, 'sampling/importance_sampling_ratio/mean': 1.003010630607605, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3307408094406128, 'clip_ratio/low_mean': 0.04029961163178086, 'clip_ratio/low_min': 0.04029961163178086, 'clip_ratio/high_mean': 0.026477832812815905, 'clip_ratio/high_max': 0.026477832812815905, 'clip_ratio/region_mean': 0.06677744444459677, 'reward_total_mean': 0.729360044002533, 'reward_meter_mean': 0.9511677026748657, 'reward_meter_std': 0.009777599945664406, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.908730149269104, 'reward_repeat_penalty_std': 0.08302231132984161, 'reward_total_composite_mean': 0.729360044002533, 'reward_total_composite_std': 0.13022364675998688, 'epoch': 0.1}

 74%|███████▍  | 2443/3300 [3:27:35<1:14:57,  5.25s/it]INFO 04-12 02:04:27 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2444/3300 [3:27:43<1:24:11,  5.90s/it]2026-04-12 02:04:34,392 | INFO | train_grpo_train | metrics_logged mode=train step=2444

                                                       
{'loss': 0.0042, 'grad_norm': 1.5168893337249756, 'learning_rate': 2.5969696969696973e-06, 'num_tokens': 5527625.0, 'completions/mean_length': 320.125, 'completions/min_length': 319.0, 'completions/max_length': 321.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 320.125, 'completions/min_terminated_length': 319.0, 'completions/max_terminated_length': 321.0, 'rewards/meter/mean': 0.9990435242652893, 'rewards/meter/std': 0.00013361620949581265, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7426470518112183, 'rewards/repeat_penalty/std': 0.062391772866249084, 'rewards/total_composite/mean': 0.74193274974823, 'rewards/total_composite/std': 0.0622798353433609, 'reward': 0.74193274974823, 'reward_std': 0.0622798316180706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016514526680111885, 'sampling/sampling_logp_difference/max': 1.009657382965088, 'sampling/importance_sampling_ratio/min': 0.41537001729011536, 'sampling/importance_sampling_ratio/mean': 1.005713939666748, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12365446705371141, 'clip_ratio/low_mean': 0.0039014052599668503, 'clip_ratio/low_min': 0.0039014052599668503, 'clip_ratio/high_mean': 0.006253673695027828, 'clip_ratio/high_max': 0.006253673695027828, 'clip_ratio/region_mean': 0.010155078954994678, 'reward_total_mean': 0.74193274974823, 'reward_meter_mean': 0.9990435242652893, 'reward_meter_std': 0.00013361620949581265, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7426470518112183, 'reward_repeat_penalty_std': 0.062391772866249084, 'reward_total_composite_mean': 0.74193274974823, 'reward_total_composite_std': 0.0622798353433609, 'epoch': 0.1}

 74%|███████▍  | 2444/3300 [3:27:43<1:24:11,  5.90s/it]INFO 04-12 02:04:34 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2445/3300 [3:27:47<1:18:20,  5.50s/it]2026-04-12 02:04:38,950 | INFO | train_grpo_train | metrics_logged mode=train step=2445

                                                       
{'loss': 0.0227, 'grad_norm': 8.448700904846191, 'learning_rate': 2.593939393939394e-06, 'num_tokens': 5529594.0, 'completions/mean_length': 68.125, 'completions/min_length': 65.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9891185760498047, 'rewards/meter/std': 0.028624890372157097, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9891185760498047, 'rewards/total_composite/std': 0.028624890372157097, 'reward': 0.9891185760498047, 'reward_std': 0.028624894097447395, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04746231809258461, 'sampling/sampling_logp_difference/max': 1.0394535064697266, 'sampling/importance_sampling_ratio/min': 0.3536478877067566, 'sampling/importance_sampling_ratio/mean': 0.9979510307312012, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2945162560790777, 'clip_ratio/low_mean': 0.0069444444961845875, 'clip_ratio/low_min': 0.0069444444961845875, 'clip_ratio/high_mean': 0.02955011453013867, 'clip_ratio/high_max': 0.02955011453013867, 'clip_ratio/region_mean': 0.03649455902632326, 'reward_total_mean': 0.9891185760498047, 'reward_meter_mean': 0.9891185760498047, 'reward_meter_std': 0.028624890372157097, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9891185760498047, 'reward_total_composite_std': 0.028624890372157097, 'epoch': 0.1}

 74%|███████▍  | 2445/3300 [3:27:47<1:18:20,  5.50s/it]INFO 04-12 02:04:39 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2446/3300 [3:27:52<1:16:13,  5.35s/it]2026-04-12 02:04:43,974 | INFO | train_grpo_train | metrics_logged mode=train step=2446

                                                       
{'loss': 0.0023, 'grad_norm': 0.9163700938224792, 'learning_rate': 2.590909090909091e-06, 'num_tokens': 5531756.0, 'completions/mean_length': 107.25, 'completions/min_length': 106.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9992789626121521, 'rewards/meter/std': 9.53831258811988e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992789626121521, 'rewards/total_composite/std': 9.53831258811988e-05, 'reward': 0.9992789626121521, 'reward_std': 9.538961603539065e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010050991550087929, 'sampling/sampling_logp_difference/max': 0.7290101051330566, 'sampling/importance_sampling_ratio/min': 0.4823862612247467, 'sampling/importance_sampling_ratio/mean': 1.0028537511825562, 'sampling/importance_sampling_ratio/max': 1.2234734296798706, 'entropy': 0.07135625509545207, 'clip_ratio/low_mean': 0.0034616037737578154, 'clip_ratio/low_min': 0.0034616037737578154, 'clip_ratio/high_mean': 0.003515693824738264, 'clip_ratio/high_max': 0.003515693824738264, 'clip_ratio/region_mean': 0.0069772975984960794, 'reward_total_mean': 0.9992789626121521, 'reward_meter_mean': 0.9992789626121521, 'reward_meter_std': 9.53831258811988e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992789626121521, 'reward_total_composite_std': 9.53831258811988e-05, 'epoch': 0.1}

 74%|███████▍  | 2446/3300 [3:27:52<1:16:13,  5.35s/it]INFO 04-12 02:04:44 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2447/3300 [3:28:03<1:36:45,  6.81s/it]2026-04-12 02:04:54,187 | INFO | train_grpo_train | metrics_logged mode=train step=2447

                                                       
{'loss': -0.0107, 'grad_norm': 1.6099560260772705, 'learning_rate': 2.5878787878787883e-06, 'num_tokens': 5537808.0, 'completions/mean_length': 447.5, 'completions/min_length': 421.0, 'completions/max_length': 472.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 447.5, 'completions/min_terminated_length': 421.0, 'completions/max_terminated_length': 472.0, 'rewards/meter/mean': 0.9982467889785767, 'rewards/meter/std': 0.0005631654057651758, 'rewards/count_adherence/mean': 0.7583333253860474, 'rewards/count_adherence/std': 0.0345032773911953, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9484989643096924, 'rewards/repeat_penalty/std': 0.03878781571984291, 'rewards/total_composite/mean': 0.7178632020950317, 'rewards/total_composite/std': 0.040422506630420685, 'reward': 0.7178632020950317, 'reward_std': 0.040422506630420685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041198987513780594, 'sampling/sampling_logp_difference/max': 1.7551989555358887, 'sampling/importance_sampling_ratio/min': 0.17287284135818481, 'sampling/importance_sampling_ratio/mean': 1.003982424736023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3355287965387106, 'clip_ratio/low_mean': 0.009800048777833581, 'clip_ratio/low_min': 0.009800048777833581, 'clip_ratio/high_mean': 0.024517936864867806, 'clip_ratio/high_max': 0.024517936864867806, 'clip_ratio/region_mean': 0.03431798564270139, 'reward_total_mean': 0.7178632020950317, 'reward_meter_mean': 0.9982467889785767, 'reward_meter_std': 0.0005631654057651758, 'reward_count_adherence_mean': 0.7583333253860474, 'reward_count_adherence_std': 0.0345032773911953, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9484989643096924, 'reward_repeat_penalty_std': 0.03878781571984291, 'reward_total_composite_mean': 0.7178632020950317, 'reward_total_composite_std': 0.040422506630420685, 'epoch': 0.1}

 74%|███████▍  | 2447/3300 [3:28:03<1:36:45,  6.81s/it]INFO 04-12 02:04:54 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2448/3300 [3:28:11<1:41:48,  7.17s/it]2026-04-12 02:05:02,187 | INFO | train_grpo_train | metrics_logged mode=train step=2448

                                                       
{'loss': 0.0097, 'grad_norm': 1.5874656438827515, 'learning_rate': 2.584848484848485e-06, 'num_tokens': 5541790.0, 'completions/mean_length': 314.75, 'completions/min_length': 306.0, 'completions/max_length': 320.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 314.75, 'completions/min_terminated_length': 306.0, 'completions/max_terminated_length': 320.0, 'rewards/meter/mean': 0.9986426830291748, 'rewards/meter/std': 0.0003120753972325474, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9333333373069763, 'rewards/repeat_penalty/std': 0.07126966118812561, 'rewards/total_composite/mean': 0.9320781230926514, 'rewards/total_composite/std': 0.0713275596499443, 'reward': 0.9320781230926514, 'reward_std': 0.0713275596499443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031515371054410934, 'sampling/sampling_logp_difference/max': 1.0676155090332031, 'sampling/importance_sampling_ratio/min': 0.34382739663124084, 'sampling/importance_sampling_ratio/mean': 1.0063222646713257, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26868183352053165, 'clip_ratio/low_mean': 0.00553075410425663, 'clip_ratio/low_min': 0.00553075410425663, 'clip_ratio/high_mean': 0.01969089114572853, 'clip_ratio/high_max': 0.01969089114572853, 'clip_ratio/region_mean': 0.02522164524998516, 'reward_total_mean': 0.9320781230926514, 'reward_meter_mean': 0.9986426830291748, 'reward_meter_std': 0.0003120753972325474, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9333333373069763, 'reward_repeat_penalty_std': 0.07126966118812561, 'reward_total_composite_mean': 0.9320781230926514, 'reward_total_composite_std': 0.0713275596499443, 'epoch': 0.1}

 74%|███████▍  | 2448/3300 [3:28:11<1:41:48,  7.17s/it]INFO 04-12 02:05:02 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2449/3300 [3:28:18<1:43:51,  7.32s/it]2026-04-12 02:05:09,860 | INFO | train_grpo_train | metrics_logged mode=train step=2449

                                                       
{'loss': 0.0056, 'grad_norm': 2.10903000831604, 'learning_rate': 2.581818181818182e-06, 'num_tokens': 5546236.0, 'completions/mean_length': 314.75, 'completions/min_length': 313.0, 'completions/max_length': 315.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 314.75, 'completions/min_terminated_length': 313.0, 'completions/max_terminated_length': 315.0, 'rewards/meter/mean': 0.9972984790802002, 'rewards/meter/std': 7.734992686891928e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.6907894611358643, 'rewards/repeat_penalty/std': 0.06560124456882477, 'rewards/total_composite/mean': 0.6889257431030273, 'rewards/total_composite/std': 0.0654601976275444, 'reward': 0.6889257431030273, 'reward_std': 0.06546018272638321, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014586514793336391, 'sampling/sampling_logp_difference/max': 1.952465534210205, 'sampling/importance_sampling_ratio/min': 0.141923725605011, 'sampling/importance_sampling_ratio/mean': 1.0018553733825684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07688481360673904, 'clip_ratio/low_mean': 0.0027777778159361333, 'clip_ratio/low_min': 0.0027777778159361333, 'clip_ratio/high_mean': 0.002396166091784835, 'clip_ratio/high_max': 0.002396166091784835, 'clip_ratio/region_mean': 0.005173943907720968, 'reward_total_mean': 0.6889257431030273, 'reward_meter_mean': 0.9972984790802002, 'reward_meter_std': 7.734992686891928e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.6907894611358643, 'reward_repeat_penalty_std': 0.06560124456882477, 'reward_total_composite_mean': 0.6889257431030273, 'reward_total_composite_std': 0.0654601976275444, 'epoch': 0.1}

 74%|███████▍  | 2449/3300 [3:28:18<1:43:51,  7.32s/it]INFO 04-12 02:05:10 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2450/3300 [3:28:23<1:31:35,  6.46s/it]2026-04-12 02:05:14,331 | INFO | train_grpo_train | metrics_logged mode=train step=2450

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.5787878787878788e-06, 'num_tokens': 5547852.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0006953292759135365, 'sampling/sampling_logp_difference/max': 0.06802265346050262, 'sampling/importance_sampling_ratio/min': 0.9342393279075623, 'sampling/importance_sampling_ratio/mean': 1.000256061553955, 'sampling/importance_sampling_ratio/max': 1.0135822296142578, 'entropy': 0.004708475491497666, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▍  | 2450/3300 [3:28:23<1:31:35,  6.46s/it]INFO 04-12 02:05:14 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:33,  3.03s/it]

 23%|██▎       | 3/13 [00:09<00:32,  3.26s/it]

 31%|███       | 4/13 [00:15<00:36,  4.10s/it]

 38%|███▊      | 5/13 [00:20<00:36,  4.59s/it]

 46%|████▌     | 6/13 [00:27<00:36,  5.20s/it]

 54%|█████▍    | 7/13 [00:32<00:31,  5.23s/it]

 62%|██████▏   | 8/13 [00:38<00:28,  5.64s/it]

 69%|██████▉   | 9/13 [00:46<00:24,  6.13s/it]

 77%|███████▋  | 10/13 [00:51<00:17,  5.79s/it]

 85%|████████▍ | 11/13 [00:56<00:11,  5.64s/it]

 92%|█████████▏| 12/13 [01:02<00:05,  5.62s/it]

100%|██████████| 13/13 [01:08<00:00,  5.76s/it]2026-04-12 02:06:29,098 | INFO | train_grpo_train | metrics_logged mode=eval step=2450

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 74.6804, 'eval_samples_per_second': 1.393, 'eval_steps_per_second': 0.174, 'eval_num_tokens': 5547852.0, 'eval_completions/mean_length': 203.57692307692307, 'eval_completions/min_length': 61.38461538461539, 'eval_completions/max_length': 394.38461538461536, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 203.57692307692307, 'eval_completions/min_terminated_length': 61.38461538461539, 'eval_completions/max_terminated_length': 394.38461538461536, 'eval_rewards/meter/mean': 0.7445399669500498, 'eval_rewards/meter/std': 0.3695275445397084, 'eval_rewards/count_adherence/mean': 0.9357927395747259, 'eval_rewards/count_adherence/std': 0.08574442823345844, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.8948483146153964, 'eval_rewards/repeat_penalty/std': 0.11098527220579293, 'eval_rewards/total_composite/mean': 0.6300999980706435, 'eval_rewards/total_composite/std': 0.3479128239246515, 'eval_reward': 0.6300999980706435, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.022716052543658476, 'eval_sampling/sampling_logp_difference/max': 0.9992542266845703, 'eval_sampling/importance_sampling_ratio/min': 0.37165725231170654, 'eval_sampling/importance_sampling_ratio/mean': 1.0057257138765776, 'eval_sampling/importance_sampling_ratio/max': 1.4525978748614972, 'eval_entropy': 0.23695378464001876, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.6300999980706435, 'eval_reward_meter_mean': 0.7445399669500498, 'eval_reward_meter_std': 0.3695275445397084, 'eval_reward_count_adherence_mean': 0.9357927395747259, 'eval_reward_count_adherence_std': 0.08574442823345844, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.8948483146153964, 'eval_reward_repeat_penalty_std': 0.11098527220579293, 'eval_reward_total_composite_mean': 0.6300999980706435, 'eval_reward_total_composite_std': 0.3479128239246515, 'epoch': 0.1}

 74%|███████▍  | 2450/3300 [3:29:38<1:31:35,  6.46s/it]

100%|██████████| 13/13 [01:08<00:00,  5.76s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 74%|███████▍  | 2451/3300 [3:29:44<6:49:27, 28.94s/it]2026-04-12 02:06:35,703 | INFO | train_grpo_train | metrics_logged mode=train step=2451

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.575757575757576e-06, 'num_tokens': 5549564.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0012545905774459243, 'sampling/sampling_logp_difference/max': 0.19019845128059387, 'sampling/importance_sampling_ratio/min': 0.8267950415611267, 'sampling/importance_sampling_ratio/mean': 1.000001311302185, 'sampling/importance_sampling_ratio/max': 1.0180944204330444, 'entropy': 0.005932128522545099, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▍  | 2451/3300 [3:29:44<6:49:27, 28.94s/it]INFO 04-12 02:06:35 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2452/3300 [3:29:49<5:08:34, 21.83s/it]2026-04-12 02:06:40,961 | INFO | train_grpo_train | metrics_logged mode=train step=2452

                                                       
{'loss': 0.1072, 'grad_norm': 8.379902839660645, 'learning_rate': 2.572727272727273e-06, 'num_tokens': 5551659.0, 'completions/mean_length': 92.875, 'completions/min_length': 85.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.875, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9926849007606506, 'rewards/meter/std': 0.0018546542851254344, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9321428537368774, 'rewards/repeat_penalty/std': 0.09529759734869003, 'rewards/total_composite/mean': 0.8898848295211792, 'rewards/total_composite/std': 0.1583697348833084, 'reward': 0.8898848295211792, 'reward_std': 0.15836970508098602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038993846625089645, 'sampling/sampling_logp_difference/max': 1.2752385139465332, 'sampling/importance_sampling_ratio/min': 0.27936431765556335, 'sampling/importance_sampling_ratio/mean': 1.0090692043304443, 'sampling/importance_sampling_ratio/max': 1.9530335664749146, 'entropy': 0.2788791488856077, 'clip_ratio/low_mean': 0.013440471375361085, 'clip_ratio/low_min': 0.013440471375361085, 'clip_ratio/high_mean': 0.01974922022782266, 'clip_ratio/high_max': 0.01974922022782266, 'clip_ratio/region_mean': 0.033189691603183746, 'reward_total_mean': 0.8898848295211792, 'reward_meter_mean': 0.9926849007606506, 'reward_meter_std': 0.0018546542851254344, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9321428537368774, 'reward_repeat_penalty_std': 0.09529759734869003, 'reward_total_composite_mean': 0.8898848295211792, 'reward_total_composite_std': 0.1583697348833084, 'epoch': 0.1}

 74%|███████▍  | 2452/3300 [3:29:49<5:08:34, 21.83s/it]INFO 04-12 02:06:41 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2453/3300 [3:29:54<3:56:22, 16.74s/it]2026-04-12 02:06:45,828 | INFO | train_grpo_train | metrics_logged mode=train step=2453

                                                       
{'loss': -0.0104, 'grad_norm': 4.428008079528809, 'learning_rate': 2.5696969696969697e-06, 'num_tokens': 5553819.0, 'completions/mean_length': 89.0, 'completions/min_length': 84.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9919699430465698, 'rewards/meter/std': 0.005432384088635445, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9670881032943726, 'rewards/total_composite/std': 0.06924090534448624, 'reward': 0.9670881032943726, 'reward_std': 0.06924089044332504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031983181834220886, 'sampling/sampling_logp_difference/max': 1.074939250946045, 'sampling/importance_sampling_ratio/min': 0.35889124870300293, 'sampling/importance_sampling_ratio/mean': 1.0001150369644165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21917428076267242, 'clip_ratio/low_mean': 0.004360465332865715, 'clip_ratio/low_min': 0.004360465332865715, 'clip_ratio/high_mean': 0.036513578495942056, 'clip_ratio/high_max': 0.036513578495942056, 'clip_ratio/region_mean': 0.04087404382880777, 'reward_total_mean': 0.9670881032943726, 'reward_meter_mean': 0.9919699430465698, 'reward_meter_std': 0.005432384088635445, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9670881032943726, 'reward_total_composite_std': 0.06924090534448624, 'epoch': 0.1}

 74%|███████▍  | 2453/3300 [3:29:54<3:56:22, 16.74s/it]INFO 04-12 02:06:46 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2454/3300 [3:29:59<3:03:17, 13.00s/it]2026-04-12 02:06:50,093 | INFO | train_grpo_train | metrics_logged mode=train step=2454

                                                       
{'loss': 0.0216, 'grad_norm': 4.521707534790039, 'learning_rate': 2.566666666666667e-06, 'num_tokens': 5555436.0, 'completions/mean_length': 40.125, 'completions/min_length': 38.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9973416328430176, 'rewards/meter/std': 0.0012014260282739997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973416328430176, 'rewards/total_composite/std': 0.0012014260282739997, 'reward': 0.9973416328430176, 'reward_std': 0.0012014324311167002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017517054453492165, 'sampling/sampling_logp_difference/max': 1.4276123046875, 'sampling/importance_sampling_ratio/min': 0.2398810088634491, 'sampling/importance_sampling_ratio/mean': 0.9969314932823181, 'sampling/importance_sampling_ratio/max': 1.7687947750091553, 'entropy': 0.11128469556570053, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0030487803742289543, 'clip_ratio/high_max': 0.0030487803742289543, 'clip_ratio/region_mean': 0.0030487803742289543, 'reward_total_mean': 0.9973416328430176, 'reward_meter_mean': 0.9973416328430176, 'reward_meter_std': 0.0012014260282739997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973416328430176, 'reward_total_composite_std': 0.0012014260282739997, 'epoch': 0.1}

 74%|███████▍  | 2454/3300 [3:29:59<3:03:17, 13.00s/it]INFO 04-12 02:06:50 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2455/3300 [3:30:03<2:27:39, 10.48s/it]2026-04-12 02:06:54,705 | INFO | train_grpo_train | metrics_logged mode=train step=2455

                                                       
{'loss': 0.2385, 'grad_norm': 2.2290995121002197, 'learning_rate': 2.5636363636363638e-06, 'num_tokens': 5556947.0, 'completions/mean_length': 36.875, 'completions/min_length': 33.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9920687675476074, 'rewards/meter/std': 0.00039521773578599095, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8681668043136597, 'rewards/total_composite/std': 0.35079243779182434, 'reward': 0.8681668043136597, 'reward_std': 0.35079240798950195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016091084107756615, 'sampling/sampling_logp_difference/max': 1.2472076416015625, 'sampling/importance_sampling_ratio/min': 0.28730595111846924, 'sampling/importance_sampling_ratio/mean': 1.0059282779693604, 'sampling/importance_sampling_ratio/max': 1.3825856447219849, 'entropy': 0.11671794764697552, 'clip_ratio/low_mean': 0.002016128972172737, 'clip_ratio/low_min': 0.002016128972172737, 'clip_ratio/high_mean': 0.007464349502697587, 'clip_ratio/high_max': 0.007464349502697587, 'clip_ratio/region_mean': 0.009480478474870324, 'reward_total_mean': 0.8681668043136597, 'reward_meter_mean': 0.9920687675476074, 'reward_meter_std': 0.00039521773578599095, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8681668043136597, 'reward_total_composite_std': 0.35079243779182434, 'epoch': 0.1}

 74%|███████▍  | 2455/3300 [3:30:03<2:27:39, 10.48s/it]INFO 04-12 02:06:54 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2456/3300 [3:30:08<2:03:01,  8.75s/it]2026-04-12 02:06:59,395 | INFO | train_grpo_train | metrics_logged mode=train step=2456

                                                       
{'loss': 0.0052, 'grad_norm': 3.4946515560150146, 'learning_rate': 2.5606060606060606e-06, 'num_tokens': 5558871.0, 'completions/mean_length': 80.5, 'completions/min_length': 79.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.5, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9981932044029236, 'rewards/meter/std': 0.0004672010545618832, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981932044029236, 'rewards/total_composite/std': 0.0004672010545618832, 'reward': 0.9981932044029236, 'reward_std': 0.0004671814094763249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02568862773478031, 'sampling/sampling_logp_difference/max': 1.0305986404418945, 'sampling/importance_sampling_ratio/min': 0.3567933142185211, 'sampling/importance_sampling_ratio/mean': 1.0071548223495483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22475284524261951, 'clip_ratio/low_mean': 0.003088302561081946, 'clip_ratio/low_min': 0.003088302561081946, 'clip_ratio/high_mean': 0.007773919845931232, 'clip_ratio/high_max': 0.007773919845931232, 'clip_ratio/region_mean': 0.010862222407013178, 'reward_total_mean': 0.9981932044029236, 'reward_meter_mean': 0.9981932044029236, 'reward_meter_std': 0.0004672010545618832, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981932044029236, 'reward_total_composite_std': 0.0004672010545618832, 'epoch': 0.1}

 74%|███████▍  | 2456/3300 [3:30:08<2:03:01,  8.75s/it]INFO 04-12 02:06:59 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2457/3300 [3:30:14<1:51:10,  7.91s/it]2026-04-12 02:07:05,398 | INFO | train_grpo_train | metrics_logged mode=train step=2457

                                                       
{'loss': 0.0114, 'grad_norm': 2.7975685596466064, 'learning_rate': 2.5575757575757574e-06, 'num_tokens': 5561940.0, 'completions/mean_length': 183.625, 'completions/min_length': 177.0, 'completions/max_length': 190.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 183.625, 'completions/min_terminated_length': 177.0, 'completions/max_terminated_length': 190.0, 'rewards/meter/mean': 0.9926738142967224, 'rewards/meter/std': 0.000535113038495183, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9318182468414307, 'rewards/repeat_penalty/std': 0.04208271950483322, 'rewards/total_composite/mean': 0.9250072836875916, 'rewards/total_composite/std': 0.042194660753011703, 'reward': 0.9250072836875916, 'reward_std': 0.042194657027721405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03491319715976715, 'sampling/sampling_logp_difference/max': 3.0987467765808105, 'sampling/importance_sampling_ratio/min': 0.045105695724487305, 'sampling/importance_sampling_ratio/mean': 0.9990267157554626, 'sampling/importance_sampling_ratio/max': 1.8111183643341064, 'entropy': 0.2139601055532694, 'clip_ratio/low_mean': 0.01427258289186284, 'clip_ratio/low_min': 0.01427258289186284, 'clip_ratio/high_mean': 0.00900993263348937, 'clip_ratio/high_max': 0.00900993263348937, 'clip_ratio/region_mean': 0.02328251552535221, 'reward_total_mean': 0.9250072836875916, 'reward_meter_mean': 0.9926738142967224, 'reward_meter_std': 0.000535113038495183, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9318182468414307, 'reward_repeat_penalty_std': 0.04208271950483322, 'reward_total_composite_mean': 0.9250072836875916, 'reward_total_composite_std': 0.042194660753011703, 'epoch': 0.1}

 74%|███████▍  | 2457/3300 [3:30:14<1:51:10,  7.91s/it]INFO 04-12 02:07:05 [block_pool.py:378] Successfully reset prefix cache

 74%|███████▍  | 2458/3300 [3:30:18<1:36:44,  6.89s/it]2026-04-12 02:07:09,887 | INFO | train_grpo_train | metrics_logged mode=train step=2458

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.5545454545454547e-06, 'num_tokens': 5563996.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00065056630410254, 'sampling/sampling_logp_difference/max': 0.08942283689975739, 'sampling/importance_sampling_ratio/min': 0.9144588112831116, 'sampling/importance_sampling_ratio/mean': 1.0000568628311157, 'sampling/importance_sampling_ratio/max': 1.0292819738388062, 'entropy': 0.005484737688675523, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 74%|███████▍  | 2458/3300 [3:30:18<1:36:44,  6.89s/it]INFO 04-12 02:07:10 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2459/3300 [3:30:24<1:32:04,  6.57s/it]2026-04-12 02:07:15,692 | INFO | train_grpo_train | metrics_logged mode=train step=2459

                                                       
{'loss': -0.0206, 'grad_norm': 6.4525146484375, 'learning_rate': 2.5515151515151515e-06, 'num_tokens': 5566684.0, 'completions/mean_length': 164.0, 'completions/min_length': 143.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 164.0, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9948143362998962, 'rewards/meter/std': 0.010059010237455368, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9674677848815918, 'rewards/total_composite/std': 0.05743533372879028, 'reward': 0.9674677848815918, 'reward_std': 0.057435326278209686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054715219885110855, 'sampling/sampling_logp_difference/max': 1.7331323623657227, 'sampling/importance_sampling_ratio/min': 0.1767299622297287, 'sampling/importance_sampling_ratio/mean': 1.0046672821044922, 'sampling/importance_sampling_ratio/max': 1.882559061050415, 'entropy': 0.37010542675852776, 'clip_ratio/low_mean': 0.009364139288663864, 'clip_ratio/low_min': 0.009364139288663864, 'clip_ratio/high_mean': 0.03573530330322683, 'clip_ratio/high_max': 0.03573530330322683, 'clip_ratio/region_mean': 0.04509944259189069, 'reward_total_mean': 0.9674677848815918, 'reward_meter_mean': 0.9948143362998962, 'reward_meter_std': 0.010059010237455368, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9674677848815918, 'reward_total_composite_std': 0.05743533372879028, 'epoch': 0.1}

 75%|███████▍  | 2459/3300 [3:30:24<1:32:04,  6.57s/it]INFO 04-12 02:07:15 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2460/3300 [3:30:29<1:23:24,  5.96s/it]2026-04-12 02:07:20,226 | INFO | train_grpo_train | metrics_logged mode=train step=2460

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.5484848484848484e-06, 'num_tokens': 5568492.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002486900775693357, 'sampling/sampling_logp_difference/max': 0.013540960848331451, 'sampling/importance_sampling_ratio/min': 0.9955151677131653, 'sampling/importance_sampling_ratio/mean': 1.0002198219299316, 'sampling/importance_sampling_ratio/max': 1.0136330127716064, 'entropy': 0.0032962941040750593, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 75%|███████▍  | 2460/3300 [3:30:29<1:23:24,  5.96s/it]INFO 04-12 02:07:20 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2461/3300 [3:30:38<1:37:00,  6.94s/it]2026-04-12 02:07:29,452 | INFO | train_grpo_train | metrics_logged mode=train step=2461

                                                       
{'loss': -0.023, 'grad_norm': 1.6912506818771362, 'learning_rate': 2.5454545454545456e-06, 'num_tokens': 5573680.0, 'completions/mean_length': 404.5, 'completions/min_length': 378.0, 'completions/max_length': 424.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 404.5, 'completions/min_terminated_length': 378.0, 'completions/max_terminated_length': 424.0, 'rewards/meter/mean': 0.9983578324317932, 'rewards/meter/std': 0.0009692087187431753, 'rewards/count_adherence/mean': 0.7980769276618958, 'rewards/count_adherence/std': 0.05723259598016739, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9470028877258301, 'rewards/repeat_penalty/std': 0.06291526556015015, 'rewards/total_composite/mean': 0.7565553188323975, 'rewards/total_composite/std': 0.09492313861846924, 'reward': 0.7565553188323975, 'reward_std': 0.09492312371730804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042006395757198334, 'sampling/sampling_logp_difference/max': 1.7684879302978516, 'sampling/importance_sampling_ratio/min': 0.17059072852134705, 'sampling/importance_sampling_ratio/mean': 1.0055227279663086, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35317912697792053, 'clip_ratio/low_mean': 0.008652225020341575, 'clip_ratio/low_min': 0.008652225020341575, 'clip_ratio/high_mean': 0.022989682853221893, 'clip_ratio/high_max': 0.022989682853221893, 'clip_ratio/region_mean': 0.03164190787356347, 'reward_total_mean': 0.7565553188323975, 'reward_meter_mean': 0.9983578324317932, 'reward_meter_std': 0.0009692087187431753, 'reward_count_adherence_mean': 0.7980769276618958, 'reward_count_adherence_std': 0.05723259598016739, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9470028877258301, 'reward_repeat_penalty_std': 0.06291526556015015, 'reward_total_composite_mean': 0.7565553188323975, 'reward_total_composite_std': 0.09492313861846924, 'epoch': 0.1}

 75%|███████▍  | 2461/3300 [3:30:38<1:37:00,  6.94s/it]INFO 04-12 02:07:29 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2462/3300 [3:30:43<1:27:44,  6.28s/it]2026-04-12 02:07:34,201 | INFO | train_grpo_train | metrics_logged mode=train step=2462

                                                       
{'loss': 0.0004, 'grad_norm': 0.51884925365448, 'learning_rate': 2.542424242424243e-06, 'num_tokens': 5575497.0, 'completions/mean_length': 67.125, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9981344938278198, 'rewards/meter/std': 2.5179730073432438e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981344938278198, 'rewards/total_composite/std': 2.5179730073432438e-05, 'reward': 0.9981344938278198, 'reward_std': 2.518050132493954e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007205577101558447, 'sampling/sampling_logp_difference/max': 0.981438159942627, 'sampling/importance_sampling_ratio/min': 0.37477174401283264, 'sampling/importance_sampling_ratio/mean': 0.9990748167037964, 'sampling/importance_sampling_ratio/max': 1.997531771659851, 'entropy': 0.017152427230030298, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9981344938278198, 'reward_meter_mean': 0.9981344938278198, 'reward_meter_std': 2.5179730073432438e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981344938278198, 'reward_total_composite_std': 2.5179730073432438e-05, 'epoch': 0.1}

 75%|███████▍  | 2462/3300 [3:30:43<1:27:44,  6.28s/it]INFO 04-12 02:07:34 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2463/3300 [3:30:48<1:21:58,  5.88s/it]2026-04-12 02:07:39,132 | INFO | train_grpo_train | metrics_logged mode=train step=2463

                                                       
{'loss': 0.0004, 'grad_norm': 0.14106905460357666, 'learning_rate': 2.5393939393939397e-06, 'num_tokens': 5577697.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9980115294456482, 'rewards/meter/std': 1.5354547940660268e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980115294456482, 'rewards/total_composite/std': 1.5354547940660268e-05, 'reward': 0.9980115294456482, 'reward_std': 1.5354515198851004e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0028446686919778585, 'sampling/sampling_logp_difference/max': 0.4193582534790039, 'sampling/importance_sampling_ratio/min': 0.6574686169624329, 'sampling/importance_sampling_ratio/mean': 1.0014915466308594, 'sampling/importance_sampling_ratio/max': 1.2301760911941528, 'entropy': 0.027404201216995716, 'clip_ratio/low_mean': 0.0012755101779475808, 'clip_ratio/low_min': 0.0012755101779475808, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0012755101779475808, 'reward_total_mean': 0.9980115294456482, 'reward_meter_mean': 0.9980115294456482, 'reward_meter_std': 1.5354547940660268e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980115294456482, 'reward_total_composite_std': 1.5354547940660268e-05, 'epoch': 0.1}

 75%|███████▍  | 2463/3300 [3:30:48<1:21:58,  5.88s/it]INFO 04-12 02:07:39 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2464/3300 [3:30:52<1:17:52,  5.59s/it]2026-04-12 02:07:44,049 | INFO | train_grpo_train | metrics_logged mode=train step=2464

                                                       
{'loss': -0.0148, 'grad_norm': 2.84199595451355, 'learning_rate': 2.536363636363637e-06, 'num_tokens': 5579505.0, 'completions/mean_length': 61.0, 'completions/min_length': 58.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9931963086128235, 'rewards/meter/std': 0.00066983891883865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9931963086128235, 'rewards/total_composite/std': 0.00066983891883865, 'reward': 0.9931963086128235, 'reward_std': 0.0006698234938085079, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013205956667661667, 'sampling/sampling_logp_difference/max': 0.7910587787628174, 'sampling/importance_sampling_ratio/min': 0.4533645212650299, 'sampling/importance_sampling_ratio/mean': 1.0031862258911133, 'sampling/importance_sampling_ratio/max': 1.462441325187683, 'entropy': 0.09446914121508598, 'clip_ratio/low_mean': 0.008408705238252878, 'clip_ratio/low_min': 0.008408705238252878, 'clip_ratio/high_mean': 0.004032257944345474, 'clip_ratio/high_max': 0.004032257944345474, 'clip_ratio/region_mean': 0.012440963182598352, 'reward_total_mean': 0.9931963086128235, 'reward_meter_mean': 0.9931963086128235, 'reward_meter_std': 0.00066983891883865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9931963086128235, 'reward_total_composite_std': 0.00066983891883865, 'epoch': 0.1}

 75%|███████▍  | 2464/3300 [3:30:53<1:17:52,  5.59s/it]INFO 04-12 02:07:44 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2465/3300 [3:31:00<1:23:53,  6.03s/it]2026-04-12 02:07:51,105 | INFO | train_grpo_train | metrics_logged mode=train step=2465

                                                       
{'loss': 0.0112, 'grad_norm': 1.982505440711975, 'learning_rate': 2.5333333333333338e-06, 'num_tokens': 5583250.0, 'completions/mean_length': 271.125, 'completions/min_length': 264.0, 'completions/max_length': 279.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 271.125, 'completions/min_terminated_length': 264.0, 'completions/max_terminated_length': 279.0, 'rewards/meter/mean': 0.998477578163147, 'rewards/meter/std': 0.0006886826595291495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9711538553237915, 'rewards/repeat_penalty/std': 0.05723259598016739, 'rewards/total_composite/mean': 0.969672441482544, 'rewards/total_composite/std': 0.057110171765089035, 'reward': 0.969672441482544, 'reward_std': 0.05711015686392784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043706730008125305, 'sampling/sampling_logp_difference/max': 1.9009828567504883, 'sampling/importance_sampling_ratio/min': 0.14942169189453125, 'sampling/importance_sampling_ratio/mean': 1.0076007843017578, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39818970672786236, 'clip_ratio/low_mean': 0.006513258791528642, 'clip_ratio/low_min': 0.006513258791528642, 'clip_ratio/high_mean': 0.028696169378235936, 'clip_ratio/high_max': 0.028696169378235936, 'clip_ratio/region_mean': 0.03520942816976458, 'reward_total_mean': 0.969672441482544, 'reward_meter_mean': 0.998477578163147, 'reward_meter_std': 0.0006886826595291495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9711538553237915, 'reward_repeat_penalty_std': 0.05723259598016739, 'reward_total_composite_mean': 0.969672441482544, 'reward_total_composite_std': 0.057110171765089035, 'epoch': 0.1}

 75%|███████▍  | 2465/3300 [3:31:00<1:23:53,  6.03s/it]INFO 04-12 02:07:51 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2466/3300 [3:31:05<1:21:24,  5.86s/it]2026-04-12 02:07:56,560 | INFO | train_grpo_train | metrics_logged mode=train step=2466

                                                       
{'loss': 0.0039, 'grad_norm': 1.754108190536499, 'learning_rate': 2.5303030303030306e-06, 'num_tokens': 5585974.0, 'completions/mean_length': 142.5, 'completions/min_length': 142.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.5, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9990931153297424, 'rewards/meter/std': 0.00018217807519249618, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9098852872848511, 'rewards/total_composite/std': 0.07382422685623169, 'reward': 0.9098852872848511, 'reward_std': 0.07382424175739288, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014981756918132305, 'sampling/sampling_logp_difference/max': 0.8543601036071777, 'sampling/importance_sampling_ratio/min': 0.42555540800094604, 'sampling/importance_sampling_ratio/mean': 1.005768060684204, 'sampling/importance_sampling_ratio/max': 1.6116867065429688, 'entropy': 0.095588818192482, 'clip_ratio/low_mean': 0.010514133609831333, 'clip_ratio/low_min': 0.010514133609831333, 'clip_ratio/high_mean': 0.004401408368721604, 'clip_ratio/high_max': 0.004401408368721604, 'clip_ratio/region_mean': 0.014915541978552938, 'reward_total_mean': 0.9098852872848511, 'reward_meter_mean': 0.9990931153297424, 'reward_meter_std': 0.00018217807519249618, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9098852872848511, 'reward_total_composite_std': 0.07382422685623169, 'epoch': 0.1}

 75%|███████▍  | 2466/3300 [3:31:05<1:21:24,  5.86s/it]INFO 04-12 02:07:56 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2467/3300 [3:31:09<1:15:37,  5.45s/it]2026-04-12 02:08:01,045 | INFO | train_grpo_train | metrics_logged mode=train step=2467

                                                       
{'loss': 0.0003, 'grad_norm': 0.4351802468299866, 'learning_rate': 2.5272727272727274e-06, 'num_tokens': 5587830.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7115719318389893, 'rewards/meter/std': 0.021074457094073296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.6950865983963013, 'rewards/total_composite/std': 0.06770216673612595, 'reward': 0.6950865983963013, 'reward_std': 0.06770216673612595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0018030240898951888, 'sampling/sampling_logp_difference/max': 0.2816805839538574, 'sampling/importance_sampling_ratio/min': 0.8967838287353516, 'sampling/importance_sampling_ratio/mean': 1.0006452798843384, 'sampling/importance_sampling_ratio/max': 1.3253552913665771, 'entropy': 0.013048171065747738, 'clip_ratio/low_mean': 0.0015625000232830644, 'clip_ratio/low_min': 0.0015625000232830644, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0015625000232830644, 'reward_total_mean': 0.6950865983963013, 'reward_meter_mean': 0.7115719318389893, 'reward_meter_std': 0.021074457094073296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.6950865983963013, 'reward_total_composite_std': 0.06770216673612595, 'epoch': 0.1}

 75%|███████▍  | 2467/3300 [3:31:10<1:15:37,  5.45s/it]INFO 04-12 02:08:01 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2468/3300 [3:31:13<1:09:39,  5.02s/it]2026-04-12 02:08:05,084 | INFO | train_grpo_train | metrics_logged mode=train step=2468

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.5242424242424247e-06, 'num_tokens': 5589294.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00015046881162561476, 'sampling/sampling_logp_difference/max': 0.0030941502191126347, 'sampling/importance_sampling_ratio/min': 0.9973318576812744, 'sampling/importance_sampling_ratio/mean': 1.0000958442687988, 'sampling/importance_sampling_ratio/max': 1.003098964691162, 'entropy': 0.0013736404580413364, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 75%|███████▍  | 2468/3300 [3:31:14<1:09:39,  5.02s/it]INFO 04-12 02:08:05 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2469/3300 [3:31:18<1:08:33,  4.95s/it]2026-04-12 02:08:09,864 | INFO | train_grpo_train | metrics_logged mode=train step=2469

                                                       
{'loss': -0.0232, 'grad_norm': 4.22209358215332, 'learning_rate': 2.5212121212121215e-06, 'num_tokens': 5591249.0, 'completions/mean_length': 90.375, 'completions/min_length': 84.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.375, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9939854145050049, 'rewards/meter/std': 0.0009144411887973547, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9691168665885925, 'rewards/total_composite/std': 0.07003802061080933, 'reward': 0.9691168665885925, 'reward_std': 0.07003801316022873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022249562665820122, 'sampling/sampling_logp_difference/max': 1.9335792064666748, 'sampling/importance_sampling_ratio/min': 0.1446296125650406, 'sampling/importance_sampling_ratio/mean': 0.9984654784202576, 'sampling/importance_sampling_ratio/max': 1.7310431003570557, 'entropy': 0.08473077462986112, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.023276447434909642, 'clip_ratio/high_max': 0.023276447434909642, 'clip_ratio/region_mean': 0.023276447434909642, 'reward_total_mean': 0.9691168665885925, 'reward_meter_mean': 0.9939854145050049, 'reward_meter_std': 0.0009144411887973547, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9691168665885925, 'reward_total_composite_std': 0.07003802061080933, 'epoch': 0.1}

 75%|███████▍  | 2469/3300 [3:31:18<1:08:33,  4.95s/it]INFO 04-12 02:08:10 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2470/3300 [3:31:24<1:11:21,  5.16s/it]2026-04-12 02:08:15,510 | INFO | train_grpo_train | metrics_logged mode=train step=2470

                                                       
{'loss': -0.0038, 'grad_norm': 2.6985630989074707, 'learning_rate': 2.5181818181818184e-06, 'num_tokens': 5594095.0, 'completions/mean_length': 163.75, 'completions/min_length': 159.0, 'completions/max_length': 168.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 163.75, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 168.0, 'rewards/meter/mean': 0.998643159866333, 'rewards/meter/std': 0.0008321683271788061, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9708901643753052, 'rewards/total_composite/std': 0.0511077418923378, 'reward': 0.9708901643753052, 'reward_std': 0.05110771581530571, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049378104507923126, 'sampling/sampling_logp_difference/max': 1.4132513999938965, 'sampling/importance_sampling_ratio/min': 0.24335075914859772, 'sampling/importance_sampling_ratio/mean': 1.012289047241211, 'sampling/importance_sampling_ratio/max': 1.9667812585830688, 'entropy': 0.4189731851220131, 'clip_ratio/low_mean': 0.010085079353302717, 'clip_ratio/low_min': 0.010085079353302717, 'clip_ratio/high_mean': 0.03505220194347203, 'clip_ratio/high_max': 0.03505220194347203, 'clip_ratio/region_mean': 0.045137281296774745, 'reward_total_mean': 0.9708901643753052, 'reward_meter_mean': 0.998643159866333, 'reward_meter_std': 0.0008321683271788061, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9708901643753052, 'reward_total_composite_std': 0.0511077418923378, 'epoch': 0.1}

 75%|███████▍  | 2470/3300 [3:31:24<1:11:21,  5.16s/it]INFO 04-12 02:08:15 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2471/3300 [3:31:28<1:06:48,  4.84s/it]2026-04-12 02:08:19,589 | INFO | train_grpo_train | metrics_logged mode=train step=2471

                                                       
{'loss': -0.0009, 'grad_norm': 1.5141384601593018, 'learning_rate': 2.5151515151515156e-06, 'num_tokens': 5595647.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9994994401931763, 'rewards/meter/std': 5.432292527984828e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994994401931763, 'rewards/total_composite/std': 5.432292527984828e-05, 'reward': 0.9994994401931763, 'reward_std': 5.4318112233886495e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019814442843198776, 'sampling/sampling_logp_difference/max': 1.097601056098938, 'sampling/importance_sampling_ratio/min': 0.33367058634757996, 'sampling/importance_sampling_ratio/mean': 1.0052924156188965, 'sampling/importance_sampling_ratio/max': 1.4859408140182495, 'entropy': 0.1270341333001852, 'clip_ratio/low_mean': 0.010714285774156451, 'clip_ratio/low_min': 0.010714285774156451, 'clip_ratio/high_mean': 0.0035714285913854837, 'clip_ratio/high_max': 0.0035714285913854837, 'clip_ratio/region_mean': 0.014285714365541935, 'reward_total_mean': 0.9994994401931763, 'reward_meter_mean': 0.9994994401931763, 'reward_meter_std': 5.432292527984828e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994994401931763, 'reward_total_composite_std': 5.432292527984828e-05, 'epoch': 0.1}

 75%|███████▍  | 2471/3300 [3:31:28<1:06:48,  4.84s/it]INFO 04-12 02:08:19 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2472/3300 [3:31:33<1:07:18,  4.88s/it]2026-04-12 02:08:24,567 | INFO | train_grpo_train | metrics_logged mode=train step=2472

                                                       
{'loss': -0.0036, 'grad_norm': 3.565403699874878, 'learning_rate': 2.5121212121212125e-06, 'num_tokens': 5598009.0, 'completions/mean_length': 105.25, 'completions/min_length': 102.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.25, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.45625758171081543, 'rewards/meter/std': 0.3081152141094208, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.38120073080062866, 'rewards/total_composite/std': 0.2558993995189667, 'reward': 0.38120073080062866, 'reward_std': 0.2558993995189667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005791721399873495, 'sampling/sampling_logp_difference/max': 0.5239999294281006, 'sampling/importance_sampling_ratio/min': 0.5921472907066345, 'sampling/importance_sampling_ratio/mean': 1.0014318227767944, 'sampling/importance_sampling_ratio/max': 1.5240379571914673, 'entropy': 0.025531375431455672, 'clip_ratio/low_mean': 0.006092437193728983, 'clip_ratio/low_min': 0.006092437193728983, 'clip_ratio/high_mean': 0.002358490601181984, 'clip_ratio/high_max': 0.002358490601181984, 'clip_ratio/region_mean': 0.008450927794910967, 'reward_total_mean': 0.38120073080062866, 'reward_meter_mean': 0.45625758171081543, 'reward_meter_std': 0.3081152141094208, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.38120073080062866, 'reward_total_composite_std': 0.2558993995189667, 'epoch': 0.1}

 75%|███████▍  | 2472/3300 [3:31:33<1:07:18,  4.88s/it]INFO 04-12 02:08:24 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2473/3300 [3:31:37<1:05:02,  4.72s/it]2026-04-12 02:08:28,918 | INFO | train_grpo_train | metrics_logged mode=train step=2473

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.5090909090909093e-06, 'num_tokens': 5599793.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002734074951149523, 'sampling/sampling_logp_difference/max': 0.016472017392516136, 'sampling/importance_sampling_ratio/min': 0.9836629033088684, 'sampling/importance_sampling_ratio/mean': 1.0000567436218262, 'sampling/importance_sampling_ratio/max': 1.010071873664856, 'entropy': 0.0028824398177675903, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 75%|███████▍  | 2473/3300 [3:31:37<1:05:02,  4.72s/it]INFO 04-12 02:08:29 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▍  | 2474/3300 [3:31:44<1:11:55,  5.22s/it]2026-04-12 02:08:35,322 | INFO | train_grpo_train | metrics_logged mode=train step=2474

                                                       
{'loss': 0.0038, 'grad_norm': 1.5029007196426392, 'learning_rate': 2.506060606060606e-06, 'num_tokens': 5603217.0, 'completions/mean_length': 221.0, 'completions/min_length': 218.0, 'completions/max_length': 223.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 221.0, 'completions/min_terminated_length': 218.0, 'completions/max_terminated_length': 223.0, 'rewards/meter/mean': 0.9701865911483765, 'rewards/meter/std': 0.02245534211397171, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7232142686843872, 'rewards/repeat_penalty/std': 0.05960877984762192, 'rewards/total_composite/mean': 0.7016157507896423, 'rewards/total_composite/std': 0.06003959849476814, 'reward': 0.7016157507896423, 'reward_std': 0.060039594769477844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02541784755885601, 'sampling/sampling_logp_difference/max': 5.142219543457031, 'sampling/importance_sampling_ratio/min': 0.0058447024784982204, 'sampling/importance_sampling_ratio/mean': 0.9988224506378174, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.079356012865901, 'clip_ratio/low_mean': 0.015850404044613242, 'clip_ratio/low_min': 0.015850404044613242, 'clip_ratio/high_mean': 0.003954203391913325, 'clip_ratio/high_max': 0.003954203391913325, 'clip_ratio/region_mean': 0.019804607436526567, 'reward_total_mean': 0.7016157507896423, 'reward_meter_mean': 0.9701865911483765, 'reward_meter_std': 0.02245534211397171, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7232142686843872, 'reward_repeat_penalty_std': 0.05960877984762192, 'reward_total_composite_mean': 0.7016157507896423, 'reward_total_composite_std': 0.06003959849476814, 'epoch': 0.1}

 75%|███████▍  | 2474/3300 [3:31:44<1:11:55,  5.22s/it]INFO 04-12 02:08:35 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2475/3300 [3:31:50<1:17:45,  5.66s/it]2026-04-12 02:08:41,981 | INFO | train_grpo_train | metrics_logged mode=train step=2475

                                                       
{'loss': 0.0005, 'grad_norm': 1.4216914176940918, 'learning_rate': 2.5030303030303034e-06, 'num_tokens': 5606940.0, 'completions/mean_length': 246.375, 'completions/min_length': 244.0, 'completions/max_length': 248.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 246.375, 'completions/min_terminated_length': 244.0, 'completions/max_terminated_length': 248.0, 'rewards/meter/mean': 0.9980918169021606, 'rewards/meter/std': 0.00017555728845763952, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7211538553237915, 'rewards/repeat_penalty/std': 0.1158415824174881, 'rewards/total_composite/mean': 0.7197900414466858, 'rewards/total_composite/std': 0.11570559442043304, 'reward': 0.7197900414466858, 'reward_std': 0.11570560187101364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020571939647197723, 'sampling/sampling_logp_difference/max': 1.042820930480957, 'sampling/importance_sampling_ratio/min': 0.35245901346206665, 'sampling/importance_sampling_ratio/mean': 1.004142165184021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14181741513311863, 'clip_ratio/low_mean': 0.006118385819718242, 'clip_ratio/low_min': 0.006118385819718242, 'clip_ratio/high_mean': 0.010605149669572711, 'clip_ratio/high_max': 0.010605149669572711, 'clip_ratio/region_mean': 0.016723535489290953, 'reward_total_mean': 0.7197900414466858, 'reward_meter_mean': 0.9980918169021606, 'reward_meter_std': 0.00017555728845763952, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7211538553237915, 'reward_repeat_penalty_std': 0.1158415824174881, 'reward_total_composite_mean': 0.7197900414466858, 'reward_total_composite_std': 0.11570559442043304, 'epoch': 0.1}

 75%|███████▌  | 2475/3300 [3:31:50<1:17:45,  5.66s/it]INFO 04-12 02:08:42 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2476/3300 [3:31:55<1:13:03,  5.32s/it]2026-04-12 02:08:46,523 | INFO | train_grpo_train | metrics_logged mode=train step=2476

                                                       
{'loss': -0.003, 'grad_norm': 3.1809544563293457, 'learning_rate': 2.5e-06, 'num_tokens': 5608790.0, 'completions/mean_length': 68.25, 'completions/min_length': 67.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9992682933807373, 'rewards/meter/std': 0.00015381410776171833, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992682933807373, 'rewards/total_composite/std': 0.00015381410776171833, 'reward': 0.9992682933807373, 'reward_std': 0.00015381253615487367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03672777861356735, 'sampling/sampling_logp_difference/max': 0.9773459434509277, 'sampling/importance_sampling_ratio/min': 0.3855144679546356, 'sampling/importance_sampling_ratio/mean': 1.0070955753326416, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3049812354147434, 'clip_ratio/low_mean': 0.014760755351744592, 'clip_ratio/low_min': 0.014760755351744592, 'clip_ratio/high_mean': 0.02171556802932173, 'clip_ratio/high_max': 0.02171556802932173, 'clip_ratio/region_mean': 0.03647632338106632, 'reward_total_mean': 0.9992682933807373, 'reward_meter_mean': 0.9992682933807373, 'reward_meter_std': 0.00015381410776171833, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992682933807373, 'reward_total_composite_std': 0.00015381410776171833, 'epoch': 0.1}

 75%|███████▌  | 2476/3300 [3:31:55<1:13:03,  5.32s/it]INFO 04-12 02:08:46 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2477/3300 [3:32:03<1:25:25,  6.23s/it]2026-04-12 02:08:54,861 | INFO | train_grpo_train | metrics_logged mode=train step=2477

                                                       
{'loss': 0.001, 'grad_norm': 2.115460157394409, 'learning_rate': 2.496969696969697e-06, 'num_tokens': 5613495.0, 'completions/mean_length': 361.125, 'completions/min_length': 339.0, 'completions/max_length': 372.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 361.125, 'completions/min_terminated_length': 339.0, 'completions/max_terminated_length': 372.0, 'rewards/meter/mean': 0.9980402588844299, 'rewards/meter/std': 0.00200492306612432, 'rewards/count_adherence/mean': 0.7857142686843872, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.06458108127117157, 'rewards/total_composite/mean': 0.714102029800415, 'rewards/total_composite/std': 0.04969966039061546, 'reward': 0.714102029800415, 'reward_std': 0.049699652940034866, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047193821519613266, 'sampling/sampling_logp_difference/max': 1.9314990043640137, 'sampling/importance_sampling_ratio/min': 0.14493077993392944, 'sampling/importance_sampling_ratio/mean': 1.0080512762069702, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3918083682656288, 'clip_ratio/low_mean': 0.015457271714694798, 'clip_ratio/low_min': 0.015457271714694798, 'clip_ratio/high_mean': 0.02343087922781706, 'clip_ratio/high_max': 0.02343087922781706, 'clip_ratio/region_mean': 0.03888815094251186, 'reward_total_mean': 0.714102029800415, 'reward_meter_mean': 0.9980402588844299, 'reward_meter_std': 0.00200492306612432, 'reward_count_adherence_mean': 0.7857142686843872, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.06458108127117157, 'reward_total_composite_mean': 0.714102029800415, 'reward_total_composite_std': 0.04969966039061546, 'epoch': 0.1}

 75%|███████▌  | 2477/3300 [3:32:03<1:25:25,  6.23s/it]INFO 04-12 02:08:55 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2478/3300 [3:32:08<1:20:20,  5.86s/it]2026-04-12 02:08:59,877 | INFO | train_grpo_train | metrics_logged mode=train step=2478

                                                       
{'loss': 0.0048, 'grad_norm': 6.063255786895752, 'learning_rate': 2.4939393939393943e-06, 'num_tokens': 5615759.0, 'completions/mean_length': 120.0, 'completions/min_length': 117.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.0, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9968112111091614, 'rewards/meter/std': 0.004492159932851791, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9968112111091614, 'rewards/total_composite/std': 0.004492159932851791, 'reward': 0.9968112111091614, 'reward_std': 0.004492142703384161, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03631357103586197, 'sampling/sampling_logp_difference/max': 0.8470320701599121, 'sampling/importance_sampling_ratio/min': 0.42868536710739136, 'sampling/importance_sampling_ratio/mean': 1.0102128982543945, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3629257455468178, 'clip_ratio/low_mean': 0.0020833334419876337, 'clip_ratio/low_min': 0.0020833334419876337, 'clip_ratio/high_mean': 0.036573544377461076, 'clip_ratio/high_max': 0.036573544377461076, 'clip_ratio/region_mean': 0.03865687781944871, 'reward_total_mean': 0.9968112111091614, 'reward_meter_mean': 0.9968112111091614, 'reward_meter_std': 0.004492159932851791, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9968112111091614, 'reward_total_composite_std': 0.004492159932851791, 'epoch': 0.1}

 75%|███████▌  | 2478/3300 [3:32:08<1:20:20,  5.86s/it]INFO 04-12 02:09:00 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2479/3300 [3:32:13<1:14:53,  5.47s/it]2026-04-12 02:09:04,442 | INFO | train_grpo_train | metrics_logged mode=train step=2479

                                                       
{'loss': 0.0172, 'grad_norm': 5.862101078033447, 'learning_rate': 2.490909090909091e-06, 'num_tokens': 5617602.0, 'completions/mean_length': 80.375, 'completions/min_length': 78.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9985413551330566, 'rewards/meter/std': 0.0006061770836822689, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9985413551330566, 'rewards/total_composite/std': 0.0006061770836822689, 'reward': 0.9985413551330566, 'reward_std': 0.0006061761523596942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03767424076795578, 'sampling/sampling_logp_difference/max': 1.3134748935699463, 'sampling/importance_sampling_ratio/min': 0.2688840925693512, 'sampling/importance_sampling_ratio/mean': 1.0086257457733154, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26199326291680336, 'clip_ratio/low_mean': 0.010923440335318446, 'clip_ratio/low_min': 0.010923440335318446, 'clip_ratio/high_mean': 0.00933689041994512, 'clip_ratio/high_max': 0.00933689041994512, 'clip_ratio/region_mean': 0.020260330755263567, 'reward_total_mean': 0.9985413551330566, 'reward_meter_mean': 0.9985413551330566, 'reward_meter_std': 0.0006061770836822689, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9985413551330566, 'reward_total_composite_std': 0.0006061770836822689, 'epoch': 0.1}

 75%|███████▌  | 2479/3300 [3:32:13<1:14:53,  5.47s/it]INFO 04-12 02:09:04 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2480/3300 [3:32:18<1:13:25,  5.37s/it]2026-04-12 02:09:09,578 | INFO | train_grpo_train | metrics_logged mode=train step=2480

                                                       
{'loss': 0.01, 'grad_norm': 2.8270602226257324, 'learning_rate': 2.487878787878788e-06, 'num_tokens': 5619995.0, 'completions/mean_length': 127.125, 'completions/min_length': 126.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.125, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.7682651281356812, 'rewards/meter/std': 0.1492764949798584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.7446907758712769, 'rewards/total_composite/std': 0.17325718700885773, 'reward': 0.7446907758712769, 'reward_std': 0.17325717210769653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03481648862361908, 'sampling/sampling_logp_difference/max': 1.9265758991241455, 'sampling/importance_sampling_ratio/min': 0.14564606547355652, 'sampling/importance_sampling_ratio/mean': 1.0069787502288818, 'sampling/importance_sampling_ratio/max': 1.9678964614868164, 'entropy': 0.25193316861987114, 'clip_ratio/low_mean': 0.009811883908696473, 'clip_ratio/low_min': 0.009811883908696473, 'clip_ratio/high_mean': 0.016681208508089185, 'clip_ratio/high_max': 0.016681208508089185, 'clip_ratio/region_mean': 0.026493092416785657, 'reward_total_mean': 0.7446907758712769, 'reward_meter_mean': 0.7682651281356812, 'reward_meter_std': 0.1492764949798584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.7446907758712769, 'reward_total_composite_std': 0.17325718700885773, 'epoch': 0.1}

 75%|███████▌  | 2480/3300 [3:32:18<1:13:25,  5.37s/it]INFO 04-12 02:09:09 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2481/3300 [3:32:22<1:09:17,  5.08s/it]2026-04-12 02:09:13,963 | INFO | train_grpo_train | metrics_logged mode=train step=2481

                                                       
{'loss': 0.0178, 'grad_norm': 5.6794843673706055, 'learning_rate': 2.4848484848484848e-06, 'num_tokens': 5621792.0, 'completions/mean_length': 65.625, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9421236515045166, 'rewards/meter/std': 0.09346310049295425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9421236515045166, 'rewards/total_composite/std': 0.09346310049295425, 'reward': 0.9421236515045166, 'reward_std': 0.09346310794353485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02599693275988102, 'sampling/sampling_logp_difference/max': 0.5891776084899902, 'sampling/importance_sampling_ratio/min': 0.5547833442687988, 'sampling/importance_sampling_ratio/mean': 1.0088194608688354, 'sampling/importance_sampling_ratio/max': 1.7557963132858276, 'entropy': 0.1946810930967331, 'clip_ratio/low_mean': 0.009328358108177781, 'clip_ratio/low_min': 0.009328358108177781, 'clip_ratio/high_mean': 0.02106643421575427, 'clip_ratio/high_max': 0.02106643421575427, 'clip_ratio/region_mean': 0.03039479232393205, 'reward_total_mean': 0.9421236515045166, 'reward_meter_mean': 0.9421236515045166, 'reward_meter_std': 0.09346310049295425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9421236515045166, 'reward_total_composite_std': 0.09346310049295425, 'epoch': 0.1}

 75%|███████▌  | 2481/3300 [3:32:22<1:09:17,  5.08s/it]INFO 04-12 02:09:14 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2482/3300 [3:32:27<1:09:01,  5.06s/it]2026-04-12 02:09:18,996 | INFO | train_grpo_train | metrics_logged mode=train step=2482

                                                       
{'loss': 0.1809, 'grad_norm': 28.66371726989746, 'learning_rate': 2.481818181818182e-06, 'num_tokens': 5623580.0, 'completions/mean_length': 60.5, 'completions/min_length': 54.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7704848051071167, 'rewards/meter/std': 0.03176296874880791, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6806069612503052, 'rewards/total_composite/std': 0.19818444550037384, 'reward': 0.6806069612503052, 'reward_std': 0.19818443059921265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.001127883093431592, 'sampling/sampling_logp_difference/max': 0.07983016967773438, 'sampling/importance_sampling_ratio/min': 0.9490934610366821, 'sampling/importance_sampling_ratio/mean': 1.0007379055023193, 'sampling/importance_sampling_ratio/max': 1.083103060722351, 'entropy': 0.007361570082139224, 'clip_ratio/low_mean': 0.0015625000232830644, 'clip_ratio/low_min': 0.0015625000232830644, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0015625000232830644, 'reward_total_mean': 0.6806069612503052, 'reward_meter_mean': 0.7704848051071167, 'reward_meter_std': 0.03176296874880791, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6806069612503052, 'reward_total_composite_std': 0.19818444550037384, 'epoch': 0.1}

 75%|███████▌  | 2482/3300 [3:32:27<1:09:01,  5.06s/it]INFO 04-12 02:09:19 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2483/3300 [3:32:32<1:06:14,  4.86s/it]2026-04-12 02:09:23,397 | INFO | train_grpo_train | metrics_logged mode=train step=2483

                                                       
{'loss': 0.0225, 'grad_norm': 3.7192165851593018, 'learning_rate': 2.478787878787879e-06, 'num_tokens': 5625374.0, 'completions/mean_length': 61.25, 'completions/min_length': 56.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.25, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9937635660171509, 'rewards/meter/std': 0.0006408250774256885, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9937635660171509, 'rewards/total_composite/std': 0.0006408250774256885, 'reward': 0.9937635660171509, 'reward_std': 0.0006408333429135382, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015063888393342495, 'sampling/sampling_logp_difference/max': 1.5699043273925781, 'sampling/importance_sampling_ratio/min': 0.20806507766246796, 'sampling/importance_sampling_ratio/mean': 0.998705267906189, 'sampling/importance_sampling_ratio/max': 1.8052740097045898, 'entropy': 0.049932122230529785, 'clip_ratio/low_mean': 0.004032257944345474, 'clip_ratio/low_min': 0.004032257944345474, 'clip_ratio/high_mean': 0.00828052987344563, 'clip_ratio/high_max': 0.00828052987344563, 'clip_ratio/region_mean': 0.012312787817791104, 'reward_total_mean': 0.9937635660171509, 'reward_meter_mean': 0.9937635660171509, 'reward_meter_std': 0.0006408250774256885, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9937635660171509, 'reward_total_composite_std': 0.0006408250774256885, 'epoch': 0.1}

 75%|███████▌  | 2483/3300 [3:32:32<1:06:14,  4.86s/it]INFO 04-12 02:09:23 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2484/3300 [3:32:36<1:04:25,  4.74s/it]2026-04-12 02:09:27,839 | INFO | train_grpo_train | metrics_logged mode=train step=2484

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.475757575757576e-06, 'num_tokens': 5627126.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00020638658315874636, 'sampling/sampling_logp_difference/max': 0.007955902256071568, 'sampling/importance_sampling_ratio/min': 0.9975246787071228, 'sampling/importance_sampling_ratio/mean': 1.0001825094223022, 'sampling/importance_sampling_ratio/max': 1.00798761844635, 'entropy': 0.002001831730012782, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 75%|███████▌  | 2484/3300 [3:32:36<1:04:25,  4.74s/it]INFO 04-12 02:09:28 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2485/3300 [3:32:45<1:19:39,  5.86s/it]2026-04-12 02:09:36,337 | INFO | train_grpo_train | metrics_logged mode=train step=2485

                                                       
{'loss': -0.0278, 'grad_norm': 1.8019484281539917, 'learning_rate': 2.472727272727273e-06, 'num_tokens': 5631659.0, 'completions/mean_length': 363.625, 'completions/min_length': 343.0, 'completions/max_length': 393.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 363.625, 'completions/min_terminated_length': 343.0, 'completions/max_terminated_length': 393.0, 'rewards/meter/mean': 0.9972596764564514, 'rewards/meter/std': 0.004136989358812571, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.04629101976752281, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9709967374801636, 'rewards/repeat_penalty/std': 0.031024247407913208, 'rewards/total_composite/mean': 0.8964345455169678, 'rewards/total_composite/std': 0.0664711743593216, 'reward': 0.8964345455169678, 'reward_std': 0.0664711743593216, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.052424363791942596, 'sampling/sampling_logp_difference/max': 2.488617181777954, 'sampling/importance_sampling_ratio/min': 0.08302469551563263, 'sampling/importance_sampling_ratio/mean': 1.0112888813018799, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4474870078265667, 'clip_ratio/low_mean': 0.01650261995382607, 'clip_ratio/low_min': 0.01650261995382607, 'clip_ratio/high_mean': 0.021971354726701975, 'clip_ratio/high_max': 0.021971354726701975, 'clip_ratio/region_mean': 0.038473974680528045, 'reward_total_mean': 0.8964345455169678, 'reward_meter_mean': 0.9972596764564514, 'reward_meter_std': 0.004136989358812571, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.04629101976752281, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9709967374801636, 'reward_repeat_penalty_std': 0.031024247407913208, 'reward_total_composite_mean': 0.8964345455169678, 'reward_total_composite_std': 0.0664711743593216, 'epoch': 0.1}

 75%|███████▌  | 2485/3300 [3:32:45<1:19:39,  5.86s/it]INFO 04-12 02:09:36 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2486/3300 [3:32:50<1:19:03,  5.83s/it]2026-04-12 02:09:42,075 | INFO | train_grpo_train | metrics_logged mode=train step=2486

                                                       
{'loss': 0.0133, 'grad_norm': 2.246661901473999, 'learning_rate': 2.46969696969697e-06, 'num_tokens': 5634271.0, 'completions/mean_length': 158.5, 'completions/min_length': 156.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.5, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.9984291791915894, 'rewards/meter/std': 0.0007291806978173554, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984291791915894, 'rewards/total_composite/std': 0.0007291806978173554, 'reward': 0.9984291791915894, 'reward_std': 0.0007291696383617818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049818817526102066, 'sampling/sampling_logp_difference/max': 1.498807430267334, 'sampling/importance_sampling_ratio/min': 0.223396435379982, 'sampling/importance_sampling_ratio/mean': 1.0025769472122192, 'sampling/importance_sampling_ratio/max': 1.7283272743225098, 'entropy': 0.45672084018588066, 'clip_ratio/low_mean': 0.005464080721139908, 'clip_ratio/low_min': 0.005464080721139908, 'clip_ratio/high_mean': 0.03322292352095246, 'clip_ratio/high_max': 0.03322292352095246, 'clip_ratio/region_mean': 0.03868700424209237, 'reward_total_mean': 0.9984291791915894, 'reward_meter_mean': 0.9984291791915894, 'reward_meter_std': 0.0007291806978173554, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984291791915894, 'reward_total_composite_std': 0.0007291806978173554, 'epoch': 0.1}

 75%|███████▌  | 2486/3300 [3:32:51<1:19:03,  5.83s/it]INFO 04-12 02:09:42 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2487/3300 [3:32:55<1:13:34,  5.43s/it]2026-04-12 02:09:46,577 | INFO | train_grpo_train | metrics_logged mode=train step=2487

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.466666666666667e-06, 'num_tokens': 5636167.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00024826714070513844, 'sampling/sampling_logp_difference/max': 0.011286328546702862, 'sampling/importance_sampling_ratio/min': 0.9940236806869507, 'sampling/importance_sampling_ratio/mean': 1.0001945495605469, 'sampling/importance_sampling_ratio/max': 1.0113502740859985, 'entropy': 0.0025851023383438587, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 75%|███████▌  | 2487/3300 [3:32:55<1:13:34,  5.43s/it]INFO 04-12 02:09:46 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2488/3300 [3:33:00<1:11:16,  5.27s/it]2026-04-12 02:09:51,459 | INFO | train_grpo_train | metrics_logged mode=train step=2488

                                                       
{'loss': -0.0124, 'grad_norm': 0.9873250126838684, 'learning_rate': 2.463636363636364e-06, 'num_tokens': 5638299.0, 'completions/mean_length': 97.5, 'completions/min_length': 94.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.5, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9964094758033752, 'rewards/meter/std': 0.004482356831431389, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9964094758033752, 'rewards/total_composite/std': 0.004482356831431389, 'reward': 0.9964094758033752, 'reward_std': 0.00448233587667346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0035261192824691534, 'sampling/sampling_logp_difference/max': 1.280348777770996, 'sampling/importance_sampling_ratio/min': 0.2779403328895569, 'sampling/importance_sampling_ratio/mean': 0.9996676445007324, 'sampling/importance_sampling_ratio/max': 1.091353416442871, 'entropy': 0.01773504843004048, 'clip_ratio/low_mean': 0.0013297871919348836, 'clip_ratio/low_min': 0.0013297871919348836, 'clip_ratio/high_mean': 0.0038265305338427424, 'clip_ratio/high_max': 0.0038265305338427424, 'clip_ratio/region_mean': 0.005156317725777626, 'reward_total_mean': 0.9964094758033752, 'reward_meter_mean': 0.9964094758033752, 'reward_meter_std': 0.004482356831431389, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9964094758033752, 'reward_total_composite_std': 0.004482356831431389, 'epoch': 0.1}

 75%|███████▌  | 2488/3300 [3:33:00<1:11:16,  5.27s/it]INFO 04-12 02:09:51 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2489/3300 [3:33:04<1:06:43,  4.94s/it]2026-04-12 02:09:55,629 | INFO | train_grpo_train | metrics_logged mode=train step=2489

                                                       
{'loss': 0.0144, 'grad_norm': 19.695940017700195, 'learning_rate': 2.4606060606060607e-06, 'num_tokens': 5639817.0, 'completions/mean_length': 23.75, 'completions/min_length': 21.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9360402822494507, 'rewards/meter/std': 0.019930923357605934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9360402822494507, 'rewards/total_composite/std': 0.019930923357605934, 'reward': 0.9360402822494507, 'reward_std': 0.019930915907025337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09426023066043854, 'sampling/sampling_logp_difference/max': 1.1065075397491455, 'sampling/importance_sampling_ratio/min': 0.3307119607925415, 'sampling/importance_sampling_ratio/mean': 1.0085498094558716, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5839281938970089, 'clip_ratio/low_mean': 0.05386740108951926, 'clip_ratio/low_min': 0.05386740108951926, 'clip_ratio/high_mean': 0.03615045174956322, 'clip_ratio/high_max': 0.03615045174956322, 'clip_ratio/region_mean': 0.09001785283908248, 'reward_total_mean': 0.9360402822494507, 'reward_meter_mean': 0.9360402822494507, 'reward_meter_std': 0.019930923357605934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9360402822494507, 'reward_total_composite_std': 0.019930923357605934, 'epoch': 0.1}

 75%|███████▌  | 2489/3300 [3:33:04<1:06:43,  4.94s/it]INFO 04-12 02:09:55 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2490/3300 [3:33:09<1:07:07,  4.97s/it]2026-04-12 02:10:00,684 | INFO | train_grpo_train | metrics_logged mode=train step=2490

                                                       
{'loss': 0.0052, 'grad_norm': 1.8480373620986938, 'learning_rate': 2.457575757575758e-06, 'num_tokens': 5642037.0, 'completions/mean_length': 107.5, 'completions/min_length': 106.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.5, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9976269602775574, 'rewards/meter/std': 0.0004338659346103668, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9727107882499695, 'rewards/total_composite/std': 0.0708698034286499, 'reward': 0.9727107882499695, 'reward_std': 0.0708698257803917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01844092458486557, 'sampling/sampling_logp_difference/max': 2.3559987545013428, 'sampling/importance_sampling_ratio/min': 0.09479877352714539, 'sampling/importance_sampling_ratio/mean': 1.003526210784912, 'sampling/importance_sampling_ratio/max': 1.686613917350769, 'entropy': 0.12307433690875769, 'clip_ratio/low_mean': 0.0022935778833925724, 'clip_ratio/low_min': 0.0022935778833925724, 'clip_ratio/high_mean': 0.008156140334904194, 'clip_ratio/high_max': 0.008156140334904194, 'clip_ratio/region_mean': 0.010449718218296766, 'reward_total_mean': 0.9727107882499695, 'reward_meter_mean': 0.9976269602775574, 'reward_meter_std': 0.0004338659346103668, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9727107882499695, 'reward_total_composite_std': 0.0708698034286499, 'epoch': 0.1}

 75%|███████▌  | 2490/3300 [3:33:09<1:07:07,  4.97s/it]INFO 04-12 02:10:00 [block_pool.py:378] Successfully reset prefix cache

 75%|███████▌  | 2491/3300 [3:33:14<1:04:48,  4.81s/it]2026-04-12 02:10:05,103 | INFO | train_grpo_train | metrics_logged mode=train step=2491

                                                       
{'loss': -0.0008, 'grad_norm': 2.2775914669036865, 'learning_rate': 2.454545454545455e-06, 'num_tokens': 5643821.0, 'completions/mean_length': 62.0, 'completions/min_length': 62.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9934873580932617, 'rewards/meter/std': 0.0007373938569799066, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9934873580932617, 'rewards/total_composite/std': 0.0007373938569799066, 'reward': 0.9934873580932617, 'reward_std': 0.000737398921046406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007004985585808754, 'sampling/sampling_logp_difference/max': 0.4697551727294922, 'sampling/importance_sampling_ratio/min': 0.6251553297042847, 'sampling/importance_sampling_ratio/mean': 1.001605749130249, 'sampling/importance_sampling_ratio/max': 1.28243088722229, 'entropy': 0.040033137891441584, 'clip_ratio/low_mean': 0.004032257944345474, 'clip_ratio/low_min': 0.004032257944345474, 'clip_ratio/high_mean': 0.002016128972172737, 'clip_ratio/high_max': 0.002016128972172737, 'clip_ratio/region_mean': 0.006048386916518211, 'reward_total_mean': 0.9934873580932617, 'reward_meter_mean': 0.9934873580932617, 'reward_meter_std': 0.0007373938569799066, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9934873580932617, 'reward_total_composite_std': 0.0007373938569799066, 'epoch': 0.1}

 75%|███████▌  | 2491/3300 [3:33:14<1:04:48,  4.81s/it]INFO 04-12 02:10:05 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2492/3300 [3:33:19<1:06:23,  4.93s/it]2026-04-12 02:10:10,322 | INFO | train_grpo_train | metrics_logged mode=train step=2492

                                                       
{'loss': -0.006, 'grad_norm': 3.0206403732299805, 'learning_rate': 2.4515151515151516e-06, 'num_tokens': 5646242.0, 'completions/mean_length': 125.625, 'completions/min_length': 123.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.625, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9930568337440491, 'rewards/meter/std': 0.00019196225912310183, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9575905799865723, 'rewards/total_composite/std': 0.06567274034023285, 'reward': 0.9575905799865723, 'reward_std': 0.06567277014255524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018952438607811928, 'sampling/sampling_logp_difference/max': 2.4076790809631348, 'sampling/importance_sampling_ratio/min': 0.09002399444580078, 'sampling/importance_sampling_ratio/mean': 1.001326322555542, 'sampling/importance_sampling_ratio/max': 1.629332184791565, 'entropy': 0.10137755703181028, 'clip_ratio/low_mean': 0.0010080644860863686, 'clip_ratio/low_min': 0.0010080644860863686, 'clip_ratio/high_mean': 0.01584376988466829, 'clip_ratio/high_max': 0.01584376988466829, 'clip_ratio/region_mean': 0.01685183437075466, 'reward_total_mean': 0.9575905799865723, 'reward_meter_mean': 0.9930568337440491, 'reward_meter_std': 0.00019196225912310183, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9575905799865723, 'reward_total_composite_std': 0.06567274034023285, 'epoch': 0.1}

 76%|███████▌  | 2492/3300 [3:33:19<1:06:23,  4.93s/it]INFO 04-12 02:10:10 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2493/3300 [3:33:25<1:10:08,  5.22s/it]2026-04-12 02:10:16,207 | INFO | train_grpo_train | metrics_logged mode=train step=2493

                                                       
{'loss': 0.0227, 'grad_norm': 3.180861234664917, 'learning_rate': 2.4484848484848485e-06, 'num_tokens': 5649068.0, 'completions/mean_length': 162.25, 'completions/min_length': 154.0, 'completions/max_length': 172.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 162.25, 'completions/min_terminated_length': 154.0, 'completions/max_terminated_length': 172.0, 'rewards/meter/mean': 0.904288649559021, 'rewards/meter/std': 0.09930223226547241, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.881544828414917, 'rewards/total_composite/std': 0.1262788623571396, 'reward': 0.881544828414917, 'reward_std': 0.1262788623571396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041033390909433365, 'sampling/sampling_logp_difference/max': 1.6256977319717407, 'sampling/importance_sampling_ratio/min': 0.19677433371543884, 'sampling/importance_sampling_ratio/mean': 1.0083729028701782, 'sampling/importance_sampling_ratio/max': 1.7914950847625732, 'entropy': 0.4163212776184082, 'clip_ratio/low_mean': 0.005151562392711639, 'clip_ratio/low_min': 0.005151562392711639, 'clip_ratio/high_mean': 0.025025799637660384, 'clip_ratio/high_max': 0.025025799637660384, 'clip_ratio/region_mean': 0.030177362030372024, 'reward_total_mean': 0.881544828414917, 'reward_meter_mean': 0.904288649559021, 'reward_meter_std': 0.09930223226547241, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.881544828414917, 'reward_total_composite_std': 0.1262788623571396, 'epoch': 0.1}

 76%|███████▌  | 2493/3300 [3:33:25<1:10:08,  5.22s/it]INFO 04-12 02:10:16 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2494/3300 [3:33:29<1:06:57,  4.98s/it]2026-04-12 02:10:20,653 | INFO | train_grpo_train | metrics_logged mode=train step=2494

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.4454545454545457e-06, 'num_tokens': 5650820.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0011890914756804705, 'sampling/sampling_logp_difference/max': 0.09505826234817505, 'sampling/importance_sampling_ratio/min': 0.9093199372291565, 'sampling/importance_sampling_ratio/mean': 1.000666856765747, 'sampling/importance_sampling_ratio/max': 1.0194705724716187, 'entropy': 0.010702198022045195, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 76%|███████▌  | 2494/3300 [3:33:29<1:06:57,  4.98s/it]INFO 04-12 02:10:20 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2495/3300 [3:33:34<1:06:26,  4.95s/it]2026-04-12 02:10:25,530 | INFO | train_grpo_train | metrics_logged mode=train step=2495

                                                       
{'loss': 0.0059, 'grad_norm': 1.8723995685577393, 'learning_rate': 2.4424242424242426e-06, 'num_tokens': 5652882.0, 'completions/mean_length': 94.75, 'completions/min_length': 93.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.75, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9930741786956787, 'rewards/meter/std': 0.0006547096418216825, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9682368040084839, 'rewards/total_composite/std': 0.07008406519889832, 'reward': 0.9682368040084839, 'reward_std': 0.07008406519889832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018359439447522163, 'sampling/sampling_logp_difference/max': 1.3887449502944946, 'sampling/importance_sampling_ratio/min': 0.24938809871673584, 'sampling/importance_sampling_ratio/mean': 1.0004006624221802, 'sampling/importance_sampling_ratio/max': 1.5056626796722412, 'entropy': 0.10162492096424103, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.014545450918376446, 'clip_ratio/high_max': 0.014545450918376446, 'clip_ratio/region_mean': 0.014545450918376446, 'reward_total_mean': 0.9682368040084839, 'reward_meter_mean': 0.9930741786956787, 'reward_meter_std': 0.0006547096418216825, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9682368040084839, 'reward_total_composite_std': 0.07008406519889832, 'epoch': 0.1}

 76%|███████▌  | 2495/3300 [3:33:34<1:06:26,  4.95s/it]INFO 04-12 02:10:25 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2496/3300 [3:33:39<1:06:26,  4.96s/it]2026-04-12 02:10:30,503 | INFO | train_grpo_train | metrics_logged mode=train step=2496

                                                       
{'loss': 0.0159, 'grad_norm': 3.207080602645874, 'learning_rate': 2.4393939393939394e-06, 'num_tokens': 5655174.0, 'completions/mean_length': 105.5, 'completions/min_length': 98.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.5, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9972423315048218, 'rewards/meter/std': 0.0021226475946605206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9473263025283813, 'rewards/total_composite/std': 0.09178309142589569, 'reward': 0.9473263025283813, 'reward_std': 0.0917830839753151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03410215303301811, 'sampling/sampling_logp_difference/max': 1.372025489807129, 'sampling/importance_sampling_ratio/min': 0.2535927891731262, 'sampling/importance_sampling_ratio/mean': 1.0094503164291382, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23731265030801296, 'clip_ratio/low_mean': 0.004629629664123058, 'clip_ratio/low_min': 0.004629629664123058, 'clip_ratio/high_mean': 0.02076363516971469, 'clip_ratio/high_max': 0.02076363516971469, 'clip_ratio/region_mean': 0.025393264833837748, 'reward_total_mean': 0.9473263025283813, 'reward_meter_mean': 0.9972423315048218, 'reward_meter_std': 0.0021226475946605206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9473263025283813, 'reward_total_composite_std': 0.09178309142589569, 'epoch': 0.1}

 76%|███████▌  | 2496/3300 [3:33:39<1:06:26,  4.96s/it]INFO 04-12 02:10:30 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2497/3300 [3:33:45<1:11:38,  5.35s/it]2026-04-12 02:10:36,778 | INFO | train_grpo_train | metrics_logged mode=train step=2497

                                                       
{'loss': 0.0681, 'grad_norm': 5.552990913391113, 'learning_rate': 2.4363636363636366e-06, 'num_tokens': 5658127.0, 'completions/mean_length': 175.125, 'completions/min_length': 166.0, 'completions/max_length': 206.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 175.125, 'completions/min_terminated_length': 166.0, 'completions/max_terminated_length': 206.0, 'rewards/meter/mean': 0.9639856219291687, 'rewards/meter/std': 0.09431485831737518, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8185431361198425, 'rewards/total_composite/std': 0.03557335212826729, 'reward': 0.8185431361198425, 'reward_std': 0.035573359578847885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0666884258389473, 'sampling/sampling_logp_difference/max': 2.1122121810913086, 'sampling/importance_sampling_ratio/min': 0.12097006291151047, 'sampling/importance_sampling_ratio/mean': 1.0070842504501343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5811385102570057, 'clip_ratio/low_mean': 0.004247572738677263, 'clip_ratio/low_min': 0.004247572738677263, 'clip_ratio/high_mean': 0.05119223310612142, 'clip_ratio/high_max': 0.05119223310612142, 'clip_ratio/region_mean': 0.055439805844798684, 'reward_total_mean': 0.8185431361198425, 'reward_meter_mean': 0.9639856219291687, 'reward_meter_std': 0.09431485831737518, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8185431361198425, 'reward_total_composite_std': 0.03557335212826729, 'epoch': 0.1}

 76%|███████▌  | 2497/3300 [3:33:45<1:11:38,  5.35s/it]INFO 04-12 02:10:37 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2498/3300 [3:33:50<1:10:13,  5.25s/it]2026-04-12 02:10:41,801 | INFO | train_grpo_train | metrics_logged mode=train step=2498

                                                       
{'loss': -0.0013, 'grad_norm': 3.1228280067443848, 'learning_rate': 2.4333333333333335e-06, 'num_tokens': 5660216.0, 'completions/mean_length': 97.125, 'completions/min_length': 95.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.125, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.991081714630127, 'rewards/meter/std': 0.004786557983607054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.949999988079071, 'rewards/repeat_penalty/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.9413519501686096, 'rewards/total_composite/std': 0.09001474827528, 'reward': 0.9413519501686096, 'reward_std': 0.09001474827528, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031947266310453415, 'sampling/sampling_logp_difference/max': 0.8252577781677246, 'sampling/importance_sampling_ratio/min': 0.43812206387519836, 'sampling/importance_sampling_ratio/mean': 1.0101373195648193, 'sampling/importance_sampling_ratio/max': 1.695669412612915, 'entropy': 0.30781341157853603, 'clip_ratio/low_mean': 0.0013020833721384406, 'clip_ratio/low_min': 0.0013020833721384406, 'clip_ratio/high_mean': 0.01925921393558383, 'clip_ratio/high_max': 0.01925921393558383, 'clip_ratio/region_mean': 0.02056129730772227, 'reward_total_mean': 0.9413519501686096, 'reward_meter_mean': 0.991081714630127, 'reward_meter_std': 0.004786557983607054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.949999988079071, 'reward_repeat_penalty_std': 0.09258200973272324, 'reward_total_composite_mean': 0.9413519501686096, 'reward_total_composite_std': 0.09001474827528, 'epoch': 0.1}

 76%|███████▌  | 2498/3300 [3:33:50<1:10:13,  5.25s/it]INFO 04-12 02:10:42 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2499/3300 [3:33:55<1:07:34,  5.06s/it]2026-04-12 02:10:46,406 | INFO | train_grpo_train | metrics_logged mode=train step=2499

                                                       
{'loss': 0.008, 'grad_norm': 3.2906627655029297, 'learning_rate': 2.4303030303030307e-06, 'num_tokens': 5661956.0, 'completions/mean_length': 67.5, 'completions/min_length': 67.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9980784058570862, 'rewards/meter/std': 0.00020892193424515426, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980784058570862, 'rewards/total_composite/std': 0.00020892193424515426, 'reward': 0.9980784058570862, 'reward_std': 0.00020891590975224972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0033059667330235243, 'sampling/sampling_logp_difference/max': 1.0325679779052734, 'sampling/importance_sampling_ratio/min': 0.35609135031700134, 'sampling/importance_sampling_ratio/mean': 0.9999598860740662, 'sampling/importance_sampling_ratio/max': 1.0832960605621338, 'entropy': 0.015344980405643582, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980784058570862, 'reward_meter_mean': 0.9980784058570862, 'reward_meter_std': 0.00020892193424515426, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980784058570862, 'reward_total_composite_std': 0.00020892193424515426, 'epoch': 0.1}

 76%|███████▌  | 2499/3300 [3:33:55<1:07:34,  5.06s/it]INFO 04-12 02:10:46 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2500/3300 [3:34:00<1:08:23,  5.13s/it]2026-04-12 02:10:51,702 | INFO | train_grpo_train | metrics_logged mode=train step=2500

                                                       
{'loss': -0.0048, 'grad_norm': 3.226564884185791, 'learning_rate': 2.4272727272727276e-06, 'num_tokens': 5664327.0, 'completions/mean_length': 131.375, 'completions/min_length': 130.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.375, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.998822808265686, 'rewards/meter/std': 0.0007065999088808894, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9096584916114807, 'rewards/total_composite/std': 0.07408842444419861, 'reward': 0.9096584916114807, 'reward_std': 0.07408840954303741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01683160476386547, 'sampling/sampling_logp_difference/max': 1.742103099822998, 'sampling/importance_sampling_ratio/min': 0.17515164613723755, 'sampling/importance_sampling_ratio/mean': 0.9981679916381836, 'sampling/importance_sampling_ratio/max': 1.8086941242218018, 'entropy': 0.05698791332542896, 'clip_ratio/low_mean': 0.011479852895718068, 'clip_ratio/low_min': 0.011479852895718068, 'clip_ratio/high_mean': 0.004727728548459709, 'clip_ratio/high_max': 0.004727728548459709, 'clip_ratio/region_mean': 0.016207581444177777, 'reward_total_mean': 0.9096584916114807, 'reward_meter_mean': 0.998822808265686, 'reward_meter_std': 0.0007065999088808894, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9096584916114807, 'reward_total_composite_std': 0.07408842444419861, 'epoch': 0.1}

 76%|███████▌  | 2500/3300 [3:34:00<1:08:23,  5.13s/it]INFO 04-12 02:10:51 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:36,  3.34s/it]

 23%|██▎       | 3/13 [00:10<00:33,  3.34s/it]

 31%|███       | 4/13 [00:15<00:38,  4.23s/it]

 38%|███▊      | 5/13 [00:21<00:36,  4.62s/it]

 46%|████▌     | 6/13 [00:27<00:35,  5.14s/it]

 54%|█████▍    | 7/13 [00:32<00:31,  5.24s/it]

 62%|██████▏   | 8/13 [00:39<00:29,  5.83s/it]

 69%|██████▉   | 9/13 [00:46<00:24,  6.23s/it]

 77%|███████▋  | 10/13 [00:51<00:17,  5.80s/it]

 85%|████████▍ | 11/13 [00:57<00:11,  5.82s/it]

 92%|█████████▏| 12/13 [01:03<00:05,  5.81s/it]

100%|██████████| 13/13 [01:08<00:00,  5.68s/it]2026-04-12 02:12:07,245 | INFO | train_grpo_train | metrics_logged mode=eval step=2500

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 75.4547, 'eval_samples_per_second': 1.378, 'eval_steps_per_second': 0.172, 'eval_num_tokens': 5664327.0, 'eval_completions/mean_length': 210.92307692307693, 'eval_completions/min_length': 62.38461538461539, 'eval_completions/max_length': 405.3076923076923, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 210.92307692307693, 'eval_completions/min_terminated_length': 62.38461538461539, 'eval_completions/max_terminated_length': 405.3076923076923, 'eval_rewards/meter/mean': 0.7411972742814285, 'eval_rewards/meter/std': 0.36808492687459177, 'eval_rewards/count_adherence/mean': 0.9494116031206571, 'eval_rewards/count_adherence/std': 0.07668151821081455, 'eval_rewards/arabic_clean/mean': 0.9711538461538461, 'eval_rewards/arabic_clean/std': 0.08158924258672275, 'eval_rewards/repeat_penalty/mean': 0.9134463484470661, 'eval_rewards/repeat_penalty/std': 0.10245848590364823, 'eval_rewards/total_composite/mean': 0.6316862404346466, 'eval_rewards/total_composite/std': 0.3585601345850871, 'eval_reward': 0.6316862404346466, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03223242281148067, 'eval_sampling/sampling_logp_difference/max': 1.1953271352327788, 'eval_sampling/importance_sampling_ratio/min': 0.31562885412803066, 'eval_sampling/importance_sampling_ratio/mean': 1.0096348799191988, 'eval_sampling/importance_sampling_ratio/max': 1.5564926862716675, 'eval_entropy': 0.37157516181468964, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.6316862404346466, 'eval_reward_meter_mean': 0.7411972742814285, 'eval_reward_meter_std': 0.36808492687459177, 'eval_reward_count_adherence_mean': 0.9494116031206571, 'eval_reward_count_adherence_std': 0.07668151821081455, 'eval_reward_arabic_clean_mean': 0.9711538461538461, 'eval_reward_arabic_clean_std': 0.08158924258672275, 'eval_reward_repeat_penalty_mean': 0.9134463484470661, 'eval_reward_repeat_penalty_std': 0.10245848590364823, 'eval_reward_total_composite_mean': 0.6316862404346466, 'eval_reward_total_composite_std': 0.3585601345850871, 'epoch': 0.1}

 76%|███████▌  | 2500/3300 [3:35:16<1:08:23,  5.13s/it]

100%|██████████| 13/13 [01:08<00:00,  5.68s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 76%|███████▌  | 2501/3300 [3:35:22<6:16:39, 28.28s/it]2026-04-12 02:12:14,014 | INFO | train_grpo_train | metrics_logged mode=train step=2501

                                                       
{'loss': 0.018, 'grad_norm': 4.538857460021973, 'learning_rate': 2.4242424242424244e-06, 'num_tokens': 5666190.0, 'completions/mean_length': 66.875, 'completions/min_length': 64.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9774847626686096, 'rewards/meter/std': 0.03591068461537361, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9774847626686096, 'rewards/total_composite/std': 0.03591068461537361, 'reward': 0.9774847626686096, 'reward_std': 0.03591068089008331, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04367325082421303, 'sampling/sampling_logp_difference/max': 0.8173356056213379, 'sampling/importance_sampling_ratio/min': 0.44160670042037964, 'sampling/importance_sampling_ratio/mean': 1.0134010314941406, 'sampling/importance_sampling_ratio/max': 1.6371313333511353, 'entropy': 0.3523239102214575, 'clip_ratio/low_mean': 0.003677265834994614, 'clip_ratio/low_min': 0.003677265834994614, 'clip_ratio/high_mean': 0.02857714961282909, 'clip_ratio/high_max': 0.02857714961282909, 'clip_ratio/region_mean': 0.0322544154478237, 'reward_total_mean': 0.9774847626686096, 'reward_meter_mean': 0.9774847626686096, 'reward_meter_std': 0.03591068461537361, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9774847626686096, 'reward_total_composite_std': 0.03591068461537361, 'epoch': 0.1}

 76%|███████▌  | 2501/3300 [3:35:23<6:16:39, 28.28s/it]INFO 04-12 02:12:14 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2502/3300 [3:35:27<4:42:31, 21.24s/it]2026-04-12 02:12:18,820 | INFO | train_grpo_train | metrics_logged mode=train step=2502

                                                       
{'loss': 0.0026, 'grad_norm': 4.3011322021484375, 'learning_rate': 2.4212121212121216e-06, 'num_tokens': 5668128.0, 'completions/mean_length': 67.25, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9222630262374878, 'rewards/meter/std': 0.14857283234596252, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9222630262374878, 'rewards/total_composite/std': 0.14857283234596252, 'reward': 0.9222630262374878, 'reward_std': 0.14857284724712372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04170745983719826, 'sampling/sampling_logp_difference/max': 0.8916559219360352, 'sampling/importance_sampling_ratio/min': 0.4099763035774231, 'sampling/importance_sampling_ratio/mean': 1.0067012310028076, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3921285644173622, 'clip_ratio/low_mean': 0.013037330703809857, 'clip_ratio/low_min': 0.013037330703809857, 'clip_ratio/high_mean': 0.029540141811594367, 'clip_ratio/high_max': 0.029540141811594367, 'clip_ratio/region_mean': 0.042577472515404224, 'reward_total_mean': 0.9222630262374878, 'reward_meter_mean': 0.9222630262374878, 'reward_meter_std': 0.14857283234596252, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9222630262374878, 'reward_total_composite_std': 0.14857283234596252, 'epoch': 0.1}

 76%|███████▌  | 2502/3300 [3:35:27<4:42:31, 21.24s/it]INFO 04-12 02:12:19 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2503/3300 [3:35:32<3:36:27, 16.30s/it]2026-04-12 02:12:23,575 | INFO | train_grpo_train | metrics_logged mode=train step=2503

                                                       
{'loss': 0.0043, 'grad_norm': 4.316682815551758, 'learning_rate': 2.4181818181818185e-06, 'num_tokens': 5669907.0, 'completions/mean_length': 71.375, 'completions/min_length': 70.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9978486895561218, 'rewards/meter/std': 0.00048536990652792156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978486895561218, 'rewards/total_composite/std': 0.00048536990652792156, 'reward': 0.9978486895561218, 'reward_std': 0.00048536164104007185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021549848839640617, 'sampling/sampling_logp_difference/max': 0.9752523899078369, 'sampling/importance_sampling_ratio/min': 0.37709715962409973, 'sampling/importance_sampling_ratio/mean': 1.0028010606765747, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14498437847942114, 'clip_ratio/low_mean': 0.008732197224162519, 'clip_ratio/low_min': 0.008732197224162519, 'clip_ratio/high_mean': 0.0035211266949772835, 'clip_ratio/high_max': 0.0035211266949772835, 'clip_ratio/region_mean': 0.012253323919139802, 'reward_total_mean': 0.9978486895561218, 'reward_meter_mean': 0.9978486895561218, 'reward_meter_std': 0.00048536990652792156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978486895561218, 'reward_total_composite_std': 0.00048536990652792156, 'epoch': 0.1}

 76%|███████▌  | 2503/3300 [3:35:32<3:36:27, 16.30s/it]INFO 04-12 02:12:23 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2504/3300 [3:35:36<2:49:02, 12.74s/it]2026-04-12 02:12:28,026 | INFO | train_grpo_train | metrics_logged mode=train step=2504

                                                       
{'loss': -0.0032, 'grad_norm': 6.806858539581299, 'learning_rate': 2.4151515151515153e-06, 'num_tokens': 5671433.0, 'completions/mean_length': 36.75, 'completions/min_length': 36.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9697108268737793, 'rewards/meter/std': 0.01697305217385292, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9697108268737793, 'rewards/total_composite/std': 0.01697305217385292, 'reward': 0.9697108268737793, 'reward_std': 0.01697305217385292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03803471848368645, 'sampling/sampling_logp_difference/max': 0.5422968864440918, 'sampling/importance_sampling_ratio/min': 0.5814112424850464, 'sampling/importance_sampling_ratio/mean': 1.0082935094833374, 'sampling/importance_sampling_ratio/max': 1.5999165773391724, 'entropy': 0.32645551674067974, 'clip_ratio/low_mean': 0.02393018058501184, 'clip_ratio/low_min': 0.02393018058501184, 'clip_ratio/high_mean': 0.02037855191156268, 'clip_ratio/high_max': 0.02037855191156268, 'clip_ratio/region_mean': 0.04430873249657452, 'reward_total_mean': 0.9697108268737793, 'reward_meter_mean': 0.9697108268737793, 'reward_meter_std': 0.01697305217385292, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9697108268737793, 'reward_total_composite_std': 0.01697305217385292, 'epoch': 0.1}

 76%|███████▌  | 2504/3300 [3:35:37<2:49:02, 12.74s/it]INFO 04-12 02:12:28 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2505/3300 [3:35:41<2:16:54, 10.33s/it]2026-04-12 02:12:32,739 | INFO | train_grpo_train | metrics_logged mode=train step=2505

                                                       
{'loss': -0.0003, 'grad_norm': 0.20346412062644958, 'learning_rate': 2.412121212121212e-06, 'num_tokens': 5673402.0, 'completions/mean_length': 67.125, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9981461763381958, 'rewards/meter/std': 1.717484337859787e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981461763381958, 'rewards/total_composite/std': 1.717484337859787e-05, 'reward': 0.9981461763381958, 'reward_std': 1.7183874660986476e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00340123544447124, 'sampling/sampling_logp_difference/max': 0.710139274597168, 'sampling/importance_sampling_ratio/min': 0.4915757179260254, 'sampling/importance_sampling_ratio/mean': 0.9994654655456543, 'sampling/importance_sampling_ratio/max': 1.091694712638855, 'entropy': 0.013536597951315343, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.0018382353009656072, 'reward_total_mean': 0.9981461763381958, 'reward_meter_mean': 0.9981461763381958, 'reward_meter_std': 1.717484337859787e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981461763381958, 'reward_total_composite_std': 1.717484337859787e-05, 'epoch': 0.1}

 76%|███████▌  | 2505/3300 [3:35:41<2:16:54, 10.33s/it]INFO 04-12 02:12:32 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2506/3300 [3:35:46<1:53:47,  8.60s/it]2026-04-12 02:12:37,291 | INFO | train_grpo_train | metrics_logged mode=train step=2506

                                                       
{'loss': -0.0096, 'grad_norm': 6.047022342681885, 'learning_rate': 2.4090909090909094e-06, 'num_tokens': 5675142.0, 'completions/mean_length': 67.5, 'completions/min_length': 65.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9976764917373657, 'rewards/meter/std': 0.004249152261763811, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976764917373657, 'rewards/total_composite/std': 0.004249152261763811, 'reward': 0.9976764917373657, 'reward_std': 0.004249148536473513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04069520905613899, 'sampling/sampling_logp_difference/max': 2.3988044261932373, 'sampling/importance_sampling_ratio/min': 0.09082647413015366, 'sampling/importance_sampling_ratio/mean': 1.0117424726486206, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3341045305132866, 'clip_ratio/low_mean': 0.003846153849735856, 'clip_ratio/low_min': 0.003846153849735856, 'clip_ratio/high_mean': 0.02579016692470759, 'clip_ratio/high_max': 0.02579016692470759, 'clip_ratio/region_mean': 0.029636320774443448, 'reward_total_mean': 0.9976764917373657, 'reward_meter_mean': 0.9976764917373657, 'reward_meter_std': 0.004249152261763811, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976764917373657, 'reward_total_composite_std': 0.004249152261763811, 'epoch': 0.1}

 76%|███████▌  | 2506/3300 [3:35:46<1:53:47,  8.60s/it]INFO 04-12 02:12:37 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2507/3300 [3:35:50<1:37:35,  7.38s/it]2026-04-12 02:12:41,841 | INFO | train_grpo_train | metrics_logged mode=train step=2507

                                                       
{'loss': 0.023, 'grad_norm': 9.360161781311035, 'learning_rate': 2.4060606060606062e-06, 'num_tokens': 5676923.0, 'completions/mean_length': 62.625, 'completions/min_length': 60.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9846850633621216, 'rewards/meter/std': 0.005829821340739727, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9846850633621216, 'rewards/total_composite/std': 0.005829821340739727, 'reward': 0.9846850633621216, 'reward_std': 0.0058298311196267605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048959724605083466, 'sampling/sampling_logp_difference/max': 4.749904632568359, 'sampling/importance_sampling_ratio/min': 0.00865252036601305, 'sampling/importance_sampling_ratio/mean': 0.9932248592376709, 'sampling/importance_sampling_ratio/max': 1.7164337635040283, 'entropy': 0.11215901840478182, 'clip_ratio/low_mean': 0.015786210540682077, 'clip_ratio/low_min': 0.015786210540682077, 'clip_ratio/high_mean': 0.008196721319109201, 'clip_ratio/high_max': 0.008196721319109201, 'clip_ratio/region_mean': 0.02398293185979128, 'reward_total_mean': 0.9846850633621216, 'reward_meter_mean': 0.9846850633621216, 'reward_meter_std': 0.005829821340739727, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9846850633621216, 'reward_total_composite_std': 0.005829821340739727, 'epoch': 0.1}

 76%|███████▌  | 2507/3300 [3:35:50<1:37:35,  7.38s/it]INFO 04-12 02:12:42 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2508/3300 [3:35:55<1:26:13,  6.53s/it]2026-04-12 02:12:46,386 | INFO | train_grpo_train | metrics_logged mode=train step=2508

                                                       
{'loss': 0.0105, 'grad_norm': 2.719905376434326, 'learning_rate': 2.403030303030303e-06, 'num_tokens': 5678652.0, 'completions/mean_length': 61.125, 'completions/min_length': 59.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9946359992027283, 'rewards/meter/std': 0.0007906173705123365, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9946359992027283, 'rewards/total_composite/std': 0.0007906173705123365, 'reward': 0.9946359992027283, 'reward_std': 0.0007906315731815994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019306229427456856, 'sampling/sampling_logp_difference/max': 1.058699131011963, 'sampling/importance_sampling_ratio/min': 0.3469068109989166, 'sampling/importance_sampling_ratio/mean': 1.002495288848877, 'sampling/importance_sampling_ratio/max': 1.7869699001312256, 'entropy': 0.10735112428665161, 'clip_ratio/low_mean': 0.010080644860863686, 'clip_ratio/low_min': 0.010080644860863686, 'clip_ratio/high_mean': 0.010315365390852094, 'clip_ratio/high_max': 0.010315365390852094, 'clip_ratio/region_mean': 0.02039601025171578, 'reward_total_mean': 0.9946359992027283, 'reward_meter_mean': 0.9946359992027283, 'reward_meter_std': 0.0007906173705123365, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9946359992027283, 'reward_total_composite_std': 0.0007906173705123365, 'epoch': 0.1}

 76%|███████▌  | 2508/3300 [3:35:55<1:26:13,  6.53s/it]INFO 04-12 02:12:46 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2509/3300 [3:35:59<1:18:15,  5.94s/it]2026-04-12 02:12:50,931 | INFO | train_grpo_train | metrics_logged mode=train step=2509

                                                       
{'loss': -0.0075, 'grad_norm': 5.611873626708984, 'learning_rate': 2.4000000000000003e-06, 'num_tokens': 5680475.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9922016859054565, 'rewards/meter/std': 0.016830753535032272, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9922016859054565, 'rewards/total_composite/std': 0.016830753535032272, 'reward': 0.9922016859054565, 'reward_std': 0.016830753535032272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006449920125305653, 'sampling/sampling_logp_difference/max': 1.5570316314697266, 'sampling/importance_sampling_ratio/min': 0.21076074242591858, 'sampling/importance_sampling_ratio/mean': 0.9988427758216858, 'sampling/importance_sampling_ratio/max': 1.0730388164520264, 'entropy': 0.01955289370380342, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.9922016859054565, 'reward_meter_mean': 0.9922016859054565, 'reward_meter_std': 0.016830753535032272, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9922016859054565, 'reward_total_composite_std': 0.016830753535032272, 'epoch': 0.1}

 76%|███████▌  | 2509/3300 [3:35:59<1:18:15,  5.94s/it]INFO 04-12 02:12:51 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2510/3300 [3:36:04<1:12:31,  5.51s/it]2026-04-12 02:12:55,436 | INFO | train_grpo_train | metrics_logged mode=train step=2510

                                                       
{'loss': -0.0001, 'grad_norm': 0.017269477248191833, 'learning_rate': 2.396969696969697e-06, 'num_tokens': 5682435.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981517195701599, 'rewards/meter/std': 1.4962344039304298e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981517195701599, 'rewards/total_composite/std': 1.4962344039304298e-06, 'reward': 0.9981517195701599, 'reward_std': 1.5052806929816143e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0025519414339214563, 'sampling/sampling_logp_difference/max': 0.5824198722839355, 'sampling/importance_sampling_ratio/min': 0.5585451126098633, 'sampling/importance_sampling_ratio/mean': 1.0006331205368042, 'sampling/importance_sampling_ratio/max': 1.0855460166931152, 'entropy': 0.014974120538681746, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981517195701599, 'reward_meter_mean': 0.9981517195701599, 'reward_meter_std': 1.4962344039304298e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981517195701599, 'reward_total_composite_std': 1.4962344039304298e-06, 'epoch': 0.1}

 76%|███████▌  | 2510/3300 [3:36:04<1:12:31,  5.51s/it]INFO 04-12 02:12:55 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2511/3300 [3:36:09<1:10:16,  5.34s/it]2026-04-12 02:13:00,401 | INFO | train_grpo_train | metrics_logged mode=train step=2511

                                                       
{'loss': 0.0007, 'grad_norm': 0.2027793526649475, 'learning_rate': 2.393939393939394e-06, 'num_tokens': 5684525.0, 'completions/mean_length': 98.25, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.25, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9980236291885376, 'rewards/meter/std': 2.7453183065517806e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980236291885376, 'rewards/total_composite/std': 2.7453183065517806e-05, 'reward': 0.9980236291885376, 'reward_std': 2.745408346527256e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009353289380669594, 'sampling/sampling_logp_difference/max': 1.3464837074279785, 'sampling/importance_sampling_ratio/min': 0.2601534426212311, 'sampling/importance_sampling_ratio/mean': 1.0008984804153442, 'sampling/importance_sampling_ratio/max': 1.3208497762680054, 'entropy': 0.0416307682171464, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0025510203558951616, 'clip_ratio/high_max': 0.0025510203558951616, 'clip_ratio/region_mean': 0.0025510203558951616, 'reward_total_mean': 0.9980236291885376, 'reward_meter_mean': 0.9980236291885376, 'reward_meter_std': 2.7453183065517806e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980236291885376, 'reward_total_composite_std': 2.7453183065517806e-05, 'epoch': 0.1}

 76%|███████▌  | 2511/3300 [3:36:09<1:10:16,  5.34s/it]INFO 04-12 02:13:00 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2512/3300 [3:36:13<1:06:52,  5.09s/it]2026-04-12 02:13:04,906 | INFO | train_grpo_train | metrics_logged mode=train step=2512

                                                       
{'loss': -0.0135, 'grad_norm': 4.541382312774658, 'learning_rate': 2.3909090909090912e-06, 'num_tokens': 5686253.0, 'completions/mean_length': 61.0, 'completions/min_length': 59.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9948880672454834, 'rewards/meter/std': 0.0007434968720190227, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8703480958938599, 'rewards/total_composite/std': 0.35167405009269714, 'reward': 0.8703480958938599, 'reward_std': 0.35167405009269714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0175700094550848, 'sampling/sampling_logp_difference/max': 0.7211494445800781, 'sampling/importance_sampling_ratio/min': 0.48619309067726135, 'sampling/importance_sampling_ratio/mean': 0.9973222613334656, 'sampling/importance_sampling_ratio/max': 1.566518783569336, 'entropy': 0.11922997329384089, 'clip_ratio/low_mean': 0.0042372881434857845, 'clip_ratio/low_min': 0.0042372881434857845, 'clip_ratio/high_mean': 0.020458750892430544, 'clip_ratio/high_max': 0.020458750892430544, 'clip_ratio/region_mean': 0.02469603903591633, 'reward_total_mean': 0.8703480958938599, 'reward_meter_mean': 0.9948880672454834, 'reward_meter_std': 0.0007434968720190227, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8703480958938599, 'reward_total_composite_std': 0.35167405009269714, 'epoch': 0.1}

 76%|███████▌  | 2512/3300 [3:36:13<1:06:52,  5.09s/it]INFO 04-12 02:13:05 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2513/3300 [3:36:19<1:09:45,  5.32s/it]2026-04-12 02:13:10,752 | INFO | train_grpo_train | metrics_logged mode=train step=2513

                                                       
{'loss': 0.0078, 'grad_norm': 3.2557570934295654, 'learning_rate': 2.387878787878788e-06, 'num_tokens': 5689270.0, 'completions/mean_length': 166.125, 'completions/min_length': 159.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.125, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.9812671542167664, 'rewards/meter/std': 0.015679193660616875, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.930555522441864, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.7611844539642334, 'rewards/total_composite/std': 0.07191365212202072, 'reward': 0.7611844539642334, 'reward_std': 0.07191364467144012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04680168256163597, 'sampling/sampling_logp_difference/max': 1.3245468139648438, 'sampling/importance_sampling_ratio/min': 0.26592347025871277, 'sampling/importance_sampling_ratio/mean': 1.0082600116729736, 'sampling/importance_sampling_ratio/max': 1.982582449913025, 'entropy': 0.37756787799298763, 'clip_ratio/low_mean': 0.015607842477038503, 'clip_ratio/low_min': 0.015607842477038503, 'clip_ratio/high_mean': 0.014292625011876225, 'clip_ratio/high_max': 0.014292625011876225, 'clip_ratio/region_mean': 0.029900467488914728, 'reward_total_mean': 0.7611844539642334, 'reward_meter_mean': 0.9812671542167664, 'reward_meter_std': 0.015679193660616875, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.930555522441864, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.7611844539642334, 'reward_total_composite_std': 0.07191365212202072, 'epoch': 0.1}

 76%|███████▌  | 2513/3300 [3:36:19<1:09:45,  5.32s/it]INFO 04-12 02:13:11 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2514/3300 [3:36:24<1:06:25,  5.07s/it]2026-04-12 02:13:15,247 | INFO | train_grpo_train | metrics_logged mode=train step=2514

                                                       
{'loss': 0.0002, 'grad_norm': 0.11680782586336136, 'learning_rate': 2.3848484848484853e-06, 'num_tokens': 5691062.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.999394416809082, 'rewards/meter/std': 3.964109055232257e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999394416809082, 'rewards/total_composite/std': 3.964109055232257e-06, 'reward': 0.999394416809082, 'reward_std': 3.96097129851114e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004099253565073013, 'sampling/sampling_logp_difference/max': 0.47015440464019775, 'sampling/importance_sampling_ratio/min': 0.6249058246612549, 'sampling/importance_sampling_ratio/mean': 1.0025361776351929, 'sampling/importance_sampling_ratio/max': 1.4564316272735596, 'entropy': 0.017428745748475194, 'clip_ratio/low_mean': 0.001953125, 'clip_ratio/low_min': 0.001953125, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.999394416809082, 'reward_meter_mean': 0.999394416809082, 'reward_meter_std': 3.964109055232257e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999394416809082, 'reward_total_composite_std': 3.964109055232257e-06, 'epoch': 0.1}

 76%|███████▌  | 2514/3300 [3:36:24<1:06:25,  5.07s/it]INFO 04-12 02:13:15 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2515/3300 [3:36:29<1:07:14,  5.14s/it]2026-04-12 02:13:20,545 | INFO | train_grpo_train | metrics_logged mode=train step=2515

                                                       
{'loss': 0.0025, 'grad_norm': 4.214909076690674, 'learning_rate': 2.381818181818182e-06, 'num_tokens': 5693436.0, 'completions/mean_length': 124.75, 'completions/min_length': 119.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.75, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9937890768051147, 'rewards/meter/std': 0.000979111879132688, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.9227913618087769, 'rewards/total_composite/std': 0.07571373134851456, 'reward': 0.9227913618087769, 'reward_std': 0.07571373879909515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021321963518857956, 'sampling/sampling_logp_difference/max': 1.0132427215576172, 'sampling/importance_sampling_ratio/min': 0.36303985118865967, 'sampling/importance_sampling_ratio/mean': 1.0026657581329346, 'sampling/importance_sampling_ratio/max': 1.6431286334991455, 'entropy': 0.14727903716266155, 'clip_ratio/low_mean': 0.0070080646546557546, 'clip_ratio/low_min': 0.0070080646546557546, 'clip_ratio/high_mean': 0.009006721433252096, 'clip_ratio/high_max': 0.009006721433252096, 'clip_ratio/region_mean': 0.01601478608790785, 'reward_total_mean': 0.9227913618087769, 'reward_meter_mean': 0.9937890768051147, 'reward_meter_std': 0.000979111879132688, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.9227913618087769, 'reward_total_composite_std': 0.07571373134851456, 'epoch': 0.1}

 76%|███████▌  | 2515/3300 [3:36:29<1:07:14,  5.14s/it]INFO 04-12 02:13:20 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▌  | 2516/3300 [3:36:34<1:06:37,  5.10s/it]2026-04-12 02:13:25,548 | INFO | train_grpo_train | metrics_logged mode=train step=2516

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.378787878787879e-06, 'num_tokens': 5695612.0, 'completions/mean_length': 106.0, 'completions/min_length': 106.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.0, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.6869902610778809, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8571428656578064, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5888487696647644, 'rewards/total_composite/std': 0.0, 'reward': 0.5888487696647644, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005320468917489052, 'sampling/sampling_logp_difference/max': 0.00885598175227642, 'sampling/importance_sampling_ratio/min': 0.9996694326400757, 'sampling/importance_sampling_ratio/mean': 1.0005320310592651, 'sampling/importance_sampling_ratio/max': 1.0088952779769897, 'entropy': 0.0042294226586818695, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.5888487696647644, 'reward_meter_mean': 0.6869902610778809, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8571428656578064, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5888487696647644, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 76%|███████▌  | 2516/3300 [3:36:34<1:06:37,  5.10s/it]INFO 04-12 02:13:25 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2517/3300 [3:36:39<1:05:10,  4.99s/it]2026-04-12 02:13:30,298 | INFO | train_grpo_train | metrics_logged mode=train step=2517

                                                       
{'loss': 0.0787, 'grad_norm': 6.815549850463867, 'learning_rate': 2.375757575757576e-06, 'num_tokens': 5697454.0, 'completions/mean_length': 67.25, 'completions/min_length': 59.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.7423630952835083, 'rewards/meter/std': 0.3543585538864136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.6705501079559326, 'rewards/total_composite/std': 0.3163512647151947, 'reward': 0.6705501079559326, 'reward_std': 0.3163512945175171, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07449356466531754, 'sampling/sampling_logp_difference/max': 1.8357524871826172, 'sampling/importance_sampling_ratio/min': 0.15949344635009766, 'sampling/importance_sampling_ratio/mean': 1.0203852653503418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45786258578300476, 'clip_ratio/low_mean': 0.02324857749044895, 'clip_ratio/low_min': 0.02324857749044895, 'clip_ratio/high_mean': 0.0541982427239418, 'clip_ratio/high_max': 0.0541982427239418, 'clip_ratio/region_mean': 0.07744682021439075, 'reward_total_mean': 0.6705501079559326, 'reward_meter_mean': 0.7423630952835083, 'reward_meter_std': 0.3543585538864136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.6705501079559326, 'reward_total_composite_std': 0.3163512647151947, 'epoch': 0.1}

 76%|███████▋  | 2517/3300 [3:36:39<1:05:10,  4.99s/it]INFO 04-12 02:13:30 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2518/3300 [3:36:43<1:01:49,  4.74s/it]2026-04-12 02:13:34,461 | INFO | train_grpo_train | metrics_logged mode=train step=2518

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.372727272727273e-06, 'num_tokens': 5698926.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005153539241291583, 'sampling/sampling_logp_difference/max': 0.0398876816034317, 'sampling/importance_sampling_ratio/min': 0.9608973860740662, 'sampling/importance_sampling_ratio/mean': 0.9997519254684448, 'sampling/importance_sampling_ratio/max': 1.0032331943511963, 'entropy': 0.007087996578775346, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 76%|███████▋  | 2518/3300 [3:36:43<1:01:49,  4.74s/it]INFO 04-12 02:13:34 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2519/3300 [3:36:48<1:04:09,  4.93s/it]2026-04-12 02:13:39,820 | INFO | train_grpo_train | metrics_logged mode=train step=2519

                                                       
{'loss': -0.0053, 'grad_norm': 1.6101531982421875, 'learning_rate': 2.36969696969697e-06, 'num_tokens': 5701507.0, 'completions/mean_length': 126.625, 'completions/min_length': 125.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.625, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9869413375854492, 'rewards/meter/std': 0.010035636834800243, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8810223340988159, 'rewards/total_composite/std': 0.06303300708532333, 'reward': 0.8810223340988159, 'reward_std': 0.06303300708532333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012142295949161053, 'sampling/sampling_logp_difference/max': 0.7789406776428223, 'sampling/importance_sampling_ratio/min': 0.458891898393631, 'sampling/importance_sampling_ratio/mean': 1.0051202774047852, 'sampling/importance_sampling_ratio/max': 1.6338825225830078, 'entropy': 0.10657109878957272, 'clip_ratio/low_mean': 0.003992063691839576, 'clip_ratio/low_min': 0.003992063691839576, 'clip_ratio/high_mean': 0.002922117244452238, 'clip_ratio/high_max': 0.002922117244452238, 'clip_ratio/region_mean': 0.006914180936291814, 'reward_total_mean': 0.8810223340988159, 'reward_meter_mean': 0.9869413375854492, 'reward_meter_std': 0.010035636834800243, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8810223340988159, 'reward_total_composite_std': 0.06303300708532333, 'epoch': 0.1}

 76%|███████▋  | 2519/3300 [3:36:48<1:04:09,  4.93s/it]INFO 04-12 02:13:40 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2520/3300 [3:36:53<1:01:34,  4.74s/it]2026-04-12 02:13:44,110 | INFO | train_grpo_train | metrics_logged mode=train step=2520

                                                       
{'loss': 0.0271, 'grad_norm': 8.599313735961914, 'learning_rate': 2.3666666666666667e-06, 'num_tokens': 5702952.0, 'completions/mean_length': 36.625, 'completions/min_length': 35.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9569580554962158, 'rewards/meter/std': 0.05064847692847252, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9569580554962158, 'rewards/total_composite/std': 0.05064847692847252, 'reward': 0.9569580554962158, 'reward_std': 0.05064847320318222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05431085824966431, 'sampling/sampling_logp_difference/max': 1.5028982162475586, 'sampling/importance_sampling_ratio/min': 0.22248442471027374, 'sampling/importance_sampling_ratio/mean': 1.0103925466537476, 'sampling/importance_sampling_ratio/max': 1.500638484954834, 'entropy': 0.44045621529221535, 'clip_ratio/low_mean': 0.013246799120679498, 'clip_ratio/low_min': 0.013246799120679498, 'clip_ratio/high_mean': 0.02055674116127193, 'clip_ratio/high_max': 0.02055674116127193, 'clip_ratio/region_mean': 0.03380354028195143, 'reward_total_mean': 0.9569580554962158, 'reward_meter_mean': 0.9569580554962158, 'reward_meter_std': 0.05064847692847252, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9569580554962158, 'reward_total_composite_std': 0.05064847692847252, 'epoch': 0.1}

 76%|███████▋  | 2520/3300 [3:36:53<1:01:34,  4.74s/it]INFO 04-12 02:13:44 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2521/3300 [3:36:57<1:00:50,  4.69s/it]2026-04-12 02:13:48,675 | INFO | train_grpo_train | metrics_logged mode=train step=2521

                                                       
{'loss': 0.004, 'grad_norm': 4.829682350158691, 'learning_rate': 2.363636363636364e-06, 'num_tokens': 5704880.0, 'completions/mean_length': 61.0, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.994844377040863, 'rewards/meter/std': 0.00046124093933030963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.994844377040863, 'rewards/total_composite/std': 0.00046124093933030963, 'reward': 0.994844377040863, 'reward_std': 0.0004612513293977827, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02323114313185215, 'sampling/sampling_logp_difference/max': 2.4490623474121094, 'sampling/importance_sampling_ratio/min': 0.08637453615665436, 'sampling/importance_sampling_ratio/mean': 0.9970283508300781, 'sampling/importance_sampling_ratio/max': 1.8434381484985352, 'entropy': 0.09736423194408417, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.006147540640085936, 'clip_ratio/high_max': 0.006147540640085936, 'clip_ratio/region_mean': 0.008196720853447914, 'reward_total_mean': 0.994844377040863, 'reward_meter_mean': 0.994844377040863, 'reward_meter_std': 0.00046124093933030963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.994844377040863, 'reward_total_composite_std': 0.00046124093933030963, 'epoch': 0.1}

 76%|███████▋  | 2521/3300 [3:36:57<1:00:50,  4.69s/it]INFO 04-12 02:13:48 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2522/3300 [3:37:02<59:51,  4.62s/it]  2026-04-12 02:13:53,129 | INFO | train_grpo_train | metrics_logged mode=train step=2522

                                                     
{'loss': 0.001, 'grad_norm': 0.5622905492782593, 'learning_rate': 2.360606060606061e-06, 'num_tokens': 5706673.0, 'completions/mean_length': 54.125, 'completions/min_length': 54.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7876288890838623, 'rewards/meter/std': 2.806982047331985e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7876288890838623, 'rewards/total_composite/std': 2.806982047331985e-05, 'reward': 0.7876288890838623, 'reward_std': 2.8069800464436412e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0021415329538285732, 'sampling/sampling_logp_difference/max': 0.7427225112915039, 'sampling/importance_sampling_ratio/min': 0.4758167266845703, 'sampling/importance_sampling_ratio/mean': 0.9992128610610962, 'sampling/importance_sampling_ratio/max': 1.0073615312576294, 'entropy': 0.0036355706688482314, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7876288890838623, 'reward_meter_mean': 0.7876288890838623, 'reward_meter_std': 2.806982047331985e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7876288890838623, 'reward_total_composite_std': 2.806982047331985e-05, 'epoch': 0.1}

 76%|███████▋  | 2522/3300 [3:37:02<59:51,  4.62s/it]INFO 04-12 02:13:53 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2523/3300 [3:37:12<1:21:02,  6.26s/it]2026-04-12 02:14:03,220 | INFO | train_grpo_train | metrics_logged mode=train step=2523

                                                       
{'loss': 0.099, 'grad_norm': 1.5801368951797485, 'learning_rate': 2.3575757575757577e-06, 'num_tokens': 5711794.0, 'completions/mean_length': 497.125, 'completions/min_length': 476.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 495.0000305175781, 'completions/min_terminated_length': 476.0, 'completions/max_terminated_length': 511.0, 'rewards/meter/mean': 0.9986308813095093, 'rewards/meter/std': 0.00048286825767718256, 'rewards/count_adherence/mean': 0.8046875, 'rewards/count_adherence/std': 0.022097086533904076, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9345512986183167, 'rewards/repeat_penalty/std': 0.06527597457170486, 'rewards/total_composite/mean': 0.751766562461853, 'rewards/total_composite/std': 0.06588190048933029, 'reward': 0.751766562461853, 'reward_std': 0.06588190793991089, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0517297238111496, 'sampling/sampling_logp_difference/max': 1.794034481048584, 'sampling/importance_sampling_ratio/min': 0.16628792881965637, 'sampling/importance_sampling_ratio/mean': 1.0117627382278442, 'sampling/importance_sampling_ratio/max': 1.9566318988800049, 'entropy': 0.43601541593670845, 'clip_ratio/low_mean': 0.009364859783090651, 'clip_ratio/low_min': 0.009364859783090651, 'clip_ratio/high_mean': 0.020694297272711992, 'clip_ratio/high_max': 0.020694297272711992, 'clip_ratio/region_mean': 0.030059157055802643, 'reward_total_mean': 0.751766562461853, 'reward_meter_mean': 0.9986308813095093, 'reward_meter_std': 0.00048286825767718256, 'reward_count_adherence_mean': 0.8046875, 'reward_count_adherence_std': 0.022097086533904076, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9345512986183167, 'reward_repeat_penalty_std': 0.06527597457170486, 'reward_total_composite_mean': 0.751766562461853, 'reward_total_composite_std': 0.06588190048933029, 'epoch': 0.1}

 76%|███████▋  | 2523/3300 [3:37:12<1:21:02,  6.26s/it]INFO 04-12 02:14:03 [block_pool.py:378] Successfully reset prefix cache

 76%|███████▋  | 2524/3300 [3:37:17<1:16:19,  5.90s/it]2026-04-12 02:14:08,290 | INFO | train_grpo_train | metrics_logged mode=train step=2524

                                                       
{'loss': 0.0218, 'grad_norm': 5.143811225891113, 'learning_rate': 2.3545454545454545e-06, 'num_tokens': 5713893.0, 'completions/mean_length': 101.375, 'completions/min_length': 96.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.375, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.9959712028503418, 'rewards/meter/std': 0.005785571411252022, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9959712028503418, 'rewards/total_composite/std': 0.005785571411252022, 'reward': 0.9959712028503418, 'reward_std': 0.005785576067864895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05520302802324295, 'sampling/sampling_logp_difference/max': 1.536201000213623, 'sampling/importance_sampling_ratio/min': 0.23741686344146729, 'sampling/importance_sampling_ratio/mean': 1.0093885660171509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4374755807220936, 'clip_ratio/low_mean': 0.008616218925453722, 'clip_ratio/low_min': 0.008616218925453722, 'clip_ratio/high_mean': 0.034931490663439035, 'clip_ratio/high_max': 0.034931490663439035, 'clip_ratio/region_mean': 0.04354770958889276, 'reward_total_mean': 0.9959712028503418, 'reward_meter_mean': 0.9959712028503418, 'reward_meter_std': 0.005785571411252022, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9959712028503418, 'reward_total_composite_std': 0.005785571411252022, 'epoch': 0.1}

 76%|███████▋  | 2524/3300 [3:37:17<1:16:19,  5.90s/it]INFO 04-12 02:14:08 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2525/3300 [3:37:23<1:19:02,  6.12s/it]2026-04-12 02:14:14,925 | INFO | train_grpo_train | metrics_logged mode=train step=2525

                                                       
{'loss': 0.0033, 'grad_norm': 2.698570966720581, 'learning_rate': 2.3515151515151517e-06, 'num_tokens': 5717394.0, 'completions/mean_length': 227.625, 'completions/min_length': 220.0, 'completions/max_length': 235.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 227.625, 'completions/min_terminated_length': 220.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.9980570673942566, 'rewards/meter/std': 0.0012863370357081294, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980570673942566, 'rewards/total_composite/std': 0.0012863370357081294, 'reward': 0.9980570673942566, 'reward_std': 0.0012863423908129334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06125364825129509, 'sampling/sampling_logp_difference/max': 1.1091375350952148, 'sampling/importance_sampling_ratio/min': 0.32984334230422974, 'sampling/importance_sampling_ratio/mean': 1.0095667839050293, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5780696794390678, 'clip_ratio/low_mean': 0.010438508819788694, 'clip_ratio/low_min': 0.010438508819788694, 'clip_ratio/high_mean': 0.03135894448496401, 'clip_ratio/high_max': 0.03135894448496401, 'clip_ratio/region_mean': 0.04179745330475271, 'reward_total_mean': 0.9980570673942566, 'reward_meter_mean': 0.9980570673942566, 'reward_meter_std': 0.0012863370357081294, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980570673942566, 'reward_total_composite_std': 0.0012863370357081294, 'epoch': 0.1}

 77%|███████▋  | 2525/3300 [3:37:23<1:19:02,  6.12s/it]INFO 04-12 02:14:15 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2526/3300 [3:37:28<1:12:19,  5.61s/it]2026-04-12 02:14:19,335 | INFO | train_grpo_train | metrics_logged mode=train step=2526

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.348484848484849e-06, 'num_tokens': 5719002.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003320652758702636, 'sampling/sampling_logp_difference/max': 0.003683246672153473, 'sampling/importance_sampling_ratio/min': 0.9999605417251587, 'sampling/importance_sampling_ratio/mean': 1.0003321170806885, 'sampling/importance_sampling_ratio/max': 1.0036901235580444, 'entropy': 0.002448402636218816, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 77%|███████▋  | 2526/3300 [3:37:28<1:12:19,  5.61s/it]INFO 04-12 02:14:19 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2527/3300 [3:37:32<1:07:55,  5.27s/it]2026-04-12 02:14:23,820 | INFO | train_grpo_train | metrics_logged mode=train step=2527

                                                       
{'loss': 0.0, 'grad_norm': 0.11079243570566177, 'learning_rate': 2.345454545454546e-06, 'num_tokens': 5720778.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993899464607239, 'rewards/meter/std': 4.941231964039616e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993899464607239, 'rewards/total_composite/std': 4.941231964039616e-06, 'reward': 0.9993899464607239, 'reward_std': 4.955206804879708e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00580887496471405, 'sampling/sampling_logp_difference/max': 0.7116183638572693, 'sampling/importance_sampling_ratio/min': 0.49084919691085815, 'sampling/importance_sampling_ratio/mean': 1.0004364252090454, 'sampling/importance_sampling_ratio/max': 1.3528425693511963, 'entropy': 0.02413589833304286, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.00390625, 'clip_ratio/high_max': 0.00390625, 'clip_ratio/region_mean': 0.0078125, 'reward_total_mean': 0.9993899464607239, 'reward_meter_mean': 0.9993899464607239, 'reward_meter_std': 4.941231964039616e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993899464607239, 'reward_total_composite_std': 4.941231964039616e-06, 'epoch': 0.1}

 77%|███████▋  | 2527/3300 [3:37:32<1:07:55,  5.27s/it]INFO 04-12 02:14:24 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2528/3300 [3:37:42<1:26:03,  6.69s/it]2026-04-12 02:14:33,813 | INFO | train_grpo_train | metrics_logged mode=train step=2528

                                                       
{'loss': 0.0113, 'grad_norm': 1.9397426843643188, 'learning_rate': 2.3424242424242427e-06, 'num_tokens': 5726703.0, 'completions/mean_length': 483.625, 'completions/min_length': 477.0, 'completions/max_length': 491.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 483.625, 'completions/min_terminated_length': 477.0, 'completions/max_terminated_length': 491.0, 'rewards/meter/mean': 0.9855457544326782, 'rewards/meter/std': 0.01022270042449236, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9074074029922485, 'rewards/repeat_penalty/std': 0.059391386806964874, 'rewards/total_composite/mean': 0.7824571132659912, 'rewards/total_composite/std': 0.05129082873463631, 'reward': 0.7824571132659912, 'reward_std': 0.051290810108184814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03337203338742256, 'sampling/sampling_logp_difference/max': 1.3847050666809082, 'sampling/importance_sampling_ratio/min': 0.25039762258529663, 'sampling/importance_sampling_ratio/mean': 1.0071146488189697, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2987423837184906, 'clip_ratio/low_mean': 0.00951099069789052, 'clip_ratio/low_min': 0.00951099069789052, 'clip_ratio/high_mean': 0.020238142693415284, 'clip_ratio/high_max': 0.020238142693415284, 'clip_ratio/region_mean': 0.029749133391305804, 'reward_total_mean': 0.7824571132659912, 'reward_meter_mean': 0.9855457544326782, 'reward_meter_std': 0.01022270042449236, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9074074029922485, 'reward_repeat_penalty_std': 0.059391386806964874, 'reward_total_composite_mean': 0.7824571132659912, 'reward_total_composite_std': 0.05129082873463631, 'epoch': 0.1}

 77%|███████▋  | 2528/3300 [3:37:42<1:26:03,  6.69s/it]INFO 04-12 02:14:34 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2529/3300 [3:37:47<1:19:58,  6.22s/it]2026-04-12 02:14:38,949 | INFO | train_grpo_train | metrics_logged mode=train step=2529

                                                       
{'loss': 0.0012, 'grad_norm': 3.176457405090332, 'learning_rate': 2.3393939393939395e-06, 'num_tokens': 5728897.0, 'completions/mean_length': 107.25, 'completions/min_length': 104.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9512283802032471, 'rewards/meter/std': 0.047763749957084656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9512283802032471, 'rewards/total_composite/std': 0.047763749957084656, 'reward': 0.9512283802032471, 'reward_std': 0.04776376485824585, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05268417298793793, 'sampling/sampling_logp_difference/max': 1.3672003746032715, 'sampling/importance_sampling_ratio/min': 0.2548193633556366, 'sampling/importance_sampling_ratio/mean': 1.0175414085388184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5495112352073193, 'clip_ratio/low_mean': 0.015321710845455527, 'clip_ratio/low_min': 0.015321710845455527, 'clip_ratio/high_mean': 0.032445638440549374, 'clip_ratio/high_max': 0.032445638440549374, 'clip_ratio/region_mean': 0.0477673492860049, 'reward_total_mean': 0.9512283802032471, 'reward_meter_mean': 0.9512283802032471, 'reward_meter_std': 0.047763749957084656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9512283802032471, 'reward_total_composite_std': 0.047763749957084656, 'epoch': 0.1}

 77%|███████▋  | 2529/3300 [3:37:47<1:19:58,  6.22s/it]INFO 04-12 02:14:39 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2530/3300 [3:37:52<1:12:35,  5.66s/it]2026-04-12 02:14:43,284 | INFO | train_grpo_train | metrics_logged mode=train step=2530

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.3363636363636367e-06, 'num_tokens': 5730521.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00033885284210555255, 'sampling/sampling_logp_difference/max': 0.004536015447229147, 'sampling/importance_sampling_ratio/min': 0.9995154738426208, 'sampling/importance_sampling_ratio/mean': 1.0003368854522705, 'sampling/importance_sampling_ratio/max': 1.0045462846755981, 'entropy': 0.002742367796599865, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 77%|███████▋  | 2530/3300 [3:37:52<1:12:35,  5.66s/it]INFO 04-12 02:14:43 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2531/3300 [3:37:56<1:09:10,  5.40s/it]2026-04-12 02:14:48,077 | INFO | train_grpo_train | metrics_logged mode=train step=2531

                                                       
{'loss': -0.0018, 'grad_norm': 3.0979435443878174, 'learning_rate': 2.3333333333333336e-06, 'num_tokens': 5732411.0, 'completions/mean_length': 79.25, 'completions/min_length': 77.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9984360933303833, 'rewards/meter/std': 0.00047055029426701367, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984360933303833, 'rewards/total_composite/std': 0.00047055029426701367, 'reward': 0.9984360933303833, 'reward_std': 0.00047055797767825425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04553651064634323, 'sampling/sampling_logp_difference/max': 1.1313624382019043, 'sampling/importance_sampling_ratio/min': 0.32259345054626465, 'sampling/importance_sampling_ratio/mean': 1.0040619373321533, 'sampling/importance_sampling_ratio/max': 1.7651348114013672, 'entropy': 0.3967149890959263, 'clip_ratio/low_mean': 0.01588775822892785, 'clip_ratio/low_min': 0.01588775822892785, 'clip_ratio/high_mean': 0.01107846642844379, 'clip_ratio/high_max': 0.01107846642844379, 'clip_ratio/region_mean': 0.02696622465737164, 'reward_total_mean': 0.9984360933303833, 'reward_meter_mean': 0.9984360933303833, 'reward_meter_std': 0.00047055029426701367, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984360933303833, 'reward_total_composite_std': 0.00047055029426701367, 'epoch': 0.1}

 77%|███████▋  | 2531/3300 [3:37:57<1:09:10,  5.40s/it]INFO 04-12 02:14:48 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2532/3300 [3:38:01<1:06:27,  5.19s/it]2026-04-12 02:14:52,790 | INFO | train_grpo_train | metrics_logged mode=train step=2532

                                                       
{'loss': 0.0168, 'grad_norm': 9.73486614227295, 'learning_rate': 2.3303030303030304e-06, 'num_tokens': 5734224.0, 'completions/mean_length': 74.625, 'completions/min_length': 63.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.3960377871990204, 'rewards/meter/std': 0.33334165811538696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.3960377871990204, 'rewards/total_composite/std': 0.33334165811538696, 'reward': 0.3960377871990204, 'reward_std': 0.33334165811538696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0764419287443161, 'sampling/sampling_logp_difference/max': 1.598515510559082, 'sampling/importance_sampling_ratio/min': 0.20219644904136658, 'sampling/importance_sampling_ratio/mean': 1.011519432067871, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5188739486038685, 'clip_ratio/low_mean': 0.032748116645962, 'clip_ratio/low_min': 0.032748116645962, 'clip_ratio/high_mean': 0.026046376209706068, 'clip_ratio/high_max': 0.026046376209706068, 'clip_ratio/region_mean': 0.05879449285566807, 'reward_total_mean': 0.3960377871990204, 'reward_meter_mean': 0.3960377871990204, 'reward_meter_std': 0.33334165811538696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.3960377871990204, 'reward_total_composite_std': 0.33334165811538696, 'epoch': 0.1}

 77%|███████▋  | 2532/3300 [3:38:01<1:06:27,  5.19s/it]INFO 04-12 02:14:53 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2533/3300 [3:38:06<1:03:07,  4.94s/it]2026-04-12 02:14:57,134 | INFO | train_grpo_train | metrics_logged mode=train step=2533

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.3272727272727277e-06, 'num_tokens': 5735968.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00032944101258181036, 'sampling/sampling_logp_difference/max': 0.003968724515289068, 'sampling/importance_sampling_ratio/min': 0.999470055103302, 'sampling/importance_sampling_ratio/mean': 1.0003273487091064, 'sampling/importance_sampling_ratio/max': 1.003976583480835, 'entropy': 0.002526927593862638, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 77%|███████▋  | 2533/3300 [3:38:06<1:03:07,  4.94s/it]INFO 04-12 02:14:57 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2534/3300 [3:38:10<1:03:00,  4.94s/it]2026-04-12 02:15:02,064 | INFO | train_grpo_train | metrics_logged mode=train step=2534

                                                       
{'loss': 0.0084, 'grad_norm': 2.3901071548461914, 'learning_rate': 2.3242424242424245e-06, 'num_tokens': 5738092.0, 'completions/mean_length': 91.5, 'completions/min_length': 86.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.5, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9918440580368042, 'rewards/meter/std': 0.003950192127376795, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9669877886772156, 'rewards/total_composite/std': 0.06943685561418533, 'reward': 0.9669877886772156, 'reward_std': 0.06943684071302414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030601778998970985, 'sampling/sampling_logp_difference/max': 0.814885139465332, 'sampling/importance_sampling_ratio/min': 0.4426901638507843, 'sampling/importance_sampling_ratio/mean': 1.0052556991577148, 'sampling/importance_sampling_ratio/max': 1.6521724462509155, 'entropy': 0.21182428859174252, 'clip_ratio/low_mean': 0.0013440860202535987, 'clip_ratio/low_min': 0.0013440860202535987, 'clip_ratio/high_mean': 0.017688892083242536, 'clip_ratio/high_max': 0.017688892083242536, 'clip_ratio/region_mean': 0.019032978103496134, 'reward_total_mean': 0.9669877886772156, 'reward_meter_mean': 0.9918440580368042, 'reward_meter_std': 0.003950192127376795, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9669877886772156, 'reward_total_composite_std': 0.06943685561418533, 'epoch': 0.1}

 77%|███████▋  | 2534/3300 [3:38:11<1:03:00,  4.94s/it]INFO 04-12 02:15:02 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2535/3300 [3:38:21<1:25:27,  6.70s/it]2026-04-12 02:15:12,891 | INFO | train_grpo_train | metrics_logged mode=train step=2535

                                                       
{'loss': 0.06, 'grad_norm': 1.6462563276290894, 'learning_rate': 2.3212121212121213e-06, 'num_tokens': 5743822.0, 'completions/mean_length': 495.25, 'completions/min_length': 479.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 492.857177734375, 'completions/min_terminated_length': 479.0, 'completions/max_terminated_length': 511.0, 'rewards/meter/mean': 0.9958187341690063, 'rewards/meter/std': 0.006666218861937523, 'rewards/count_adherence/mean': 0.8046875, 'rewards/count_adherence/std': 0.022097086533904076, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9601762294769287, 'rewards/repeat_penalty/std': 0.0010186029830947518, 'rewards/total_composite/mean': 0.7694027423858643, 'rewards/total_composite/std': 0.0214486513286829, 'reward': 0.7694027423858643, 'reward_std': 0.021448642015457153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05616435408592224, 'sampling/sampling_logp_difference/max': 2.1436595916748047, 'sampling/importance_sampling_ratio/min': 0.11722506582736969, 'sampling/importance_sampling_ratio/mean': 1.014970302581787, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4710547477006912, 'clip_ratio/low_mean': 0.006211606087163091, 'clip_ratio/low_min': 0.006211606087163091, 'clip_ratio/high_mean': 0.023648475529626012, 'clip_ratio/high_max': 0.023648475529626012, 'clip_ratio/region_mean': 0.029860081616789103, 'reward_total_mean': 0.7694027423858643, 'reward_meter_mean': 0.9958187341690063, 'reward_meter_std': 0.006666218861937523, 'reward_count_adherence_mean': 0.8046875, 'reward_count_adherence_std': 0.022097086533904076, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9601762294769287, 'reward_repeat_penalty_std': 0.0010186029830947518, 'reward_total_composite_mean': 0.7694027423858643, 'reward_total_composite_std': 0.0214486513286829, 'epoch': 0.1}

 77%|███████▋  | 2535/3300 [3:38:21<1:25:27,  6.70s/it]INFO 04-12 02:15:13 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2536/3300 [3:38:29<1:30:26,  7.10s/it]2026-04-12 02:15:20,924 | INFO | train_grpo_train | metrics_logged mode=train step=2536

                                                       
{'loss': 0.0168, 'grad_norm': 3.1345582008361816, 'learning_rate': 2.318181818181818e-06, 'num_tokens': 5748428.0, 'completions/mean_length': 334.75, 'completions/min_length': 306.0, 'completions/max_length': 353.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 334.75, 'completions/min_terminated_length': 306.0, 'completions/max_terminated_length': 353.0, 'rewards/meter/mean': 0.9545778036117554, 'rewards/meter/std': 0.0988415852189064, 'rewards/count_adherence/mean': 0.8636363744735718, 'rewards/count_adherence/std': 0.04859296977519989, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9794891476631165, 'rewards/repeat_penalty/std': 0.028372056782245636, 'rewards/total_composite/mean': 0.8046990633010864, 'rewards/total_composite/std': 0.07172781974077225, 'reward': 0.8046990633010864, 'reward_std': 0.07172780483961105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08371323347091675, 'sampling/sampling_logp_difference/max': 4.665629863739014, 'sampling/importance_sampling_ratio/min': 0.00941331684589386, 'sampling/importance_sampling_ratio/mean': 1.0162872076034546, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.806172601878643, 'clip_ratio/low_mean': 0.013427280820906162, 'clip_ratio/low_min': 0.013427280820906162, 'clip_ratio/high_mean': 0.052931731566786766, 'clip_ratio/high_max': 0.052931731566786766, 'clip_ratio/region_mean': 0.06635901238769293, 'reward_total_mean': 0.8046990633010864, 'reward_meter_mean': 0.9545778036117554, 'reward_meter_std': 0.0988415852189064, 'reward_count_adherence_mean': 0.8636363744735718, 'reward_count_adherence_std': 0.04859296977519989, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9794891476631165, 'reward_repeat_penalty_std': 0.028372056782245636, 'reward_total_composite_mean': 0.8046990633010864, 'reward_total_composite_std': 0.07172781974077225, 'epoch': 0.1}

 77%|███████▋  | 2536/3300 [3:38:29<1:30:26,  7.10s/it]INFO 04-12 02:15:21 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2537/3300 [3:38:34<1:19:50,  6.28s/it]2026-04-12 02:15:25,284 | INFO | train_grpo_train | metrics_logged mode=train step=2537

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.3151515151515154e-06, 'num_tokens': 5750148.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00031052454141899943, 'sampling/sampling_logp_difference/max': 0.0035831150598824024, 'sampling/importance_sampling_ratio/min': 0.999996542930603, 'sampling/importance_sampling_ratio/mean': 1.0003106594085693, 'sampling/importance_sampling_ratio/max': 1.0035895109176636, 'entropy': 0.0026191577198915184, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 77%|███████▋  | 2537/3300 [3:38:34<1:19:50,  6.28s/it]INFO 04-12 02:15:25 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2538/3300 [3:38:38<1:13:53,  5.82s/it]2026-04-12 02:15:30,026 | INFO | train_grpo_train | metrics_logged mode=train step=2538

                                                       
{'loss': 0.0261, 'grad_norm': 8.2296781539917, 'learning_rate': 2.3121212121212123e-06, 'num_tokens': 5751904.0, 'completions/mean_length': 70.5, 'completions/min_length': 67.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9975142478942871, 'rewards/meter/std': 0.003199664643034339, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975142478942871, 'rewards/total_composite/std': 0.003199664643034339, 'reward': 0.9975142478942871, 'reward_std': 0.0031996748875826597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06007351353764534, 'sampling/sampling_logp_difference/max': 1.1276092529296875, 'sampling/importance_sampling_ratio/min': 0.3238064646720886, 'sampling/importance_sampling_ratio/mean': 1.0132718086242676, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.577742449939251, 'clip_ratio/low_mean': 0.02077368483878672, 'clip_ratio/low_min': 0.02077368483878672, 'clip_ratio/high_mean': 0.027109916205517948, 'clip_ratio/high_max': 0.027109916205517948, 'clip_ratio/region_mean': 0.04788360104430467, 'reward_total_mean': 0.9975142478942871, 'reward_meter_mean': 0.9975142478942871, 'reward_meter_std': 0.003199664643034339, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975142478942871, 'reward_total_composite_std': 0.003199664643034339, 'epoch': 0.1}

 77%|███████▋  | 2538/3300 [3:38:39<1:13:53,  5.82s/it]INFO 04-12 02:15:30 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2539/3300 [3:38:43<1:09:48,  5.50s/it]2026-04-12 02:15:34,799 | INFO | train_grpo_train | metrics_logged mode=train step=2539

                                                       
{'loss': -0.0026, 'grad_norm': 3.5843658447265625, 'learning_rate': 2.309090909090909e-06, 'num_tokens': 5753708.0, 'completions/mean_length': 72.5, 'completions/min_length': 71.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9982837438583374, 'rewards/meter/std': 0.0006137907621450722, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982837438583374, 'rewards/total_composite/std': 0.0006137907621450722, 'reward': 0.9982837438583374, 'reward_std': 0.0006137940799817443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023249905556440353, 'sampling/sampling_logp_difference/max': 1.0794258117675781, 'sampling/importance_sampling_ratio/min': 0.33979055285453796, 'sampling/importance_sampling_ratio/mean': 1.0038199424743652, 'sampling/importance_sampling_ratio/max': 1.6161738634109497, 'entropy': 0.16330011747777462, 'clip_ratio/low_mean': 0.005116959102451801, 'clip_ratio/low_min': 0.005116959102451801, 'clip_ratio/high_mean': 0.017147069913335145, 'clip_ratio/high_max': 0.017147069913335145, 'clip_ratio/region_mean': 0.022264029015786946, 'reward_total_mean': 0.9982837438583374, 'reward_meter_mean': 0.9982837438583374, 'reward_meter_std': 0.0006137907621450722, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982837438583374, 'reward_total_composite_std': 0.0006137907621450722, 'epoch': 0.1}

 77%|███████▋  | 2539/3300 [3:38:43<1:09:48,  5.50s/it]INFO 04-12 02:15:35 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2540/3300 [3:38:50<1:12:46,  5.75s/it]2026-04-12 02:15:41,106 | INFO | train_grpo_train | metrics_logged mode=train step=2540

                                                       
{'loss': -0.0035, 'grad_norm': 2.297759771347046, 'learning_rate': 2.306060606060606e-06, 'num_tokens': 5756765.0, 'completions/mean_length': 193.125, 'completions/min_length': 184.0, 'completions/max_length': 203.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 193.125, 'completions/min_terminated_length': 184.0, 'completions/max_terminated_length': 203.0, 'rewards/meter/mean': 0.9978808760643005, 'rewards/meter/std': 0.001249481923878193, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978808760643005, 'rewards/total_composite/std': 0.001249481923878193, 'reward': 0.9978808760643005, 'reward_std': 0.0012494820402935147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05693472549319267, 'sampling/sampling_logp_difference/max': 1.3195972442626953, 'sampling/importance_sampling_ratio/min': 0.2672429084777832, 'sampling/importance_sampling_ratio/mean': 1.0168569087982178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5925077982246876, 'clip_ratio/low_mean': 0.024065676843747497, 'clip_ratio/low_min': 0.024065676843747497, 'clip_ratio/high_mean': 0.017286017769947648, 'clip_ratio/high_max': 0.017286017769947648, 'clip_ratio/region_mean': 0.041351694613695145, 'reward_total_mean': 0.9978808760643005, 'reward_meter_mean': 0.9978808760643005, 'reward_meter_std': 0.001249481923878193, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978808760643005, 'reward_total_composite_std': 0.001249481923878193, 'epoch': 0.1}

 77%|███████▋  | 2540/3300 [3:38:50<1:12:46,  5.75s/it]INFO 04-12 02:15:41 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2541/3300 [3:38:55<1:10:06,  5.54s/it]2026-04-12 02:15:46,175 | INFO | train_grpo_train | metrics_logged mode=train step=2541

                                                       
{'loss': 0.0017, 'grad_norm': 5.599558353424072, 'learning_rate': 2.303030303030303e-06, 'num_tokens': 5758973.0, 'completions/mean_length': 107.0, 'completions/min_length': 105.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.0, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9983425140380859, 'rewards/meter/std': 0.0005217316211201251, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9733964800834656, 'rewards/total_composite/std': 0.07076215744018555, 'reward': 0.9733964800834656, 'reward_std': 0.07076217234134674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02940228208899498, 'sampling/sampling_logp_difference/max': 1.1934094429016113, 'sampling/importance_sampling_ratio/min': 0.3031857907772064, 'sampling/importance_sampling_ratio/mean': 1.00425386428833, 'sampling/importance_sampling_ratio/max': 1.613302230834961, 'entropy': 0.19336522929370403, 'clip_ratio/low_mean': 0.002358490601181984, 'clip_ratio/low_min': 0.002358490601181984, 'clip_ratio/high_mean': 0.025725040584802628, 'clip_ratio/high_max': 0.025725040584802628, 'clip_ratio/region_mean': 0.02808353118598461, 'reward_total_mean': 0.9733964800834656, 'reward_meter_mean': 0.9983425140380859, 'reward_meter_std': 0.0005217316211201251, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9733964800834656, 'reward_total_composite_std': 0.07076215744018555, 'epoch': 0.1}

 77%|███████▋  | 2541/3300 [3:38:55<1:10:06,  5.54s/it]INFO 04-12 02:15:46 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2542/3300 [3:38:59<1:05:58,  5.22s/it]2026-04-12 02:15:50,649 | INFO | train_grpo_train | metrics_logged mode=train step=2542

                                                       
{'loss': 0.0001, 'grad_norm': 0.0943424329161644, 'learning_rate': 2.3000000000000004e-06, 'num_tokens': 5760901.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993921518325806, 'rewards/meter/std': 5.212498763285112e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993921518325806, 'rewards/total_composite/std': 5.212498763285112e-06, 'reward': 0.9993921518325806, 'reward_std': 5.226743269304279e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004389709793031216, 'sampling/sampling_logp_difference/max': 0.464327335357666, 'sampling/importance_sampling_ratio/min': 0.777456521987915, 'sampling/importance_sampling_ratio/mean': 1.0001180171966553, 'sampling/importance_sampling_ratio/max': 1.5909435749053955, 'entropy': 0.02692276705056429, 'clip_ratio/low_mean': 0.001953125, 'clip_ratio/low_min': 0.001953125, 'clip_ratio/high_mean': 0.0078125, 'clip_ratio/high_max': 0.0078125, 'clip_ratio/region_mean': 0.009765625, 'reward_total_mean': 0.9993921518325806, 'reward_meter_mean': 0.9993921518325806, 'reward_meter_std': 5.212498763285112e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993921518325806, 'reward_total_composite_std': 5.212498763285112e-06, 'epoch': 0.1}

 77%|███████▋  | 2542/3300 [3:38:59<1:05:58,  5.22s/it]INFO 04-12 02:15:50 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2543/3300 [3:39:07<1:17:50,  6.17s/it]2026-04-12 02:15:59,029 | INFO | train_grpo_train | metrics_logged mode=train step=2543

                                                       
{'loss': 0.0244, 'grad_norm': 2.6460602283477783, 'learning_rate': 2.2969696969696973e-06, 'num_tokens': 5765713.0, 'completions/mean_length': 359.5, 'completions/min_length': 338.0, 'completions/max_length': 384.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 359.5, 'completions/min_terminated_length': 338.0, 'completions/max_terminated_length': 384.0, 'rewards/meter/mean': 0.9972070455551147, 'rewards/meter/std': 0.0018466432811692357, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.05175493285059929, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9864766001701355, 'rewards/repeat_penalty/std': 0.025052649900317192, 'rewards/total_composite/mean': 0.7973617315292358, 'rewards/total_composite/std': 0.32495424151420593, 'reward': 0.7973617315292358, 'reward_std': 0.32495421171188354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07025956362485886, 'sampling/sampling_logp_difference/max': 1.513051986694336, 'sampling/importance_sampling_ratio/min': 0.22023677825927734, 'sampling/importance_sampling_ratio/mean': 1.0157252550125122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6531303524971008, 'clip_ratio/low_mean': 0.0061848959885537624, 'clip_ratio/low_min': 0.0061848959885537624, 'clip_ratio/high_mean': 0.03669871832244098, 'clip_ratio/high_max': 0.03669871832244098, 'clip_ratio/region_mean': 0.042883614310994744, 'reward_total_mean': 0.7973617315292358, 'reward_meter_mean': 0.9972070455551147, 'reward_meter_std': 0.0018466432811692357, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.05175493285059929, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9864766001701355, 'reward_repeat_penalty_std': 0.025052649900317192, 'reward_total_composite_mean': 0.7973617315292358, 'reward_total_composite_std': 0.32495424151420593, 'epoch': 0.1}

 77%|███████▋  | 2543/3300 [3:39:08<1:17:50,  6.17s/it]INFO 04-12 02:15:59 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2544/3300 [3:39:12<1:10:40,  5.61s/it]2026-04-12 02:16:03,331 | INFO | train_grpo_train | metrics_logged mode=train step=2544

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.293939393939394e-06, 'num_tokens': 5767393.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003962568298447877, 'sampling/sampling_logp_difference/max': 0.008148963563144207, 'sampling/importance_sampling_ratio/min': 0.9980019927024841, 'sampling/importance_sampling_ratio/mean': 1.000386118888855, 'sampling/importance_sampling_ratio/max': 1.0081822872161865, 'entropy': 0.0035496088094078004, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 77%|███████▋  | 2544/3300 [3:39:12<1:10:40,  5.61s/it]INFO 04-12 02:16:03 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2545/3300 [3:39:16<1:06:03,  5.25s/it]2026-04-12 02:16:07,739 | INFO | train_grpo_train | metrics_logged mode=train step=2545

                                                       
{'loss': -0.0032, 'grad_norm': 3.4628183841705322, 'learning_rate': 2.2909090909090913e-06, 'num_tokens': 5769250.0, 'completions/mean_length': 61.125, 'completions/min_length': 60.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9939138889312744, 'rewards/meter/std': 0.0031664494890719652, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9939138889312744, 'rewards/total_composite/std': 0.0031664494890719652, 'reward': 0.9939138889312744, 'reward_std': 0.0031664466951042414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020268075168132782, 'sampling/sampling_logp_difference/max': 1.9380130767822266, 'sampling/importance_sampling_ratio/min': 0.14398977160453796, 'sampling/importance_sampling_ratio/mean': 1.004980444908142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09053080063313246, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.004065309185534716, 'clip_ratio/high_max': 0.004065309185534716, 'clip_ratio/region_mean': 0.006114489398896694, 'reward_total_mean': 0.9939138889312744, 'reward_meter_mean': 0.9939138889312744, 'reward_meter_std': 0.0031664494890719652, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9939138889312744, 'reward_total_composite_std': 0.0031664494890719652, 'epoch': 0.1}

 77%|███████▋  | 2545/3300 [3:39:16<1:06:03,  5.25s/it]INFO 04-12 02:16:07 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2546/3300 [3:39:21<1:03:05,  5.02s/it]2026-04-12 02:16:12,228 | INFO | train_grpo_train | metrics_logged mode=train step=2546

                                                       
{'loss': -0.0005, 'grad_norm': 0.5305520296096802, 'learning_rate': 2.287878787878788e-06, 'num_tokens': 5771130.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981376528739929, 'rewards/meter/std': 2.376116935920436e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981376528739929, 'rewards/total_composite/std': 2.376116935920436e-05, 'reward': 0.9981376528739929, 'reward_std': 2.3771623091306537e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004880668129771948, 'sampling/sampling_logp_difference/max': 0.47992897033691406, 'sampling/importance_sampling_ratio/min': 0.6188273429870605, 'sampling/importance_sampling_ratio/mean': 0.9998449087142944, 'sampling/importance_sampling_ratio/max': 1.1664998531341553, 'entropy': 0.025023984955623746, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.9981376528739929, 'reward_meter_mean': 0.9981376528739929, 'reward_meter_std': 2.376116935920436e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981376528739929, 'reward_total_composite_std': 2.376116935920436e-05, 'epoch': 0.1}

 77%|███████▋  | 2546/3300 [3:39:21<1:03:05,  5.02s/it]INFO 04-12 02:16:12 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2547/3300 [3:39:28<1:13:39,  5.87s/it]2026-04-12 02:16:20,075 | INFO | train_grpo_train | metrics_logged mode=train step=2547

                                                       
{'loss': 0.0142, 'grad_norm': 3.3049399852752686, 'learning_rate': 2.284848484848485e-06, 'num_tokens': 5775349.0, 'completions/mean_length': 322.375, 'completions/min_length': 287.0, 'completions/max_length': 343.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 322.375, 'completions/min_terminated_length': 287.0, 'completions/max_terminated_length': 343.0, 'rewards/meter/mean': 0.9897097945213318, 'rewards/meter/std': 0.02135160192847252, 'rewards/count_adherence/mean': 0.8020833134651184, 'rewards/count_adherence/std': 0.0431290864944458, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6906000375747681, 'rewards/total_composite/std': 0.2835569977760315, 'reward': 0.6906000375747681, 'reward_std': 0.2835569977760315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09224426746368408, 'sampling/sampling_logp_difference/max': 1.8807573318481445, 'sampling/importance_sampling_ratio/min': 0.15247458219528198, 'sampling/importance_sampling_ratio/mean': 1.0231871604919434, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9364352449774742, 'clip_ratio/low_mean': 0.008633634075522423, 'clip_ratio/low_min': 0.008633634075522423, 'clip_ratio/high_mean': 0.05134916538372636, 'clip_ratio/high_max': 0.05134916538372636, 'clip_ratio/region_mean': 0.05998279945924878, 'reward_total_mean': 0.6906000375747681, 'reward_meter_mean': 0.9897097945213318, 'reward_meter_std': 0.02135160192847252, 'reward_count_adherence_mean': 0.8020833134651184, 'reward_count_adherence_std': 0.0431290864944458, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6906000375747681, 'reward_total_composite_std': 0.2835569977760315, 'epoch': 0.1}

 77%|███████▋  | 2547/3300 [3:39:29<1:13:39,  5.87s/it]INFO 04-12 02:16:20 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2548/3300 [3:39:33<1:07:43,  5.40s/it]2026-04-12 02:16:24,394 | INFO | train_grpo_train | metrics_logged mode=train step=2548

                                                       
{'loss': 0.0725, 'grad_norm': 13.814697265625, 'learning_rate': 2.281818181818182e-06, 'num_tokens': 5776993.0, 'completions/mean_length': 42.5, 'completions/min_length': 40.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9308160543441772, 'rewards/meter/std': 0.01748719997704029, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9308160543441772, 'rewards/total_composite/std': 0.01748719997704029, 'reward': 0.9308160543441772, 'reward_std': 0.0174871776252985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0993395447731018, 'sampling/sampling_logp_difference/max': 2.4624288082122803, 'sampling/importance_sampling_ratio/min': 0.08522769808769226, 'sampling/importance_sampling_ratio/mean': 0.9965495467185974, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5177558250725269, 'clip_ratio/low_mean': 0.00866977241821587, 'clip_ratio/low_min': 0.00866977241821587, 'clip_ratio/high_mean': 0.06528556300327182, 'clip_ratio/high_max': 0.06528556300327182, 'clip_ratio/region_mean': 0.07395533542148769, 'reward_total_mean': 0.9308160543441772, 'reward_meter_mean': 0.9308160543441772, 'reward_meter_std': 0.01748719997704029, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9308160543441772, 'reward_total_composite_std': 0.01748719997704029, 'epoch': 0.1}

 77%|███████▋  | 2548/3300 [3:39:33<1:07:43,  5.40s/it]INFO 04-12 02:16:24 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2549/3300 [3:39:39<1:09:57,  5.59s/it]2026-04-12 02:16:30,418 | INFO | train_grpo_train | metrics_logged mode=train step=2549

                                                       
{'loss': 0.0034, 'grad_norm': 1.8126007318496704, 'learning_rate': 2.278787878787879e-06, 'num_tokens': 5780008.0, 'completions/mean_length': 191.875, 'completions/min_length': 187.0, 'completions/max_length': 196.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 191.875, 'completions/min_terminated_length': 187.0, 'completions/max_terminated_length': 196.0, 'rewards/meter/mean': 0.9984524250030518, 'rewards/meter/std': 0.0005822303937748075, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984524250030518, 'rewards/total_composite/std': 0.0005822303937748075, 'reward': 0.9984524250030518, 'reward_std': 0.0005822345847263932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.055033281445503235, 'sampling/sampling_logp_difference/max': 1.1893186569213867, 'sampling/importance_sampling_ratio/min': 0.3044286072254181, 'sampling/importance_sampling_ratio/mean': 1.0145859718322754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4880506880581379, 'clip_ratio/low_mean': 0.014413215219974518, 'clip_ratio/low_min': 0.014413215219974518, 'clip_ratio/high_mean': 0.023415555711835623, 'clip_ratio/high_max': 0.023415555711835623, 'clip_ratio/region_mean': 0.03782877093181014, 'reward_total_mean': 0.9984524250030518, 'reward_meter_mean': 0.9984524250030518, 'reward_meter_std': 0.0005822303937748075, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984524250030518, 'reward_total_composite_std': 0.0005822303937748075, 'epoch': 0.1}

 77%|███████▋  | 2549/3300 [3:39:39<1:09:57,  5.59s/it]INFO 04-12 02:16:30 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2550/3300 [3:39:43<1:06:17,  5.30s/it]2026-04-12 02:16:35,053 | INFO | train_grpo_train | metrics_logged mode=train step=2550

                                                       
{'loss': 0.0457, 'grad_norm': 4.8445634841918945, 'learning_rate': 2.275757575757576e-06, 'num_tokens': 5781794.0, 'completions/mean_length': 68.25, 'completions/min_length': 64.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.980836033821106, 'rewards/meter/std': 0.014489145018160343, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.980836033821106, 'rewards/total_composite/std': 0.014489145018160343, 'reward': 0.980836033821106, 'reward_std': 0.014489145949482918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05946962162852287, 'sampling/sampling_logp_difference/max': 1.5740957260131836, 'sampling/importance_sampling_ratio/min': 0.20719482004642487, 'sampling/importance_sampling_ratio/mean': 1.006621241569519, 'sampling/importance_sampling_ratio/max': 1.8987267017364502, 'entropy': 0.5948160253465176, 'clip_ratio/low_mean': 0.017614107578992844, 'clip_ratio/low_min': 0.017614107578992844, 'clip_ratio/high_mean': 0.020960328169167042, 'clip_ratio/high_max': 0.020960328169167042, 'clip_ratio/region_mean': 0.038574435748159885, 'reward_total_mean': 0.980836033821106, 'reward_meter_mean': 0.980836033821106, 'reward_meter_std': 0.014489145018160343, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.980836033821106, 'reward_total_composite_std': 0.014489145018160343, 'epoch': 0.1}

 77%|███████▋  | 2550/3300 [3:39:44<1:06:17,  5.30s/it]INFO 04-12 02:16:35 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:36,  3.36s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.44s/it]

 31%|███       | 4/13 [00:17<00:43,  4.81s/it]

 38%|███▊      | 5/13 [00:22<00:39,  4.91s/it]

 46%|████▌     | 6/13 [00:28<00:38,  5.43s/it]

 54%|█████▍    | 7/13 [00:34<00:32,  5.43s/it]

 62%|██████▏   | 8/13 [00:41<00:29,  5.86s/it]

 69%|██████▉   | 9/13 [00:48<00:25,  6.25s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.82s/it]

 85%|████████▍ | 11/13 [00:58<00:11,  5.68s/it]

 92%|█████████▏| 12/13 [01:04<00:05,  5.79s/it]

100%|██████████| 13/13 [01:09<00:00,  5.63s/it]2026-04-12 02:17:51,356 | INFO | train_grpo_train | metrics_logged mode=eval step=2550

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 76.2084, 'eval_samples_per_second': 1.365, 'eval_steps_per_second': 0.171, 'eval_num_tokens': 5781794.0, 'eval_completions/mean_length': 210.5, 'eval_completions/min_length': 62.92307692307692, 'eval_completions/max_length': 409.0769230769231, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 207.71153963529147, 'eval_completions/min_terminated_length': 62.92307692307692, 'eval_completions/max_terminated_length': 399.2307692307692, 'eval_rewards/meter/mean': 0.768360605606666, 'eval_rewards/meter/std': 0.34562337971650636, 'eval_rewards/count_adherence/mean': 0.9420714332507207, 'eval_rewards/count_adherence/std': 0.08270380875239006, 'eval_rewards/arabic_clean/mean': 0.9711538461538461, 'eval_rewards/arabic_clean/std': 0.08158924258672275, 'eval_rewards/repeat_penalty/mean': 0.9319783999369695, 'eval_rewards/repeat_penalty/std': 0.09335804444092971, 'eval_rewards/total_composite/mean': 0.6745542379525992, 'eval_rewards/total_composite/std': 0.3481578013071647, 'eval_reward': 0.6745542379525992, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03692194346625071, 'eval_sampling/sampling_logp_difference/max': 1.2396357609675481, 'eval_sampling/importance_sampling_ratio/min': 0.29530371954807866, 'eval_sampling/importance_sampling_ratio/mean': 1.0112521740106435, 'eval_sampling/importance_sampling_ratio/max': 1.5434642296570997, 'eval_entropy': 0.42294850601599765, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.6745542379525992, 'eval_reward_meter_mean': 0.768360605606666, 'eval_reward_meter_std': 0.34562337971650636, 'eval_reward_count_adherence_mean': 0.9420714332507207, 'eval_reward_count_adherence_std': 0.08270380875239006, 'eval_reward_arabic_clean_mean': 0.9711538461538461, 'eval_reward_arabic_clean_std': 0.08158924258672275, 'eval_reward_repeat_penalty_mean': 0.9319783999369695, 'eval_reward_repeat_penalty_std': 0.09335804444092971, 'eval_reward_total_composite_mean': 0.6745542379525992, 'eval_reward_total_composite_std': 0.3481578013071647, 'epoch': 0.1}

 77%|███████▋  | 2550/3300 [3:41:00<1:06:17,  5.30s/it]

100%|██████████| 13/13 [01:09<00:00,  5.63s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 77%|███████▋  | 2551/3300 [3:41:08<6:04:05, 29.17s/it]2026-04-12 02:17:59,907 | INFO | train_grpo_train | metrics_logged mode=train step=2551

                                                       
{'loss': 0.004, 'grad_norm': 1.7443753480911255, 'learning_rate': 2.2727272727272728e-06, 'num_tokens': 5784519.0, 'completions/mean_length': 166.625, 'completions/min_length': 166.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.625, 'completions/min_terminated_length': 166.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.998813271522522, 'rewards/meter/std': 0.00025403356994502246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998813271522522, 'rewards/total_composite/std': 0.00025403356994502246, 'reward': 0.998813271522522, 'reward_std': 0.00025402315077371895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021267293021082878, 'sampling/sampling_logp_difference/max': 1.1385552883148193, 'sampling/importance_sampling_ratio/min': 0.32028141617774963, 'sampling/importance_sampling_ratio/mean': 1.0034915208816528, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1492440328001976, 'clip_ratio/low_mean': 0.005239521153271198, 'clip_ratio/low_min': 0.005239521153271198, 'clip_ratio/high_mean': 0.013518144143745303, 'clip_ratio/high_max': 0.013518144143745303, 'clip_ratio/region_mean': 0.0187576652970165, 'reward_total_mean': 0.998813271522522, 'reward_meter_mean': 0.998813271522522, 'reward_meter_std': 0.00025403356994502246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998813271522522, 'reward_total_composite_std': 0.00025403356994502246, 'epoch': 0.1}

 77%|███████▋  | 2551/3300 [3:41:08<6:04:05, 29.17s/it]INFO 04-12 02:18:00 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2552/3300 [3:41:13<4:33:32, 21.94s/it]2026-04-12 02:18:04,984 | INFO | train_grpo_train | metrics_logged mode=train step=2552

                                                       
{'loss': 0.0005, 'grad_norm': 0.3295544683933258, 'learning_rate': 2.2696969696969696e-06, 'num_tokens': 5786536.0, 'completions/mean_length': 98.125, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.125, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9980067014694214, 'rewards/meter/std': 2.103978840750642e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980067014694214, 'rewards/total_composite/std': 2.103978840750642e-05, 'reward': 0.9980067014694214, 'reward_std': 2.104529266944155e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008558080531656742, 'sampling/sampling_logp_difference/max': 1.609175205230713, 'sampling/importance_sampling_ratio/min': 0.20005254447460175, 'sampling/importance_sampling_ratio/mean': 1.0038623809814453, 'sampling/importance_sampling_ratio/max': 1.583804726600647, 'entropy': 0.05678324867039919, 'clip_ratio/low_mean': 0.0012755101779475808, 'clip_ratio/low_min': 0.0012755101779475808, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0012755101779475808, 'reward_total_mean': 0.9980067014694214, 'reward_meter_mean': 0.9980067014694214, 'reward_meter_std': 2.103978840750642e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980067014694214, 'reward_total_composite_std': 2.103978840750642e-05, 'epoch': 0.1}

 77%|███████▋  | 2552/3300 [3:41:13<4:33:32, 21.94s/it]INFO 04-12 02:18:05 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2553/3300 [3:41:18<3:27:55, 16.70s/it]2026-04-12 02:18:09,456 | INFO | train_grpo_train | metrics_logged mode=train step=2553

                                                       
{'loss': 0.0792, 'grad_norm': 12.867728233337402, 'learning_rate': 2.266666666666667e-06, 'num_tokens': 5788162.0, 'completions/mean_length': 45.25, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.5925682783126831, 'rewards/meter/std': 0.4251895844936371, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5925682783126831, 'rewards/total_composite/std': 0.4251895844936371, 'reward': 0.5925682783126831, 'reward_std': 0.4251895546913147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1222197413444519, 'sampling/sampling_logp_difference/max': 4.768388271331787, 'sampling/importance_sampling_ratio/min': 0.008494059555232525, 'sampling/importance_sampling_ratio/mean': 0.9826568365097046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5326411537826061, 'clip_ratio/low_mean': 0.028055555652827024, 'clip_ratio/low_min': 0.028055555652827024, 'clip_ratio/high_mean': 0.08001901814714074, 'clip_ratio/high_max': 0.08001901814714074, 'clip_ratio/region_mean': 0.10807457379996777, 'reward_total_mean': 0.5925682783126831, 'reward_meter_mean': 0.5925682783126831, 'reward_meter_std': 0.4251895844936371, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5925682783126831, 'reward_total_composite_std': 0.4251895844936371, 'epoch': 0.1}

 77%|███████▋  | 2553/3300 [3:41:18<3:27:55, 16.70s/it]INFO 04-12 02:18:09 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2554/3300 [3:41:23<2:44:35, 13.24s/it]2026-04-12 02:18:14,618 | INFO | train_grpo_train | metrics_logged mode=train step=2554

                                                       
{'loss': -0.0001, 'grad_norm': 3.188581705093384, 'learning_rate': 2.2636363636363637e-06, 'num_tokens': 5790271.0, 'completions/mean_length': 89.625, 'completions/min_length': 88.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.625, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.9921834468841553, 'rewards/meter/std': 0.003569816704839468, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9921834468841553, 'rewards/total_composite/std': 0.003569816704839468, 'reward': 0.9921834468841553, 'reward_std': 0.0035698034334927797, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032123863697052, 'sampling/sampling_logp_difference/max': 0.9250731468200684, 'sampling/importance_sampling_ratio/min': 0.39650243520736694, 'sampling/importance_sampling_ratio/mean': 1.0064231157302856, 'sampling/importance_sampling_ratio/max': 1.3844112157821655, 'entropy': 0.2214542292058468, 'clip_ratio/low_mean': 0.009753433521836996, 'clip_ratio/low_min': 0.009753433521836996, 'clip_ratio/high_mean': 0.011097930022515357, 'clip_ratio/high_max': 0.011097930022515357, 'clip_ratio/region_mean': 0.020851363544352353, 'reward_total_mean': 0.9921834468841553, 'reward_meter_mean': 0.9921834468841553, 'reward_meter_std': 0.003569816704839468, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9921834468841553, 'reward_total_composite_std': 0.003569816704839468, 'epoch': 0.1}

 77%|███████▋  | 2554/3300 [3:41:23<2:44:35, 13.24s/it]INFO 04-12 02:18:14 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2555/3300 [3:41:28<2:13:57, 10.79s/it]2026-04-12 02:18:19,694 | INFO | train_grpo_train | metrics_logged mode=train step=2555

                                                       
{'loss': 0.0004, 'grad_norm': 0.22581298649311066, 'learning_rate': 2.260606060606061e-06, 'num_tokens': 5792325.0, 'completions/mean_length': 98.75, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.75, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9993069171905518, 'rewards/meter/std': 1.8531554815126583e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993069171905518, 'rewards/total_composite/std': 1.8531554815126583e-05, 'reward': 0.9993069171905518, 'reward_std': 1.8528446162235923e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008686289191246033, 'sampling/sampling_logp_difference/max': 0.6104934215545654, 'sampling/importance_sampling_ratio/min': 0.5430828332901001, 'sampling/importance_sampling_ratio/mean': 1.0017414093017578, 'sampling/importance_sampling_ratio/max': 1.5615042448043823, 'entropy': 0.042933082208037376, 'clip_ratio/low_mean': 0.005063388962298632, 'clip_ratio/low_min': 0.005063388962298632, 'clip_ratio/high_mean': 0.0025510203558951616, 'clip_ratio/high_max': 0.0025510203558951616, 'clip_ratio/region_mean': 0.007614409318193793, 'reward_total_mean': 0.9993069171905518, 'reward_meter_mean': 0.9993069171905518, 'reward_meter_std': 1.8531554815126583e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993069171905518, 'reward_total_composite_std': 1.8531554815126583e-05, 'epoch': 0.1}

 77%|███████▋  | 2555/3300 [3:41:28<2:13:57, 10.79s/it]INFO 04-12 02:18:19 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2556/3300 [3:41:34<1:57:01,  9.44s/it]2026-04-12 02:18:25,974 | INFO | train_grpo_train | metrics_logged mode=train step=2556

                                                       
{'loss': 0.0125, 'grad_norm': 2.697803020477295, 'learning_rate': 2.2575757575757578e-06, 'num_tokens': 5795614.0, 'completions/mean_length': 199.125, 'completions/min_length': 188.0, 'completions/max_length': 209.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 199.125, 'completions/min_terminated_length': 188.0, 'completions/max_terminated_length': 209.0, 'rewards/meter/mean': 0.9469772577285767, 'rewards/meter/std': 0.0619339719414711, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9090909361839294, 'rewards/repeat_penalty/std': 0.08416546136140823, 'rewards/total_composite/mean': 0.8613824248313904, 'rewards/total_composite/std': 0.10320668667554855, 'reward': 0.8613824248313904, 'reward_std': 0.10320668667554855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04738626256585121, 'sampling/sampling_logp_difference/max': 1.4838252067565918, 'sampling/importance_sampling_ratio/min': 0.2267685979604721, 'sampling/importance_sampling_ratio/mean': 1.012196660041809, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47996126115322113, 'clip_ratio/low_mean': 0.008094559656456113, 'clip_ratio/low_min': 0.008094559656456113, 'clip_ratio/high_mean': 0.0364626687951386, 'clip_ratio/high_max': 0.0364626687951386, 'clip_ratio/region_mean': 0.04455722845159471, 'reward_total_mean': 0.8613824248313904, 'reward_meter_mean': 0.9469772577285767, 'reward_meter_std': 0.0619339719414711, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9090909361839294, 'reward_repeat_penalty_std': 0.08416546136140823, 'reward_total_composite_mean': 0.8613824248313904, 'reward_total_composite_std': 0.10320668667554855, 'epoch': 0.1}

 77%|███████▋  | 2556/3300 [3:41:34<1:57:01,  9.44s/it]INFO 04-12 02:18:26 [block_pool.py:378] Successfully reset prefix cache

 77%|███████▋  | 2557/3300 [3:41:39<1:39:37,  8.05s/it]2026-04-12 02:18:30,774 | INFO | train_grpo_train | metrics_logged mode=train step=2557

                                                       
{'loss': 0.0071, 'grad_norm': 4.737760066986084, 'learning_rate': 2.254545454545455e-06, 'num_tokens': 5797542.0, 'completions/mean_length': 68.0, 'completions/min_length': 64.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.974303126335144, 'rewards/meter/std': 0.020303523167967796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.974303126335144, 'rewards/total_composite/std': 0.020303523167967796, 'reward': 0.974303126335144, 'reward_std': 0.020303526893258095, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04797229915857315, 'sampling/sampling_logp_difference/max': 1.6128530502319336, 'sampling/importance_sampling_ratio/min': 0.19931812584400177, 'sampling/importance_sampling_ratio/mean': 1.0142627954483032, 'sampling/importance_sampling_ratio/max': 1.927941083908081, 'entropy': 0.4355851337313652, 'clip_ratio/low_mean': 0.01822700910270214, 'clip_ratio/low_min': 0.01822700910270214, 'clip_ratio/high_mean': 0.012197049451060593, 'clip_ratio/high_max': 0.012197049451060593, 'clip_ratio/region_mean': 0.030424058553762734, 'reward_total_mean': 0.974303126335144, 'reward_meter_mean': 0.974303126335144, 'reward_meter_std': 0.020303523167967796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.974303126335144, 'reward_total_composite_std': 0.020303523167967796, 'epoch': 0.1}

 77%|███████▋  | 2557/3300 [3:41:39<1:39:37,  8.05s/it]INFO 04-12 02:18:31 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2558/3300 [3:41:45<1:29:42,  7.25s/it]2026-04-12 02:18:36,180 | INFO | train_grpo_train | metrics_logged mode=train step=2558

                                                       
{'loss': 0.0093, 'grad_norm': 3.3955912590026855, 'learning_rate': 2.251515151515152e-06, 'num_tokens': 5799904.0, 'completions/mean_length': 136.25, 'completions/min_length': 133.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.25, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.975925624370575, 'rewards/meter/std': 0.038366202265024185, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9227904081344604, 'rewards/total_composite/std': 0.0701058879494667, 'reward': 0.9227904081344604, 'reward_std': 0.07010588049888611, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05260032042860985, 'sampling/sampling_logp_difference/max': 1.8960437774658203, 'sampling/importance_sampling_ratio/min': 0.1501615196466446, 'sampling/importance_sampling_ratio/mean': 1.006500005722046, 'sampling/importance_sampling_ratio/max': 1.85683012008667, 'entropy': 0.4538475573062897, 'clip_ratio/low_mean': 0.022996979532763362, 'clip_ratio/low_min': 0.022996979532763362, 'clip_ratio/high_mean': 0.020191184477880597, 'clip_ratio/high_max': 0.020191184477880597, 'clip_ratio/region_mean': 0.04318816401064396, 'reward_total_mean': 0.9227904081344604, 'reward_meter_mean': 0.975925624370575, 'reward_meter_std': 0.038366202265024185, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9227904081344604, 'reward_total_composite_std': 0.0701058879494667, 'epoch': 0.1}

 78%|███████▊  | 2558/3300 [3:41:45<1:29:42,  7.25s/it]INFO 04-12 02:18:36 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2559/3300 [3:41:49<1:19:11,  6.41s/it]2026-04-12 02:18:40,624 | INFO | train_grpo_train | metrics_logged mode=train step=2559

                                                       
{'loss': -0.0096, 'grad_norm': 12.212129592895508, 'learning_rate': 2.2484848484848487e-06, 'num_tokens': 5801541.0, 'completions/mean_length': 44.625, 'completions/min_length': 40.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8007308840751648, 'rewards/meter/std': 0.19306603074073792, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8007308840751648, 'rewards/total_composite/std': 0.19306603074073792, 'reward': 0.8007308840751648, 'reward_std': 0.19306600093841553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09867885708808899, 'sampling/sampling_logp_difference/max': 1.36256742477417, 'sampling/importance_sampling_ratio/min': 0.25600266456604004, 'sampling/importance_sampling_ratio/mean': 0.9941558837890625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5555657297372818, 'clip_ratio/low_mean': 0.030968211824074388, 'clip_ratio/low_min': 0.030968211824074388, 'clip_ratio/high_mean': 0.04882641462609172, 'clip_ratio/high_max': 0.04882641462609172, 'clip_ratio/region_mean': 0.0797946264501661, 'reward_total_mean': 0.8007308840751648, 'reward_meter_mean': 0.8007308840751648, 'reward_meter_std': 0.19306603074073792, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8007308840751648, 'reward_total_composite_std': 0.19306603074073792, 'epoch': 0.1}

 78%|███████▊  | 2559/3300 [3:41:49<1:19:11,  6.41s/it]INFO 04-12 02:18:40 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2560/3300 [3:41:57<1:25:31,  6.93s/it]2026-04-12 02:18:48,782 | INFO | train_grpo_train | metrics_logged mode=train step=2560

                                                       
{'loss': -0.0056, 'grad_norm': 1.910988211631775, 'learning_rate': 2.2454545454545455e-06, 'num_tokens': 5806083.0, 'completions/mean_length': 347.75, 'completions/min_length': 337.0, 'completions/max_length': 361.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 347.75, 'completions/min_terminated_length': 337.0, 'completions/max_terminated_length': 361.0, 'rewards/meter/mean': 0.9986293315887451, 'rewards/meter/std': 0.00039406708674505353, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9489378929138184, 'rewards/repeat_penalty/std': 0.05824849754571915, 'rewards/total_composite/mean': 0.7107186317443848, 'rewards/total_composite/std': 0.043464019894599915, 'reward': 0.7107186317443848, 'reward_std': 0.04346400871872902, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.061371464282274246, 'sampling/sampling_logp_difference/max': 14.583121299743652, 'sampling/importance_sampling_ratio/min': 4.6412063170464535e-07, 'sampling/importance_sampling_ratio/mean': 1.0138837099075317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4555246904492378, 'clip_ratio/low_mean': 0.014391222270205617, 'clip_ratio/low_min': 0.014391222270205617, 'clip_ratio/high_mean': 0.014688471332192421, 'clip_ratio/high_max': 0.014688471332192421, 'clip_ratio/region_mean': 0.029079693602398038, 'reward_total_mean': 0.7107186317443848, 'reward_meter_mean': 0.9986293315887451, 'reward_meter_std': 0.00039406708674505353, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9489378929138184, 'reward_repeat_penalty_std': 0.05824849754571915, 'reward_total_composite_mean': 0.7107186317443848, 'reward_total_composite_std': 0.043464019894599915, 'epoch': 0.1}

 78%|███████▊  | 2560/3300 [3:41:57<1:25:31,  6.93s/it]INFO 04-12 02:18:49 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2561/3300 [3:42:02<1:17:18,  6.28s/it]2026-04-12 02:18:53,526 | INFO | train_grpo_train | metrics_logged mode=train step=2561

                                                       
{'loss': -0.0045, 'grad_norm': 3.3982653617858887, 'learning_rate': 2.2424242424242428e-06, 'num_tokens': 5807836.0, 'completions/mean_length': 67.125, 'completions/min_length': 65.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9521244764328003, 'rewards/meter/std': 0.04941366985440254, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9521244764328003, 'rewards/total_composite/std': 0.04941366985440254, 'reward': 0.9521244764328003, 'reward_std': 0.049413666129112244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03345394507050514, 'sampling/sampling_logp_difference/max': 2.0841522216796875, 'sampling/importance_sampling_ratio/min': 0.12441255152225494, 'sampling/importance_sampling_ratio/mean': 1.0056248903274536, 'sampling/importance_sampling_ratio/max': 1.7526792287826538, 'entropy': 0.28918380104005337, 'clip_ratio/low_mean': 0.007634902372956276, 'clip_ratio/low_min': 0.007634902372956276, 'clip_ratio/high_mean': 0.022279977099969983, 'clip_ratio/high_max': 0.022279977099969983, 'clip_ratio/region_mean': 0.02991487947292626, 'reward_total_mean': 0.9521244764328003, 'reward_meter_mean': 0.9521244764328003, 'reward_meter_std': 0.04941366985440254, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9521244764328003, 'reward_total_composite_std': 0.04941366985440254, 'epoch': 0.1}

 78%|███████▊  | 2561/3300 [3:42:02<1:17:18,  6.28s/it]INFO 04-12 02:18:53 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2562/3300 [3:42:06<1:09:08,  5.62s/it]2026-04-12 02:18:57,620 | INFO | train_grpo_train | metrics_logged mode=train step=2562

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.2393939393939396e-06, 'num_tokens': 5809388.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00015674103633500636, 'sampling/sampling_logp_difference/max': 0.007963716052472591, 'sampling/importance_sampling_ratio/min': 0.9975297451019287, 'sampling/importance_sampling_ratio/mean': 1.0001332759857178, 'sampling/importance_sampling_ratio/max': 1.0079954862594604, 'entropy': 0.002722353514400311, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2562/3300 [3:42:06<1:09:08,  5.62s/it]INFO 04-12 02:18:57 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2563/3300 [3:42:11<1:05:52,  5.36s/it]2026-04-12 02:19:02,375 | INFO | train_grpo_train | metrics_logged mode=train step=2563

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.2363636363636364e-06, 'num_tokens': 5811516.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00030092071392573416, 'sampling/sampling_logp_difference/max': 0.00446719815954566, 'sampling/importance_sampling_ratio/min': 0.999937891960144, 'sampling/importance_sampling_ratio/mean': 1.000300645828247, 'sampling/importance_sampling_ratio/max': 1.0044771432876587, 'entropy': 0.0025233181950170547, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2563/3300 [3:42:11<1:05:52,  5.36s/it]INFO 04-12 02:19:02 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2564/3300 [3:42:15<1:01:52,  5.04s/it]2026-04-12 02:19:06,674 | INFO | train_grpo_train | metrics_logged mode=train step=2564

                                                       
{'loss': 0.0027, 'grad_norm': 1.6162538528442383, 'learning_rate': 2.2333333333333333e-06, 'num_tokens': 5813045.0, 'completions/mean_length': 37.125, 'completions/min_length': 37.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9993616938591003, 'rewards/meter/std': 0.0003300250100437552, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993616938591003, 'rewards/total_composite/std': 0.0003300250100437552, 'reward': 0.9993616938591003, 'reward_std': 0.00033003126736730337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01185247115790844, 'sampling/sampling_logp_difference/max': 1.2676396369934082, 'sampling/importance_sampling_ratio/min': 0.28149527311325073, 'sampling/importance_sampling_ratio/mean': 0.9978834986686707, 'sampling/importance_sampling_ratio/max': 1.1445908546447754, 'entropy': 0.059151682537049055, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0033783784601837397, 'reward_total_mean': 0.9993616938591003, 'reward_meter_mean': 0.9993616938591003, 'reward_meter_std': 0.0003300250100437552, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993616938591003, 'reward_total_composite_std': 0.0003300250100437552, 'epoch': 0.1}

 78%|███████▊  | 2564/3300 [3:42:15<1:01:52,  5.04s/it]INFO 04-12 02:19:06 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2565/3300 [3:42:19<57:49,  4.72s/it]  2026-04-12 02:19:10,637 | INFO | train_grpo_train | metrics_logged mode=train step=2565

                                                     
{'loss': -0.0009, 'grad_norm': 0.15634091198444366, 'learning_rate': 2.2303030303030305e-06, 'num_tokens': 5814621.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957231283187866, 'rewards/meter/std': 5.816265183966607e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957231283187866, 'rewards/total_composite/std': 5.816265183966607e-05, 'reward': 0.9957231283187866, 'reward_std': 5.816265183966607e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0017511904006823897, 'sampling/sampling_logp_difference/max': 0.1276249885559082, 'sampling/importance_sampling_ratio/min': 0.8801833987236023, 'sampling/importance_sampling_ratio/mean': 0.99995356798172, 'sampling/importance_sampling_ratio/max': 1.0548250675201416, 'entropy': 0.007859309727791697, 'clip_ratio/low_mean': 0.004310344811528921, 'clip_ratio/low_min': 0.004310344811528921, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004310344811528921, 'reward_total_mean': 0.9957231283187866, 'reward_meter_mean': 0.9957231283187866, 'reward_meter_std': 5.816265183966607e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957231283187866, 'reward_total_composite_std': 5.816265183966607e-05, 'epoch': 0.1}

 78%|███████▊  | 2565/3300 [3:42:19<57:49,  4.72s/it]INFO 04-12 02:19:10 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2566/3300 [3:42:24<58:00,  4.74s/it]2026-04-12 02:19:15,433 | INFO | train_grpo_train | metrics_logged mode=train step=2566

                                                     
{'loss': -0.0106, 'grad_norm': 1.9842983484268188, 'learning_rate': 2.2272727272727274e-06, 'num_tokens': 5816563.0, 'completions/mean_length': 78.75, 'completions/min_length': 77.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.75, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.99830561876297, 'rewards/meter/std': 0.001100848545320332, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99830561876297, 'rewards/total_composite/std': 0.001100848545320332, 'reward': 0.99830561876297, 'reward_std': 0.0011008477304130793, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02984953485429287, 'sampling/sampling_logp_difference/max': 0.9692831039428711, 'sampling/importance_sampling_ratio/min': 0.37935489416122437, 'sampling/importance_sampling_ratio/mean': 1.0084333419799805, 'sampling/importance_sampling_ratio/max': 1.649073600769043, 'entropy': 0.22580487374216318, 'clip_ratio/low_mean': 0.001623376621864736, 'clip_ratio/low_min': 0.001623376621864736, 'clip_ratio/high_mean': 0.014201456913724542, 'clip_ratio/high_max': 0.014201456913724542, 'clip_ratio/region_mean': 0.015824833535589278, 'reward_total_mean': 0.99830561876297, 'reward_meter_mean': 0.99830561876297, 'reward_meter_std': 0.001100848545320332, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99830561876297, 'reward_total_composite_std': 0.001100848545320332, 'epoch': 0.1}

 78%|███████▊  | 2566/3300 [3:42:24<58:00,  4.74s/it]INFO 04-12 02:19:15 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2567/3300 [3:42:29<58:08,  4.76s/it]2026-04-12 02:19:20,234 | INFO | train_grpo_train | metrics_logged mode=train step=2567

                                                     
{'loss': 0.03, 'grad_norm': 5.494159698486328, 'learning_rate': 2.224242424242424e-06, 'num_tokens': 5818390.0, 'completions/mean_length': 70.375, 'completions/min_length': 67.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.375, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9767506122589111, 'rewards/meter/std': 0.017817148938775063, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9767506122589111, 'rewards/total_composite/std': 0.017817148938775063, 'reward': 0.9767506122589111, 'reward_std': 0.017817148938775063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05213453993201256, 'sampling/sampling_logp_difference/max': 2.3027992248535156, 'sampling/importance_sampling_ratio/min': 0.09997859597206116, 'sampling/importance_sampling_ratio/mean': 1.0100852251052856, 'sampling/importance_sampling_ratio/max': 1.9086910486221313, 'entropy': 0.41449311561882496, 'clip_ratio/low_mean': 0.015345451422035694, 'clip_ratio/low_min': 0.015345451422035694, 'clip_ratio/high_mean': 0.030598473269492388, 'clip_ratio/high_max': 0.030598473269492388, 'clip_ratio/region_mean': 0.04594392469152808, 'reward_total_mean': 0.9767506122589111, 'reward_meter_mean': 0.9767506122589111, 'reward_meter_std': 0.017817148938775063, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9767506122589111, 'reward_total_composite_std': 0.017817148938775063, 'epoch': 0.1}

 78%|███████▊  | 2567/3300 [3:42:29<58:08,  4.76s/it]INFO 04-12 02:19:20 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2568/3300 [3:42:33<58:23,  4.79s/it]2026-04-12 02:19:25,087 | INFO | train_grpo_train | metrics_logged mode=train step=2568

                                                     
{'loss': -0.0085, 'grad_norm': 4.675406455993652, 'learning_rate': 2.2212121212121214e-06, 'num_tokens': 5820304.0, 'completions/mean_length': 72.25, 'completions/min_length': 68.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.842863917350769, 'rewards/meter/std': 0.26622849702835083, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.842863917350769, 'rewards/total_composite/std': 0.26622849702835083, 'reward': 0.842863917350769, 'reward_std': 0.26622846722602844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06255535036325455, 'sampling/sampling_logp_difference/max': 1.408411979675293, 'sampling/importance_sampling_ratio/min': 0.2445313036441803, 'sampling/importance_sampling_ratio/mean': 1.005039095878601, 'sampling/importance_sampling_ratio/max': 1.7533190250396729, 'entropy': 0.4912409372627735, 'clip_ratio/low_mean': 0.008893084479495883, 'clip_ratio/low_min': 0.008893084479495883, 'clip_ratio/high_mean': 0.03215251048095524, 'clip_ratio/high_max': 0.03215251048095524, 'clip_ratio/region_mean': 0.041045594960451126, 'reward_total_mean': 0.842863917350769, 'reward_meter_mean': 0.842863917350769, 'reward_meter_std': 0.26622849702835083, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.842863917350769, 'reward_total_composite_std': 0.26622849702835083, 'epoch': 0.1}

 78%|███████▊  | 2568/3300 [3:42:34<58:23,  4.79s/it]INFO 04-12 02:19:25 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2569/3300 [3:42:38<56:20,  4.63s/it]2026-04-12 02:19:29,332 | INFO | train_grpo_train | metrics_logged mode=train step=2569

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.2181818181818187e-06, 'num_tokens': 5821728.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002904376306105405, 'sampling/sampling_logp_difference/max': 0.0029266122728586197, 'sampling/importance_sampling_ratio/min': 0.9983823895454407, 'sampling/importance_sampling_ratio/mean': 1.000272274017334, 'sampling/importance_sampling_ratio/max': 1.0029308795928955, 'entropy': 0.002900286461226642, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2569/3300 [3:42:38<56:20,  4.63s/it]INFO 04-12 02:19:29 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2570/3300 [3:42:42<56:00,  4.60s/it]2026-04-12 02:19:33,885 | INFO | train_grpo_train | metrics_logged mode=train step=2570

                                                     
{'loss': 0.011, 'grad_norm': 12.921791076660156, 'learning_rate': 2.2151515151515155e-06, 'num_tokens': 5823592.0, 'completions/mean_length': 43.0, 'completions/min_length': 39.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8914755582809448, 'rewards/meter/std': 0.08936496078968048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8914755582809448, 'rewards/total_composite/std': 0.08936496078968048, 'reward': 0.8914755582809448, 'reward_std': 0.08936495333909988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09845032542943954, 'sampling/sampling_logp_difference/max': 1.851944923400879, 'sampling/importance_sampling_ratio/min': 0.15693165361881256, 'sampling/importance_sampling_ratio/mean': 0.9992160201072693, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.562019981443882, 'clip_ratio/low_mean': 0.035002983175218105, 'clip_ratio/low_min': 0.035002983175218105, 'clip_ratio/high_mean': 0.05010954383760691, 'clip_ratio/high_max': 0.05010954383760691, 'clip_ratio/region_mean': 0.08511252701282501, 'reward_total_mean': 0.8914755582809448, 'reward_meter_mean': 0.8914755582809448, 'reward_meter_std': 0.08936496078968048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8914755582809448, 'reward_total_composite_std': 0.08936496078968048, 'epoch': 0.1}

 78%|███████▊  | 2570/3300 [3:42:42<56:00,  4.60s/it]INFO 04-12 02:19:34 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2571/3300 [3:42:47<55:58,  4.61s/it]2026-04-12 02:19:38,498 | INFO | train_grpo_train | metrics_logged mode=train step=2571

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.2121212121212124e-06, 'num_tokens': 5824992.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00020078742818441242, 'sampling/sampling_logp_difference/max': 0.0017774300649762154, 'sampling/importance_sampling_ratio/min': 1.0, 'sampling/importance_sampling_ratio/mean': 1.0002009868621826, 'sampling/importance_sampling_ratio/max': 1.0017789602279663, 'entropy': 0.0015226418327074498, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2571/3300 [3:42:47<55:58,  4.61s/it]INFO 04-12 02:19:38 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2572/3300 [3:42:51<54:35,  4.50s/it]2026-04-12 02:19:42,746 | INFO | train_grpo_train | metrics_logged mode=train step=2572

                                                     
{'loss': -0.0009, 'grad_norm': 1.0375055074691772, 'learning_rate': 2.209090909090909e-06, 'num_tokens': 5826320.0, 'completions/mean_length': 37.0, 'completions/min_length': 37.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9995728731155396, 'rewards/meter/std': 3.974059291067533e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9995728731155396, 'rewards/total_composite/std': 3.974059291067533e-05, 'reward': 0.9995728731155396, 'reward_std': 3.974912760895677e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009617374278604984, 'sampling/sampling_logp_difference/max': 0.6554868221282959, 'sampling/importance_sampling_ratio/min': 0.5191892385482788, 'sampling/importance_sampling_ratio/mean': 1.0013068914413452, 'sampling/importance_sampling_ratio/max': 1.1907140016555786, 'entropy': 0.054603309370577335, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0033783784601837397, 'reward_total_mean': 0.9995728731155396, 'reward_meter_mean': 0.9995728731155396, 'reward_meter_std': 3.974059291067533e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9995728731155396, 'reward_total_composite_std': 3.974059291067533e-05, 'epoch': 0.1}

 78%|███████▊  | 2572/3300 [3:42:51<54:35,  4.50s/it]INFO 04-12 02:19:42 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2573/3300 [3:42:56<55:09,  4.55s/it]2026-04-12 02:19:47,424 | INFO | train_grpo_train | metrics_logged mode=train step=2573

                                                     
{'loss': 0.0037, 'grad_norm': 5.262681484222412, 'learning_rate': 2.2060606060606064e-06, 'num_tokens': 5828013.0, 'completions/mean_length': 67.625, 'completions/min_length': 64.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9842870235443115, 'rewards/meter/std': 0.03939536586403847, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8596113920211792, 'rewards/total_composite/std': 0.34952229261398315, 'reward': 0.8596113920211792, 'reward_std': 0.34952226281166077, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06369344890117645, 'sampling/sampling_logp_difference/max': 2.000176429748535, 'sampling/importance_sampling_ratio/min': 0.13531140983104706, 'sampling/importance_sampling_ratio/mean': 1.0002013444900513, 'sampling/importance_sampling_ratio/max': 1.7594947814941406, 'entropy': 0.5100602991878986, 'clip_ratio/low_mean': 0.005514706019312143, 'clip_ratio/low_min': 0.005514706019312143, 'clip_ratio/high_mean': 0.04813995969016105, 'clip_ratio/high_max': 0.04813995969016105, 'clip_ratio/region_mean': 0.05365466570947319, 'reward_total_mean': 0.8596113920211792, 'reward_meter_mean': 0.9842870235443115, 'reward_meter_std': 0.03939536586403847, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8596113920211792, 'reward_total_composite_std': 0.34952229261398315, 'epoch': 0.1}

 78%|███████▊  | 2573/3300 [3:42:56<55:09,  4.55s/it]INFO 04-12 02:19:47 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2574/3300 [3:43:01<55:46,  4.61s/it]2026-04-12 02:19:52,171 | INFO | train_grpo_train | metrics_logged mode=train step=2574

                                                     
{'loss': -0.0004, 'grad_norm': 1.3739620447158813, 'learning_rate': 2.2030303030303033e-06, 'num_tokens': 5829911.0, 'completions/mean_length': 72.25, 'completions/min_length': 71.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9990666508674622, 'rewards/meter/std': 0.00011669577361317351, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990666508674622, 'rewards/total_composite/std': 0.00011669577361317351, 'reward': 0.9990666508674622, 'reward_std': 0.00011668946535792202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01342670526355505, 'sampling/sampling_logp_difference/max': 1.0127067565917969, 'sampling/importance_sampling_ratio/min': 0.3632344603538513, 'sampling/importance_sampling_ratio/mean': 1.0047719478607178, 'sampling/importance_sampling_ratio/max': 1.3253597021102905, 'entropy': 0.09299392439424992, 'clip_ratio/low_mean': 0.0017605633474886417, 'clip_ratio/low_min': 0.0017605633474886417, 'clip_ratio/high_mean': 0.008656773250550032, 'clip_ratio/high_max': 0.008656773250550032, 'clip_ratio/region_mean': 0.010417336598038673, 'reward_total_mean': 0.9990666508674622, 'reward_meter_mean': 0.9990666508674622, 'reward_meter_std': 0.00011669577361317351, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990666508674622, 'reward_total_composite_std': 0.00011669577361317351, 'epoch': 0.1}

 78%|███████▊  | 2574/3300 [3:43:01<55:46,  4.61s/it]INFO 04-12 02:19:52 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2575/3300 [3:43:06<58:32,  4.84s/it]2026-04-12 02:19:57,561 | INFO | train_grpo_train | metrics_logged mode=train step=2575

                                                     
{'loss': -0.001, 'grad_norm': 0.6759248971939087, 'learning_rate': 2.2e-06, 'num_tokens': 5832393.0, 'completions/mean_length': 132.25, 'completions/min_length': 130.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.25, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9992468357086182, 'rewards/meter/std': 8.659085870021954e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992468357086182, 'rewards/total_composite/std': 8.659085870021954e-05, 'reward': 0.9992468357086182, 'reward_std': 8.659218292450532e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010475593619048595, 'sampling/sampling_logp_difference/max': 1.35489821434021, 'sampling/importance_sampling_ratio/min': 0.2579735517501831, 'sampling/importance_sampling_ratio/mean': 1.0009084939956665, 'sampling/importance_sampling_ratio/max': 1.3609951734542847, 'entropy': 0.08119754772633314, 'clip_ratio/low_mean': 0.004734848625957966, 'clip_ratio/low_min': 0.004734848625957966, 'clip_ratio/high_mean': 0.006622972548939288, 'clip_ratio/high_max': 0.006622972548939288, 'clip_ratio/region_mean': 0.011357821174897254, 'reward_total_mean': 0.9992468357086182, 'reward_meter_mean': 0.9992468357086182, 'reward_meter_std': 8.659085870021954e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992468357086182, 'reward_total_composite_std': 8.659085870021954e-05, 'epoch': 0.1}

 78%|███████▊  | 2575/3300 [3:43:06<58:32,  4.84s/it]INFO 04-12 02:19:57 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2576/3300 [3:43:11<58:04,  4.81s/it]2026-04-12 02:20:02,300 | INFO | train_grpo_train | metrics_logged mode=train step=2576

                                                     
{'loss': 0.0001, 'grad_norm': 1.7544798851013184, 'learning_rate': 2.196969696969697e-06, 'num_tokens': 5834484.0, 'completions/mean_length': 80.375, 'completions/min_length': 79.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9987189769744873, 'rewards/meter/std': 0.00029780645854771137, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987189769744873, 'rewards/total_composite/std': 0.00029780645854771137, 'reward': 0.9987189769744873, 'reward_std': 0.00029780055047012866, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023812301456928253, 'sampling/sampling_logp_difference/max': 0.7268044948577881, 'sampling/importance_sampling_ratio/min': 0.48345139622688293, 'sampling/importance_sampling_ratio/mean': 1.005149483680725, 'sampling/importance_sampling_ratio/max': 1.3658860921859741, 'entropy': 0.21419258043169975, 'clip_ratio/low_mean': 0.0062509768176823854, 'clip_ratio/low_min': 0.0062509768176823854, 'clip_ratio/high_mean': 0.014005606528371572, 'clip_ratio/high_max': 0.014005606528371572, 'clip_ratio/region_mean': 0.020256583346053958, 'reward_total_mean': 0.9987189769744873, 'reward_meter_mean': 0.9987189769744873, 'reward_meter_std': 0.00029780645854771137, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987189769744873, 'reward_total_composite_std': 0.00029780645854771137, 'epoch': 0.1}

 78%|███████▊  | 2576/3300 [3:43:11<58:04,  4.81s/it]INFO 04-12 02:20:02 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2577/3300 [3:43:15<57:05,  4.74s/it]2026-04-12 02:20:06,858 | INFO | train_grpo_train | metrics_logged mode=train step=2577

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.193939393939394e-06, 'num_tokens': 5836412.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0015179180772975087, 'sampling/sampling_logp_difference/max': 0.12468904256820679, 'sampling/importance_sampling_ratio/min': 0.882771372795105, 'sampling/importance_sampling_ratio/mean': 1.0000942945480347, 'sampling/importance_sampling_ratio/max': 1.0703659057617188, 'entropy': 0.015041560400277376, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2577/3300 [3:43:15<57:05,  4.74s/it]INFO 04-12 02:20:07 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2578/3300 [3:43:20<57:39,  4.79s/it]2026-04-12 02:20:11,777 | INFO | train_grpo_train | metrics_logged mode=train step=2578

                                                     
{'loss': 0.0004, 'grad_norm': 0.25849685072898865, 'learning_rate': 2.190909090909091e-06, 'num_tokens': 5838518.0, 'completions/mean_length': 98.25, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.25, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9980080127716064, 'rewards/meter/std': 1.598181188455783e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980080127716064, 'rewards/total_composite/std': 1.598181188455783e-05, 'reward': 0.9980080127716064, 'reward_std': 1.5980675016180612e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007937485352158546, 'sampling/sampling_logp_difference/max': 1.1361374855041504, 'sampling/importance_sampling_ratio/min': 0.3210567235946655, 'sampling/importance_sampling_ratio/mean': 1.0036529302597046, 'sampling/importance_sampling_ratio/max': 1.5758888721466064, 'entropy': 0.0589327453635633, 'clip_ratio/low_mean': 0.0038007627008482814, 'clip_ratio/low_min': 0.0038007627008482814, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0038007627008482814, 'reward_total_mean': 0.9980080127716064, 'reward_meter_mean': 0.9980080127716064, 'reward_meter_std': 1.598181188455783e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980080127716064, 'reward_total_composite_std': 1.598181188455783e-05, 'epoch': 0.1}

 78%|███████▊  | 2578/3300 [3:43:20<57:39,  4.79s/it]INFO 04-12 02:20:12 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2579/3300 [3:43:25<58:36,  4.88s/it]2026-04-12 02:20:16,857 | INFO | train_grpo_train | metrics_logged mode=train step=2579

                                                     
{'loss': 0.0013, 'grad_norm': 3.5709714889526367, 'learning_rate': 2.187878787878788e-06, 'num_tokens': 5840674.0, 'completions/mean_length': 104.5, 'completions/min_length': 98.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.5, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9814869165420532, 'rewards/meter/std': 0.012266860343515873, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9814869165420532, 'rewards/total_composite/std': 0.012266860343515873, 'reward': 0.9814869165420532, 'reward_std': 0.012266861274838448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05761587247252464, 'sampling/sampling_logp_difference/max': 1.8346548080444336, 'sampling/importance_sampling_ratio/min': 0.15966860949993134, 'sampling/importance_sampling_ratio/mean': 1.0061341524124146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47100016474723816, 'clip_ratio/low_mean': 0.01786035276018083, 'clip_ratio/low_min': 0.01786035276018083, 'clip_ratio/high_mean': 0.02378739172127098, 'clip_ratio/high_max': 0.02378739172127098, 'clip_ratio/region_mean': 0.04164774448145181, 'reward_total_mean': 0.9814869165420532, 'reward_meter_mean': 0.9814869165420532, 'reward_meter_std': 0.012266860343515873, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9814869165420532, 'reward_total_composite_std': 0.012266860343515873, 'epoch': 0.1}

 78%|███████▊  | 2579/3300 [3:43:25<58:36,  4.88s/it]INFO 04-12 02:20:17 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2580/3300 [3:43:30<57:21,  4.78s/it]2026-04-12 02:20:21,409 | INFO | train_grpo_train | metrics_logged mode=train step=2580

                                                     
{'loss': 0.009, 'grad_norm': 4.734373569488525, 'learning_rate': 2.184848484848485e-06, 'num_tokens': 5842442.0, 'completions/mean_length': 59.0, 'completions/min_length': 56.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9932022094726562, 'rewards/meter/std': 0.002190361265093088, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9932022094726562, 'rewards/total_composite/std': 0.002190361265093088, 'reward': 0.9932022094726562, 'reward_std': 0.002190381521359086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03438284620642662, 'sampling/sampling_logp_difference/max': 1.2286577224731445, 'sampling/importance_sampling_ratio/min': 0.2926851809024811, 'sampling/importance_sampling_ratio/mean': 1.0038007497787476, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.25036873295903206, 'clip_ratio/low_mean': 0.008511104388162494, 'clip_ratio/low_min': 0.008511104388162494, 'clip_ratio/high_mean': 0.01871931995265186, 'clip_ratio/high_max': 0.01871931995265186, 'clip_ratio/region_mean': 0.027230424340814352, 'reward_total_mean': 0.9932022094726562, 'reward_meter_mean': 0.9932022094726562, 'reward_meter_std': 0.002190361265093088, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9932022094726562, 'reward_total_composite_std': 0.002190361265093088, 'epoch': 0.1}

 78%|███████▊  | 2580/3300 [3:43:30<57:21,  4.78s/it]INFO 04-12 02:20:21 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2581/3300 [3:43:34<56:44,  4.74s/it]2026-04-12 02:20:26,044 | INFO | train_grpo_train | metrics_logged mode=train step=2581

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.181818181818182e-06, 'num_tokens': 5844282.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005618593422695994, 'sampling/sampling_logp_difference/max': 0.03692399710416794, 'sampling/importance_sampling_ratio/min': 0.9637494087219238, 'sampling/importance_sampling_ratio/mean': 0.9999182224273682, 'sampling/importance_sampling_ratio/max': 1.0165749788284302, 'entropy': 0.006089928210712969, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2581/3300 [3:43:35<56:44,  4.74s/it]INFO 04-12 02:20:26 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2582/3300 [3:43:39<56:29,  4.72s/it]2026-04-12 02:20:30,733 | INFO | train_grpo_train | metrics_logged mode=train step=2582

                                                     
{'loss': -0.0011, 'grad_norm': 1.6909281015396118, 'learning_rate': 2.1787878787878788e-06, 'num_tokens': 5846329.0, 'completions/mean_length': 72.875, 'completions/min_length': 72.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.875, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9989848136901855, 'rewards/meter/std': 0.00030317477649077773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989848136901855, 'rewards/total_composite/std': 0.00030317477649077773, 'reward': 0.9989848136901855, 'reward_std': 0.0003031773376278579, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014997649937868118, 'sampling/sampling_logp_difference/max': 1.249094009399414, 'sampling/importance_sampling_ratio/min': 0.2867645025253296, 'sampling/importance_sampling_ratio/mean': 1.0025291442871094, 'sampling/importance_sampling_ratio/max': 1.3127074241638184, 'entropy': 0.0825590007007122, 'clip_ratio/low_mean': 0.0017123287543654442, 'clip_ratio/low_min': 0.0017123287543654442, 'clip_ratio/high_mean': 0.0017361111240461469, 'clip_ratio/high_max': 0.0017361111240461469, 'clip_ratio/region_mean': 0.003448439878411591, 'reward_total_mean': 0.9989848136901855, 'reward_meter_mean': 0.9989848136901855, 'reward_meter_std': 0.00030317477649077773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989848136901855, 'reward_total_composite_std': 0.00030317477649077773, 'epoch': 0.1}

 78%|███████▊  | 2582/3300 [3:43:39<56:29,  4.72s/it]INFO 04-12 02:20:30 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2583/3300 [3:43:44<58:41,  4.91s/it]2026-04-12 02:20:36,090 | INFO | train_grpo_train | metrics_logged mode=train step=2583

                                                     
{'loss': 0.0175, 'grad_norm': 4.086141109466553, 'learning_rate': 2.175757575757576e-06, 'num_tokens': 5848580.0, 'completions/mean_length': 120.375, 'completions/min_length': 116.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.375, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.989915132522583, 'rewards/meter/std': 0.006506068166345358, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9723613262176514, 'rewards/total_composite/std': 0.052766378968954086, 'reward': 0.9723613262176514, 'reward_std': 0.05276636406779289, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059543609619140625, 'sampling/sampling_logp_difference/max': 1.5191116333007812, 'sampling/importance_sampling_ratio/min': 0.21890626847743988, 'sampling/importance_sampling_ratio/mean': 0.9998201727867126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4293239116668701, 'clip_ratio/low_mean': 0.005081300623714924, 'clip_ratio/low_min': 0.005081300623714924, 'clip_ratio/high_mean': 0.057334170676767826, 'clip_ratio/high_max': 0.057334170676767826, 'clip_ratio/region_mean': 0.06241547130048275, 'reward_total_mean': 0.9723613262176514, 'reward_meter_mean': 0.989915132522583, 'reward_meter_std': 0.006506068166345358, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9723613262176514, 'reward_total_composite_std': 0.052766378968954086, 'epoch': 0.1}

 78%|███████▊  | 2583/3300 [3:43:45<58:41,  4.91s/it]INFO 04-12 02:20:36 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2584/3300 [3:43:49<57:00,  4.78s/it]2026-04-12 02:20:40,552 | INFO | train_grpo_train | metrics_logged mode=train step=2584

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.172727272727273e-06, 'num_tokens': 5850268.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00024031523207668215, 'sampling/sampling_logp_difference/max': 0.002918010577559471, 'sampling/importance_sampling_ratio/min': 0.9997970461845398, 'sampling/importance_sampling_ratio/mean': 1.0002394914627075, 'sampling/importance_sampling_ratio/max': 1.0029222965240479, 'entropy': 0.0018527817592257634, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2584/3300 [3:43:49<57:00,  4.78s/it]INFO 04-12 02:20:40 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2585/3300 [3:43:54<56:42,  4.76s/it]2026-04-12 02:20:45,269 | INFO | train_grpo_train | metrics_logged mode=train step=2585

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.16969696969697e-06, 'num_tokens': 5852236.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002460446848999709, 'sampling/sampling_logp_difference/max': 0.0034597045741975307, 'sampling/importance_sampling_ratio/min': 0.9998869299888611, 'sampling/importance_sampling_ratio/mean': 1.0002455711364746, 'sampling/importance_sampling_ratio/max': 1.0034657716751099, 'entropy': 0.0019878629682352766, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2585/3300 [3:43:54<56:42,  4.76s/it]INFO 04-12 02:20:45 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2586/3300 [3:44:02<1:10:07,  5.89s/it]2026-04-12 02:20:53,807 | INFO | train_grpo_train | metrics_logged mode=train step=2586

                                                       
{'loss': -0.0205, 'grad_norm': 1.9402741193771362, 'learning_rate': 2.166666666666667e-06, 'num_tokens': 5856712.0, 'completions/mean_length': 365.5, 'completions/min_length': 342.0, 'completions/max_length': 391.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 365.5, 'completions/min_terminated_length': 342.0, 'completions/max_terminated_length': 391.0, 'rewards/meter/mean': 0.9985870122909546, 'rewards/meter/std': 0.000464948097942397, 'rewards/count_adherence/mean': 0.7916666269302368, 'rewards/count_adherence/std': 0.04454353079199791, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9663312435150146, 'rewards/repeat_penalty/std': 0.039662934839725494, 'rewards/total_composite/mean': 0.7631400227546692, 'rewards/total_composite/std': 0.03845896199345589, 'reward': 0.7631400227546692, 'reward_std': 0.03845897316932678, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05639910697937012, 'sampling/sampling_logp_difference/max': 2.606635570526123, 'sampling/importance_sampling_ratio/min': 0.07378236204385757, 'sampling/importance_sampling_ratio/mean': 1.0101393461227417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4673043563961983, 'clip_ratio/low_mean': 0.01918130088597536, 'clip_ratio/low_min': 0.01918130088597536, 'clip_ratio/high_mean': 0.012192239752039313, 'clip_ratio/high_max': 0.012192239752039313, 'clip_ratio/region_mean': 0.031373540638014674, 'reward_total_mean': 0.7631400227546692, 'reward_meter_mean': 0.9985870122909546, 'reward_meter_std': 0.000464948097942397, 'reward_count_adherence_mean': 0.7916666269302368, 'reward_count_adherence_std': 0.04454353079199791, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9663312435150146, 'reward_repeat_penalty_std': 0.039662934839725494, 'reward_total_composite_mean': 0.7631400227546692, 'reward_total_composite_std': 0.03845896199345589, 'epoch': 0.1}

 78%|███████▊  | 2586/3300 [3:44:02<1:10:07,  5.89s/it]INFO 04-12 02:20:54 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2587/3300 [3:44:07<1:07:16,  5.66s/it]2026-04-12 02:20:58,926 | INFO | train_grpo_train | metrics_logged mode=train step=2587

                                                       
{'loss': -0.0067, 'grad_norm': 6.472557067871094, 'learning_rate': 2.163636363636364e-06, 'num_tokens': 5858927.0, 'completions/mean_length': 100.875, 'completions/min_length': 97.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.979087233543396, 'rewards/meter/std': 0.056263912469148636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.979087233543396, 'rewards/total_composite/std': 0.056263912469148636, 'reward': 0.979087233543396, 'reward_std': 0.056263916194438934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04917147755622864, 'sampling/sampling_logp_difference/max': 3.7527036666870117, 'sampling/importance_sampling_ratio/min': 0.023454248905181885, 'sampling/importance_sampling_ratio/mean': 1.0070948600769043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4231472760438919, 'clip_ratio/low_mean': 0.005154639016836882, 'clip_ratio/low_min': 0.005154639016836882, 'clip_ratio/high_mean': 0.038169488427229226, 'clip_ratio/high_max': 0.038169488427229226, 'clip_ratio/region_mean': 0.04332412744406611, 'reward_total_mean': 0.979087233543396, 'reward_meter_mean': 0.979087233543396, 'reward_meter_std': 0.056263912469148636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.979087233543396, 'reward_total_composite_std': 0.056263912469148636, 'epoch': 0.1}

 78%|███████▊  | 2587/3300 [3:44:07<1:07:16,  5.66s/it]INFO 04-12 02:20:59 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2588/3300 [3:44:14<1:09:15,  5.84s/it]2026-04-12 02:21:05,168 | INFO | train_grpo_train | metrics_logged mode=train step=2588

                                                       
{'loss': 0.0085, 'grad_norm': 1.698296308517456, 'learning_rate': 2.1606060606060606e-06, 'num_tokens': 5862067.0, 'completions/mean_length': 202.5, 'completions/min_length': 198.0, 'completions/max_length': 205.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 202.5, 'completions/min_terminated_length': 198.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.9954801797866821, 'rewards/meter/std': 0.008464212529361248, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9772727489471436, 'rewards/repeat_penalty/std': 0.04208271950483322, 'rewards/total_composite/mean': 0.9727857112884521, 'rewards/total_composite/std': 0.04096178710460663, 'reward': 0.9727857112884521, 'reward_std': 0.04096178337931633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02448193170130253, 'sampling/sampling_logp_difference/max': 1.4660539627075195, 'sampling/importance_sampling_ratio/min': 0.23083457350730896, 'sampling/importance_sampling_ratio/mean': 1.0026134252548218, 'sampling/importance_sampling_ratio/max': 1.8658126592636108, 'entropy': 0.16349555738270283, 'clip_ratio/low_mean': 0.0030577474972233176, 'clip_ratio/low_min': 0.0030577474972233176, 'clip_ratio/high_mean': 0.019755184242967516, 'clip_ratio/high_max': 0.019755184242967516, 'clip_ratio/region_mean': 0.022812931740190834, 'reward_total_mean': 0.9727857112884521, 'reward_meter_mean': 0.9954801797866821, 'reward_meter_std': 0.008464212529361248, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9772727489471436, 'reward_repeat_penalty_std': 0.04208271950483322, 'reward_total_composite_mean': 0.9727857112884521, 'reward_total_composite_std': 0.04096178710460663, 'epoch': 0.1}

 78%|███████▊  | 2588/3300 [3:44:14<1:09:15,  5.84s/it]INFO 04-12 02:21:05 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2589/3300 [3:44:18<1:03:59,  5.40s/it]2026-04-12 02:21:09,552 | INFO | train_grpo_train | metrics_logged mode=train step=2589

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.157575757575758e-06, 'num_tokens': 5863619.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00021978352742735296, 'sampling/sampling_logp_difference/max': 0.0021914218086749315, 'sampling/importance_sampling_ratio/min': 0.9999104142189026, 'sampling/importance_sampling_ratio/mean': 1.0002191066741943, 'sampling/importance_sampling_ratio/max': 1.0021939277648926, 'entropy': 0.0016017362504499033, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 78%|███████▊  | 2589/3300 [3:44:18<1:03:59,  5.40s/it]INFO 04-12 02:21:09 [block_pool.py:378] Successfully reset prefix cache

 78%|███████▊  | 2590/3300 [3:44:23<1:01:34,  5.20s/it]2026-04-12 02:21:14,295 | INFO | train_grpo_train | metrics_logged mode=train step=2590

                                                       
{'loss': 0.0105, 'grad_norm': 6.6089911460876465, 'learning_rate': 2.1545454545454547e-06, 'num_tokens': 5865559.0, 'completions/mean_length': 70.5, 'completions/min_length': 68.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9882650971412659, 'rewards/meter/std': 0.010569445788860321, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9882650971412659, 'rewards/total_composite/std': 0.010569445788860321, 'reward': 0.9882650971412659, 'reward_std': 0.010569446720182896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05746662616729736, 'sampling/sampling_logp_difference/max': 2.042877435684204, 'sampling/importance_sampling_ratio/min': 0.12965509295463562, 'sampling/importance_sampling_ratio/mean': 1.0161855220794678, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.514446884393692, 'clip_ratio/low_mean': 0.013986697886139154, 'clip_ratio/low_min': 0.013986697886139154, 'clip_ratio/high_mean': 0.03218326787464321, 'clip_ratio/high_max': 0.03218326787464321, 'clip_ratio/region_mean': 0.04616996576078236, 'reward_total_mean': 0.9882650971412659, 'reward_meter_mean': 0.9882650971412659, 'reward_meter_std': 0.010569445788860321, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9882650971412659, 'reward_total_composite_std': 0.010569445788860321, 'epoch': 0.1}

 78%|███████▊  | 2590/3300 [3:44:23<1:01:34,  5.20s/it]INFO 04-12 02:21:14 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2591/3300 [3:44:27<57:34,  4.87s/it]  2026-04-12 02:21:18,403 | INFO | train_grpo_train | metrics_logged mode=train step=2591

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.1515151515151515e-06, 'num_tokens': 5866935.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003589413536246866, 'sampling/sampling_logp_difference/max': 0.0062316060066223145, 'sampling/importance_sampling_ratio/min': 0.9997910261154175, 'sampling/importance_sampling_ratio/mean': 1.0003573894500732, 'sampling/importance_sampling_ratio/max': 1.0062510967254639, 'entropy': 0.002733113680733368, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.1}

 79%|███████▊  | 2591/3300 [3:44:27<57:34,  4.87s/it]INFO 04-12 02:21:18 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2592/3300 [3:44:31<56:35,  4.80s/it]2026-04-12 02:21:23,017 | INFO | train_grpo_train | metrics_logged mode=train step=2592

                                                     
{'loss': 0.0079, 'grad_norm': 5.403855800628662, 'learning_rate': 2.148484848484849e-06, 'num_tokens': 5868615.0, 'completions/mean_length': 59.0, 'completions/min_length': 57.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.0, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9917755126953125, 'rewards/meter/std': 0.003381570801138878, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9917755126953125, 'rewards/total_composite/std': 0.003381570801138878, 'reward': 0.9917755126953125, 'reward_std': 0.0033815728966146708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03163475543260574, 'sampling/sampling_logp_difference/max': 1.120924949645996, 'sampling/importance_sampling_ratio/min': 0.3259781301021576, 'sampling/importance_sampling_ratio/mean': 1.003715991973877, 'sampling/importance_sampling_ratio/max': 1.8206579685211182, 'entropy': 0.2373839858919382, 'clip_ratio/low_mean': 0.014869472477585077, 'clip_ratio/low_min': 0.014869472477585077, 'clip_ratio/high_mean': 0.019220190355554223, 'clip_ratio/high_max': 0.019220190355554223, 'clip_ratio/region_mean': 0.0340896628331393, 'reward_total_mean': 0.9917755126953125, 'reward_meter_mean': 0.9917755126953125, 'reward_meter_std': 0.003381570801138878, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9917755126953125, 'reward_total_composite_std': 0.003381570801138878, 'epoch': 0.1}

 79%|███████▊  | 2592/3300 [3:44:32<56:35,  4.80s/it]INFO 04-12 02:21:23 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2593/3300 [3:44:38<1:03:18,  5.37s/it]2026-04-12 02:21:29,735 | INFO | train_grpo_train | metrics_logged mode=train step=2593

                                                       
{'loss': 0.0705, 'grad_norm': 3.471240282058716, 'learning_rate': 2.1454545454545456e-06, 'num_tokens': 5871770.0, 'completions/mean_length': 216.375, 'completions/min_length': 197.0, 'completions/max_length': 239.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 216.375, 'completions/min_terminated_length': 197.0, 'completions/max_terminated_length': 239.0, 'rewards/meter/mean': 0.990897536277771, 'rewards/meter/std': 0.006994582246989012, 'rewards/count_adherence/mean': 0.9464285969734192, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9820512533187866, 'rewards/repeat_penalty/std': 0.033347416669130325, 'rewards/total_composite/mean': 0.921344518661499, 'rewards/total_composite/std': 0.0833897739648819, 'reward': 0.921344518661499, 'reward_std': 0.0833897814154625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07251539081335068, 'sampling/sampling_logp_difference/max': 1.7865657806396484, 'sampling/importance_sampling_ratio/min': 0.16753454506397247, 'sampling/importance_sampling_ratio/mean': 1.0076571702957153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5770430602133274, 'clip_ratio/low_mean': 0.0256543830037117, 'clip_ratio/low_min': 0.0256543830037117, 'clip_ratio/high_mean': 0.03610773291438818, 'clip_ratio/high_max': 0.03610773291438818, 'clip_ratio/region_mean': 0.06176211591809988, 'reward_total_mean': 0.921344518661499, 'reward_meter_mean': 0.990897536277771, 'reward_meter_std': 0.006994582246989012, 'reward_count_adherence_mean': 0.9464285969734192, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9820512533187866, 'reward_repeat_penalty_std': 0.033347416669130325, 'reward_total_composite_mean': 0.921344518661499, 'reward_total_composite_std': 0.0833897739648819, 'epoch': 0.1}

 79%|███████▊  | 2593/3300 [3:44:38<1:03:18,  5.37s/it]INFO 04-12 02:21:29 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2594/3300 [3:44:43<1:00:31,  5.14s/it]2026-04-12 02:21:34,343 | INFO | train_grpo_train | metrics_logged mode=train step=2594

                                                       
{'loss': 0.0196, 'grad_norm': 5.423774719238281, 'learning_rate': 2.1424242424242425e-06, 'num_tokens': 5873489.0, 'completions/mean_length': 58.875, 'completions/min_length': 57.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.992070198059082, 'rewards/meter/std': 0.0043101259507238865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.992070198059082, 'rewards/total_composite/std': 0.0043101259507238865, 'reward': 0.992070198059082, 'reward_std': 0.004310120362788439, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03451910987496376, 'sampling/sampling_logp_difference/max': 1.3982751369476318, 'sampling/importance_sampling_ratio/min': 0.24702267348766327, 'sampling/importance_sampling_ratio/mean': 0.9979619383811951, 'sampling/importance_sampling_ratio/max': 1.3894490003585815, 'entropy': 0.27704715728759766, 'clip_ratio/low_mean': 0.010563877876847982, 'clip_ratio/low_min': 0.010563877876847982, 'clip_ratio/high_mean': 0.025721082463860512, 'clip_ratio/high_max': 0.025721082463860512, 'clip_ratio/region_mean': 0.036284960340708494, 'reward_total_mean': 0.992070198059082, 'reward_meter_mean': 0.992070198059082, 'reward_meter_std': 0.0043101259507238865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.992070198059082, 'reward_total_composite_std': 0.0043101259507238865, 'epoch': 0.1}

 79%|███████▊  | 2594/3300 [3:44:43<1:00:31,  5.14s/it]INFO 04-12 02:21:34 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2595/3300 [3:44:50<1:08:57,  5.87s/it]2026-04-12 02:21:41,909 | INFO | train_grpo_train | metrics_logged mode=train step=2595

                                                       
{'loss': -0.0043, 'grad_norm': 2.122164726257324, 'learning_rate': 2.1393939393939393e-06, 'num_tokens': 5877533.0, 'completions/mean_length': 299.5, 'completions/min_length': 293.0, 'completions/max_length': 307.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 299.5, 'completions/min_terminated_length': 293.0, 'completions/max_terminated_length': 307.0, 'rewards/meter/mean': 0.9690572023391724, 'rewards/meter/std': 0.0723068043589592, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.970588207244873, 'rewards/repeat_penalty/std': 0.03144249692559242, 'rewards/total_composite/mean': 0.8223901987075806, 'rewards/total_composite/std': 0.05969366803765297, 'reward': 0.8223901987075806, 'reward_std': 0.05969364941120148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04698567092418671, 'sampling/sampling_logp_difference/max': 1.767293930053711, 'sampling/importance_sampling_ratio/min': 0.1707945466041565, 'sampling/importance_sampling_ratio/mean': 1.0036664009094238, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3467197176069021, 'clip_ratio/low_mean': 0.0235387550201267, 'clip_ratio/low_min': 0.0235387550201267, 'clip_ratio/high_mean': 0.012042732443660498, 'clip_ratio/high_max': 0.012042732443660498, 'clip_ratio/region_mean': 0.0355814874637872, 'reward_total_mean': 0.8223901987075806, 'reward_meter_mean': 0.9690572023391724, 'reward_meter_std': 0.0723068043589592, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.970588207244873, 'reward_repeat_penalty_std': 0.03144249692559242, 'reward_total_composite_mean': 0.8223901987075806, 'reward_total_composite_std': 0.05969366803765297, 'epoch': 0.1}

 79%|███████▊  | 2595/3300 [3:44:50<1:08:57,  5.87s/it]INFO 04-12 02:21:42 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2596/3300 [3:45:01<1:24:05,  7.17s/it]2026-04-12 02:21:52,099 | INFO | train_grpo_train | metrics_logged mode=train step=2596

                                                       
{'loss': 0.112, 'grad_norm': 1.1440000534057617, 'learning_rate': 2.1363636363636365e-06, 'num_tokens': 5882790.0, 'completions/mean_length': 501.125, 'completions/min_length': 470.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 499.5714416503906, 'completions/min_terminated_length': 470.0, 'completions/max_terminated_length': 510.0, 'rewards/meter/mean': 0.9981234073638916, 'rewards/meter/std': 0.0008658914593979716, 'rewards/count_adherence/mean': 0.8046875, 'rewards/count_adherence/std': 0.022097086533904076, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9128260612487793, 'rewards/repeat_penalty/std': 0.08217112720012665, 'rewards/total_composite/mean': 0.7341695427894592, 'rewards/total_composite/std': 0.07871917635202408, 'reward': 0.7341695427894592, 'reward_std': 0.07871916890144348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048293329775333405, 'sampling/sampling_logp_difference/max': 1.6189045906066895, 'sampling/importance_sampling_ratio/min': 0.19811558723449707, 'sampling/importance_sampling_ratio/mean': 1.011509895324707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34851858019828796, 'clip_ratio/low_mean': 0.005782836233265698, 'clip_ratio/low_min': 0.005782836233265698, 'clip_ratio/high_mean': 0.01985564432106912, 'clip_ratio/high_max': 0.01985564432106912, 'clip_ratio/region_mean': 0.02563848055433482, 'reward_total_mean': 0.7341695427894592, 'reward_meter_mean': 0.9981234073638916, 'reward_meter_std': 0.0008658914593979716, 'reward_count_adherence_mean': 0.8046875, 'reward_count_adherence_std': 0.022097086533904076, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9128260612487793, 'reward_repeat_penalty_std': 0.08217112720012665, 'reward_total_composite_mean': 0.7341695427894592, 'reward_total_composite_std': 0.07871917635202408, 'epoch': 0.1}

 79%|███████▊  | 2596/3300 [3:45:01<1:24:05,  7.17s/it]INFO 04-12 02:21:52 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2597/3300 [3:45:06<1:17:44,  6.64s/it]2026-04-12 02:21:57,496 | INFO | train_grpo_train | metrics_logged mode=train step=2597

                                                       
{'loss': -0.0172, 'grad_norm': 3.617664098739624, 'learning_rate': 2.133333333333334e-06, 'num_tokens': 5884995.0, 'completions/mean_length': 122.625, 'completions/min_length': 113.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.625, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.9898524284362793, 'rewards/meter/std': 0.005112205166369677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8836444616317749, 'rewards/total_composite/std': 0.06303286552429199, 'reward': 0.8836444616317749, 'reward_std': 0.0630328580737114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034280113875865936, 'sampling/sampling_logp_difference/max': 1.1297893524169922, 'sampling/importance_sampling_ratio/min': 0.32310131192207336, 'sampling/importance_sampling_ratio/mean': 1.013502597808838, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3435132773593068, 'clip_ratio/low_mean': 0.01861410157289356, 'clip_ratio/low_min': 0.01861410157289356, 'clip_ratio/high_mean': 0.010750473011285067, 'clip_ratio/high_max': 0.010750473011285067, 'clip_ratio/region_mean': 0.029364574584178627, 'reward_total_mean': 0.8836444616317749, 'reward_meter_mean': 0.9898524284362793, 'reward_meter_std': 0.005112205166369677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8836444616317749, 'reward_total_composite_std': 0.06303286552429199, 'epoch': 0.1}

 79%|███████▊  | 2597/3300 [3:45:06<1:17:44,  6.64s/it]INFO 04-12 02:21:57 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▊  | 2598/3300 [3:45:10<1:08:58,  5.89s/it]2026-04-12 02:22:01,661 | INFO | train_grpo_train | metrics_logged mode=train step=2598

                                                       
{'loss': -0.0118, 'grad_norm': 7.509369850158691, 'learning_rate': 2.1303030303030306e-06, 'num_tokens': 5886546.0, 'completions/mean_length': 28.875, 'completions/min_length': 28.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9953750371932983, 'rewards/meter/std': 0.0010425866348668933, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953750371932983, 'rewards/total_composite/std': 0.0010425866348668933, 'reward': 0.9953750371932983, 'reward_std': 0.0010425925720483065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00547828758135438, 'sampling/sampling_logp_difference/max': 1.0384702682495117, 'sampling/importance_sampling_ratio/min': 0.35399580001831055, 'sampling/importance_sampling_ratio/mean': 0.9976253509521484, 'sampling/importance_sampling_ratio/max': 1.1243258714675903, 'entropy': 0.006841129012173042, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9953750371932983, 'reward_meter_mean': 0.9953750371932983, 'reward_meter_std': 0.0010425866348668933, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953750371932983, 'reward_total_composite_std': 0.0010425866348668933, 'epoch': 0.1}

 79%|███████▊  | 2598/3300 [3:45:10<1:08:58,  5.89s/it]INFO 04-12 02:22:01 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2599/3300 [3:45:15<1:05:55,  5.64s/it]2026-04-12 02:22:06,709 | INFO | train_grpo_train | metrics_logged mode=train step=2599

                                                       
{'loss': 0.0041, 'grad_norm': 1.446718692779541, 'learning_rate': 2.1272727272727275e-06, 'num_tokens': 5888682.0, 'completions/mean_length': 98.0, 'completions/min_length': 97.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9992596507072449, 'rewards/meter/std': 0.0001470481656724587, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992596507072449, 'rewards/total_composite/std': 0.0001470481656724587, 'reward': 0.9992596507072449, 'reward_std': 0.000147052516695112, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011022298596799374, 'sampling/sampling_logp_difference/max': 0.9700416326522827, 'sampling/importance_sampling_ratio/min': 0.3790672719478607, 'sampling/importance_sampling_ratio/mean': 0.9985703229904175, 'sampling/importance_sampling_ratio/max': 1.4146671295166016, 'entropy': 0.06834368547424674, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007614409434609115, 'clip_ratio/high_max': 0.007614409434609115, 'clip_ratio/region_mean': 0.007614409434609115, 'reward_total_mean': 0.9992596507072449, 'reward_meter_mean': 0.9992596507072449, 'reward_meter_std': 0.0001470481656724587, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992596507072449, 'reward_total_composite_std': 0.0001470481656724587, 'epoch': 0.1}

 79%|███████▉  | 2599/3300 [3:45:15<1:05:55,  5.64s/it]INFO 04-12 02:22:06 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2600/3300 [3:45:19<1:00:52,  5.22s/it]2026-04-12 02:22:10,941 | INFO | train_grpo_train | metrics_logged mode=train step=2600

                                                       
{'loss': 0.0094, 'grad_norm': 5.452374458312988, 'learning_rate': 2.1242424242424243e-06, 'num_tokens': 5890094.0, 'completions/mean_length': 34.5, 'completions/min_length': 34.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9899568557739258, 'rewards/meter/std': 0.008095722645521164, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9899568557739258, 'rewards/total_composite/std': 0.008095722645521164, 'reward': 0.9899568557739258, 'reward_std': 0.00809571985155344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02594204805791378, 'sampling/sampling_logp_difference/max': 1.6147489547729492, 'sampling/importance_sampling_ratio/min': 0.19894060492515564, 'sampling/importance_sampling_ratio/mean': 0.9940556287765503, 'sampling/importance_sampling_ratio/max': 1.2015459537506104, 'entropy': 0.1272981008514762, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007352941203862429, 'clip_ratio/high_max': 0.007352941203862429, 'clip_ratio/region_mean': 0.007352941203862429, 'reward_total_mean': 0.9899568557739258, 'reward_meter_mean': 0.9899568557739258, 'reward_meter_std': 0.008095722645521164, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9899568557739258, 'reward_total_composite_std': 0.008095722645521164, 'epoch': 0.1}

 79%|███████▉  | 2600/3300 [3:45:19<1:00:52,  5.22s/it]INFO 04-12 02:22:11 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:39,  3.56s/it]

 23%|██▎       | 3/13 [00:10<00:36,  3.64s/it]

 31%|███       | 4/13 [00:16<00:40,  4.52s/it]

 38%|███▊      | 5/13 [00:22<00:39,  4.92s/it]

 46%|████▌     | 6/13 [00:28<00:37,  5.34s/it]

 54%|█████▍    | 7/13 [00:34<00:32,  5.40s/it]

 62%|██████▏   | 8/13 [00:41<00:29,  5.86s/it]

 69%|██████▉   | 9/13 [00:48<00:25,  6.32s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.96s/it]

 85%|████████▍ | 11/13 [00:59<00:11,  5.95s/it]

 92%|█████████▏| 12/13 [01:05<00:05,  5.98s/it]

100%|██████████| 13/13 [01:10<00:00,  5.77s/it]2026-04-12 02:23:28,407 | INFO | train_grpo_train | metrics_logged mode=eval step=2600

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 77.3652, 'eval_samples_per_second': 1.344, 'eval_steps_per_second': 0.168, 'eval_num_tokens': 5890094.0, 'eval_completions/mean_length': 209.5096153846154, 'eval_completions/min_length': 60.92307692307692, 'eval_completions/max_length': 414.3076923076923, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 203.01236314039963, 'eval_completions/min_terminated_length': 60.92307692307692, 'eval_completions/max_terminated_length': 392.0, 'eval_rewards/meter/mean': 0.7689516819440402, 'eval_rewards/meter/std': 0.3567994758486748, 'eval_rewards/count_adherence/mean': 0.9526959611819341, 'eval_rewards/count_adherence/std': 0.0694849301989262, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.92968055835137, 'eval_rewards/repeat_penalty/std': 0.09778400797110337, 'eval_rewards/total_composite/mean': 0.6867790864064143, 'eval_rewards/total_composite/std': 0.3467414522400269, 'eval_reward': 0.6867790864064143, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03625713331768146, 'eval_sampling/sampling_logp_difference/max': 1.1111029111422026, 'eval_sampling/importance_sampling_ratio/min': 0.3349975932102937, 'eval_sampling/importance_sampling_ratio/mean': 1.0097874861497145, 'eval_sampling/importance_sampling_ratio/max': 1.5790529526196992, 'eval_entropy': 0.40257097207582915, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.6867790864064143, 'eval_reward_meter_mean': 0.7689516819440402, 'eval_reward_meter_std': 0.3567994758486748, 'eval_reward_count_adherence_mean': 0.9526959611819341, 'eval_reward_count_adherence_std': 0.0694849301989262, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.92968055835137, 'eval_reward_repeat_penalty_std': 0.09778400797110337, 'eval_reward_total_composite_mean': 0.6867790864064143, 'eval_reward_total_composite_std': 0.3467414522400269, 'epoch': 0.1}

 79%|███████▉  | 2600/3300 [3:46:37<1:00:52,  5.22s/it]

100%|██████████| 13/13 [01:10<00:00,  5.77s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 79%|███████▉  | 2601/3300 [3:46:44<5:37:44, 28.99s/it]2026-04-12 02:23:35,405 | INFO | train_grpo_train | metrics_logged mode=train step=2601

                                                       
{'loss': -0.0003, 'grad_norm': 1.6252681016921997, 'learning_rate': 2.1212121212121216e-06, 'num_tokens': 5891951.0, 'completions/mean_length': 72.125, 'completions/min_length': 72.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.125, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9991148710250854, 'rewards/meter/std': 0.00010878306056838483, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991148710250854, 'rewards/total_composite/std': 0.00010878306056838483, 'reward': 0.9991148710250854, 'reward_std': 0.00010877339082071558, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010965629480779171, 'sampling/sampling_logp_difference/max': 1.061844825744629, 'sampling/importance_sampling_ratio/min': 0.3458172678947449, 'sampling/importance_sampling_ratio/mean': 1.0020719766616821, 'sampling/importance_sampling_ratio/max': 1.4661310911178589, 'entropy': 0.06924279127269983, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0017123287543654442, 'clip_ratio/high_max': 0.0017123287543654442, 'clip_ratio/region_mean': 0.0017123287543654442, 'reward_total_mean': 0.9991148710250854, 'reward_meter_mean': 0.9991148710250854, 'reward_meter_std': 0.00010878306056838483, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991148710250854, 'reward_total_composite_std': 0.00010878306056838483, 'epoch': 0.1}

 79%|███████▉  | 2601/3300 [3:46:44<5:37:44, 28.99s/it]INFO 04-12 02:23:35 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2602/3300 [3:46:49<4:13:10, 21.76s/it]2026-04-12 02:23:40,301 | INFO | train_grpo_train | metrics_logged mode=train step=2602

                                                       
{'loss': -0.0044, 'grad_norm': 8.062291145324707, 'learning_rate': 2.1181818181818184e-06, 'num_tokens': 5893679.0, 'completions/mean_length': 68.0, 'completions/min_length': 65.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9672091007232666, 'rewards/meter/std': 0.05921924114227295, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9672091007232666, 'rewards/total_composite/std': 0.05921924114227295, 'reward': 0.9672091007232666, 'reward_std': 0.059219252318143845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03210901468992233, 'sampling/sampling_logp_difference/max': 0.8111929893493652, 'sampling/importance_sampling_ratio/min': 0.444327712059021, 'sampling/importance_sampling_ratio/mean': 1.0095655918121338, 'sampling/importance_sampling_ratio/max': 1.7304493188858032, 'entropy': 0.24536396749317646, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.020244444953277707, 'clip_ratio/high_max': 0.020244444953277707, 'clip_ratio/region_mean': 0.022138384403660893, 'reward_total_mean': 0.9672091007232666, 'reward_meter_mean': 0.9672091007232666, 'reward_meter_std': 0.05921924114227295, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9672091007232666, 'reward_total_composite_std': 0.05921924114227295, 'epoch': 0.1}

 79%|███████▉  | 2602/3300 [3:46:49<4:13:10, 21.76s/it]INFO 04-12 02:23:40 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2603/3300 [3:46:54<3:13:42, 16.67s/it]2026-04-12 02:23:45,112 | INFO | train_grpo_train | metrics_logged mode=train step=2603

                                                       
{'loss': -0.002, 'grad_norm': 0.387520432472229, 'learning_rate': 2.1151515151515152e-06, 'num_tokens': 5895490.0, 'completions/mean_length': 72.375, 'completions/min_length': 72.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9991236925125122, 'rewards/meter/std': 0.00013291009236127138, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991236925125122, 'rewards/total_composite/std': 0.00013291009236127138, 'reward': 0.9991236925125122, 'reward_std': 0.00013291911454871297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00656279968097806, 'sampling/sampling_logp_difference/max': 0.30100059509277344, 'sampling/importance_sampling_ratio/min': 0.8066705465316772, 'sampling/importance_sampling_ratio/mean': 1.0055519342422485, 'sampling/importance_sampling_ratio/max': 1.3512102365493774, 'entropy': 0.05477497586980462, 'clip_ratio/low_mean': 0.0034246575087308884, 'clip_ratio/low_min': 0.0034246575087308884, 'clip_ratio/high_mean': 0.0017123287543654442, 'clip_ratio/high_max': 0.0017123287543654442, 'clip_ratio/region_mean': 0.0051369862630963326, 'reward_total_mean': 0.9991236925125122, 'reward_meter_mean': 0.9991236925125122, 'reward_meter_std': 0.00013291009236127138, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991236925125122, 'reward_total_composite_std': 0.00013291009236127138, 'epoch': 0.1}

 79%|███████▉  | 2603/3300 [3:46:54<3:13:42, 16.67s/it]INFO 04-12 02:23:45 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2604/3300 [3:46:58<2:31:49, 13.09s/it]2026-04-12 02:23:49,831 | INFO | train_grpo_train | metrics_logged mode=train step=2604

                                                       
{'loss': 0.0218, 'grad_norm': 4.805464267730713, 'learning_rate': 2.1121212121212125e-06, 'num_tokens': 5897338.0, 'completions/mean_length': 68.0, 'completions/min_length': 67.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9956408739089966, 'rewards/meter/std': 0.004533765371888876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9956408739089966, 'rewards/total_composite/std': 0.004533765371888876, 'reward': 0.9956408739089966, 'reward_std': 0.0045337737537920475, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015114509500563145, 'sampling/sampling_logp_difference/max': 1.2838521003723145, 'sampling/importance_sampling_ratio/min': 0.2769683301448822, 'sampling/importance_sampling_ratio/mean': 1.0015579462051392, 'sampling/importance_sampling_ratio/max': 1.9683966636657715, 'entropy': 0.07368437433615327, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.007380377617664635, 'clip_ratio/high_max': 0.007380377617664635, 'clip_ratio/region_mean': 0.010852599865756929, 'reward_total_mean': 0.9956408739089966, 'reward_meter_mean': 0.9956408739089966, 'reward_meter_std': 0.004533765371888876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9956408739089966, 'reward_total_composite_std': 0.004533765371888876, 'epoch': 0.1}

 79%|███████▉  | 2604/3300 [3:46:58<2:31:49, 13.09s/it]INFO 04-12 02:23:50 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2605/3300 [3:47:03<2:02:13, 10.55s/it]2026-04-12 02:23:54,454 | INFO | train_grpo_train | metrics_logged mode=train step=2605

                                                       
{'loss': -0.0034, 'grad_norm': 3.96225905418396, 'learning_rate': 2.1090909090909093e-06, 'num_tokens': 5899077.0, 'completions/mean_length': 59.375, 'completions/min_length': 56.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9955159425735474, 'rewards/meter/std': 0.0011379423085600138, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955159425735474, 'rewards/total_composite/std': 0.0011379423085600138, 'reward': 0.9955159425735474, 'reward_std': 0.001137925311923027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022009698674082756, 'sampling/sampling_logp_difference/max': 0.7918639183044434, 'sampling/importance_sampling_ratio/min': 0.45299965143203735, 'sampling/importance_sampling_ratio/mean': 1.0065569877624512, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18665459845215082, 'clip_ratio/low_mean': 0.00836864416487515, 'clip_ratio/low_min': 0.00836864416487515, 'clip_ratio/high_mean': 0.018881317228078842, 'clip_ratio/high_max': 0.018881317228078842, 'clip_ratio/region_mean': 0.027249961392953992, 'reward_total_mean': 0.9955159425735474, 'reward_meter_mean': 0.9955159425735474, 'reward_meter_std': 0.0011379423085600138, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955159425735474, 'reward_total_composite_std': 0.0011379423085600138, 'epoch': 0.1}

 79%|███████▉  | 2605/3300 [3:47:03<2:02:13, 10.55s/it]INFO 04-12 02:23:54 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2606/3300 [3:47:08<1:44:09,  9.01s/it]2026-04-12 02:23:59,854 | INFO | train_grpo_train | metrics_logged mode=train step=2606

                                                       
{'loss': 0.015, 'grad_norm': 3.778212070465088, 'learning_rate': 2.106060606060606e-06, 'num_tokens': 5901663.0, 'completions/mean_length': 137.25, 'completions/min_length': 135.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.25, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.997505784034729, 'rewards/meter/std': 0.002905552973970771, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997505784034729, 'rewards/total_composite/std': 0.002905552973970771, 'reward': 0.997505784034729, 'reward_std': 0.0029055566992610693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05034590885043144, 'sampling/sampling_logp_difference/max': 1.172276496887207, 'sampling/importance_sampling_ratio/min': 0.30966120958328247, 'sampling/importance_sampling_ratio/mean': 1.0051594972610474, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44699474796652794, 'clip_ratio/low_mean': 0.010766045656055212, 'clip_ratio/low_min': 0.010766045656055212, 'clip_ratio/high_mean': 0.04019012441858649, 'clip_ratio/high_max': 0.04019012441858649, 'clip_ratio/region_mean': 0.050956170074641705, 'reward_total_mean': 0.997505784034729, 'reward_meter_mean': 0.997505784034729, 'reward_meter_std': 0.002905552973970771, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997505784034729, 'reward_total_composite_std': 0.002905552973970771, 'epoch': 0.1}

 79%|███████▉  | 2606/3300 [3:47:08<1:44:09,  9.01s/it]INFO 04-12 02:24:00 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2607/3300 [3:47:14<1:34:05,  8.15s/it]2026-04-12 02:24:05,994 | INFO | train_grpo_train | metrics_logged mode=train step=2607

                                                       
{'loss': 0.0094, 'grad_norm': 2.1808922290802, 'learning_rate': 2.103030303030303e-06, 'num_tokens': 5904846.0, 'completions/mean_length': 195.875, 'completions/min_length': 190.0, 'completions/max_length': 199.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 195.875, 'completions/min_terminated_length': 190.0, 'completions/max_terminated_length': 199.0, 'rewards/meter/mean': 0.9988412261009216, 'rewards/meter/std': 0.0004946960834786296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9711034893989563, 'rewards/total_composite/std': 0.05153718218207359, 'reward': 0.9711034893989563, 'reward_std': 0.05153718218207359, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04786364734172821, 'sampling/sampling_logp_difference/max': 1.6877174377441406, 'sampling/importance_sampling_ratio/min': 0.18494117259979248, 'sampling/importance_sampling_ratio/mean': 1.0106346607208252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3686945028603077, 'clip_ratio/low_mean': 0.0025254478096030653, 'clip_ratio/low_min': 0.0025254478096030653, 'clip_ratio/high_mean': 0.023096036864444613, 'clip_ratio/high_max': 0.023096036864444613, 'clip_ratio/region_mean': 0.02562148467404768, 'reward_total_mean': 0.9711034893989563, 'reward_meter_mean': 0.9988412261009216, 'reward_meter_std': 0.0004946960834786296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9711034893989563, 'reward_total_composite_std': 0.05153718218207359, 'epoch': 0.1}

 79%|███████▉  | 2607/3300 [3:47:14<1:34:05,  8.15s/it]INFO 04-12 02:24:06 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2608/3300 [3:47:20<1:23:31,  7.24s/it]2026-04-12 02:24:11,128 | INFO | train_grpo_train | metrics_logged mode=train step=2608

                                                       
{'loss': -0.0114, 'grad_norm': 2.9761064052581787, 'learning_rate': 2.1000000000000002e-06, 'num_tokens': 5907083.0, 'completions/mean_length': 118.625, 'completions/min_length': 115.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.625, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9947597980499268, 'rewards/meter/std': 0.010325920768082142, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9947597980499268, 'rewards/total_composite/std': 0.010325920768082142, 'reward': 0.9947597980499268, 'reward_std': 0.010325928218662739, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03754591941833496, 'sampling/sampling_logp_difference/max': 1.3222484588623047, 'sampling/importance_sampling_ratio/min': 0.26653534173965454, 'sampling/importance_sampling_ratio/mean': 1.0031919479370117, 'sampling/importance_sampling_ratio/max': 1.8022807836532593, 'entropy': 0.283309917896986, 'clip_ratio/low_mean': 0.0021739129442721605, 'clip_ratio/low_min': 0.0021739129442721605, 'clip_ratio/high_mean': 0.020967748598195612, 'clip_ratio/high_max': 0.020967748598195612, 'clip_ratio/region_mean': 0.023141661542467773, 'reward_total_mean': 0.9947597980499268, 'reward_meter_mean': 0.9947597980499268, 'reward_meter_std': 0.010325920768082142, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9947597980499268, 'reward_total_composite_std': 0.010325920768082142, 'epoch': 0.1}

 79%|███████▉  | 2608/3300 [3:47:20<1:23:31,  7.24s/it]INFO 04-12 02:24:11 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2609/3300 [3:47:24<1:14:01,  6.43s/it]2026-04-12 02:24:15,654 | INFO | train_grpo_train | metrics_logged mode=train step=2609

                                                       
{'loss': 0.0134, 'grad_norm': 5.4968791007995605, 'learning_rate': 2.096969696969697e-06, 'num_tokens': 5908961.0, 'completions/mean_length': 68.75, 'completions/min_length': 66.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9976909756660461, 'rewards/meter/std': 0.003610314568504691, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976909756660461, 'rewards/total_composite/std': 0.003610314568504691, 'reward': 0.9976909756660461, 'reward_std': 0.00361032597720623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03931521996855736, 'sampling/sampling_logp_difference/max': 1.6529312133789062, 'sampling/importance_sampling_ratio/min': 0.19148778915405273, 'sampling/importance_sampling_ratio/mean': 1.0021377801895142, 'sampling/importance_sampling_ratio/max': 1.6728876829147339, 'entropy': 0.3395673893392086, 'clip_ratio/low_mean': 0.008980331476777792, 'clip_ratio/low_min': 0.008980331476777792, 'clip_ratio/high_mean': 0.029262988013215363, 'clip_ratio/high_max': 0.029262988013215363, 'clip_ratio/region_mean': 0.038243319489993155, 'reward_total_mean': 0.9976909756660461, 'reward_meter_mean': 0.9976909756660461, 'reward_meter_std': 0.003610314568504691, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976909756660461, 'reward_total_composite_std': 0.003610314568504691, 'epoch': 0.1}

 79%|███████▉  | 2609/3300 [3:47:24<1:14:01,  6.43s/it]INFO 04-12 02:24:15 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2610/3300 [3:47:30<1:12:55,  6.34s/it]2026-04-12 02:24:21,797 | INFO | train_grpo_train | metrics_logged mode=train step=2610

                                                       
{'loss': -0.01, 'grad_norm': 2.755105495452881, 'learning_rate': 2.093939393939394e-06, 'num_tokens': 5911873.0, 'completions/mean_length': 191.0, 'completions/min_length': 170.0, 'completions/max_length': 205.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 191.0, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.9434575438499451, 'rewards/meter/std': 0.07025814801454544, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9431818127632141, 'rewards/repeat_penalty/std': 0.08328413218259811, 'rewards/total_composite/mean': 0.832134485244751, 'rewards/total_composite/std': 0.10688581317663193, 'reward': 0.832134485244751, 'reward_std': 0.10688581317663193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.057438045740127563, 'sampling/sampling_logp_difference/max': 2.1229028701782227, 'sampling/importance_sampling_ratio/min': 0.11968369781970978, 'sampling/importance_sampling_ratio/mean': 1.007143259048462, 'sampling/importance_sampling_ratio/max': 1.9386733770370483, 'entropy': 0.49758507683873177, 'clip_ratio/low_mean': 0.021181484800763428, 'clip_ratio/low_min': 0.021181484800763428, 'clip_ratio/high_mean': 0.013105392456054688, 'clip_ratio/high_max': 0.013105392456054688, 'clip_ratio/region_mean': 0.034286877256818116, 'reward_total_mean': 0.832134485244751, 'reward_meter_mean': 0.9434575438499451, 'reward_meter_std': 0.07025814801454544, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625820279121399, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9431818127632141, 'reward_repeat_penalty_std': 0.08328413218259811, 'reward_total_composite_mean': 0.832134485244751, 'reward_total_composite_std': 0.10688581317663193, 'epoch': 0.1}

 79%|███████▉  | 2610/3300 [3:47:30<1:12:55,  6.34s/it]INFO 04-12 02:24:22 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2611/3300 [3:47:37<1:14:09,  6.46s/it]2026-04-12 02:24:28,522 | INFO | train_grpo_train | metrics_logged mode=train step=2611

                                                       
{'loss': 0.005, 'grad_norm': 1.116234540939331, 'learning_rate': 2.090909090909091e-06, 'num_tokens': 5915555.0, 'completions/mean_length': 247.25, 'completions/min_length': 245.0, 'completions/max_length': 249.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 247.25, 'completions/min_terminated_length': 245.0, 'completions/max_terminated_length': 249.0, 'rewards/meter/mean': 0.9987331032752991, 'rewards/meter/std': 0.00010533389286138117, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7788461446762085, 'rewards/repeat_penalty/std': 0.04929768666625023, 'rewards/total_composite/mean': 0.7778571844100952, 'rewards/total_composite/std': 0.04919851943850517, 'reward': 0.7778571844100952, 'reward_std': 0.04919853433966637, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01725967787206173, 'sampling/sampling_logp_difference/max': 1.344933032989502, 'sampling/importance_sampling_ratio/min': 0.2605571746826172, 'sampling/importance_sampling_ratio/mean': 1.0057003498077393, 'sampling/importance_sampling_ratio/max': 1.6523418426513672, 'entropy': 0.13410179316997528, 'clip_ratio/low_mean': 0.007064696401357651, 'clip_ratio/low_min': 0.007064696401357651, 'clip_ratio/high_mean': 0.00305292836856097, 'clip_ratio/high_max': 0.00305292836856097, 'clip_ratio/region_mean': 0.01011762476991862, 'reward_total_mean': 0.7778571844100952, 'reward_meter_mean': 0.9987331032752991, 'reward_meter_std': 0.00010533389286138117, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7788461446762085, 'reward_repeat_penalty_std': 0.04929768666625023, 'reward_total_composite_mean': 0.7778571844100952, 'reward_total_composite_std': 0.04919851943850517, 'epoch': 0.1}

 79%|███████▉  | 2611/3300 [3:47:37<1:14:09,  6.46s/it]INFO 04-12 02:24:28 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2612/3300 [3:47:41<1:07:23,  5.88s/it]2026-04-12 02:24:33,044 | INFO | train_grpo_train | metrics_logged mode=train step=2612

                                                       
{'loss': -0.02, 'grad_norm': 3.173931121826172, 'learning_rate': 2.087878787878788e-06, 'num_tokens': 5917441.0, 'completions/mean_length': 57.75, 'completions/min_length': 56.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.995871365070343, 'rewards/meter/std': 0.0009862068109214306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.995871365070343, 'rewards/total_composite/std': 0.0009862068109214306, 'reward': 0.995871365070343, 'reward_std': 0.0009862068109214306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01807764358818531, 'sampling/sampling_logp_difference/max': 0.9005880355834961, 'sampling/importance_sampling_ratio/min': 0.4063306748867035, 'sampling/importance_sampling_ratio/mean': 1.0038785934448242, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08830610243603587, 'clip_ratio/low_mean': 0.011121554300189018, 'clip_ratio/low_min': 0.011121554300189018, 'clip_ratio/high_mean': 0.004204352619126439, 'clip_ratio/high_max': 0.004204352619126439, 'clip_ratio/region_mean': 0.015325906919315457, 'reward_total_mean': 0.995871365070343, 'reward_meter_mean': 0.995871365070343, 'reward_meter_std': 0.0009862068109214306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.995871365070343, 'reward_total_composite_std': 0.0009862068109214306, 'epoch': 0.1}

 79%|███████▉  | 2612/3300 [3:47:42<1:07:23,  5.88s/it]INFO 04-12 02:24:33 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2613/3300 [3:47:49<1:14:44,  6.53s/it]2026-04-12 02:24:41,089 | INFO | train_grpo_train | metrics_logged mode=train step=2613

                                                       
{'loss': 0.0013, 'grad_norm': 2.3518784046173096, 'learning_rate': 2.0848484848484852e-06, 'num_tokens': 5922023.0, 'completions/mean_length': 343.75, 'completions/min_length': 326.0, 'completions/max_length': 360.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 343.75, 'completions/min_terminated_length': 326.0, 'completions/max_terminated_length': 360.0, 'rewards/meter/mean': 0.990763247013092, 'rewards/meter/std': 0.004047461785376072, 'rewards/count_adherence/mean': 0.759615421295166, 'rewards/count_adherence/std': 0.027196412906050682, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9213085174560547, 'rewards/repeat_penalty/std': 0.04907441511750221, 'rewards/total_composite/mean': 0.6937090158462524, 'rewards/total_composite/std': 0.049949031323194504, 'reward': 0.6937090158462524, 'reward_std': 0.04994901269674301, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05348372459411621, 'sampling/sampling_logp_difference/max': 1.539764404296875, 'sampling/importance_sampling_ratio/min': 0.21443161368370056, 'sampling/importance_sampling_ratio/mean': 1.0127009153366089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5114922970533371, 'clip_ratio/low_mean': 0.01867919461801648, 'clip_ratio/low_min': 0.01867919461801648, 'clip_ratio/high_mean': 0.021714599570259452, 'clip_ratio/high_max': 0.021714599570259452, 'clip_ratio/region_mean': 0.04039379418827593, 'reward_total_mean': 0.6937090158462524, 'reward_meter_mean': 0.990763247013092, 'reward_meter_std': 0.004047461785376072, 'reward_count_adherence_mean': 0.759615421295166, 'reward_count_adherence_std': 0.027196412906050682, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9213085174560547, 'reward_repeat_penalty_std': 0.04907441511750221, 'reward_total_composite_mean': 0.6937090158462524, 'reward_total_composite_std': 0.049949031323194504, 'epoch': 0.1}

 79%|███████▉  | 2613/3300 [3:47:50<1:14:44,  6.53s/it]INFO 04-12 02:24:41 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2614/3300 [3:47:55<1:09:51,  6.11s/it]2026-04-12 02:24:46,227 | INFO | train_grpo_train | metrics_logged mode=train step=2614

                                                       
{'loss': 0.0036, 'grad_norm': 4.099353790283203, 'learning_rate': 2.081818181818182e-06, 'num_tokens': 5924333.0, 'completions/mean_length': 118.75, 'completions/min_length': 116.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.75, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9893277883529663, 'rewards/meter/std': 0.016893420368433, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.953779935836792, 'rewards/total_composite/std': 0.06425921618938446, 'reward': 0.953779935836792, 'reward_std': 0.06425920873880386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028960946947336197, 'sampling/sampling_logp_difference/max': 1.3063240051269531, 'sampling/importance_sampling_ratio/min': 0.2708137333393097, 'sampling/importance_sampling_ratio/mean': 1.0034631490707397, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20165002904832363, 'clip_ratio/low_mean': 0.009516695979982615, 'clip_ratio/low_min': 0.009516695979982615, 'clip_ratio/high_mean': 0.02407875331118703, 'clip_ratio/high_max': 0.02407875331118703, 'clip_ratio/region_mean': 0.03359544929116964, 'reward_total_mean': 0.953779935836792, 'reward_meter_mean': 0.9893277883529663, 'reward_meter_std': 0.016893420368433, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.953779935836792, 'reward_total_composite_std': 0.06425921618938446, 'epoch': 0.1}

 79%|███████▉  | 2614/3300 [3:47:55<1:09:51,  6.11s/it]INFO 04-12 02:24:46 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2615/3300 [3:47:59<1:04:45,  5.67s/it]2026-04-12 02:24:50,876 | INFO | train_grpo_train | metrics_logged mode=train step=2615

                                                       
{'loss': -0.0063, 'grad_norm': 2.5716392993927, 'learning_rate': 2.078787878787879e-06, 'num_tokens': 5926114.0, 'completions/mean_length': 69.625, 'completions/min_length': 67.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9630937576293945, 'rewards/meter/std': 0.028052091598510742, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9630937576293945, 'rewards/total_composite/std': 0.028052091598510742, 'reward': 0.9630937576293945, 'reward_std': 0.028052086010575294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04609367623925209, 'sampling/sampling_logp_difference/max': 1.8025195598602295, 'sampling/importance_sampling_ratio/min': 0.16488294303417206, 'sampling/importance_sampling_ratio/mean': 1.0128580331802368, 'sampling/importance_sampling_ratio/max': 1.9397474527359009, 'entropy': 0.3094298355281353, 'clip_ratio/low_mean': 0.010907472460530698, 'clip_ratio/low_min': 0.010907472460530698, 'clip_ratio/high_mean': 0.015257316990755498, 'clip_ratio/high_max': 0.015257316990755498, 'clip_ratio/region_mean': 0.026164789451286197, 'reward_total_mean': 0.9630937576293945, 'reward_meter_mean': 0.9630937576293945, 'reward_meter_std': 0.028052091598510742, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9630937576293945, 'reward_total_composite_std': 0.028052091598510742, 'epoch': 0.11}

 79%|███████▉  | 2615/3300 [3:47:59<1:04:45,  5.67s/it]INFO 04-12 02:24:51 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2616/3300 [3:48:05<1:04:14,  5.64s/it]2026-04-12 02:24:56,425 | INFO | train_grpo_train | metrics_logged mode=train step=2616

                                                       
{'loss': -0.0043, 'grad_norm': 1.3603968620300293, 'learning_rate': 2.075757575757576e-06, 'num_tokens': 5928878.0, 'completions/mean_length': 159.5, 'completions/min_length': 156.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.5, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9918752908706665, 'rewards/meter/std': 0.01707914099097252, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.07120776921510696, 'rewards/total_composite/mean': 0.8684781789779663, 'rewards/total_composite/std': 0.07916970551013947, 'reward': 0.8684781789779663, 'reward_std': 0.07916972041130066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00969278160482645, 'sampling/sampling_logp_difference/max': 0.5411806106567383, 'sampling/importance_sampling_ratio/min': 0.582060694694519, 'sampling/importance_sampling_ratio/mean': 1.0036389827728271, 'sampling/importance_sampling_ratio/max': 1.4536888599395752, 'entropy': 0.09610292688012123, 'clip_ratio/low_mean': 0.0007812500116415322, 'clip_ratio/low_min': 0.0007812500116415322, 'clip_ratio/high_mean': 0.003906250058207661, 'clip_ratio/high_max': 0.003906250058207661, 'clip_ratio/region_mean': 0.004687500069849193, 'reward_total_mean': 0.8684781789779663, 'reward_meter_mean': 0.9918752908706665, 'reward_meter_std': 0.01707914099097252, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.07120776921510696, 'reward_total_composite_mean': 0.8684781789779663, 'reward_total_composite_std': 0.07916970551013947, 'epoch': 0.11}

 79%|███████▉  | 2616/3300 [3:48:05<1:04:14,  5.64s/it]INFO 04-12 02:24:56 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2617/3300 [3:48:10<1:02:54,  5.53s/it]2026-04-12 02:25:01,705 | INFO | train_grpo_train | metrics_logged mode=train step=2617

                                                       
{'loss': -0.0021, 'grad_norm': 1.5142834186553955, 'learning_rate': 2.072727272727273e-06, 'num_tokens': 5931264.0, 'completions/mean_length': 127.25, 'completions/min_length': 124.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.25, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9976711869239807, 'rewards/meter/std': 0.0005347781116142869, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9442130923271179, 'rewards/total_composite/std': 0.07360443472862244, 'reward': 0.9442130923271179, 'reward_std': 0.07360443472862244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011169607751071453, 'sampling/sampling_logp_difference/max': 0.7878756523132324, 'sampling/importance_sampling_ratio/min': 0.4548099637031555, 'sampling/importance_sampling_ratio/mean': 1.001734733581543, 'sampling/importance_sampling_ratio/max': 1.5947275161743164, 'entropy': 0.11067105457186699, 'clip_ratio/low_mean': 0.002922236453741789, 'clip_ratio/low_min': 0.002922236453741789, 'clip_ratio/high_mean': 0.00300038093701005, 'clip_ratio/high_max': 0.00300038093701005, 'clip_ratio/region_mean': 0.005922617390751839, 'reward_total_mean': 0.9442130923271179, 'reward_meter_mean': 0.9976711869239807, 'reward_meter_std': 0.0005347781116142869, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9442130923271179, 'reward_total_composite_std': 0.07360443472862244, 'epoch': 0.11}

 79%|███████▉  | 2617/3300 [3:48:10<1:02:54,  5.53s/it]INFO 04-12 02:25:01 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2618/3300 [3:48:14<58:55,  5.18s/it]  2026-04-12 02:25:06,093 | INFO | train_grpo_train | metrics_logged mode=train step=2618

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.06969696969697e-06, 'num_tokens': 5932856.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00027682288782671094, 'sampling/sampling_logp_difference/max': 0.002793335122987628, 'sampling/importance_sampling_ratio/min': 1.0, 'sampling/importance_sampling_ratio/mean': 1.000277042388916, 'sampling/importance_sampling_ratio/max': 1.002797245979309, 'entropy': 0.002262885362142697, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 79%|███████▉  | 2618/3300 [3:48:15<58:55,  5.18s/it]INFO 04-12 02:25:06 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2619/3300 [3:48:19<58:09,  5.12s/it]2026-04-12 02:25:11,072 | INFO | train_grpo_train | metrics_logged mode=train step=2619

                                                     
{'loss': -0.01, 'grad_norm': 3.8841686248779297, 'learning_rate': 2.0666666666666666e-06, 'num_tokens': 5935062.0, 'completions/mean_length': 90.75, 'completions/min_length': 87.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.75, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9945749640464783, 'rewards/meter/std': 0.00539820222184062, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9945749640464783, 'rewards/total_composite/std': 0.00539820222184062, 'reward': 0.9945749640464783, 'reward_std': 0.005398200359195471, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038317468017339706, 'sampling/sampling_logp_difference/max': 2.6888628005981445, 'sampling/importance_sampling_ratio/min': 0.06795818358659744, 'sampling/importance_sampling_ratio/mean': 1.0023733377456665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22051831148564816, 'clip_ratio/low_mean': 0.011174242943525314, 'clip_ratio/low_min': 0.011174242943525314, 'clip_ratio/high_mean': 0.030287093366496265, 'clip_ratio/high_max': 0.030287093366496265, 'clip_ratio/region_mean': 0.04146133631002158, 'reward_total_mean': 0.9945749640464783, 'reward_meter_mean': 0.9945749640464783, 'reward_meter_std': 0.00539820222184062, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9945749640464783, 'reward_total_composite_std': 0.00539820222184062, 'epoch': 0.11}

 79%|███████▉  | 2619/3300 [3:48:20<58:09,  5.12s/it]INFO 04-12 02:25:11 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2620/3300 [3:48:27<1:05:50,  5.81s/it]2026-04-12 02:25:18,482 | INFO | train_grpo_train | metrics_logged mode=train step=2620

                                                       
{'loss': 0.0003, 'grad_norm': 1.9527921676635742, 'learning_rate': 2.063636363636364e-06, 'num_tokens': 5939163.0, 'completions/mean_length': 298.625, 'completions/min_length': 293.0, 'completions/max_length': 304.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 298.625, 'completions/min_terminated_length': 293.0, 'completions/max_terminated_length': 304.0, 'rewards/meter/mean': 0.9977849721908569, 'rewards/meter/std': 0.0011817796621471643, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9632352590560913, 'rewards/repeat_penalty/std': 0.04376610368490219, 'rewards/total_composite/mean': 0.9610700011253357, 'rewards/total_composite/std': 0.042887065559625626, 'reward': 0.9610700011253357, 'reward_std': 0.042887069284915924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0335632786154747, 'sampling/sampling_logp_difference/max': 2.3920373916625977, 'sampling/importance_sampling_ratio/min': 0.09144318848848343, 'sampling/importance_sampling_ratio/mean': 1.0045593976974487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2537948340177536, 'clip_ratio/low_mean': 0.009193609352223575, 'clip_ratio/low_min': 0.009193609352223575, 'clip_ratio/high_mean': 0.016332621686160564, 'clip_ratio/high_max': 0.016332621686160564, 'clip_ratio/region_mean': 0.02552623103838414, 'reward_total_mean': 0.9610700011253357, 'reward_meter_mean': 0.9977849721908569, 'reward_meter_std': 0.0011817796621471643, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9632352590560913, 'reward_repeat_penalty_std': 0.04376610368490219, 'reward_total_composite_mean': 0.9610700011253357, 'reward_total_composite_std': 0.042887065559625626, 'epoch': 0.11}

 79%|███████▉  | 2620/3300 [3:48:27<1:05:50,  5.81s/it]INFO 04-12 02:25:18 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2621/3300 [3:48:35<1:13:49,  6.52s/it]2026-04-12 02:25:26,677 | INFO | train_grpo_train | metrics_logged mode=train step=2621

                                                       
{'loss': 0.0049, 'grad_norm': 1.2172192335128784, 'learning_rate': 2.0606060606060607e-06, 'num_tokens': 5943660.0, 'completions/mean_length': 349.125, 'completions/min_length': 319.0, 'completions/max_length': 355.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 349.125, 'completions/min_terminated_length': 319.0, 'completions/max_terminated_length': 355.0, 'rewards/meter/mean': 0.9986586570739746, 'rewards/meter/std': 0.0001423863577656448, 'rewards/count_adherence/mean': 0.987500011920929, 'rewards/count_adherence/std': 0.0353553481400013, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8010836243629456, 'rewards/repeat_penalty/std': 0.058759476989507675, 'rewards/total_composite/mean': 0.7889921069145203, 'rewards/total_composite/std': 0.04864564165472984, 'reward': 0.7889921069145203, 'reward_std': 0.04864564538002014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02395707368850708, 'sampling/sampling_logp_difference/max': 1.3942632675170898, 'sampling/importance_sampling_ratio/min': 0.24801570177078247, 'sampling/importance_sampling_ratio/mean': 1.0064207315444946, 'sampling/importance_sampling_ratio/max': 1.901794672012329, 'entropy': 0.18446156568825245, 'clip_ratio/low_mean': 0.006353942735586315, 'clip_ratio/low_min': 0.006353942735586315, 'clip_ratio/high_mean': 0.004175611422397196, 'clip_ratio/high_max': 0.004175611422397196, 'clip_ratio/region_mean': 0.010529554157983512, 'reward_total_mean': 0.7889921069145203, 'reward_meter_mean': 0.9986586570739746, 'reward_meter_std': 0.0001423863577656448, 'reward_count_adherence_mean': 0.987500011920929, 'reward_count_adherence_std': 0.0353553481400013, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8010836243629456, 'reward_repeat_penalty_std': 0.058759476989507675, 'reward_total_composite_mean': 0.7889921069145203, 'reward_total_composite_std': 0.04864564165472984, 'epoch': 0.11}

 79%|███████▉  | 2621/3300 [3:48:35<1:13:49,  6.52s/it]INFO 04-12 02:25:26 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2622/3300 [3:48:40<1:07:22,  5.96s/it]2026-04-12 02:25:31,330 | INFO | train_grpo_train | metrics_logged mode=train step=2622

                                                       
{'loss': -0.0037, 'grad_norm': 5.598334789276123, 'learning_rate': 2.0575757575757576e-06, 'num_tokens': 5945475.0, 'completions/mean_length': 59.875, 'completions/min_length': 58.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.875, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9960442185401917, 'rewards/meter/std': 0.0010061725042760372, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9960442185401917, 'rewards/total_composite/std': 0.0010061725042760372, 'reward': 0.9960442185401917, 'reward_std': 0.0010061829816550016, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036090608686208725, 'sampling/sampling_logp_difference/max': 1.571319580078125, 'sampling/importance_sampling_ratio/min': 0.20777083933353424, 'sampling/importance_sampling_ratio/mean': 0.9990103244781494, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1663585538044572, 'clip_ratio/low_mean': 0.006287686061114073, 'clip_ratio/low_min': 0.006287686061114073, 'clip_ratio/high_mean': 0.016635986510664225, 'clip_ratio/high_max': 0.016635986510664225, 'clip_ratio/region_mean': 0.022923672571778297, 'reward_total_mean': 0.9960442185401917, 'reward_meter_mean': 0.9960442185401917, 'reward_meter_std': 0.0010061725042760372, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9960442185401917, 'reward_total_composite_std': 0.0010061725042760372, 'epoch': 0.11}

 79%|███████▉  | 2622/3300 [3:48:40<1:07:22,  5.96s/it]INFO 04-12 02:25:31 [block_pool.py:378] Successfully reset prefix cache

 79%|███████▉  | 2623/3300 [3:48:45<1:03:50,  5.66s/it]2026-04-12 02:25:36,273 | INFO | train_grpo_train | metrics_logged mode=train step=2623

                                                       
{'loss': 0.003, 'grad_norm': 3.970113515853882, 'learning_rate': 2.054545454545455e-06, 'num_tokens': 5947273.0, 'completions/mean_length': 71.75, 'completions/min_length': 71.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.75, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9978034496307373, 'rewards/meter/std': 0.004221748560667038, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978034496307373, 'rewards/total_composite/std': 0.004221748560667038, 'reward': 0.9978034496307373, 'reward_std': 0.004221739247441292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020419280976057053, 'sampling/sampling_logp_difference/max': 0.9410266876220703, 'sampling/importance_sampling_ratio/min': 0.39022698998451233, 'sampling/importance_sampling_ratio/mean': 1.0029164552688599, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12103518098592758, 'clip_ratio/low_mean': 0.0017605633474886417, 'clip_ratio/low_min': 0.0017605633474886417, 'clip_ratio/high_mean': 0.0052327855955809355, 'clip_ratio/high_max': 0.0052327855955809355, 'clip_ratio/region_mean': 0.006993348943069577, 'reward_total_mean': 0.9978034496307373, 'reward_meter_mean': 0.9978034496307373, 'reward_meter_std': 0.004221748560667038, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978034496307373, 'reward_total_composite_std': 0.004221748560667038, 'epoch': 0.11}

 79%|███████▉  | 2623/3300 [3:48:45<1:03:50,  5.66s/it]INFO 04-12 02:25:36 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2624/3300 [3:48:49<1:00:36,  5.38s/it]2026-04-12 02:25:41,005 | INFO | train_grpo_train | metrics_logged mode=train step=2624

                                                       
{'loss': 0.0286, 'grad_norm': 9.467141151428223, 'learning_rate': 2.0515151515151517e-06, 'num_tokens': 5948959.0, 'completions/mean_length': 64.75, 'completions/min_length': 62.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.017361678183078766, 'rewards/meter/std': 0.014113202691078186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.017361678183078766, 'rewards/total_composite/std': 0.014113202691078186, 'reward': 0.017361678183078766, 'reward_std': 0.014113202691078186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08726956695318222, 'sampling/sampling_logp_difference/max': 2.049853563308716, 'sampling/importance_sampling_ratio/min': 0.12875375151634216, 'sampling/importance_sampling_ratio/mean': 0.9971011877059937, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3771820291876793, 'clip_ratio/low_mean': 0.041920434683561325, 'clip_ratio/low_min': 0.041920434683561325, 'clip_ratio/high_mean': 0.011576013872399926, 'clip_ratio/high_max': 0.011576013872399926, 'clip_ratio/region_mean': 0.05349644855596125, 'reward_total_mean': 0.017361678183078766, 'reward_meter_mean': 0.017361678183078766, 'reward_meter_std': 0.014113202691078186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.017361678183078766, 'reward_total_composite_std': 0.014113202691078186, 'epoch': 0.11}

 80%|███████▉  | 2624/3300 [3:48:50<1:00:36,  5.38s/it]INFO 04-12 02:25:41 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2625/3300 [3:48:54<56:47,  5.05s/it]  2026-04-12 02:25:45,284 | INFO | train_grpo_train | metrics_logged mode=train step=2625

                                                     
{'loss': -0.0006, 'grad_norm': 1.3292933702468872, 'learning_rate': 2.0484848484848485e-06, 'num_tokens': 5950439.0, 'completions/mean_length': 37.0, 'completions/min_length': 37.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9995855093002319, 'rewards/meter/std': 5.365296237869188e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9995855093002319, 'rewards/total_composite/std': 5.365296237869188e-05, 'reward': 0.9995855093002319, 'reward_std': 5.3658961405744776e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004437229596078396, 'sampling/sampling_logp_difference/max': 0.6433219909667969, 'sampling/importance_sampling_ratio/min': 0.5255436897277832, 'sampling/importance_sampling_ratio/mean': 1.000502109527588, 'sampling/importance_sampling_ratio/max': 1.0283234119415283, 'entropy': 0.022779989056289196, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9995855093002319, 'reward_meter_mean': 0.9995855093002319, 'reward_meter_std': 5.365296237869188e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9995855093002319, 'reward_total_composite_std': 5.365296237869188e-05, 'epoch': 0.11}

 80%|███████▉  | 2625/3300 [3:48:54<56:47,  5.05s/it]INFO 04-12 02:25:45 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2626/3300 [3:48:59<57:20,  5.10s/it]2026-04-12 02:25:50,510 | INFO | train_grpo_train | metrics_logged mode=train step=2626

                                                     
{'loss': 0.0077, 'grad_norm': 2.536959409713745, 'learning_rate': 2.0454545454545457e-06, 'num_tokens': 5952538.0, 'completions/mean_length': 102.375, 'completions/min_length': 100.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.375, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9648023843765259, 'rewards/meter/std': 0.04039134085178375, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9648023843765259, 'rewards/total_composite/std': 0.04039134085178375, 'reward': 0.9648023843765259, 'reward_std': 0.04039134085178375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03456467390060425, 'sampling/sampling_logp_difference/max': 1.0130319595336914, 'sampling/importance_sampling_ratio/min': 0.3631163537502289, 'sampling/importance_sampling_ratio/mean': 1.0097938776016235, 'sampling/importance_sampling_ratio/max': 1.6411556005477905, 'entropy': 0.3144999463111162, 'clip_ratio/low_mean': 0.008605769136920571, 'clip_ratio/low_min': 0.008605769136920571, 'clip_ratio/high_mean': 0.009794845129363239, 'clip_ratio/high_max': 0.009794845129363239, 'clip_ratio/region_mean': 0.01840061426628381, 'reward_total_mean': 0.9648023843765259, 'reward_meter_mean': 0.9648023843765259, 'reward_meter_std': 0.04039134085178375, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9648023843765259, 'reward_total_composite_std': 0.04039134085178375, 'epoch': 0.11}

 80%|███████▉  | 2626/3300 [3:48:59<57:20,  5.10s/it]INFO 04-12 02:25:50 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2627/3300 [3:49:04<55:38,  4.96s/it]2026-04-12 02:25:55,132 | INFO | train_grpo_train | metrics_logged mode=train step=2627

                                                     
{'loss': 0.2115, 'grad_norm': 6.214305877685547, 'learning_rate': 2.0424242424242426e-06, 'num_tokens': 5954150.0, 'completions/mean_length': 60.5, 'completions/min_length': 36.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9953294396400452, 'rewards/meter/std': 0.010616513900458813, 'rewards/count_adherence/mean': 0.25, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.24602118134498596, 'rewards/total_composite/std': 0.4556134343147278, 'reward': 0.24602118134498596, 'reward_std': 0.4556134045124054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03473228961229324, 'sampling/sampling_logp_difference/max': 1.8016653060913086, 'sampling/importance_sampling_ratio/min': 0.16502384841442108, 'sampling/importance_sampling_ratio/mean': 1.0146692991256714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2536655478179455, 'clip_ratio/low_mean': 0.012954836362041533, 'clip_ratio/low_min': 0.012954836362041533, 'clip_ratio/high_mean': 0.012847222620621324, 'clip_ratio/high_max': 0.012847222620621324, 'clip_ratio/region_mean': 0.025802058982662857, 'reward_total_mean': 0.24602118134498596, 'reward_meter_mean': 0.9953294396400452, 'reward_meter_std': 0.010616513900458813, 'reward_count_adherence_mean': 0.25, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.24602118134498596, 'reward_total_composite_std': 0.4556134343147278, 'epoch': 0.11}

 80%|███████▉  | 2627/3300 [3:49:04<55:38,  4.96s/it]INFO 04-12 02:25:55 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2628/3300 [3:49:10<1:00:50,  5.43s/it]2026-04-12 02:26:01,666 | INFO | train_grpo_train | metrics_logged mode=train step=2628

                                                       
{'loss': 0.014, 'grad_norm': 3.508739709854126, 'learning_rate': 2.03939393939394e-06, 'num_tokens': 5957532.0, 'completions/mean_length': 239.75, 'completions/min_length': 233.0, 'completions/max_length': 243.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 239.75, 'completions/min_terminated_length': 233.0, 'completions/max_terminated_length': 243.0, 'rewards/meter/mean': 0.9953933954238892, 'rewards/meter/std': 0.00770949712023139, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9953933954238892, 'rewards/total_composite/std': 0.00770949712023139, 'reward': 0.9953933954238892, 'reward_std': 0.007709511090070009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07987385243177414, 'sampling/sampling_logp_difference/max': 3.092388153076172, 'sampling/importance_sampling_ratio/min': 0.04539341852068901, 'sampling/importance_sampling_ratio/mean': 1.0158778429031372, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7806448265910149, 'clip_ratio/low_mean': 0.00927626434713602, 'clip_ratio/low_min': 0.00927626434713602, 'clip_ratio/high_mean': 0.0423777480609715, 'clip_ratio/high_max': 0.0423777480609715, 'clip_ratio/region_mean': 0.05165401240810752, 'reward_total_mean': 0.9953933954238892, 'reward_meter_mean': 0.9953933954238892, 'reward_meter_std': 0.00770949712023139, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9953933954238892, 'reward_total_composite_std': 0.00770949712023139, 'epoch': 0.11}

 80%|███████▉  | 2628/3300 [3:49:10<1:00:50,  5.43s/it]INFO 04-12 02:26:01 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2629/3300 [3:49:18<1:08:37,  6.14s/it]2026-04-12 02:26:09,446 | INFO | train_grpo_train | metrics_logged mode=train step=2629

                                                       
{'loss': 0.014, 'grad_norm': 1.9617708921432495, 'learning_rate': 2.0363636363636367e-06, 'num_tokens': 5961811.0, 'completions/mean_length': 317.875, 'completions/min_length': 300.0, 'completions/max_length': 337.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 317.875, 'completions/min_terminated_length': 300.0, 'completions/max_terminated_length': 337.0, 'rewards/meter/mean': 0.9979603886604309, 'rewards/meter/std': 0.0011684981873258948, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9593137502670288, 'rewards/repeat_penalty/std': 0.06067804619669914, 'rewards/total_composite/mean': 0.927051305770874, 'rewards/total_composite/std': 0.07678526639938354, 'reward': 0.927051305770874, 'reward_std': 0.07678526639938354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054092179983854294, 'sampling/sampling_logp_difference/max': 1.6163597106933594, 'sampling/importance_sampling_ratio/min': 0.19862042367458344, 'sampling/importance_sampling_ratio/mean': 1.0117024183273315, 'sampling/importance_sampling_ratio/max': 1.913344383239746, 'entropy': 0.4582854099571705, 'clip_ratio/low_mean': 0.012730259681120515, 'clip_ratio/low_min': 0.012730259681120515, 'clip_ratio/high_mean': 0.018142351880669594, 'clip_ratio/high_max': 0.018142351880669594, 'clip_ratio/region_mean': 0.03087261156179011, 'reward_total_mean': 0.927051305770874, 'reward_meter_mean': 0.9979603886604309, 'reward_meter_std': 0.0011684981873258948, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9593137502670288, 'reward_repeat_penalty_std': 0.06067804619669914, 'reward_total_composite_mean': 0.927051305770874, 'reward_total_composite_std': 0.07678526639938354, 'epoch': 0.11}

 80%|███████▉  | 2629/3300 [3:49:18<1:08:37,  6.14s/it]INFO 04-12 02:26:09 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2630/3300 [3:49:22<1:02:38,  5.61s/it]2026-04-12 02:26:13,830 | INFO | train_grpo_train | metrics_logged mode=train step=2630

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.0333333333333335e-06, 'num_tokens': 5963587.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0030186243820935488, 'sampling/sampling_logp_difference/max': 0.3742567002773285, 'sampling/importance_sampling_ratio/min': 0.6878003478050232, 'sampling/importance_sampling_ratio/mean': 0.9990300536155701, 'sampling/importance_sampling_ratio/max': 1.1619328260421753, 'entropy': 0.015519737149588764, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 80%|███████▉  | 2630/3300 [3:49:22<1:02:38,  5.61s/it]INFO 04-12 02:26:14 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2631/3300 [3:49:28<1:01:54,  5.55s/it]2026-04-12 02:26:19,247 | INFO | train_grpo_train | metrics_logged mode=train step=2631

                                                       
{'loss': 0.0062, 'grad_norm': 1.6037280559539795, 'learning_rate': 2.0303030303030303e-06, 'num_tokens': 5966253.0, 'completions/mean_length': 143.25, 'completions/min_length': 142.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.25, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9989732503890991, 'rewards/meter/std': 0.00014036186621524394, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8392857313156128, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.838421106338501, 'rewards/total_composite/std': 0.05040955916047096, 'reward': 0.838421106338501, 'reward_std': 0.050409551709890366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013544322922825813, 'sampling/sampling_logp_difference/max': 1.0127596855163574, 'sampling/importance_sampling_ratio/min': 0.36321523785591125, 'sampling/importance_sampling_ratio/mean': 1.008257508277893, 'sampling/importance_sampling_ratio/max': 1.676928997039795, 'entropy': 0.09556546807289124, 'clip_ratio/low_mean': 0.0017241379246115685, 'clip_ratio/low_min': 0.0017241379246115685, 'clip_ratio/high_mean': 0.006118966557551175, 'clip_ratio/high_max': 0.006118966557551175, 'clip_ratio/region_mean': 0.007843104482162744, 'reward_total_mean': 0.838421106338501, 'reward_meter_mean': 0.9989732503890991, 'reward_meter_std': 0.00014036186621524394, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8392857313156128, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.838421106338501, 'reward_total_composite_std': 0.05040955916047096, 'epoch': 0.11}

 80%|███████▉  | 2631/3300 [3:49:28<1:01:54,  5.55s/it]INFO 04-12 02:26:19 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2632/3300 [3:49:32<58:22,  5.24s/it]  2026-04-12 02:26:23,765 | INFO | train_grpo_train | metrics_logged mode=train step=2632

                                                     
{'loss': -0.007, 'grad_norm': 5.14970064163208, 'learning_rate': 2.0272727272727276e-06, 'num_tokens': 5968023.0, 'completions/mean_length': 70.25, 'completions/min_length': 69.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9124512672424316, 'rewards/meter/std': 0.0726349726319313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9124512672424316, 'rewards/total_composite/std': 0.0726349726319313, 'reward': 0.9124512672424316, 'reward_std': 0.0726349726319313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04469385743141174, 'sampling/sampling_logp_difference/max': 1.9476492404937744, 'sampling/importance_sampling_ratio/min': 0.14260892570018768, 'sampling/importance_sampling_ratio/mean': 1.0102332830429077, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3492947220802307, 'clip_ratio/low_mean': 0.023193509317934513, 'clip_ratio/low_min': 0.023193509317934513, 'clip_ratio/high_mean': 0.015901052742265165, 'clip_ratio/high_max': 0.015901052742265165, 'clip_ratio/region_mean': 0.03909456206019968, 'reward_total_mean': 0.9124512672424316, 'reward_meter_mean': 0.9124512672424316, 'reward_meter_std': 0.0726349726319313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9124512672424316, 'reward_total_composite_std': 0.0726349726319313, 'epoch': 0.11}

 80%|███████▉  | 2632/3300 [3:49:32<58:22,  5.24s/it]INFO 04-12 02:26:23 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2633/3300 [3:49:37<58:19,  5.25s/it]2026-04-12 02:26:29,024 | INFO | train_grpo_train | metrics_logged mode=train step=2633

                                                     
{'loss': 0.0122, 'grad_norm': 2.2837815284729004, 'learning_rate': 2.0242424242424244e-06, 'num_tokens': 5970465.0, 'completions/mean_length': 136.25, 'completions/min_length': 133.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.25, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.9715265035629272, 'rewards/meter/std': 0.011229481548070908, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.910714328289032, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.8844000101089478, 'rewards/total_composite/std': 0.06688975542783737, 'reward': 0.8844000101089478, 'reward_std': 0.06688975542783737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03914301469922066, 'sampling/sampling_logp_difference/max': 1.8302927017211914, 'sampling/importance_sampling_ratio/min': 0.16036662459373474, 'sampling/importance_sampling_ratio/mean': 1.0096205472946167, 'sampling/importance_sampling_ratio/max': 1.7595460414886475, 'entropy': 0.3410081285983324, 'clip_ratio/low_mean': 0.010007655015215278, 'clip_ratio/low_min': 0.010007655015215278, 'clip_ratio/high_mean': 0.011128432815894485, 'clip_ratio/high_max': 0.011128432815894485, 'clip_ratio/region_mean': 0.021136087831109762, 'reward_total_mean': 0.8844000101089478, 'reward_meter_mean': 0.9715265035629272, 'reward_meter_std': 0.011229481548070908, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.910714328289032, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.8844000101089478, 'reward_total_composite_std': 0.06688975542783737, 'epoch': 0.11}

 80%|███████▉  | 2633/3300 [3:49:38<58:19,  5.25s/it]INFO 04-12 02:26:29 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2634/3300 [3:49:42<55:57,  5.04s/it]2026-04-12 02:26:33,591 | INFO | train_grpo_train | metrics_logged mode=train step=2634

                                                     
{'loss': 0.0021, 'grad_norm': 3.726565361022949, 'learning_rate': 2.0212121212121212e-06, 'num_tokens': 5972297.0, 'completions/mean_length': 71.0, 'completions/min_length': 68.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9706488251686096, 'rewards/meter/std': 0.049138810485601425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9706488251686096, 'rewards/total_composite/std': 0.049138810485601425, 'reward': 0.9706488251686096, 'reward_std': 0.049138810485601425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07419008761644363, 'sampling/sampling_logp_difference/max': 2.089165449142456, 'sampling/importance_sampling_ratio/min': 0.12379039824008942, 'sampling/importance_sampling_ratio/mean': 1.0056947469711304, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5871141031384468, 'clip_ratio/low_mean': 0.0086805559694767, 'clip_ratio/low_min': 0.0086805559694767, 'clip_ratio/high_mean': 0.04050721391104162, 'clip_ratio/high_max': 0.04050721391104162, 'clip_ratio/region_mean': 0.04918776988051832, 'reward_total_mean': 0.9706488251686096, 'reward_meter_mean': 0.9706488251686096, 'reward_meter_std': 0.049138810485601425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9706488251686096, 'reward_total_composite_std': 0.049138810485601425, 'epoch': 0.11}

 80%|███████▉  | 2634/3300 [3:49:42<55:57,  5.04s/it]INFO 04-12 02:26:33 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2635/3300 [3:49:47<54:08,  4.89s/it]2026-04-12 02:26:38,107 | INFO | train_grpo_train | metrics_logged mode=train step=2635

                                                     
{'loss': 0.0001, 'grad_norm': 0.053679000586271286, 'learning_rate': 2.0181818181818185e-06, 'num_tokens': 5974177.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993990659713745, 'rewards/meter/std': 1.0326285746486974e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993990659713745, 'rewards/total_composite/std': 1.0326285746486974e-06, 'reward': 0.9993990659713745, 'reward_std': 1.023619461193448e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002415498485788703, 'sampling/sampling_logp_difference/max': 0.5120803117752075, 'sampling/importance_sampling_ratio/min': 0.9006934762001038, 'sampling/importance_sampling_ratio/mean': 1.0016577243804932, 'sampling/importance_sampling_ratio/max': 1.6687591075897217, 'entropy': 0.014239851152524352, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993990659713745, 'reward_meter_mean': 0.9993990659713745, 'reward_meter_std': 1.0326285746486974e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993990659713745, 'reward_total_composite_std': 1.0326285746486974e-06, 'epoch': 0.11}

 80%|███████▉  | 2635/3300 [3:49:47<54:08,  4.89s/it]INFO 04-12 02:26:38 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2636/3300 [3:49:51<54:16,  4.90s/it]2026-04-12 02:26:43,055 | INFO | train_grpo_train | metrics_logged mode=train step=2636

                                                     
{'loss': -0.0168, 'grad_norm': 2.7606544494628906, 'learning_rate': 2.0151515151515153e-06, 'num_tokens': 5976300.0, 'completions/mean_length': 90.375, 'completions/min_length': 87.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.375, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9972773790359497, 'rewards/meter/std': 0.0004967614077031612, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972773790359497, 'rewards/total_composite/std': 0.0004967614077031612, 'reward': 0.9972773790359497, 'reward_std': 0.0004967702552676201, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025844357907772064, 'sampling/sampling_logp_difference/max': 1.573763370513916, 'sampling/importance_sampling_ratio/min': 0.20726369321346283, 'sampling/importance_sampling_ratio/mean': 1.0045125484466553, 'sampling/importance_sampling_ratio/max': 1.9276334047317505, 'entropy': 0.12905630934983492, 'clip_ratio/low_mean': 0.009945004130713642, 'clip_ratio/low_min': 0.009945004130713642, 'clip_ratio/high_mean': 0.008138859295286238, 'clip_ratio/high_max': 0.008138859295286238, 'clip_ratio/region_mean': 0.01808386342599988, 'reward_total_mean': 0.9972773790359497, 'reward_meter_mean': 0.9972773790359497, 'reward_meter_std': 0.0004967614077031612, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972773790359497, 'reward_total_composite_std': 0.0004967614077031612, 'epoch': 0.11}

 80%|███████▉  | 2636/3300 [3:49:52<54:16,  4.90s/it]INFO 04-12 02:26:43 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2637/3300 [3:49:57<55:01,  4.98s/it]2026-04-12 02:26:48,212 | INFO | train_grpo_train | metrics_logged mode=train step=2637

                                                     
{'loss': -0.0064, 'grad_norm': 2.6201107501983643, 'learning_rate': 2.012121212121212e-06, 'num_tokens': 5978451.0, 'completions/mean_length': 117.875, 'completions/min_length': 113.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.875, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9981694221496582, 'rewards/meter/std': 0.0008605656330473721, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981694221496582, 'rewards/total_composite/std': 0.0008605656330473721, 'reward': 0.9981694221496582, 'reward_std': 0.0008605782641097903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044327329844236374, 'sampling/sampling_logp_difference/max': 1.3086605072021484, 'sampling/importance_sampling_ratio/min': 0.27018171548843384, 'sampling/importance_sampling_ratio/mean': 1.0066555738449097, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33568100817501545, 'clip_ratio/low_mean': 0.010715189506299794, 'clip_ratio/low_min': 0.010715189506299794, 'clip_ratio/high_mean': 0.007327559753321111, 'clip_ratio/high_max': 0.007327559753321111, 'clip_ratio/region_mean': 0.018042749259620905, 'reward_total_mean': 0.9981694221496582, 'reward_meter_mean': 0.9981694221496582, 'reward_meter_std': 0.0008605656330473721, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981694221496582, 'reward_total_composite_std': 0.0008605656330473721, 'epoch': 0.11}

 80%|███████▉  | 2637/3300 [3:49:57<55:01,  4.98s/it]INFO 04-12 02:26:48 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2638/3300 [3:50:01<52:31,  4.76s/it]2026-04-12 02:26:52,460 | INFO | train_grpo_train | metrics_logged mode=train step=2638

                                                     
{'loss': 0.0217, 'grad_norm': 6.669706344604492, 'learning_rate': 2.009090909090909e-06, 'num_tokens': 5979937.0, 'completions/mean_length': 35.75, 'completions/min_length': 34.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9986928701400757, 'rewards/meter/std': 0.002228210447356105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9986928701400757, 'rewards/total_composite/std': 0.002228210447356105, 'reward': 0.9986928701400757, 'reward_std': 0.002228200202807784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02272917330265045, 'sampling/sampling_logp_difference/max': 0.7570033073425293, 'sampling/importance_sampling_ratio/min': 0.4690699875354767, 'sampling/importance_sampling_ratio/mean': 1.0115097761154175, 'sampling/importance_sampling_ratio/max': 1.4079704284667969, 'entropy': 0.1990708950906992, 'clip_ratio/low_mean': 0.003289473708719015, 'clip_ratio/low_min': 0.003289473708719015, 'clip_ratio/high_mean': 0.017471515806391835, 'clip_ratio/high_max': 0.017471515806391835, 'clip_ratio/region_mean': 0.02076098951511085, 'reward_total_mean': 0.9986928701400757, 'reward_meter_mean': 0.9986928701400757, 'reward_meter_std': 0.002228210447356105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9986928701400757, 'reward_total_composite_std': 0.002228210447356105, 'epoch': 0.11}

 80%|███████▉  | 2638/3300 [3:50:01<52:31,  4.76s/it]INFO 04-12 02:26:52 [block_pool.py:378] Successfully reset prefix cache

 80%|███████▉  | 2639/3300 [3:50:06<53:25,  4.85s/it]2026-04-12 02:26:57,522 | INFO | train_grpo_train | metrics_logged mode=train step=2639

                                                     
{'loss': -0.0005, 'grad_norm': 1.0811501741409302, 'learning_rate': 2.0060606060606062e-06, 'num_tokens': 5982303.0, 'completions/mean_length': 106.75, 'completions/min_length': 104.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.75, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9990361332893372, 'rewards/meter/std': 0.00013772756210528314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990361332893372, 'rewards/total_composite/std': 0.00013772756210528314, 'reward': 0.9990361332893372, 'reward_std': 0.00013771916565019637, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014354676008224487, 'sampling/sampling_logp_difference/max': 1.5451383590698242, 'sampling/importance_sampling_ratio/min': 0.21328236162662506, 'sampling/importance_sampling_ratio/mean': 1.0037332773208618, 'sampling/importance_sampling_ratio/max': 1.4171427488327026, 'entropy': 0.09720816975459456, 'clip_ratio/low_mean': 0.003537735901772976, 'clip_ratio/low_min': 0.003537735901772976, 'clip_ratio/high_mean': 0.004762336844578385, 'clip_ratio/high_max': 0.004762336844578385, 'clip_ratio/region_mean': 0.008300072746351361, 'reward_total_mean': 0.9990361332893372, 'reward_meter_mean': 0.9990361332893372, 'reward_meter_std': 0.00013772756210528314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990361332893372, 'reward_total_composite_std': 0.00013772756210528314, 'epoch': 0.11}

 80%|███████▉  | 2639/3300 [3:50:06<53:25,  4.85s/it]INFO 04-12 02:26:57 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2640/3300 [3:50:10<52:11,  4.74s/it]2026-04-12 02:27:02,016 | INFO | train_grpo_train | metrics_logged mode=train step=2640

                                                     
{'loss': 0.0003, 'grad_norm': 1.591088891029358, 'learning_rate': 2.0030303030303035e-06, 'num_tokens': 5984115.0, 'completions/mean_length': 71.5, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9991612434387207, 'rewards/meter/std': 0.00017454942280892283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991612434387207, 'rewards/total_composite/std': 0.00017454942280892283, 'reward': 0.9991612434387207, 'reward_std': 0.00017454590124543756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008529691025614738, 'sampling/sampling_logp_difference/max': 0.8294587135314941, 'sampling/importance_sampling_ratio/min': 0.4362854063510895, 'sampling/importance_sampling_ratio/mean': 1.001218557357788, 'sampling/importance_sampling_ratio/max': 1.096482753753662, 'entropy': 0.04928429890424013, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0017361111240461469, 'clip_ratio/high_max': 0.0017361111240461469, 'clip_ratio/region_mean': 0.0017361111240461469, 'reward_total_mean': 0.9991612434387207, 'reward_meter_mean': 0.9991612434387207, 'reward_meter_std': 0.00017454942280892283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991612434387207, 'reward_total_composite_std': 0.00017454942280892283, 'epoch': 0.11}

 80%|████████  | 2640/3300 [3:50:11<52:11,  4.74s/it]INFO 04-12 02:27:02 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2641/3300 [3:50:15<50:53,  4.63s/it]2026-04-12 02:27:06,390 | INFO | train_grpo_train | metrics_logged mode=train step=2641

                                                     
{'loss': 0.0002, 'grad_norm': 0.12622658908367157, 'learning_rate': 2.0000000000000003e-06, 'num_tokens': 5985907.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993976354598999, 'rewards/meter/std': 3.0407550184463616e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993976354598999, 'rewards/total_composite/std': 3.0407550184463616e-06, 'reward': 0.9993976354598999, 'reward_std': 3.030640073120594e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004023827612400055, 'sampling/sampling_logp_difference/max': 0.6670032739639282, 'sampling/importance_sampling_ratio/min': 0.5132443308830261, 'sampling/importance_sampling_ratio/mean': 0.9996961951255798, 'sampling/importance_sampling_ratio/max': 1.6674405336380005, 'entropy': 0.01789803896099329, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.00390625, 'reward_total_mean': 0.9993976354598999, 'reward_meter_mean': 0.9993976354598999, 'reward_meter_std': 3.0407550184463616e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993976354598999, 'reward_total_composite_std': 3.0407550184463616e-06, 'epoch': 0.11}

 80%|████████  | 2641/3300 [3:50:15<50:53,  4.63s/it]INFO 04-12 02:27:06 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2642/3300 [3:50:19<49:54,  4.55s/it]2026-04-12 02:27:10,755 | INFO | train_grpo_train | metrics_logged mode=train step=2642

                                                     
{'loss': 0.0, 'grad_norm': 0.3109838664531708, 'learning_rate': 1.996969696969697e-06, 'num_tokens': 5987731.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993950724601746, 'rewards/meter/std': 5.620834144792752e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993950724601746, 'rewards/total_composite/std': 5.620834144792752e-06, 'reward': 0.9993950724601746, 'reward_std': 5.6233166105812415e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00695427879691124, 'sampling/sampling_logp_difference/max': 0.8221545219421387, 'sampling/importance_sampling_ratio/min': 0.43948376178741455, 'sampling/importance_sampling_ratio/mean': 0.9969460964202881, 'sampling/importance_sampling_ratio/max': 1.4890631437301636, 'entropy': 0.01494286535307765, 'clip_ratio/low_mean': 0.00390625, 'clip_ratio/low_min': 0.00390625, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.005859375, 'reward_total_mean': 0.9993950724601746, 'reward_meter_mean': 0.9993950724601746, 'reward_meter_std': 5.620834144792752e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993950724601746, 'reward_total_composite_std': 5.620834144792752e-06, 'epoch': 0.11}

 80%|████████  | 2642/3300 [3:50:19<49:54,  4.55s/it]INFO 04-12 02:27:10 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2643/3300 [3:50:23<48:50,  4.46s/it]2026-04-12 02:27:14,995 | INFO | train_grpo_train | metrics_logged mode=train step=2643

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.993939393939394e-06, 'num_tokens': 5989459.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00024456807295791805, 'sampling/sampling_logp_difference/max': 0.003122988622635603, 'sampling/importance_sampling_ratio/min': 0.9998295307159424, 'sampling/importance_sampling_ratio/mean': 1.0002434253692627, 'sampling/importance_sampling_ratio/max': 1.003127932548523, 'entropy': 0.0020185734319966286, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 80%|████████  | 2643/3300 [3:50:23<48:50,  4.46s/it]INFO 04-12 02:27:15 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2644/3300 [3:50:28<49:53,  4.56s/it]2026-04-12 02:27:19,801 | INFO | train_grpo_train | metrics_logged mode=train step=2644

                                                     
{'loss': -0.0016, 'grad_norm': 0.6552127003669739, 'learning_rate': 1.9909090909090913e-06, 'num_tokens': 5991639.0, 'completions/mean_length': 97.5, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.5, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9979605674743652, 'rewards/meter/std': 0.00010473600559635088, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979605674743652, 'rewards/total_composite/std': 0.00010473600559635088, 'reward': 0.9979605674743652, 'reward_std': 0.00010474038572283462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011263424530625343, 'sampling/sampling_logp_difference/max': 1.1425704956054688, 'sampling/importance_sampling_ratio/min': 0.31899797916412354, 'sampling/importance_sampling_ratio/mean': 0.9999778270721436, 'sampling/importance_sampling_ratio/max': 1.362816333770752, 'entropy': 0.07244368549436331, 'clip_ratio/low_mean': 0.0012886597542092204, 'clip_ratio/low_min': 0.0012886597542092204, 'clip_ratio/high_mean': 0.006403850042261183, 'clip_ratio/high_max': 0.006403850042261183, 'clip_ratio/region_mean': 0.007692509796470404, 'reward_total_mean': 0.9979605674743652, 'reward_meter_mean': 0.9979605674743652, 'reward_meter_std': 0.00010473600559635088, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979605674743652, 'reward_total_composite_std': 0.00010473600559635088, 'epoch': 0.11}

 80%|████████  | 2644/3300 [3:50:28<49:53,  4.56s/it]INFO 04-12 02:27:20 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2645/3300 [3:50:33<49:12,  4.51s/it]2026-04-12 02:27:24,174 | INFO | train_grpo_train | metrics_logged mode=train step=2645

                                                     
{'loss': -0.0, 'grad_norm': 0.473066508769989, 'learning_rate': 1.987878787878788e-06, 'num_tokens': 5993422.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981163740158081, 'rewards/meter/std': 1.9677801901707426e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981163740158081, 'rewards/total_composite/std': 1.9677801901707426e-05, 'reward': 0.9981163740158081, 'reward_std': 1.96640412468696e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004463972058147192, 'sampling/sampling_logp_difference/max': 0.4455909729003906, 'sampling/importance_sampling_ratio/min': 0.6404457092285156, 'sampling/importance_sampling_ratio/mean': 1.000008225440979, 'sampling/importance_sampling_ratio/max': 1.1160743236541748, 'entropy': 0.04345296695828438, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.9981163740158081, 'reward_meter_mean': 0.9981163740158081, 'reward_meter_std': 1.9677801901707426e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981163740158081, 'reward_total_composite_std': 1.9677801901707426e-05, 'epoch': 0.11}

 80%|████████  | 2645/3300 [3:50:33<49:12,  4.51s/it]INFO 04-12 02:27:24 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2646/3300 [3:50:37<47:19,  4.34s/it]2026-04-12 02:27:28,133 | INFO | train_grpo_train | metrics_logged mode=train step=2646

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.984848484848485e-06, 'num_tokens': 5994830.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002391562593402341, 'sampling/sampling_logp_difference/max': 0.0038578114472329617, 'sampling/importance_sampling_ratio/min': 0.999919056892395, 'sampling/importance_sampling_ratio/mean': 1.0002386569976807, 'sampling/importance_sampling_ratio/max': 1.003865361213684, 'entropy': 0.0018622480711201206, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 80%|████████  | 2646/3300 [3:50:37<47:19,  4.34s/it]INFO 04-12 02:27:28 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2647/3300 [3:50:43<52:53,  4.86s/it]2026-04-12 02:27:34,209 | INFO | train_grpo_train | metrics_logged mode=train step=2647

                                                     
{'loss': 0.0201, 'grad_norm': 2.459423303604126, 'learning_rate': 1.981818181818182e-06, 'num_tokens': 5997841.0, 'completions/mean_length': 192.375, 'completions/min_length': 185.0, 'completions/max_length': 199.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 192.375, 'completions/min_terminated_length': 185.0, 'completions/max_terminated_length': 199.0, 'rewards/meter/mean': 0.9985500574111938, 'rewards/meter/std': 0.0007222077692858875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9708031415939331, 'rewards/total_composite/std': 0.05117257311940193, 'reward': 0.9708031415939331, 'reward_std': 0.05117259919643402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04225311800837517, 'sampling/sampling_logp_difference/max': 1.1989145278930664, 'sampling/importance_sampling_ratio/min': 0.30152130126953125, 'sampling/importance_sampling_ratio/mean': 1.0101455450057983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3521938771009445, 'clip_ratio/low_mean': 0.006954187294468284, 'clip_ratio/low_min': 0.006954187294468284, 'clip_ratio/high_mean': 0.023607419105246663, 'clip_ratio/high_max': 0.023607419105246663, 'clip_ratio/region_mean': 0.030561606399714947, 'reward_total_mean': 0.9708031415939331, 'reward_meter_mean': 0.9985500574111938, 'reward_meter_std': 0.0007222077692858875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9708031415939331, 'reward_total_composite_std': 0.05117257311940193, 'epoch': 0.11}

 80%|████████  | 2647/3300 [3:50:43<52:53,  4.86s/it]INFO 04-12 02:27:34 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2648/3300 [3:50:48<54:23,  5.01s/it]2026-04-12 02:27:39,553 | INFO | train_grpo_train | metrics_logged mode=train step=2648

                                                     
{'loss': -0.0035, 'grad_norm': 2.53367280960083, 'learning_rate': 1.978787878787879e-06, 'num_tokens': 6000235.0, 'completions/mean_length': 131.25, 'completions/min_length': 130.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.25, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9992580413818359, 'rewards/meter/std': 5.2671864978037775e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9814149141311646, 'rewards/total_composite/std': 0.050484977662563324, 'reward': 0.9814149141311646, 'reward_std': 0.050484977662563324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014524316415190697, 'sampling/sampling_logp_difference/max': 1.7293694019317627, 'sampling/importance_sampling_ratio/min': 0.1773962378501892, 'sampling/importance_sampling_ratio/mean': 0.9994350075721741, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08133039530366659, 'clip_ratio/low_mean': 0.000961538462433964, 'clip_ratio/low_min': 0.000961538462433964, 'clip_ratio/high_mean': 0.01428450463572517, 'clip_ratio/high_max': 0.01428450463572517, 'clip_ratio/region_mean': 0.015246043098159134, 'reward_total_mean': 0.9814149141311646, 'reward_meter_mean': 0.9992580413818359, 'reward_meter_std': 5.2671864978037775e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9814149141311646, 'reward_total_composite_std': 0.050484977662563324, 'epoch': 0.11}

 80%|████████  | 2648/3300 [3:50:48<54:23,  5.01s/it]INFO 04-12 02:27:39 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2649/3300 [3:50:52<52:46,  4.86s/it]2026-04-12 02:27:44,086 | INFO | train_grpo_train | metrics_logged mode=train step=2649

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.975757575757576e-06, 'num_tokens': 6002123.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00039579044096171856, 'sampling/sampling_logp_difference/max': 0.016826428472995758, 'sampling/importance_sampling_ratio/min': 0.9833143353462219, 'sampling/importance_sampling_ratio/mean': 1.0002467632293701, 'sampling/importance_sampling_ratio/max': 1.0139360427856445, 'entropy': 0.004253393795806915, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 80%|████████  | 2649/3300 [3:50:53<52:46,  4.86s/it]INFO 04-12 02:27:44 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2650/3300 [3:50:57<52:02,  4.80s/it]2026-04-12 02:27:48,754 | INFO | train_grpo_train | metrics_logged mode=train step=2650

                                                     
{'loss': 0.0008, 'grad_norm': 0.25559931993484497, 'learning_rate': 1.9727272727272727e-06, 'num_tokens': 6003906.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981404542922974, 'rewards/meter/std': 1.4544022633344866e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981404542922974, 'rewards/total_composite/std': 1.4544022633344866e-05, 'reward': 0.9981404542922974, 'reward_std': 1.4552377251675352e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005564166698604822, 'sampling/sampling_logp_difference/max': 0.6460444927215576, 'sampling/importance_sampling_ratio/min': 0.5241148471832275, 'sampling/importance_sampling_ratio/mean': 0.9998576641082764, 'sampling/importance_sampling_ratio/max': 1.3914573192596436, 'entropy': 0.036395941860973835, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007519222097471356, 'clip_ratio/high_max': 0.007519222097471356, 'clip_ratio/region_mean': 0.007519222097471356, 'reward_total_mean': 0.9981404542922974, 'reward_meter_mean': 0.9981404542922974, 'reward_meter_std': 1.4544022633344866e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981404542922974, 'reward_total_composite_std': 1.4544022633344866e-05, 'epoch': 0.11}

 80%|████████  | 2650/3300 [3:50:57<52:02,  4.80s/it]INFO 04-12 02:27:49 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:39,  3.59s/it]

 23%|██▎       | 3/13 [00:10<00:35,  3.59s/it]

 31%|███       | 4/13 [00:16<00:38,  4.32s/it]

 38%|███▊      | 5/13 [00:21<00:36,  4.55s/it]

 46%|████▌     | 6/13 [00:27<00:35,  5.07s/it]

 54%|█████▍    | 7/13 [00:32<00:30,  5.01s/it]

 62%|██████▏   | 8/13 [00:38<00:27,  5.50s/it]

 69%|██████▉   | 9/13 [00:46<00:24,  6.01s/it]

 77%|███████▋  | 10/13 [00:50<00:16,  5.58s/it]

 85%|████████▍ | 11/13 [00:55<00:10,  5.48s/it]

 92%|█████████▏| 12/13 [01:01<00:05,  5.51s/it]

100%|██████████| 13/13 [01:06<00:00,  5.42s/it]2026-04-12 02:29:02,253 | INFO | train_grpo_train | metrics_logged mode=eval step=2650

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 73.4017, 'eval_samples_per_second': 1.417, 'eval_steps_per_second': 0.177, 'eval_num_tokens': 6003906.0, 'eval_completions/mean_length': 202.45192307692307, 'eval_completions/min_length': 61.0, 'eval_completions/max_length': 391.2307692307692, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 199.20054978590744, 'eval_completions/min_terminated_length': 61.0, 'eval_completions/max_terminated_length': 385.53846153846155, 'eval_rewards/meter/mean': 0.8157925880872287, 'eval_rewards/meter/std': 0.30426162918313193, 'eval_rewards/count_adherence/mean': 0.9371546369332534, 'eval_rewards/count_adherence/std': 0.08564070440255679, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.9287464526983408, 'eval_rewards/repeat_penalty/std': 0.09337464204201332, 'eval_rewards/total_composite/mean': 0.7124902972808251, 'eval_rewards/total_composite/std': 0.31389044053279436, 'eval_reward': 0.7124902972808251, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03128652188640375, 'eval_sampling/sampling_logp_difference/max': 1.0985317963820238, 'eval_sampling/importance_sampling_ratio/min': 0.3371729736144726, 'eval_sampling/importance_sampling_ratio/mean': 1.0089522416775043, 'eval_sampling/importance_sampling_ratio/max': 1.4974250793457031, 'eval_entropy': 0.33797027514531064, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7124902972808251, 'eval_reward_meter_mean': 0.8157925880872287, 'eval_reward_meter_std': 0.30426162918313193, 'eval_reward_count_adherence_mean': 0.9371546369332534, 'eval_reward_count_adherence_std': 0.08564070440255679, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.9287464526983408, 'eval_reward_repeat_penalty_std': 0.09337464204201332, 'eval_reward_total_composite_mean': 0.7124902972808251, 'eval_reward_total_composite_std': 0.31389044053279436, 'epoch': 0.11}

 80%|████████  | 2650/3300 [3:52:11<52:02,  4.80s/it]

100%|██████████| 13/13 [01:06<00:00,  5.42s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 80%|████████  | 2651/3300 [3:52:21<5:07:35, 28.44s/it]2026-04-12 02:29:12,331 | INFO | train_grpo_train | metrics_logged mode=train step=2651

                                                       
{'loss': -0.0237, 'grad_norm': 2.0326192378997803, 'learning_rate': 1.96969696969697e-06, 'num_tokens': 6008199.0, 'completions/mean_length': 316.625, 'completions/min_length': 300.0, 'completions/max_length': 341.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 316.625, 'completions/min_terminated_length': 300.0, 'completions/max_terminated_length': 341.0, 'rewards/meter/mean': 0.9982768297195435, 'rewards/meter/std': 0.0005552266375161707, 'rewards/count_adherence/mean': 0.930555522441864, 'rewards/count_adherence/std': 0.05750546231865883, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9622548818588257, 'rewards/repeat_penalty/std': 0.05441712588071823, 'rewards/total_composite/mean': 0.892192542552948, 'rewards/total_composite/std': 0.047268956899642944, 'reward': 0.892192542552948, 'reward_std': 0.047268956899642944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04812028631567955, 'sampling/sampling_logp_difference/max': 2.127549171447754, 'sampling/importance_sampling_ratio/min': 0.11912889778614044, 'sampling/importance_sampling_ratio/mean': 1.0096842050552368, 'sampling/importance_sampling_ratio/max': 1.9213883876800537, 'entropy': 0.3517002984881401, 'clip_ratio/low_mean': 0.024150803219527006, 'clip_ratio/low_min': 0.024150803219527006, 'clip_ratio/high_mean': 0.005865102633833885, 'clip_ratio/high_max': 0.005865102633833885, 'clip_ratio/region_mean': 0.03001590585336089, 'reward_total_mean': 0.892192542552948, 'reward_meter_mean': 0.9982768297195435, 'reward_meter_std': 0.0005552266375161707, 'reward_count_adherence_mean': 0.930555522441864, 'reward_count_adherence_std': 0.05750546231865883, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9622548818588257, 'reward_repeat_penalty_std': 0.05441712588071823, 'reward_total_composite_mean': 0.892192542552948, 'reward_total_composite_std': 0.047268956899642944, 'epoch': 0.11}

 80%|████████  | 2651/3300 [3:52:21<5:07:35, 28.44s/it]INFO 04-12 02:29:12 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2652/3300 [3:52:26<3:53:10, 21.59s/it]2026-04-12 02:29:17,948 | INFO | train_grpo_train | metrics_logged mode=train step=2652

                                                       
{'loss': -0.004, 'grad_norm': 2.323451519012451, 'learning_rate': 1.9666666666666668e-06, 'num_tokens': 6010904.0, 'completions/mean_length': 142.125, 'completions/min_length': 140.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.125, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9992616176605225, 'rewards/meter/std': 9.245655382983387e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.874354362487793, 'rewards/total_composite/std': 0.050480589270591736, 'reward': 0.874354362487793, 'reward_std': 0.050480592995882034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01351867988705635, 'sampling/sampling_logp_difference/max': 1.036015510559082, 'sampling/importance_sampling_ratio/min': 0.3548658490180969, 'sampling/importance_sampling_ratio/mean': 1.0047122240066528, 'sampling/importance_sampling_ratio/max': 1.3792026042938232, 'entropy': 0.09845188539475203, 'clip_ratio/low_mean': 0.004389182257000357, 'clip_ratio/low_min': 0.004389182257000357, 'clip_ratio/high_mean': 0.0017361111240461469, 'clip_ratio/high_max': 0.0017361111240461469, 'clip_ratio/region_mean': 0.006125293381046504, 'reward_total_mean': 0.874354362487793, 'reward_meter_mean': 0.9992616176605225, 'reward_meter_std': 9.245655382983387e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.874354362487793, 'reward_total_composite_std': 0.050480589270591736, 'epoch': 0.11}

 80%|████████  | 2652/3300 [3:52:26<3:53:10, 21.59s/it]INFO 04-12 02:29:18 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2653/3300 [3:52:35<3:11:57, 17.80s/it]2026-04-12 02:29:26,900 | INFO | train_grpo_train | metrics_logged mode=train step=2653

                                                       
{'loss': -0.0161, 'grad_norm': 1.9425232410430908, 'learning_rate': 1.9636363636363636e-06, 'num_tokens': 6015870.0, 'completions/mean_length': 414.75, 'completions/min_length': 393.0, 'completions/max_length': 425.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 414.75, 'completions/min_terminated_length': 393.0, 'completions/max_terminated_length': 425.0, 'rewards/meter/mean': 0.9989011883735657, 'rewards/meter/std': 0.0002041146653937176, 'rewards/count_adherence/mean': 0.7767857313156128, 'rewards/count_adherence/std': 0.02525380253791809, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9577068090438843, 'rewards/repeat_penalty/std': 0.017178824171423912, 'rewards/total_composite/mean': 0.7432065010070801, 'rewards/total_composite/std': 0.030130181461572647, 'reward': 0.7432065010070801, 'reward_std': 0.03013017773628235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04433398321270943, 'sampling/sampling_logp_difference/max': 1.5436639785766602, 'sampling/importance_sampling_ratio/min': 0.21359705924987793, 'sampling/importance_sampling_ratio/mean': 1.0107157230377197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3583802431821823, 'clip_ratio/low_mean': 0.0025445292703807354, 'clip_ratio/low_min': 0.0025445292703807354, 'clip_ratio/high_mean': 0.02485937112942338, 'clip_ratio/high_max': 0.02485937112942338, 'clip_ratio/region_mean': 0.027403900399804115, 'reward_total_mean': 0.7432065010070801, 'reward_meter_mean': 0.9989011883735657, 'reward_meter_std': 0.0002041146653937176, 'reward_count_adherence_mean': 0.7767857313156128, 'reward_count_adherence_std': 0.02525380253791809, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9577068090438843, 'reward_repeat_penalty_std': 0.017178824171423912, 'reward_total_composite_mean': 0.7432065010070801, 'reward_total_composite_std': 0.030130181461572647, 'epoch': 0.11}

 80%|████████  | 2653/3300 [3:52:35<3:11:57, 17.80s/it]INFO 04-12 02:29:27 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2654/3300 [3:52:40<2:29:25, 13.88s/it]2026-04-12 02:29:31,628 | INFO | train_grpo_train | metrics_logged mode=train step=2654

                                                       
{'loss': -0.0037, 'grad_norm': 1.9117748737335205, 'learning_rate': 1.960606060606061e-06, 'num_tokens': 6017843.0, 'completions/mean_length': 72.625, 'completions/min_length': 70.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9993941783905029, 'rewards/meter/std': 0.00015994130808394402, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993941783905029, 'rewards/total_composite/std': 0.00015994130808394402, 'reward': 0.9993941783905029, 'reward_std': 0.0001599277020432055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01229790784418583, 'sampling/sampling_logp_difference/max': 0.7019014358520508, 'sampling/importance_sampling_ratio/min': 0.4956419765949249, 'sampling/importance_sampling_ratio/mean': 1.0008054971694946, 'sampling/importance_sampling_ratio/max': 1.5019214153289795, 'entropy': 0.06974478112533689, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.0017361111240461469, 'clip_ratio/high_max': 0.0017361111240461469, 'clip_ratio/region_mean': 0.0035218254197388887, 'reward_total_mean': 0.9993941783905029, 'reward_meter_mean': 0.9993941783905029, 'reward_meter_std': 0.00015994130808394402, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993941783905029, 'reward_total_composite_std': 0.00015994130808394402, 'epoch': 0.11}

 80%|████████  | 2654/3300 [3:52:40<2:29:25, 13.88s/it]INFO 04-12 02:29:31 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2655/3300 [3:52:45<1:59:21, 11.10s/it]2026-04-12 02:29:36,257 | INFO | train_grpo_train | metrics_logged mode=train step=2655

                                                       
{'loss': -0.014, 'grad_norm': 9.298754692077637, 'learning_rate': 1.9575757575757577e-06, 'num_tokens': 6019661.0, 'completions/mean_length': 71.25, 'completions/min_length': 67.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9782952070236206, 'rewards/meter/std': 0.013127398677170277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9782952070236206, 'rewards/total_composite/std': 0.013127398677170277, 'reward': 0.9782952070236206, 'reward_std': 0.01312740333378315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06088637933135033, 'sampling/sampling_logp_difference/max': 2.4444236755371094, 'sampling/importance_sampling_ratio/min': 0.08677612990140915, 'sampling/importance_sampling_ratio/mean': 1.0034326314926147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4243501238524914, 'clip_ratio/low_mean': 0.024847375229001045, 'clip_ratio/low_min': 0.024847375229001045, 'clip_ratio/high_mean': 0.017125858925282955, 'clip_ratio/high_max': 0.017125858925282955, 'clip_ratio/region_mean': 0.041973234154284, 'reward_total_mean': 0.9782952070236206, 'reward_meter_mean': 0.9782952070236206, 'reward_meter_std': 0.013127398677170277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9782952070236206, 'reward_total_composite_std': 0.013127398677170277, 'epoch': 0.11}

 80%|████████  | 2655/3300 [3:52:45<1:59:21, 11.10s/it]INFO 04-12 02:29:36 [block_pool.py:378] Successfully reset prefix cache

 80%|████████  | 2656/3300 [3:52:49<1:37:54,  9.12s/it]2026-04-12 02:29:40,753 | INFO | train_grpo_train | metrics_logged mode=train step=2656

                                                       
{'loss': 0.0, 'grad_norm': 0.48045873641967773, 'learning_rate': 1.954545454545455e-06, 'num_tokens': 6021436.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981427788734436, 'rewards/meter/std': 1.307969432673417e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981427788734436, 'rewards/total_composite/std': 1.307969432673417e-05, 'reward': 0.9981427788734436, 'reward_std': 1.3079693417239469e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004913026466965675, 'sampling/sampling_logp_difference/max': 0.5399925708770752, 'sampling/importance_sampling_ratio/min': 0.5827525854110718, 'sampling/importance_sampling_ratio/mean': 1.0033403635025024, 'sampling/importance_sampling_ratio/max': 1.6235767602920532, 'entropy': 0.029891937039792538, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.9981427788734436, 'reward_meter_mean': 0.9981427788734436, 'reward_meter_std': 1.307969432673417e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981427788734436, 'reward_total_composite_std': 1.307969432673417e-05, 'epoch': 0.11}

 80%|████████  | 2656/3300 [3:52:49<1:37:54,  9.12s/it]INFO 04-12 02:29:40 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2657/3300 [3:52:54<1:23:20,  7.78s/it]2026-04-12 02:29:45,393 | INFO | train_grpo_train | metrics_logged mode=train step=2657

                                                       
{'loss': 0.0029, 'grad_norm': 6.030494213104248, 'learning_rate': 1.9515151515151518e-06, 'num_tokens': 6023229.0, 'completions/mean_length': 70.125, 'completions/min_length': 67.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9570232629776001, 'rewards/meter/std': 0.08651266247034073, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9570232629776001, 'rewards/total_composite/std': 0.08651266247034073, 'reward': 0.9570232629776001, 'reward_std': 0.08651266247034073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05514438450336456, 'sampling/sampling_logp_difference/max': 1.165654182434082, 'sampling/importance_sampling_ratio/min': 0.3117186725139618, 'sampling/importance_sampling_ratio/mean': 1.0104464292526245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44630854949355125, 'clip_ratio/low_mean': 0.0074352502124384046, 'clip_ratio/low_min': 0.0074352502124384046, 'clip_ratio/high_mean': 0.031580700539052486, 'clip_ratio/high_max': 0.031580700539052486, 'clip_ratio/region_mean': 0.03901595075149089, 'reward_total_mean': 0.9570232629776001, 'reward_meter_mean': 0.9570232629776001, 'reward_meter_std': 0.08651266247034073, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9570232629776001, 'reward_total_composite_std': 0.08651266247034073, 'epoch': 0.11}

 81%|████████  | 2657/3300 [3:52:54<1:23:20,  7.78s/it]INFO 04-12 02:29:45 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2658/3300 [3:52:58<1:12:34,  6.78s/it]2026-04-12 02:29:49,857 | INFO | train_grpo_train | metrics_logged mode=train step=2658

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.9484848484848486e-06, 'num_tokens': 6025013.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00031924626091495156, 'sampling/sampling_logp_difference/max': 0.01690489798784256, 'sampling/importance_sampling_ratio/min': 0.9832371473312378, 'sampling/importance_sampling_ratio/mean': 1.0001012086868286, 'sampling/importance_sampling_ratio/max': 1.0096546411514282, 'entropy': 0.0046398533741012216, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 81%|████████  | 2658/3300 [3:52:58<1:12:34,  6.78s/it]INFO 04-12 02:29:50 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2659/3300 [3:53:02<1:03:48,  5.97s/it]2026-04-12 02:29:53,938 | INFO | train_grpo_train | metrics_logged mode=train step=2659

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.945454545454546e-06, 'num_tokens': 6026461.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00031867477810010314, 'sampling/sampling_logp_difference/max': 0.00753726065158844, 'sampling/importance_sampling_ratio/min': 0.9997876882553101, 'sampling/importance_sampling_ratio/mean': 1.0003173351287842, 'sampling/importance_sampling_ratio/max': 1.0075657367706299, 'entropy': 0.002239607958472334, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 81%|████████  | 2659/3300 [3:53:02<1:03:48,  5.97s/it]INFO 04-12 02:29:54 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2660/3300 [3:53:07<59:42,  5.60s/it]  2026-04-12 02:29:58,661 | INFO | train_grpo_train | metrics_logged mode=train step=2660

                                                     
{'loss': 0.0104, 'grad_norm': 11.615287780761719, 'learning_rate': 1.9424242424242427e-06, 'num_tokens': 6028445.0, 'completions/mean_length': 65.0, 'completions/min_length': 63.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9398962259292603, 'rewards/meter/std': 0.01433228887617588, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9398962259292603, 'rewards/total_composite/std': 0.01433228887617588, 'reward': 0.9398962259292603, 'reward_std': 0.014332284219563007, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0898202583193779, 'sampling/sampling_logp_difference/max': 2.166830062866211, 'sampling/importance_sampling_ratio/min': 0.22292140126228333, 'sampling/importance_sampling_ratio/mean': 1.0091454982757568, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42190510779619217, 'clip_ratio/low_mean': 0.0330266528762877, 'clip_ratio/low_min': 0.0330266528762877, 'clip_ratio/high_mean': 0.03413416654802859, 'clip_ratio/high_max': 0.03413416654802859, 'clip_ratio/region_mean': 0.06716081942431629, 'reward_total_mean': 0.9398962259292603, 'reward_meter_mean': 0.9398962259292603, 'reward_meter_std': 0.01433228887617588, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9398962259292603, 'reward_total_composite_std': 0.01433228887617588, 'epoch': 0.11}

 81%|████████  | 2660/3300 [3:53:07<59:42,  5.60s/it]INFO 04-12 02:29:58 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2661/3300 [3:53:12<56:02,  5.26s/it]2026-04-12 02:30:03,138 | INFO | train_grpo_train | metrics_logged mode=train step=2661

                                                     
{'loss': -0.0063, 'grad_norm': 3.7002193927764893, 'learning_rate': 1.9393939393939395e-06, 'num_tokens': 6030265.0, 'completions/mean_length': 60.5, 'completions/min_length': 57.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9970978498458862, 'rewards/meter/std': 0.00044736076961271465, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970978498458862, 'rewards/total_composite/std': 0.00044736076961271465, 'reward': 0.9970978498458862, 'reward_std': 0.0004473589942790568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011318504810333252, 'sampling/sampling_logp_difference/max': 0.8053483963012695, 'sampling/importance_sampling_ratio/min': 0.642192542552948, 'sampling/importance_sampling_ratio/mean': 1.0017904043197632, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.053432762157171965, 'clip_ratio/low_mean': 0.004385964944958687, 'clip_ratio/low_min': 0.004385964944958687, 'clip_ratio/high_mean': 0.008196720853447914, 'clip_ratio/high_max': 0.008196720853447914, 'clip_ratio/region_mean': 0.012582685798406601, 'reward_total_mean': 0.9970978498458862, 'reward_meter_mean': 0.9970978498458862, 'reward_meter_std': 0.00044736076961271465, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970978498458862, 'reward_total_composite_std': 0.00044736076961271465, 'epoch': 0.11}

 81%|████████  | 2661/3300 [3:53:12<56:02,  5.26s/it]INFO 04-12 02:30:03 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2662/3300 [3:53:16<53:01,  4.99s/it]2026-04-12 02:30:07,485 | INFO | train_grpo_train | metrics_logged mode=train step=2662

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.9363636363636363e-06, 'num_tokens': 6031961.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002394376788288355, 'sampling/sampling_logp_difference/max': 0.002065029926598072, 'sampling/importance_sampling_ratio/min': 0.9995288252830505, 'sampling/importance_sampling_ratio/mean': 1.0002366304397583, 'sampling/importance_sampling_ratio/max': 1.0020672082901, 'entropy': 0.0019165669655194506, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 81%|████████  | 2662/3300 [3:53:16<53:01,  4.99s/it]INFO 04-12 02:30:07 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2663/3300 [3:53:21<52:55,  4.98s/it]2026-04-12 02:30:12,464 | INFO | train_grpo_train | metrics_logged mode=train step=2663

                                                     
{'loss': -0.0089, 'grad_norm': 3.3872175216674805, 'learning_rate': 1.9333333333333336e-06, 'num_tokens': 6034070.0, 'completions/mean_length': 90.625, 'completions/min_length': 85.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.625, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9972575306892395, 'rewards/meter/std': 0.0004964250256307423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972575306892395, 'rewards/total_composite/std': 0.0004964250256307423, 'reward': 0.9972575306892395, 'reward_std': 0.0004964345716871321, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02105124667286873, 'sampling/sampling_logp_difference/max': 1.149244785308838, 'sampling/importance_sampling_ratio/min': 0.31687599420547485, 'sampling/importance_sampling_ratio/mean': 0.9998160600662231, 'sampling/importance_sampling_ratio/max': 1.9014232158660889, 'entropy': 0.10070738103240728, 'clip_ratio/low_mean': 0.007012551999650896, 'clip_ratio/low_min': 0.007012551999650896, 'clip_ratio/high_mean': 0.008197940303944051, 'clip_ratio/high_max': 0.008197940303944051, 'clip_ratio/region_mean': 0.015210492303594947, 'reward_total_mean': 0.9972575306892395, 'reward_meter_mean': 0.9972575306892395, 'reward_meter_std': 0.0004964250256307423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972575306892395, 'reward_total_composite_std': 0.0004964250256307423, 'epoch': 0.11}

 81%|████████  | 2663/3300 [3:53:21<52:55,  4.98s/it]INFO 04-12 02:30:12 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2664/3300 [3:53:26<51:49,  4.89s/it]2026-04-12 02:30:17,128 | INFO | train_grpo_train | metrics_logged mode=train step=2664

                                                     
{'loss': -0.0067, 'grad_norm': 1.5907872915267944, 'learning_rate': 1.9303030303030304e-06, 'num_tokens': 6035802.0, 'completions/mean_length': 70.5, 'completions/min_length': 69.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9993501305580139, 'rewards/meter/std': 0.0001369249657727778, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993501305580139, 'rewards/total_composite/std': 0.0001369249657727778, 'reward': 0.9993501305580139, 'reward_std': 0.0001369154779240489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019528387114405632, 'sampling/sampling_logp_difference/max': 0.8889780044555664, 'sampling/importance_sampling_ratio/min': 0.4110756516456604, 'sampling/importance_sampling_ratio/mean': 1.007887363433838, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10792513377964497, 'clip_ratio/low_mean': 0.014288630336523056, 'clip_ratio/low_min': 0.014288630336523056, 'clip_ratio/high_mean': 0.0017361111240461469, 'clip_ratio/high_max': 0.0017361111240461469, 'clip_ratio/region_mean': 0.016024741460569203, 'reward_total_mean': 0.9993501305580139, 'reward_meter_mean': 0.9993501305580139, 'reward_meter_std': 0.0001369249657727778, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993501305580139, 'reward_total_composite_std': 0.0001369249657727778, 'epoch': 0.11}

 81%|████████  | 2664/3300 [3:53:26<51:49,  4.89s/it]INFO 04-12 02:30:17 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2665/3300 [3:53:30<50:54,  4.81s/it]2026-04-12 02:30:21,766 | INFO | train_grpo_train | metrics_logged mode=train step=2665

                                                     
{'loss': -0.0018, 'grad_norm': 12.907752990722656, 'learning_rate': 1.9272727272727273e-06, 'num_tokens': 6037606.0, 'completions/mean_length': 64.5, 'completions/min_length': 61.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8035573959350586, 'rewards/meter/std': 0.2991119921207428, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8035573959350586, 'rewards/total_composite/std': 0.2991119921207428, 'reward': 0.8035573959350586, 'reward_std': 0.2991119921207428, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06501687318086624, 'sampling/sampling_logp_difference/max': 1.923792839050293, 'sampling/importance_sampling_ratio/min': 0.14605195820331573, 'sampling/importance_sampling_ratio/mean': 1.003564715385437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.25452338345348835, 'clip_ratio/low_mean': 0.012102970853447914, 'clip_ratio/low_min': 0.012102970853447914, 'clip_ratio/high_mean': 0.044199546333402395, 'clip_ratio/high_max': 0.044199546333402395, 'clip_ratio/region_mean': 0.05630251718685031, 'reward_total_mean': 0.8035573959350586, 'reward_meter_mean': 0.8035573959350586, 'reward_meter_std': 0.2991119921207428, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8035573959350586, 'reward_total_composite_std': 0.2991119921207428, 'epoch': 0.11}

 81%|████████  | 2665/3300 [3:53:30<50:54,  4.81s/it]INFO 04-12 02:30:22 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2666/3300 [3:53:35<49:47,  4.71s/it]2026-04-12 02:30:26,241 | INFO | train_grpo_train | metrics_logged mode=train step=2666

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.924242424242424e-06, 'num_tokens': 6039238.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00027814743225462735, 'sampling/sampling_logp_difference/max': 0.00260642496868968, 'sampling/importance_sampling_ratio/min': 0.9999665021896362, 'sampling/importance_sampling_ratio/mean': 1.000278115272522, 'sampling/importance_sampling_ratio/max': 1.0026098489761353, 'entropy': 0.002076218879665248, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 81%|████████  | 2666/3300 [3:53:35<49:47,  4.71s/it]INFO 04-12 02:30:26 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2667/3300 [3:53:40<50:27,  4.78s/it]2026-04-12 02:30:31,186 | INFO | train_grpo_train | metrics_logged mode=train step=2667

                                                     
{'loss': 0.0, 'grad_norm': 0.08655895292758942, 'learning_rate': 1.9212121212121213e-06, 'num_tokens': 6041406.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9993359446525574, 'rewards/meter/std': 2.59008470493427e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993359446525574, 'rewards/total_composite/std': 2.59008470493427e-06, 'reward': 0.9993359446525574, 'reward_std': 2.6060056370624807e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0031785282772034407, 'sampling/sampling_logp_difference/max': 0.29243287444114685, 'sampling/importance_sampling_ratio/min': 0.74644535779953, 'sampling/importance_sampling_ratio/mean': 1.0002384185791016, 'sampling/importance_sampling_ratio/max': 1.1448330879211426, 'entropy': 0.021847927011549473, 'clip_ratio/low_mean': 0.0038265305338427424, 'clip_ratio/low_min': 0.0038265305338427424, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0038265305338427424, 'reward_total_mean': 0.9993359446525574, 'reward_meter_mean': 0.9993359446525574, 'reward_meter_std': 2.59008470493427e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993359446525574, 'reward_total_composite_std': 2.59008470493427e-06, 'epoch': 0.11}

 81%|████████  | 2667/3300 [3:53:40<50:27,  4.78s/it]INFO 04-12 02:30:31 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2668/3300 [3:53:44<48:37,  4.62s/it]2026-04-12 02:30:35,417 | INFO | train_grpo_train | metrics_logged mode=train step=2668

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.918181818181818e-06, 'num_tokens': 6043074.0, 'completions/mean_length': 36.5, 'completions/min_length': 36.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9996045231819153, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9996045231819153, 'rewards/total_composite/std': 0.0, 'reward': 0.9996045231819153, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.005902368109673262, 'sampling/sampling_logp_difference/max': 0.8077952861785889, 'sampling/importance_sampling_ratio/min': 0.44583994150161743, 'sampling/importance_sampling_ratio/mean': 0.9983490109443665, 'sampling/importance_sampling_ratio/max': 1.073022723197937, 'entropy': 0.03415517252869904, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9996045231819153, 'reward_meter_mean': 0.9996045231819153, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9996045231819153, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 81%|████████  | 2668/3300 [3:53:44<48:37,  4.62s/it]INFO 04-12 02:30:35 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2669/3300 [3:53:49<51:27,  4.89s/it]2026-04-12 02:30:40,951 | INFO | train_grpo_train | metrics_logged mode=train step=2669

                                                     
{'loss': -0.0093, 'grad_norm': 5.236100673675537, 'learning_rate': 1.9151515151515154e-06, 'num_tokens': 6045151.0, 'completions/mean_length': 102.625, 'completions/min_length': 98.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.625, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9903435707092285, 'rewards/meter/std': 0.007074963301420212, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9903435707092285, 'rewards/total_composite/std': 0.007074963301420212, 'reward': 0.9903435707092285, 'reward_std': 0.00707497913390398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04934058338403702, 'sampling/sampling_logp_difference/max': 1.2319703102111816, 'sampling/importance_sampling_ratio/min': 0.2917172312736511, 'sampling/importance_sampling_ratio/mean': 1.001228928565979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39514047652482986, 'clip_ratio/low_mean': 0.006191797088831663, 'clip_ratio/low_min': 0.006191797088831663, 'clip_ratio/high_mean': 0.041350313229486346, 'clip_ratio/high_max': 0.041350313229486346, 'clip_ratio/region_mean': 0.04754211031831801, 'reward_total_mean': 0.9903435707092285, 'reward_meter_mean': 0.9903435707092285, 'reward_meter_std': 0.007074963301420212, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9903435707092285, 'reward_total_composite_std': 0.007074963301420212, 'epoch': 0.11}

 81%|████████  | 2669/3300 [3:53:49<51:27,  4.89s/it]INFO 04-12 02:30:41 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2670/3300 [3:53:54<50:22,  4.80s/it]2026-04-12 02:30:45,533 | INFO | train_grpo_train | metrics_logged mode=train step=2670

                                                     
{'loss': -0.0, 'grad_norm': 0.37225642800331116, 'learning_rate': 1.9121212121212123e-06, 'num_tokens': 6046947.0, 'completions/mean_length': 66.5, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981483817100525, 'rewards/meter/std': 1.0958180610032286e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981483817100525, 'rewards/total_composite/std': 1.0958180610032286e-05, 'reward': 0.9981483817100525, 'reward_std': 1.0958180610032286e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006186656653881073, 'sampling/sampling_logp_difference/max': 1.6149165630340576, 'sampling/importance_sampling_ratio/min': 0.1989072561264038, 'sampling/importance_sampling_ratio/mean': 1.0005658864974976, 'sampling/importance_sampling_ratio/max': 1.4169889688491821, 'entropy': 0.016892601503059268, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.9981483817100525, 'reward_meter_mean': 0.9981483817100525, 'reward_meter_std': 1.0958180610032286e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981483817100525, 'reward_total_composite_std': 1.0958180610032286e-05, 'epoch': 0.11}

 81%|████████  | 2670/3300 [3:53:54<50:22,  4.80s/it]INFO 04-12 02:30:45 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2671/3300 [3:53:58<49:22,  4.71s/it]2026-04-12 02:30:50,034 | INFO | train_grpo_train | metrics_logged mode=train step=2671

                                                     
{'loss': 0.0007, 'grad_norm': 7.428125381469727, 'learning_rate': 1.9090909090909095e-06, 'num_tokens': 6048617.0, 'completions/mean_length': 60.75, 'completions/min_length': 60.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.75, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9970404505729675, 'rewards/meter/std': 0.00046239994117058814, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970404505729675, 'rewards/total_composite/std': 0.00046239994117058814, 'reward': 0.9970404505729675, 'reward_std': 0.0004624156281352043, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014924735762178898, 'sampling/sampling_logp_difference/max': 0.8970005512237549, 'sampling/importance_sampling_ratio/min': 0.4077909588813782, 'sampling/importance_sampling_ratio/mean': 1.0006794929504395, 'sampling/importance_sampling_ratio/max': 1.6209132671356201, 'entropy': 0.07036425499245524, 'clip_ratio/low_mean': 0.014549180399626493, 'clip_ratio/low_min': 0.014549180399626493, 'clip_ratio/high_mean': 0.004098360426723957, 'clip_ratio/high_max': 0.004098360426723957, 'clip_ratio/region_mean': 0.01864754082635045, 'reward_total_mean': 0.9970404505729675, 'reward_meter_mean': 0.9970404505729675, 'reward_meter_std': 0.00046239994117058814, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970404505729675, 'reward_total_composite_std': 0.00046239994117058814, 'epoch': 0.11}

 81%|████████  | 2671/3300 [3:53:59<49:22,  4.71s/it]INFO 04-12 02:30:50 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2672/3300 [3:54:07<59:55,  5.72s/it]2026-04-12 02:30:58,127 | INFO | train_grpo_train | metrics_logged mode=train step=2672

                                                     
{'loss': 0.0098, 'grad_norm': 6.031638145446777, 'learning_rate': 1.9060606060606064e-06, 'num_tokens': 6052766.0, 'completions/mean_length': 316.625, 'completions/min_length': 309.0, 'completions/max_length': 328.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 316.625, 'completions/min_terminated_length': 309.0, 'completions/max_terminated_length': 328.0, 'rewards/meter/mean': 0.9499752521514893, 'rewards/meter/std': 0.06226501986384392, 'rewards/count_adherence/mean': 0.8409091234207153, 'rewards/count_adherence/std': 0.04208274558186531, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.8695175647735596, 'rewards/repeat_penalty/std': 0.058811455965042114, 'rewards/total_composite/mean': 0.6005241870880127, 'rewards/total_composite/std': 0.24652278423309326, 'reward': 0.6005241870880127, 'reward_std': 0.24652279913425446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04524172842502594, 'sampling/sampling_logp_difference/max': 1.6112018823623657, 'sampling/importance_sampling_ratio/min': 0.19964751601219177, 'sampling/importance_sampling_ratio/mean': 1.0106778144836426, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46788543090224266, 'clip_ratio/low_mean': 0.0011718750465661287, 'clip_ratio/low_min': 0.0011718750465661287, 'clip_ratio/high_mean': 0.022553268587216735, 'clip_ratio/high_max': 0.022553268587216735, 'clip_ratio/region_mean': 0.023725143633782864, 'reward_total_mean': 0.6005241870880127, 'reward_meter_mean': 0.9499752521514893, 'reward_meter_std': 0.06226501986384392, 'reward_count_adherence_mean': 0.8409091234207153, 'reward_count_adherence_std': 0.04208274558186531, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.8695175647735596, 'reward_repeat_penalty_std': 0.058811455965042114, 'reward_total_composite_mean': 0.6005241870880127, 'reward_total_composite_std': 0.24652278423309326, 'epoch': 0.11}

 81%|████████  | 2672/3300 [3:54:07<59:55,  5.72s/it]INFO 04-12 02:30:58 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2673/3300 [3:54:11<56:49,  5.44s/it]2026-04-12 02:31:02,893 | INFO | train_grpo_train | metrics_logged mode=train step=2673

                                                     
{'loss': -0.0033, 'grad_norm': 1.5127347707748413, 'learning_rate': 1.9030303030303032e-06, 'num_tokens': 6054615.0, 'completions/mean_length': 79.125, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.99886155128479, 'rewards/meter/std': 0.00016637769294902682, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99886155128479, 'rewards/total_composite/std': 0.00016637769294902682, 'reward': 0.99886155128479, 'reward_std': 0.00016637658700346947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0236604493111372, 'sampling/sampling_logp_difference/max': 0.8233041763305664, 'sampling/importance_sampling_ratio/min': 0.43897879123687744, 'sampling/importance_sampling_ratio/mean': 1.0049008131027222, 'sampling/importance_sampling_ratio/max': 1.7465615272521973, 'entropy': 0.12998810317367315, 'clip_ratio/low_mean': 0.00795196380931884, 'clip_ratio/low_min': 0.00795196380931884, 'clip_ratio/high_mean': 0.006329114083200693, 'clip_ratio/high_max': 0.006329114083200693, 'clip_ratio/region_mean': 0.014281077892519534, 'reward_total_mean': 0.99886155128479, 'reward_meter_mean': 0.99886155128479, 'reward_meter_std': 0.00016637769294902682, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99886155128479, 'reward_total_composite_std': 0.00016637769294902682, 'epoch': 0.11}

 81%|████████  | 2673/3300 [3:54:11<56:49,  5.44s/it]INFO 04-12 02:31:03 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2674/3300 [3:54:18<1:01:55,  5.94s/it]2026-04-12 02:31:09,995 | INFO | train_grpo_train | metrics_logged mode=train step=2674

                                                       
{'loss': 0.0073, 'grad_norm': 2.1606380939483643, 'learning_rate': 1.9000000000000002e-06, 'num_tokens': 6058386.0, 'completions/mean_length': 268.375, 'completions/min_length': 262.0, 'completions/max_length': 272.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 268.375, 'completions/min_terminated_length': 262.0, 'completions/max_terminated_length': 272.0, 'rewards/meter/mean': 0.9988313317298889, 'rewards/meter/std': 0.0005680599715560675, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.942307710647583, 'rewards/repeat_penalty/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.9412178993225098, 'rewards/total_composite/std': 0.0545286163687706, 'reward': 0.9412178993225098, 'reward_std': 0.0545286163687706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043261099606752396, 'sampling/sampling_logp_difference/max': 1.4657955169677734, 'sampling/importance_sampling_ratio/min': 0.23089425265789032, 'sampling/importance_sampling_ratio/mean': 1.0066032409667969, 'sampling/importance_sampling_ratio/max': 1.7729755640029907, 'entropy': 0.33767288736999035, 'clip_ratio/low_mean': 0.012097747880034149, 'clip_ratio/low_min': 0.012097747880034149, 'clip_ratio/high_mean': 0.012176374206319451, 'clip_ratio/high_max': 0.012176374206319451, 'clip_ratio/region_mean': 0.0242741220863536, 'reward_total_mean': 0.9412178993225098, 'reward_meter_mean': 0.9988313317298889, 'reward_meter_std': 0.0005680599715560675, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.942307710647583, 'reward_repeat_penalty_std': 0.054392825812101364, 'reward_total_composite_mean': 0.9412178993225098, 'reward_total_composite_std': 0.0545286163687706, 'epoch': 0.11}

 81%|████████  | 2674/3300 [3:54:18<1:01:55,  5.94s/it]INFO 04-12 02:31:10 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2675/3300 [3:54:27<1:08:41,  6.59s/it]2026-04-12 02:31:18,132 | INFO | train_grpo_train | metrics_logged mode=train step=2675

                                                       
{'loss': 0.0064, 'grad_norm': 1.5513638257980347, 'learning_rate': 1.896969696969697e-06, 'num_tokens': 6062752.0, 'completions/mean_length': 340.75, 'completions/min_length': 337.0, 'completions/max_length': 346.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 340.75, 'completions/min_terminated_length': 337.0, 'completions/max_terminated_length': 346.0, 'rewards/meter/mean': 0.9985557794570923, 'rewards/meter/std': 0.0006976545555517077, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9117647409439087, 'rewards/repeat_penalty/std': 0.05446000397205353, 'rewards/total_composite/mean': 0.8193838000297546, 'rewards/total_composite/std': 0.04855939745903015, 'reward': 0.8193838000297546, 'reward_std': 0.04855942353606224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040726643055677414, 'sampling/sampling_logp_difference/max': 1.2356281280517578, 'sampling/importance_sampling_ratio/min': 0.2906521260738373, 'sampling/importance_sampling_ratio/mean': 1.0090699195861816, 'sampling/importance_sampling_ratio/max': 1.8999286890029907, 'entropy': 0.3350350670516491, 'clip_ratio/low_mean': 0.009543466847389936, 'clip_ratio/low_min': 0.009543466847389936, 'clip_ratio/high_mean': 0.014709851937368512, 'clip_ratio/high_max': 0.014709851937368512, 'clip_ratio/region_mean': 0.02425331878475845, 'reward_total_mean': 0.8193838000297546, 'reward_meter_mean': 0.9985557794570923, 'reward_meter_std': 0.0006976545555517077, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9117647409439087, 'reward_repeat_penalty_std': 0.05446000397205353, 'reward_total_composite_mean': 0.8193838000297546, 'reward_total_composite_std': 0.04855939745903015, 'epoch': 0.11}

 81%|████████  | 2675/3300 [3:54:27<1:08:41,  6.59s/it]INFO 04-12 02:31:18 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2676/3300 [3:54:31<1:01:46,  5.94s/it]2026-04-12 02:31:22,540 | INFO | train_grpo_train | metrics_logged mode=train step=2676

                                                       
{'loss': -0.0079, 'grad_norm': 3.5680060386657715, 'learning_rate': 1.8939393939393941e-06, 'num_tokens': 6064157.0, 'completions/mean_length': 34.625, 'completions/min_length': 34.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9994817972183228, 'rewards/meter/std': 0.0001635835214983672, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994817972183228, 'rewards/total_composite/std': 0.0001635835214983672, 'reward': 0.9994817972183228, 'reward_std': 0.00016358528228010982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015038218349218369, 'sampling/sampling_logp_difference/max': 0.7005414962768555, 'sampling/importance_sampling_ratio/min': 0.49631646275520325, 'sampling/importance_sampling_ratio/mean': 1.0051720142364502, 'sampling/importance_sampling_ratio/max': 1.5120035409927368, 'entropy': 0.07462144270539284, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.0071428571827709675, 'clip_ratio/high_max': 0.0071428571827709675, 'clip_ratio/region_mean': 0.010819327784702182, 'reward_total_mean': 0.9994817972183228, 'reward_meter_mean': 0.9994817972183228, 'reward_meter_std': 0.0001635835214983672, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994817972183228, 'reward_total_composite_std': 0.0001635835214983672, 'epoch': 0.11}

 81%|████████  | 2676/3300 [3:54:31<1:01:46,  5.94s/it]INFO 04-12 02:31:22 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2677/3300 [3:54:37<1:02:23,  6.01s/it]2026-04-12 02:31:28,713 | INFO | train_grpo_train | metrics_logged mode=train step=2677

                                                       
{'loss': 0.0237, 'grad_norm': 3.107592821121216, 'learning_rate': 1.890909090909091e-06, 'num_tokens': 6067073.0, 'completions/mean_length': 175.5, 'completions/min_length': 167.0, 'completions/max_length': 184.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 175.5, 'completions/min_terminated_length': 167.0, 'completions/max_terminated_length': 184.0, 'rewards/meter/mean': 0.9970433712005615, 'rewards/meter/std': 0.0006079506129026413, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9772727489471436, 'rewards/repeat_penalty/std': 0.04208271950483322, 'rewards/total_composite/mean': 0.974378228187561, 'rewards/total_composite/std': 0.041835807263851166, 'reward': 0.974378228187561, 'reward_std': 0.041835810989141464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033983275294303894, 'sampling/sampling_logp_difference/max': 2.001898765563965, 'sampling/importance_sampling_ratio/min': 0.135078564286232, 'sampling/importance_sampling_ratio/mean': 1.0057967901229858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19678651168942451, 'clip_ratio/low_mean': 0.007593599148094654, 'clip_ratio/low_min': 0.007593599148094654, 'clip_ratio/high_mean': 0.029550255741924047, 'clip_ratio/high_max': 0.029550255741924047, 'clip_ratio/region_mean': 0.0371438548900187, 'reward_total_mean': 0.974378228187561, 'reward_meter_mean': 0.9970433712005615, 'reward_meter_std': 0.0006079506129026413, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9772727489471436, 'reward_repeat_penalty_std': 0.04208271950483322, 'reward_total_composite_mean': 0.974378228187561, 'reward_total_composite_std': 0.041835807263851166, 'epoch': 0.11}

 81%|████████  | 2677/3300 [3:54:37<1:02:23,  6.01s/it]INFO 04-12 02:31:28 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2678/3300 [3:54:42<59:21,  5.73s/it]  2026-04-12 02:31:33,775 | INFO | train_grpo_train | metrics_logged mode=train step=2678

                                                     
{'loss': -0.0038, 'grad_norm': 2.5906176567077637, 'learning_rate': 1.887878787878788e-06, 'num_tokens': 6069055.0, 'completions/mean_length': 91.75, 'completions/min_length': 88.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.75, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9974672198295593, 'rewards/meter/std': 0.0002358894416829571, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974672198295593, 'rewards/total_composite/std': 0.0002358894416829571, 'reward': 0.9974672198295593, 'reward_std': 0.00023587615578435361, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025346064940094948, 'sampling/sampling_logp_difference/max': 1.5407707691192627, 'sampling/importance_sampling_ratio/min': 0.2142159342765808, 'sampling/importance_sampling_ratio/mean': 1.0014500617980957, 'sampling/importance_sampling_ratio/max': 1.6885485649108887, 'entropy': 0.12407996132969856, 'clip_ratio/low_mean': 0.004108626628294587, 'clip_ratio/low_min': 0.004108626628294587, 'clip_ratio/high_mean': 0.014815408736467361, 'clip_ratio/high_max': 0.014815408736467361, 'clip_ratio/region_mean': 0.01892403536476195, 'reward_total_mean': 0.9974672198295593, 'reward_meter_mean': 0.9974672198295593, 'reward_meter_std': 0.0002358894416829571, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974672198295593, 'reward_total_composite_std': 0.0002358894416829571, 'epoch': 0.11}

 81%|████████  | 2678/3300 [3:54:42<59:21,  5.73s/it]INFO 04-12 02:31:34 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2679/3300 [3:54:47<56:04,  5.42s/it]2026-04-12 02:31:38,467 | INFO | train_grpo_train | metrics_logged mode=train step=2679

                                                     
{'loss': 0.0065, 'grad_norm': 3.563011646270752, 'learning_rate': 1.884848484848485e-06, 'num_tokens': 6070908.0, 'completions/mean_length': 67.625, 'completions/min_length': 64.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9295022487640381, 'rewards/meter/std': 0.1213918924331665, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9295022487640381, 'rewards/total_composite/std': 0.1213918924331665, 'reward': 0.9295022487640381, 'reward_std': 0.12139188498258591, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034729912877082825, 'sampling/sampling_logp_difference/max': 1.1639995574951172, 'sampling/importance_sampling_ratio/min': 0.31223487854003906, 'sampling/importance_sampling_ratio/mean': 1.0032422542572021, 'sampling/importance_sampling_ratio/max': 1.9054324626922607, 'entropy': 0.2862533424049616, 'clip_ratio/low_mean': 0.007462686393409967, 'clip_ratio/low_min': 0.007462686393409967, 'clip_ratio/high_mean': 0.021945541491732, 'clip_ratio/high_max': 0.021945541491732, 'clip_ratio/region_mean': 0.02940822788514197, 'reward_total_mean': 0.9295022487640381, 'reward_meter_mean': 0.9295022487640381, 'reward_meter_std': 0.1213918924331665, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9295022487640381, 'reward_total_composite_std': 0.1213918924331665, 'epoch': 0.11}

 81%|████████  | 2679/3300 [3:54:47<56:04,  5.42s/it]INFO 04-12 02:31:38 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2680/3300 [3:54:52<54:24,  5.27s/it]2026-04-12 02:31:43,375 | INFO | train_grpo_train | metrics_logged mode=train step=2680

                                                     
{'loss': -0.0195, 'grad_norm': 8.263280868530273, 'learning_rate': 1.8818181818181819e-06, 'num_tokens': 6072994.0, 'completions/mean_length': 89.75, 'completions/min_length': 83.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.75, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9924962520599365, 'rewards/meter/std': 0.013577880337834358, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924962520599365, 'rewards/total_composite/std': 0.013577880337834358, 'reward': 0.9924962520599365, 'reward_std': 0.01357787661254406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030782943591475487, 'sampling/sampling_logp_difference/max': 1.3811171054840088, 'sampling/importance_sampling_ratio/min': 0.2512976825237274, 'sampling/importance_sampling_ratio/mean': 0.9978711605072021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12778305634856224, 'clip_ratio/low_mean': 0.0015060240402817726, 'clip_ratio/low_min': 0.0015060240402817726, 'clip_ratio/high_mean': 0.024966524448245764, 'clip_ratio/high_max': 0.024966524448245764, 'clip_ratio/region_mean': 0.026472548488527536, 'reward_total_mean': 0.9924962520599365, 'reward_meter_mean': 0.9924962520599365, 'reward_meter_std': 0.013577880337834358, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924962520599365, 'reward_total_composite_std': 0.013577880337834358, 'epoch': 0.11}

 81%|████████  | 2680/3300 [3:54:52<54:24,  5.27s/it]INFO 04-12 02:31:43 [block_pool.py:378] Successfully reset prefix cache

 81%|████████  | 2681/3300 [3:54:57<54:39,  5.30s/it]2026-04-12 02:31:48,752 | INFO | train_grpo_train | metrics_logged mode=train step=2681

                                                     
{'loss': -0.0005, 'grad_norm': 2.0896339416503906, 'learning_rate': 1.878787878787879e-06, 'num_tokens': 6075627.0, 'completions/mean_length': 132.125, 'completions/min_length': 131.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.125, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9991657733917236, 'rewards/meter/std': 0.0002873947087209672, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991657733917236, 'rewards/total_composite/std': 0.0002873947087209672, 'reward': 0.9991657733917236, 'reward_std': 0.00028738551191054285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015285594388842583, 'sampling/sampling_logp_difference/max': 1.4336557388305664, 'sampling/importance_sampling_ratio/min': 0.23843567073345184, 'sampling/importance_sampling_ratio/mean': 0.9993587136268616, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.058492994867265224, 'clip_ratio/low_mean': 0.006650691735558212, 'clip_ratio/low_min': 0.006650691735558212, 'clip_ratio/high_mean': 0.004713488393463194, 'clip_ratio/high_max': 0.004713488393463194, 'clip_ratio/region_mean': 0.011364180129021406, 'reward_total_mean': 0.9991657733917236, 'reward_meter_mean': 0.9991657733917236, 'reward_meter_std': 0.0002873947087209672, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991657733917236, 'reward_total_composite_std': 0.0002873947087209672, 'epoch': 0.11}

 81%|████████  | 2681/3300 [3:54:57<54:39,  5.30s/it]INFO 04-12 02:31:48 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2682/3300 [3:55:02<53:17,  5.17s/it]2026-04-12 02:31:53,634 | INFO | train_grpo_train | metrics_logged mode=train step=2682

                                                     
{'loss': 0.0047, 'grad_norm': 2.7295045852661133, 'learning_rate': 1.8757575757575757e-06, 'num_tokens': 6077792.0, 'completions/mean_length': 92.625, 'completions/min_length': 91.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.625, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9971524477005005, 'rewards/meter/std': 0.0005293982103466988, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971524477005005, 'rewards/total_composite/std': 0.0005293982103466988, 'reward': 0.9971524477005005, 'reward_std': 0.0005293957656249404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03304116055369377, 'sampling/sampling_logp_difference/max': 1.4015769958496094, 'sampling/importance_sampling_ratio/min': 0.24620838463306427, 'sampling/importance_sampling_ratio/mean': 1.0033948421478271, 'sampling/importance_sampling_ratio/max': 1.9131104946136475, 'entropy': 0.1409629676491022, 'clip_ratio/low_mean': 0.021640333347022533, 'clip_ratio/low_min': 0.021640333347022533, 'clip_ratio/high_mean': 0.010841309442184865, 'clip_ratio/high_max': 0.010841309442184865, 'clip_ratio/region_mean': 0.0324816427892074, 'reward_total_mean': 0.9971524477005005, 'reward_meter_mean': 0.9971524477005005, 'reward_meter_std': 0.0005293982103466988, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971524477005005, 'reward_total_composite_std': 0.0005293982103466988, 'epoch': 0.11}

 81%|████████▏ | 2682/3300 [3:55:02<53:17,  5.17s/it]INFO 04-12 02:31:53 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2683/3300 [3:55:06<50:35,  4.92s/it]2026-04-12 02:31:57,961 | INFO | train_grpo_train | metrics_logged mode=train step=2683

                                                     
{'loss': -0.0069, 'grad_norm': 13.13068962097168, 'learning_rate': 1.872727272727273e-06, 'num_tokens': 6079318.0, 'completions/mean_length': 42.75, 'completions/min_length': 42.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.953346848487854, 'rewards/meter/std': 0.009581836871802807, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.953346848487854, 'rewards/total_composite/std': 0.009581836871802807, 'reward': 0.953346848487854, 'reward_std': 0.009581834077835083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06227937340736389, 'sampling/sampling_logp_difference/max': 1.1528511047363281, 'sampling/importance_sampling_ratio/min': 0.3157352805137634, 'sampling/importance_sampling_ratio/mean': 1.0015714168548584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28468130715191364, 'clip_ratio/low_mean': 0.02360188332386315, 'clip_ratio/low_min': 0.02360188332386315, 'clip_ratio/high_mean': 0.014336680993437767, 'clip_ratio/high_max': 0.014336680993437767, 'clip_ratio/region_mean': 0.037938564317300916, 'reward_total_mean': 0.953346848487854, 'reward_meter_mean': 0.953346848487854, 'reward_meter_std': 0.009581836871802807, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.953346848487854, 'reward_total_composite_std': 0.009581836871802807, 'epoch': 0.11}

 81%|████████▏ | 2683/3300 [3:55:06<50:35,  4.92s/it]INFO 04-12 02:31:58 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2684/3300 [3:55:11<48:19,  4.71s/it]2026-04-12 02:32:02,174 | INFO | train_grpo_train | metrics_logged mode=train step=2684

                                                     
{'loss': -0.0002, 'grad_norm': 0.11564414203166962, 'learning_rate': 1.86969696969697e-06, 'num_tokens': 6081094.0, 'completions/mean_length': 41.0, 'completions/min_length': 41.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9985654354095459, 'rewards/meter/std': 4.657397312257672e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9985654354095459, 'rewards/total_composite/std': 4.657397312257672e-06, 'reward': 0.9985654354095459, 'reward_std': 4.660256763600046e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0033583450131118298, 'sampling/sampling_logp_difference/max': 0.5027565956115723, 'sampling/importance_sampling_ratio/min': 0.6048609614372253, 'sampling/importance_sampling_ratio/mean': 1.0005881786346436, 'sampling/importance_sampling_ratio/max': 1.0496618747711182, 'entropy': 0.01764180068857968, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9985654354095459, 'reward_meter_mean': 0.9985654354095459, 'reward_meter_std': 4.657397312257672e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9985654354095459, 'reward_total_composite_std': 4.657397312257672e-06, 'epoch': 0.11}

 81%|████████▏ | 2684/3300 [3:55:11<48:19,  4.71s/it]INFO 04-12 02:32:02 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2685/3300 [3:55:15<47:26,  4.63s/it]2026-04-12 02:32:06,620 | INFO | train_grpo_train | metrics_logged mode=train step=2685

                                                     
{'loss': -0.0036, 'grad_norm': 5.177779197692871, 'learning_rate': 1.8666666666666669e-06, 'num_tokens': 6082730.0, 'completions/mean_length': 60.5, 'completions/min_length': 57.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9971067905426025, 'rewards/meter/std': 0.0004163467965554446, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971067905426025, 'rewards/total_composite/std': 0.0004163467965554446, 'reward': 0.9971067905426025, 'reward_std': 0.0004163416160736233, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014101680368185043, 'sampling/sampling_logp_difference/max': 0.9058688879013062, 'sampling/importance_sampling_ratio/min': 0.4041905701160431, 'sampling/importance_sampling_ratio/mean': 1.0009123086929321, 'sampling/importance_sampling_ratio/max': 1.7392834424972534, 'entropy': 0.06518607307225466, 'clip_ratio/low_mean': 0.004385964944958687, 'clip_ratio/low_min': 0.004385964944958687, 'clip_ratio/high_mean': 0.006147540640085936, 'clip_ratio/high_max': 0.006147540640085936, 'clip_ratio/region_mean': 0.010533505585044622, 'reward_total_mean': 0.9971067905426025, 'reward_meter_mean': 0.9971067905426025, 'reward_meter_std': 0.0004163467965554446, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971067905426025, 'reward_total_composite_std': 0.0004163467965554446, 'epoch': 0.11}

 81%|████████▏ | 2685/3300 [3:55:15<47:26,  4.63s/it]INFO 04-12 02:32:06 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2686/3300 [3:55:19<45:54,  4.49s/it]2026-04-12 02:32:10,774 | INFO | train_grpo_train | metrics_logged mode=train step=2686

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.863636363636364e-06, 'num_tokens': 6084138.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005363853415474296, 'sampling/sampling_logp_difference/max': 0.028025653213262558, 'sampling/importance_sampling_ratio/min': 0.9723634123802185, 'sampling/importance_sampling_ratio/mean': 0.999911904335022, 'sampling/importance_sampling_ratio/max': 1.010701298713684, 'entropy': 0.0043548993999138474, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 81%|████████▏ | 2686/3300 [3:55:19<45:54,  4.49s/it]INFO 04-12 02:32:11 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2687/3300 [3:55:29<1:02:40,  6.13s/it]2026-04-12 02:32:20,748 | INFO | train_grpo_train | metrics_logged mode=train step=2687

                                                       
{'loss': -0.0234, 'grad_norm': 1.300607681274414, 'learning_rate': 1.8606060606060607e-06, 'num_tokens': 6090080.0, 'completions/mean_length': 472.75, 'completions/min_length': 433.0, 'completions/max_length': 491.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 472.75, 'completions/min_terminated_length': 433.0, 'completions/max_terminated_length': 491.0, 'rewards/meter/mean': 0.998869776725769, 'rewards/meter/std': 0.00026758189778774977, 'rewards/count_adherence/mean': 0.7734375, 'rewards/count_adherence/std': 0.04650149121880531, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.946873664855957, 'rewards/repeat_penalty/std': 0.06902104616165161, 'rewards/total_composite/mean': 0.7318886518478394, 'rewards/total_composite/std': 0.0725080594420433, 'reward': 0.7318886518478394, 'reward_std': 0.07250804454088211, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03919937461614609, 'sampling/sampling_logp_difference/max': 1.6993675231933594, 'sampling/importance_sampling_ratio/min': 0.1827991008758545, 'sampling/importance_sampling_ratio/mean': 1.009418249130249, 'sampling/importance_sampling_ratio/max': 1.9875471591949463, 'entropy': 0.35818639770150185, 'clip_ratio/low_mean': 0.0056649468606337905, 'clip_ratio/low_min': 0.0056649468606337905, 'clip_ratio/high_mean': 0.020963466726243496, 'clip_ratio/high_max': 0.020963466726243496, 'clip_ratio/region_mean': 0.026628413586877286, 'reward_total_mean': 0.7318886518478394, 'reward_meter_mean': 0.998869776725769, 'reward_meter_std': 0.00026758189778774977, 'reward_count_adherence_mean': 0.7734375, 'reward_count_adherence_std': 0.04650149121880531, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.946873664855957, 'reward_repeat_penalty_std': 0.06902104616165161, 'reward_total_composite_mean': 0.7318886518478394, 'reward_total_composite_std': 0.0725080594420433, 'epoch': 0.11}

 81%|████████▏ | 2687/3300 [3:55:29<1:02:40,  6.13s/it]INFO 04-12 02:32:20 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2688/3300 [3:55:35<1:00:23,  5.92s/it]2026-04-12 02:32:26,174 | INFO | train_grpo_train | metrics_logged mode=train step=2688

                                                       
{'loss': 0.0052, 'grad_norm': 3.9790732860565186, 'learning_rate': 1.8575757575757578e-06, 'num_tokens': 6092497.0, 'completions/mean_length': 140.125, 'completions/min_length': 137.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.125, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9988832473754883, 'rewards/meter/std': 0.0008235845598392189, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988832473754883, 'rewards/total_composite/std': 0.0008235845598392189, 'reward': 0.9988832473754883, 'reward_std': 0.0008235886925831437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0495881512761116, 'sampling/sampling_logp_difference/max': 2.3660759925842285, 'sampling/importance_sampling_ratio/min': 0.09384826570749283, 'sampling/importance_sampling_ratio/mean': 1.00178861618042, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39379359781742096, 'clip_ratio/low_mean': 0.008036528481170535, 'clip_ratio/low_min': 0.008036528481170535, 'clip_ratio/high_mean': 0.04208028828725219, 'clip_ratio/high_max': 0.04208028828725219, 'clip_ratio/region_mean': 0.05011681676842272, 'reward_total_mean': 0.9988832473754883, 'reward_meter_mean': 0.9988832473754883, 'reward_meter_std': 0.0008235845598392189, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988832473754883, 'reward_total_composite_std': 0.0008235845598392189, 'epoch': 0.11}

 81%|████████▏ | 2688/3300 [3:55:35<1:00:23,  5.92s/it]INFO 04-12 02:32:26 [block_pool.py:378] Successfully reset prefix cache

 81%|████████▏ | 2689/3300 [3:55:40<58:18,  5.73s/it]  2026-04-12 02:32:31,440 | INFO | train_grpo_train | metrics_logged mode=train step=2689

                                                     
{'loss': 0.0078, 'grad_norm': 3.656802177429199, 'learning_rate': 1.8545454545454546e-06, 'num_tokens': 6095097.0, 'completions/mean_length': 132.0, 'completions/min_length': 130.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.0, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9993307590484619, 'rewards/meter/std': 7.24185592844151e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9636418223381042, 'rewards/total_composite/std': 0.06610792875289917, 'reward': 0.9636418223381042, 'reward_std': 0.06610793620347977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011180154979228973, 'sampling/sampling_logp_difference/max': 0.804922342300415, 'sampling/importance_sampling_ratio/min': 0.4471226632595062, 'sampling/importance_sampling_ratio/mean': 0.9992504715919495, 'sampling/importance_sampling_ratio/max': 1.8010199069976807, 'entropy': 0.06141908373683691, 'clip_ratio/low_mean': 0.004699247889220715, 'clip_ratio/low_min': 0.004699247889220715, 'clip_ratio/high_mean': 0.00855186500120908, 'clip_ratio/high_max': 0.00855186500120908, 'clip_ratio/region_mean': 0.013251112890429795, 'reward_total_mean': 0.9636418223381042, 'reward_meter_mean': 0.9993307590484619, 'reward_meter_std': 7.24185592844151e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9636418223381042, 'reward_total_composite_std': 0.06610792875289917, 'epoch': 0.11}

 81%|████████▏ | 2689/3300 [3:55:40<58:18,  5.73s/it]INFO 04-12 02:32:31 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2690/3300 [3:55:44<54:14,  5.34s/it]2026-04-12 02:32:35,867 | INFO | train_grpo_train | metrics_logged mode=train step=2690

                                                     
{'loss': 0.0005, 'grad_norm': 0.08232353627681732, 'learning_rate': 1.8515151515151517e-06, 'num_tokens': 6097105.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981546401977539, 'rewards/meter/std': 6.638248123636004e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981546401977539, 'rewards/total_composite/std': 6.638248123636004e-06, 'reward': 0.9981546401977539, 'reward_std': 6.6471684476709925e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0016760394209995866, 'sampling/sampling_logp_difference/max': 0.5546226501464844, 'sampling/importance_sampling_ratio/min': 0.5742889642715454, 'sampling/importance_sampling_ratio/mean': 0.9995981454849243, 'sampling/importance_sampling_ratio/max': 1.0263193845748901, 'entropy': 0.007268769608344883, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981546401977539, 'reward_meter_mean': 0.9981546401977539, 'reward_meter_std': 6.638248123636004e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981546401977539, 'reward_total_composite_std': 6.638248123636004e-06, 'epoch': 0.11}

 82%|████████▏ | 2690/3300 [3:55:44<54:14,  5.34s/it]INFO 04-12 02:32:36 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2691/3300 [3:55:50<55:20,  5.45s/it]2026-04-12 02:32:41,593 | INFO | train_grpo_train | metrics_logged mode=train step=2691

                                                     
{'loss': 0.0031, 'grad_norm': 2.4885315895080566, 'learning_rate': 1.8484848484848487e-06, 'num_tokens': 6099876.0, 'completions/mean_length': 166.375, 'completions/min_length': 165.0, 'completions/max_length': 168.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.375, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 168.0, 'rewards/meter/mean': 0.9992619752883911, 'rewards/meter/std': 0.00024342280812561512, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9715026617050171, 'rewards/total_composite/std': 0.051356665790081024, 'reward': 0.9715026617050171, 'reward_std': 0.05135667324066162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015284853056073189, 'sampling/sampling_logp_difference/max': 1.3873584270477295, 'sampling/importance_sampling_ratio/min': 0.24973411858081818, 'sampling/importance_sampling_ratio/mean': 1.0019655227661133, 'sampling/importance_sampling_ratio/max': 1.893328070640564, 'entropy': 0.07314449176192284, 'clip_ratio/low_mean': 0.0015015150420367718, 'clip_ratio/low_min': 0.0015015150420367718, 'clip_ratio/high_mean': 0.009009360568597913, 'clip_ratio/high_max': 0.009009360568597913, 'clip_ratio/region_mean': 0.010510875610634685, 'reward_total_mean': 0.9715026617050171, 'reward_meter_mean': 0.9992619752883911, 'reward_meter_std': 0.00024342280812561512, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9715026617050171, 'reward_total_composite_std': 0.051356665790081024, 'epoch': 0.11}

 82%|████████▏ | 2691/3300 [3:55:50<55:20,  5.45s/it]INFO 04-12 02:32:41 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2692/3300 [3:55:54<52:22,  5.17s/it]2026-04-12 02:32:46,098 | INFO | train_grpo_train | metrics_logged mode=train step=2692

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.8454545454545455e-06, 'num_tokens': 6101556.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002783673407975584, 'sampling/sampling_logp_difference/max': 0.003993465099483728, 'sampling/importance_sampling_ratio/min': 1.0, 'sampling/importance_sampling_ratio/mean': 1.0002787113189697, 'sampling/importance_sampling_ratio/max': 1.004001498222351, 'entropy': 0.0021011842327425256, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2692/3300 [3:55:55<52:22,  5.17s/it]INFO 04-12 02:32:46 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2693/3300 [3:55:59<50:57,  5.04s/it]2026-04-12 02:32:50,831 | INFO | train_grpo_train | metrics_logged mode=train step=2693

                                                     
{'loss': 0.0009, 'grad_norm': 1.4883558750152588, 'learning_rate': 1.8424242424242426e-06, 'num_tokens': 6103724.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9992461204528809, 'rewards/meter/std': 0.00010268352343700826, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992461204528809, 'rewards/total_composite/std': 0.00010268352343700826, 'reward': 0.9992461204528809, 'reward_std': 0.00010269100312143564, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005682831630110741, 'sampling/sampling_logp_difference/max': 1.3028526306152344, 'sampling/importance_sampling_ratio/min': 0.6744450330734253, 'sampling/importance_sampling_ratio/mean': 1.0015344619750977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.01480487291701138, 'clip_ratio/low_mean': 0.0038265305338427424, 'clip_ratio/low_min': 0.0038265305338427424, 'clip_ratio/high_mean': 0.0025510203558951616, 'clip_ratio/high_max': 0.0025510203558951616, 'clip_ratio/region_mean': 0.006377550889737904, 'reward_total_mean': 0.9992461204528809, 'reward_meter_mean': 0.9992461204528809, 'reward_meter_std': 0.00010268352343700826, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992461204528809, 'reward_total_composite_std': 0.00010268352343700826, 'epoch': 0.11}

 82%|████████▏ | 2693/3300 [3:55:59<50:57,  5.04s/it]INFO 04-12 02:32:51 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2694/3300 [3:56:04<49:16,  4.88s/it]2026-04-12 02:32:55,338 | INFO | train_grpo_train | metrics_logged mode=train step=2694

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.8393939393939394e-06, 'num_tokens': 6105620.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002477174566593021, 'sampling/sampling_logp_difference/max': 0.006054490804672241, 'sampling/importance_sampling_ratio/min': 0.9999150037765503, 'sampling/importance_sampling_ratio/mean': 1.000247597694397, 'sampling/importance_sampling_ratio/max': 1.0060728788375854, 'entropy': 0.0020191115036141127, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2694/3300 [3:56:04<49:16,  4.88s/it]INFO 04-12 02:32:55 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2695/3300 [3:56:08<48:06,  4.77s/it]2026-04-12 02:32:59,861 | INFO | train_grpo_train | metrics_logged mode=train step=2695

                                                     
{'loss': 0.064, 'grad_norm': 12.76325798034668, 'learning_rate': 1.8363636363636365e-06, 'num_tokens': 6107356.0, 'completions/mean_length': 63.0, 'completions/min_length': 58.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6803799867630005, 'rewards/meter/std': 0.3893775939941406, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6803799867630005, 'rewards/total_composite/std': 0.3893775939941406, 'reward': 0.6803799867630005, 'reward_std': 0.3893775939941406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09513663500547409, 'sampling/sampling_logp_difference/max': 2.7315380573272705, 'sampling/importance_sampling_ratio/min': 0.06511905789375305, 'sampling/importance_sampling_ratio/mean': 1.008878231048584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.356488361954689, 'clip_ratio/low_mean': 0.03159340703859925, 'clip_ratio/low_min': 0.03159340703859925, 'clip_ratio/high_mean': 0.0434071971103549, 'clip_ratio/high_max': 0.0434071971103549, 'clip_ratio/region_mean': 0.07500060414895415, 'reward_total_mean': 0.6803799867630005, 'reward_meter_mean': 0.6803799867630005, 'reward_meter_std': 0.3893775939941406, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6803799867630005, 'reward_total_composite_std': 0.3893775939941406, 'epoch': 0.11}

 82%|████████▏ | 2695/3300 [3:56:08<48:06,  4.77s/it]INFO 04-12 02:33:00 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2696/3300 [3:56:13<47:42,  4.74s/it]2026-04-12 02:33:04,526 | INFO | train_grpo_train | metrics_logged mode=train step=2696

                                                     
{'loss': 0.0176, 'grad_norm': 9.303409576416016, 'learning_rate': 1.8333333333333333e-06, 'num_tokens': 6109361.0, 'completions/mean_length': 66.625, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9331727027893066, 'rewards/meter/std': 0.008176974020898342, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9331727027893066, 'rewards/total_composite/std': 0.008176974020898342, 'reward': 0.9331727027893066, 'reward_std': 0.008176976814866066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036672260612249374, 'sampling/sampling_logp_difference/max': 0.9801895618438721, 'sampling/importance_sampling_ratio/min': 0.3752399682998657, 'sampling/importance_sampling_ratio/mean': 0.9995071291923523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18041662964969873, 'clip_ratio/low_mean': 0.01482684025540948, 'clip_ratio/low_min': 0.01482684025540948, 'clip_ratio/high_mean': 0.016798642929643393, 'clip_ratio/high_max': 0.016798642929643393, 'clip_ratio/region_mean': 0.03162548318505287, 'reward_total_mean': 0.9331727027893066, 'reward_meter_mean': 0.9331727027893066, 'reward_meter_std': 0.008176974020898342, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9331727027893066, 'reward_total_composite_std': 0.008176974020898342, 'epoch': 0.11}

 82%|████████▏ | 2696/3300 [3:56:13<47:42,  4.74s/it]INFO 04-12 02:33:04 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2697/3300 [3:56:19<50:39,  5.04s/it]2026-04-12 02:33:10,273 | INFO | train_grpo_train | metrics_logged mode=train step=2697

                                                     
{'loss': 0.0029, 'grad_norm': 0.7284796237945557, 'learning_rate': 1.8303030303030305e-06, 'num_tokens': 6112233.0, 'completions/mean_length': 168.0, 'completions/min_length': 166.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.0, 'completions/min_terminated_length': 166.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9992363452911377, 'rewards/meter/std': 9.161681373370811e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992363452911377, 'rewards/total_composite/std': 9.161681373370811e-05, 'reward': 0.9992363452911377, 'reward_std': 9.160472109215334e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007861536927521229, 'sampling/sampling_logp_difference/max': 0.685643196105957, 'sampling/importance_sampling_ratio/min': 0.5037661194801331, 'sampling/importance_sampling_ratio/mean': 1.0003851652145386, 'sampling/importance_sampling_ratio/max': 1.376589298248291, 'entropy': 0.0484649995341897, 'clip_ratio/low_mean': 0.002214635896962136, 'clip_ratio/low_min': 0.002214635896962136, 'clip_ratio/high_mean': 0.003729202609974891, 'clip_ratio/high_max': 0.003729202609974891, 'clip_ratio/region_mean': 0.005943838506937027, 'reward_total_mean': 0.9992363452911377, 'reward_meter_mean': 0.9992363452911377, 'reward_meter_std': 9.161681373370811e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992363452911377, 'reward_total_composite_std': 9.161681373370811e-05, 'epoch': 0.11}

 82%|████████▏ | 2697/3300 [3:56:19<50:39,  5.04s/it]INFO 04-12 02:33:10 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2698/3300 [3:56:23<49:31,  4.94s/it]2026-04-12 02:33:14,963 | INFO | train_grpo_train | metrics_logged mode=train step=2698

                                                     
{'loss': -0.0019, 'grad_norm': 1.4812349081039429, 'learning_rate': 1.8272727272727276e-06, 'num_tokens': 6114086.0, 'completions/mean_length': 71.625, 'completions/min_length': 70.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.999380350112915, 'rewards/meter/std': 0.00012280534429010004, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999380350112915, 'rewards/total_composite/std': 0.00012280534429010004, 'reward': 0.999380350112915, 'reward_std': 0.0001228114851983264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012132828123867512, 'sampling/sampling_logp_difference/max': 1.2307639122009277, 'sampling/importance_sampling_ratio/min': 0.29206937551498413, 'sampling/importance_sampling_ratio/mean': 0.9971534013748169, 'sampling/importance_sampling_ratio/max': 1.2945717573165894, 'entropy': 0.051547068171203136, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.00870500784367323, 'clip_ratio/high_max': 0.00870500784367323, 'clip_ratio/region_mean': 0.012276436435058713, 'reward_total_mean': 0.999380350112915, 'reward_meter_mean': 0.999380350112915, 'reward_meter_std': 0.00012280534429010004, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999380350112915, 'reward_total_composite_std': 0.00012280534429010004, 'epoch': 0.11}

 82%|████████▏ | 2698/3300 [3:56:23<49:31,  4.94s/it]INFO 04-12 02:33:15 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2699/3300 [3:56:28<47:12,  4.71s/it]2026-04-12 02:33:19,154 | INFO | train_grpo_train | metrics_logged mode=train step=2699

                                                     
{'loss': 0.0187, 'grad_norm': 21.836759567260742, 'learning_rate': 1.8242424242424244e-06, 'num_tokens': 6115559.0, 'completions/mean_length': 34.125, 'completions/min_length': 34.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9915792942047119, 'rewards/meter/std': 0.0012008182238787413, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9915792942047119, 'rewards/total_composite/std': 0.0012008182238787413, 'reward': 0.9915792942047119, 'reward_std': 0.001200812985189259, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024363644421100616, 'sampling/sampling_logp_difference/max': 1.109255313873291, 'sampling/importance_sampling_ratio/min': 0.3298044800758362, 'sampling/importance_sampling_ratio/mean': 1.0031808614730835, 'sampling/importance_sampling_ratio/max': 1.8209221363067627, 'entropy': 0.18179897591471672, 'clip_ratio/low_mean': 0.01460084063000977, 'clip_ratio/low_min': 0.01460084063000977, 'clip_ratio/high_mean': 0.007352941203862429, 'clip_ratio/high_max': 0.007352941203862429, 'clip_ratio/region_mean': 0.0219537818338722, 'reward_total_mean': 0.9915792942047119, 'reward_meter_mean': 0.9915792942047119, 'reward_meter_std': 0.0012008182238787413, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9915792942047119, 'reward_total_composite_std': 0.0012008182238787413, 'epoch': 0.11}

 82%|████████▏ | 2699/3300 [3:56:28<47:12,  4.71s/it]INFO 04-12 02:33:19 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2700/3300 [3:56:32<46:55,  4.69s/it]2026-04-12 02:33:23,802 | INFO | train_grpo_train | metrics_logged mode=train step=2700

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.8212121212121215e-06, 'num_tokens': 6117340.0, 'completions/mean_length': 71.625, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.99944007396698, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99944007396698, 'rewards/total_composite/std': 0.0, 'reward': 0.99944007396698, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.004740848205983639, 'sampling/sampling_logp_difference/max': 0.2854344844818115, 'sampling/importance_sampling_ratio/min': 0.7516875863075256, 'sampling/importance_sampling_ratio/mean': 1.002069115638733, 'sampling/importance_sampling_ratio/max': 1.2885301113128662, 'entropy': 0.03858046652749181, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.99944007396698, 'reward_meter_mean': 0.99944007396698, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99944007396698, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2700/3300 [3:56:32<46:55,  4.69s/it]INFO 04-12 02:33:24 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:36,  3.35s/it]

 23%|██▎       | 3/13 [00:10<00:35,  3.51s/it]

 31%|███       | 4/13 [00:16<00:38,  4.29s/it]

 38%|███▊      | 5/13 [00:21<00:36,  4.61s/it]

 46%|████▌     | 6/13 [00:27<00:37,  5.29s/it]

 54%|█████▍    | 7/13 [00:33<00:31,  5.31s/it]

 62%|██████▏   | 8/13 [00:40<00:29,  5.82s/it]

 69%|██████▉   | 9/13 [00:47<00:24,  6.15s/it]

 77%|███████▋  | 10/13 [00:52<00:17,  5.81s/it]

 85%|████████▍ | 11/13 [00:57<00:11,  5.73s/it]

 92%|█████████▏| 12/13 [01:03<00:05,  5.69s/it]

100%|██████████| 13/13 [01:08<00:00,  5.54s/it]2026-04-12 02:34:38,749 | INFO | train_grpo_train | metrics_logged mode=eval step=2700

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 74.8526, 'eval_samples_per_second': 1.389, 'eval_steps_per_second': 0.174, 'eval_num_tokens': 6117340.0, 'eval_completions/mean_length': 205.91346153846155, 'eval_completions/min_length': 61.07692307692308, 'eval_completions/max_length': 402.0, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 205.91346153846155, 'eval_completions/min_terminated_length': 61.07692307692308, 'eval_completions/max_terminated_length': 402.0, 'eval_rewards/meter/mean': 0.818766135435838, 'eval_rewards/meter/std': 0.28128915272939664, 'eval_rewards/count_adherence/mean': 0.9412473761118375, 'eval_rewards/count_adherence/std': 0.08066883425299938, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.917067867058974, 'eval_rewards/repeat_penalty/std': 0.09873221929256733, 'eval_rewards/total_composite/mean': 0.7056683852122381, 'eval_rewards/total_composite/std': 0.3022650325527558, 'eval_reward': 0.7056683852122381, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.026865268770891886, 'eval_sampling/sampling_logp_difference/max': 1.0165979678814228, 'eval_sampling/importance_sampling_ratio/min': 0.36844213765401107, 'eval_sampling/importance_sampling_ratio/mean': 1.00816364471729, 'eval_sampling/importance_sampling_ratio/max': 1.4687065803087676, 'eval_entropy': 0.2846654344063539, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7056683852122381, 'eval_reward_meter_mean': 0.818766135435838, 'eval_reward_meter_std': 0.28128915272939664, 'eval_reward_count_adherence_mean': 0.9412473761118375, 'eval_reward_count_adherence_std': 0.08066883425299938, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.917067867058974, 'eval_reward_repeat_penalty_std': 0.09873221929256733, 'eval_reward_total_composite_mean': 0.7056683852122381, 'eval_reward_total_composite_std': 0.3022650325527558, 'epoch': 0.11}

 82%|████████▏ | 2700/3300 [3:57:47<46:55,  4.69s/it]

100%|██████████| 13/13 [01:08<00:00,  5.54s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 82%|████████▏ | 2701/3300 [3:57:54<4:38:01, 27.85s/it]2026-04-12 02:34:45,680 | INFO | train_grpo_train | metrics_logged mode=train step=2701

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.8181818181818183e-06, 'num_tokens': 6119068.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0014518008101731539, 'sampling/sampling_logp_difference/max': 0.2512565851211548, 'sampling/importance_sampling_ratio/min': 0.7778228521347046, 'sampling/importance_sampling_ratio/mean': 0.9999621510505676, 'sampling/importance_sampling_ratio/max': 1.0299800634384155, 'entropy': 0.012467616703361273, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2701/3300 [3:57:54<4:38:01, 27.85s/it]INFO 04-12 02:34:45 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2702/3300 [3:57:59<3:28:28, 20.92s/it]2026-04-12 02:34:50,424 | INFO | train_grpo_train | metrics_logged mode=train step=2702

                                                       
{'loss': -0.001, 'grad_norm': 3.3459887504577637, 'learning_rate': 1.8151515151515153e-06, 'num_tokens': 6120768.0, 'completions/mean_length': 67.5, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.937604546546936, 'rewards/meter/std': 0.136740043759346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.937604546546936, 'rewards/total_composite/std': 0.136740043759346, 'reward': 0.937604546546936, 'reward_std': 0.1367400586605072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025002291426062584, 'sampling/sampling_logp_difference/max': 1.505655288696289, 'sampling/importance_sampling_ratio/min': 0.22187186777591705, 'sampling/importance_sampling_ratio/mean': 1.0090194940567017, 'sampling/importance_sampling_ratio/max': 1.6750621795654297, 'entropy': 0.19097852148115635, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.007548359571956098, 'clip_ratio/high_max': 0.007548359571956098, 'clip_ratio/region_mean': 0.00941403117030859, 'reward_total_mean': 0.937604546546936, 'reward_meter_mean': 0.937604546546936, 'reward_meter_std': 0.136740043759346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.937604546546936, 'reward_total_composite_std': 0.136740043759346, 'epoch': 0.11}

 82%|████████▏ | 2702/3300 [3:57:59<3:28:28, 20.92s/it]INFO 04-12 02:34:50 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2703/3300 [3:58:05<2:45:00, 16.58s/it]2026-04-12 02:34:56,901 | INFO | train_grpo_train | metrics_logged mode=train step=2703

                                                       
{'loss': 0.0104, 'grad_norm': 2.7519359588623047, 'learning_rate': 1.8121212121212124e-06, 'num_tokens': 6124022.0, 'completions/mean_length': 213.75, 'completions/min_length': 207.0, 'completions/max_length': 218.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 213.75, 'completions/min_terminated_length': 207.0, 'completions/max_terminated_length': 218.0, 'rewards/meter/mean': 0.9976348876953125, 'rewards/meter/std': 0.0003981303598266095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9711538553237915, 'rewards/repeat_penalty/std': 0.039811473339796066, 'rewards/total_composite/mean': 0.9688572883605957, 'rewards/total_composite/std': 0.039728421717882156, 'reward': 0.9688572883605957, 'reward_std': 0.03972842916846275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034267134964466095, 'sampling/sampling_logp_difference/max': 1.2862721681594849, 'sampling/importance_sampling_ratio/min': 0.276298850774765, 'sampling/importance_sampling_ratio/mean': 1.0066944360733032, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.220721784979105, 'clip_ratio/low_mean': 0.016337704844772816, 'clip_ratio/low_min': 0.016337704844772816, 'clip_ratio/high_mean': 0.019398832926526666, 'clip_ratio/high_max': 0.019398832926526666, 'clip_ratio/region_mean': 0.03573653777129948, 'reward_total_mean': 0.9688572883605957, 'reward_meter_mean': 0.9976348876953125, 'reward_meter_std': 0.0003981303598266095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9711538553237915, 'reward_repeat_penalty_std': 0.039811473339796066, 'reward_total_composite_mean': 0.9688572883605957, 'reward_total_composite_std': 0.039728421717882156, 'epoch': 0.11}

 82%|████████▏ | 2703/3300 [3:58:05<2:45:00, 16.58s/it]INFO 04-12 02:34:57 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2704/3300 [3:58:10<2:09:03, 12.99s/it]2026-04-12 02:35:01,506 | INFO | train_grpo_train | metrics_logged mode=train step=2704

                                                       
{'loss': -0.0003, 'grad_norm': 0.004991916939616203, 'learning_rate': 1.8090909090909092e-06, 'num_tokens': 6125799.0, 'completions/mean_length': 67.125, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9981493353843689, 'rewards/meter/std': 8.239712769864127e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981493353843689, 'rewards/total_composite/std': 8.239712769864127e-06, 'reward': 0.9981493353843689, 'reward_std': 8.248746780736838e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0028709208127111197, 'sampling/sampling_logp_difference/max': 0.8200433254241943, 'sampling/importance_sampling_ratio/min': 0.44041258096694946, 'sampling/importance_sampling_ratio/mean': 0.9988438487052917, 'sampling/importance_sampling_ratio/max': 1.0250744819641113, 'entropy': 0.007660032075364143, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.003703906899318099, 'reward_total_mean': 0.9981493353843689, 'reward_meter_mean': 0.9981493353843689, 'reward_meter_std': 8.239712769864127e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981493353843689, 'reward_total_composite_std': 8.239712769864127e-06, 'epoch': 0.11}

 82%|████████▏ | 2704/3300 [3:58:10<2:09:03, 12.99s/it]INFO 04-12 02:35:01 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2705/3300 [3:58:14<1:43:26, 10.43s/it]2026-04-12 02:35:05,965 | INFO | train_grpo_train | metrics_logged mode=train step=2705

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.8060606060606063e-06, 'num_tokens': 6127407.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00024437991669401526, 'sampling/sampling_logp_difference/max': 0.002418494550511241, 'sampling/importance_sampling_ratio/min': 0.9996000528335571, 'sampling/importance_sampling_ratio/mean': 1.0002423524856567, 'sampling/importance_sampling_ratio/max': 1.002421498298645, 'entropy': 0.0020572251814883202, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2705/3300 [3:58:14<1:43:26, 10.43s/it]INFO 04-12 02:35:06 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2706/3300 [3:58:18<1:24:27,  8.53s/it]2026-04-12 02:35:10,063 | INFO | train_grpo_train | metrics_logged mode=train step=2706

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.803030303030303e-06, 'num_tokens': 6128951.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002643416519276798, 'sampling/sampling_logp_difference/max': 0.011745302006602287, 'sampling/importance_sampling_ratio/min': 0.9924525022506714, 'sampling/importance_sampling_ratio/mean': 1.0001347064971924, 'sampling/importance_sampling_ratio/max': 1.0118145942687988, 'entropy': 0.002913154661655426, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2706/3300 [3:58:19<1:24:27,  8.53s/it]INFO 04-12 02:35:10 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2707/3300 [3:58:23<1:11:47,  7.26s/it]2026-04-12 02:35:14,369 | INFO | train_grpo_train | metrics_logged mode=train step=2707

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.8000000000000001e-06, 'num_tokens': 6130687.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00021008482144679874, 'sampling/sampling_logp_difference/max': 0.0022089765407145023, 'sampling/importance_sampling_ratio/min': 0.9998108148574829, 'sampling/importance_sampling_ratio/mean': 1.0002092123031616, 'sampling/importance_sampling_ratio/max': 1.0022114515304565, 'entropy': 0.001665496762143448, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2707/3300 [3:58:23<1:11:47,  7.26s/it]INFO 04-12 02:35:14 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2708/3300 [3:58:28<1:05:11,  6.61s/it]2026-04-12 02:35:19,444 | INFO | train_grpo_train | metrics_logged mode=train step=2708

                                                       
{'loss': -0.0031, 'grad_norm': 1.1375532150268555, 'learning_rate': 1.796969696969697e-06, 'num_tokens': 6132899.0, 'completions/mean_length': 106.5, 'completions/min_length': 105.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.5, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.999268651008606, 'rewards/meter/std': 0.00017192552331835032, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999268651008606, 'rewards/total_composite/std': 0.00017192552331835032, 'reward': 0.999268651008606, 'reward_std': 0.00017192917584907264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008916010148823261, 'sampling/sampling_logp_difference/max': 0.8204793930053711, 'sampling/importance_sampling_ratio/min': 0.44022059440612793, 'sampling/importance_sampling_ratio/mean': 1.0018383264541626, 'sampling/importance_sampling_ratio/max': 1.3604763746261597, 'entropy': 0.053598769940435886, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.004673305433243513, 'clip_ratio/high_max': 0.004673305433243513, 'clip_ratio/region_mean': 0.004673305433243513, 'reward_total_mean': 0.999268651008606, 'reward_meter_mean': 0.999268651008606, 'reward_meter_std': 0.00017192552331835032, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999268651008606, 'reward_total_composite_std': 0.00017192552331835032, 'epoch': 0.11}

 82%|████████▏ | 2708/3300 [3:58:28<1:05:11,  6.61s/it]INFO 04-12 02:35:19 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2709/3300 [3:58:33<1:00:38,  6.16s/it]2026-04-12 02:35:24,551 | INFO | train_grpo_train | metrics_logged mode=train step=2709

                                                       
{'loss': 0.0051, 'grad_norm': 3.0748214721679688, 'learning_rate': 1.793939393939394e-06, 'num_tokens': 6135215.0, 'completions/mean_length': 102.5, 'completions/min_length': 98.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.5, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.998659610748291, 'rewards/meter/std': 0.0006201548385433853, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998659610748291, 'rewards/total_composite/std': 0.0006201548385433853, 'reward': 0.998659610748291, 'reward_std': 0.0006201544310897589, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03828549385070801, 'sampling/sampling_logp_difference/max': 1.1598763465881348, 'sampling/importance_sampling_ratio/min': 0.35953280329704285, 'sampling/importance_sampling_ratio/mean': 1.0111318826675415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2976728491485119, 'clip_ratio/low_mean': 0.010879121022298932, 'clip_ratio/low_min': 0.010879121022298932, 'clip_ratio/high_mean': 0.01722477504517883, 'clip_ratio/high_max': 0.01722477504517883, 'clip_ratio/region_mean': 0.028103896067477763, 'reward_total_mean': 0.998659610748291, 'reward_meter_mean': 0.998659610748291, 'reward_meter_std': 0.0006201548385433853, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998659610748291, 'reward_total_composite_std': 0.0006201548385433853, 'epoch': 0.11}

 82%|████████▏ | 2709/3300 [3:58:33<1:00:38,  6.16s/it]INFO 04-12 02:35:24 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2710/3300 [3:58:38<58:20,  5.93s/it]  2026-04-12 02:35:29,967 | INFO | train_grpo_train | metrics_logged mode=train step=2710

                                                     
{'loss': 0.0027, 'grad_norm': 1.891419529914856, 'learning_rate': 1.7909090909090908e-06, 'num_tokens': 6137547.0, 'completions/mean_length': 142.5, 'completions/min_length': 141.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.5, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.999262809753418, 'rewards/meter/std': 0.00011097361129941419, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8392857313156128, 'rewards/repeat_penalty/std': 0.09155284613370895, 'rewards/total_composite/mean': 0.8386699557304382, 'rewards/total_composite/std': 0.0915173664689064, 'reward': 0.8386699557304382, 'reward_std': 0.0915173590183258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010055704973638058, 'sampling/sampling_logp_difference/max': 0.6665773391723633, 'sampling/importance_sampling_ratio/min': 0.5134629607200623, 'sampling/importance_sampling_ratio/mean': 1.0032413005828857, 'sampling/importance_sampling_ratio/max': 1.472945213317871, 'entropy': 0.0695994533598423, 'clip_ratio/low_mean': 0.003502913983538747, 'clip_ratio/low_min': 0.003502913983538747, 'clip_ratio/high_mean': 0.001754407538101077, 'clip_ratio/high_max': 0.001754407538101077, 'clip_ratio/region_mean': 0.005257321521639824, 'reward_total_mean': 0.8386699557304382, 'reward_meter_mean': 0.999262809753418, 'reward_meter_std': 0.00011097361129941419, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8392857313156128, 'reward_repeat_penalty_std': 0.09155284613370895, 'reward_total_composite_mean': 0.8386699557304382, 'reward_total_composite_std': 0.0915173664689064, 'epoch': 0.11}

 82%|████████▏ | 2710/3300 [3:58:38<58:20,  5.93s/it]INFO 04-12 02:35:30 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2711/3300 [3:58:43<54:00,  5.50s/it]2026-04-12 02:35:34,455 | INFO | train_grpo_train | metrics_logged mode=train step=2711

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.787878787878788e-06, 'num_tokens': 6139309.0, 'completions/mean_length': 66.25, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0032224601600319147, 'sampling/sampling_logp_difference/max': 0.3331582546234131, 'sampling/importance_sampling_ratio/min': 0.729793906211853, 'sampling/importance_sampling_ratio/mean': 1.0023350715637207, 'sampling/importance_sampling_ratio/max': 1.395367980003357, 'entropy': 0.016254089423455298, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2711/3300 [3:58:43<54:00,  5.50s/it]INFO 04-12 02:35:34 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2712/3300 [3:58:47<50:34,  5.16s/it]2026-04-12 02:35:38,818 | INFO | train_grpo_train | metrics_logged mode=train step=2712

                                                     
{'loss': -0.0068, 'grad_norm': 3.4030656814575195, 'learning_rate': 1.7848484848484851e-06, 'num_tokens': 6141147.0, 'completions/mean_length': 60.75, 'completions/min_length': 59.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.75, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9972513318061829, 'rewards/meter/std': 0.0001800779573386535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972513318061829, 'rewards/total_composite/std': 0.0001800779573386535, 'reward': 0.9972513318061829, 'reward_std': 0.00018008207553066313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0070107607170939445, 'sampling/sampling_logp_difference/max': 0.9685866832733154, 'sampling/importance_sampling_ratio/min': 0.7316676378250122, 'sampling/importance_sampling_ratio/mean': 1.0053867101669312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.02895409008488059, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0020491802133619785, 'reward_total_mean': 0.9972513318061829, 'reward_meter_mean': 0.9972513318061829, 'reward_meter_std': 0.0001800779573386535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972513318061829, 'reward_total_composite_std': 0.0001800779573386535, 'epoch': 0.11}

 82%|████████▏ | 2712/3300 [3:58:47<50:34,  5.16s/it]INFO 04-12 02:35:39 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2713/3300 [3:58:52<48:11,  4.93s/it]2026-04-12 02:35:43,197 | INFO | train_grpo_train | metrics_logged mode=train step=2713

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.781818181818182e-06, 'num_tokens': 6142939.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0010938982013612986, 'sampling/sampling_logp_difference/max': 0.2828308939933777, 'sampling/importance_sampling_ratio/min': 0.7536472678184509, 'sampling/importance_sampling_ratio/mean': 0.9994007349014282, 'sampling/importance_sampling_ratio/max': 1.023779034614563, 'entropy': 0.004417360032675788, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2713/3300 [3:58:52<48:11,  4.93s/it]INFO 04-12 02:35:43 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2714/3300 [3:58:56<47:26,  4.86s/it]2026-04-12 02:35:47,898 | INFO | train_grpo_train | metrics_logged mode=train step=2714

                                                     
{'loss': 0.0083, 'grad_norm': 2.3612427711486816, 'learning_rate': 1.778787878787879e-06, 'num_tokens': 6144770.0, 'completions/mean_length': 78.875, 'completions/min_length': 77.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.875, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9988387227058411, 'rewards/meter/std': 0.0002311533025931567, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988387227058411, 'rewards/total_composite/std': 0.0002311533025931567, 'reward': 0.9988387227058411, 'reward_std': 0.00023116436204873025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028173334896564484, 'sampling/sampling_logp_difference/max': 1.179065227508545, 'sampling/importance_sampling_ratio/min': 0.3075661063194275, 'sampling/importance_sampling_ratio/mean': 1.0010093450546265, 'sampling/importance_sampling_ratio/max': 1.6622776985168457, 'entropy': 0.15530052781105042, 'clip_ratio/low_mean': 0.0015432098880410194, 'clip_ratio/low_min': 0.0015432098880410194, 'clip_ratio/high_mean': 0.016068846685811877, 'clip_ratio/high_max': 0.016068846685811877, 'clip_ratio/region_mean': 0.017612056573852897, 'reward_total_mean': 0.9988387227058411, 'reward_meter_mean': 0.9988387227058411, 'reward_meter_std': 0.0002311533025931567, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988387227058411, 'reward_total_composite_std': 0.0002311533025931567, 'epoch': 0.11}

 82%|████████▏ | 2714/3300 [3:58:56<47:26,  4.86s/it]INFO 04-12 02:35:48 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2715/3300 [3:59:02<49:36,  5.09s/it]2026-04-12 02:35:53,521 | INFO | train_grpo_train | metrics_logged mode=train step=2715

                                                     
{'loss': 0.0025, 'grad_norm': 0.7956473231315613, 'learning_rate': 1.775757575757576e-06, 'num_tokens': 6147492.0, 'completions/mean_length': 159.25, 'completions/min_length': 159.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.25, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9979335069656372, 'rewards/meter/std': 1.7607557310839184e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9027777910232544, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.9009125232696533, 'rewards/total_composite/std': 0.03921113535761833, 'reward': 0.9009125232696533, 'reward_std': 0.03921113163232803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009351929649710655, 'sampling/sampling_logp_difference/max': 0.5963006019592285, 'sampling/importance_sampling_ratio/min': 0.5508456826210022, 'sampling/importance_sampling_ratio/mean': 1.0039114952087402, 'sampling/importance_sampling_ratio/max': 1.4207649230957031, 'entropy': 0.08590239379554987, 'clip_ratio/low_mean': 0.0015674135065637529, 'clip_ratio/low_min': 0.0015674135065637529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0015674135065637529, 'reward_total_mean': 0.9009125232696533, 'reward_meter_mean': 0.9979335069656372, 'reward_meter_std': 1.7607557310839184e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9027777910232544, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.9009125232696533, 'reward_total_composite_std': 0.03921113535761833, 'epoch': 0.11}

 82%|████████▏ | 2715/3300 [3:59:02<49:36,  5.09s/it]INFO 04-12 02:35:53 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2716/3300 [3:59:11<1:01:49,  6.35s/it]2026-04-12 02:36:02,824 | INFO | train_grpo_train | metrics_logged mode=train step=2716

                                                       
{'loss': -0.0565, 'grad_norm': 1.8313432931900024, 'learning_rate': 1.7727272727272729e-06, 'num_tokens': 6152509.0, 'completions/mean_length': 438.125, 'completions/min_length': 361.0, 'completions/max_length': 459.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 438.125, 'completions/min_terminated_length': 361.0, 'completions/max_terminated_length': 459.0, 'rewards/meter/mean': 0.9989956617355347, 'rewards/meter/std': 0.00016425758076366037, 'rewards/count_adherence/mean': 0.7666666507720947, 'rewards/count_adherence/std': 0.07126966118812561, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9524672031402588, 'rewards/repeat_penalty/std': 0.05135857313871384, 'rewards/total_composite/mean': 0.7314679622650146, 'rewards/total_composite/std': 0.09445607662200928, 'reward': 0.7314679622650146, 'reward_std': 0.09445606172084808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04289248585700989, 'sampling/sampling_logp_difference/max': 1.8779144287109375, 'sampling/importance_sampling_ratio/min': 0.15290868282318115, 'sampling/importance_sampling_ratio/mean': 1.0078824758529663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3533891662955284, 'clip_ratio/low_mean': 0.008343940833583474, 'clip_ratio/low_min': 0.008343940833583474, 'clip_ratio/high_mean': 0.014018616639077663, 'clip_ratio/high_max': 0.014018616639077663, 'clip_ratio/region_mean': 0.022362557472661138, 'reward_total_mean': 0.7314679622650146, 'reward_meter_mean': 0.9989956617355347, 'reward_meter_std': 0.00016425758076366037, 'reward_count_adherence_mean': 0.7666666507720947, 'reward_count_adherence_std': 0.07126966118812561, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9524672031402588, 'reward_repeat_penalty_std': 0.05135857313871384, 'reward_total_composite_mean': 0.7314679622650146, 'reward_total_composite_std': 0.09445607662200928, 'epoch': 0.11}

 82%|████████▏ | 2716/3300 [3:59:11<1:01:49,  6.35s/it]INFO 04-12 02:36:03 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2717/3300 [3:59:17<58:35,  6.03s/it]  2026-04-12 02:36:08,102 | INFO | train_grpo_train | metrics_logged mode=train step=2717

                                                     
{'loss': 0.0033, 'grad_norm': 1.0715020895004272, 'learning_rate': 1.76969696969697e-06, 'num_tokens': 6155040.0, 'completions/mean_length': 131.375, 'completions/min_length': 130.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.375, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9993519186973572, 'rewards/meter/std': 4.195739529677667e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993519186973572, 'rewards/total_composite/std': 4.195739529677667e-05, 'reward': 0.9993519186973572, 'reward_std': 4.196614827378653e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011481878347694874, 'sampling/sampling_logp_difference/max': 1.0313186645507812, 'sampling/importance_sampling_ratio/min': 0.35653650760650635, 'sampling/importance_sampling_ratio/mean': 0.9986187815666199, 'sampling/importance_sampling_ratio/max': 1.3561691045761108, 'entropy': 0.05451698787510395, 'clip_ratio/low_mean': 0.0056604581186547875, 'clip_ratio/low_min': 0.0056604581186547875, 'clip_ratio/high_mean': 0.00957868475234136, 'clip_ratio/high_max': 0.00957868475234136, 'clip_ratio/region_mean': 0.015239142870996147, 'reward_total_mean': 0.9993519186973572, 'reward_meter_mean': 0.9993519186973572, 'reward_meter_std': 4.195739529677667e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993519186973572, 'reward_total_composite_std': 4.195739529677667e-05, 'epoch': 0.11}

 82%|████████▏ | 2717/3300 [3:59:17<58:35,  6.03s/it]INFO 04-12 02:36:08 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2718/3300 [3:59:24<1:02:46,  6.47s/it]2026-04-12 02:36:15,603 | INFO | train_grpo_train | metrics_logged mode=train step=2718

                                                       
{'loss': 0.0052, 'grad_norm': 3.4733986854553223, 'learning_rate': 1.7666666666666668e-06, 'num_tokens': 6159098.0, 'completions/mean_length': 307.25, 'completions/min_length': 298.0, 'completions/max_length': 313.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 307.25, 'completions/min_terminated_length': 298.0, 'completions/max_terminated_length': 313.0, 'rewards/meter/mean': 0.9986767768859863, 'rewards/meter/std': 0.0004087568959221244, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9632352590560913, 'rewards/repeat_penalty/std': 0.04376610368490219, 'rewards/total_composite/mean': 0.8657643795013428, 'rewards/total_composite/std': 0.039332009851932526, 'reward': 0.8657643795013428, 'reward_std': 0.03933200240135193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059995777904987335, 'sampling/sampling_logp_difference/max': 3.847691059112549, 'sampling/importance_sampling_ratio/min': 0.02132892794907093, 'sampling/importance_sampling_ratio/mean': 1.010833740234375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5477991737425327, 'clip_ratio/low_mean': 0.023148135747760534, 'clip_ratio/low_min': 0.023148135747760534, 'clip_ratio/high_mean': 0.027308928780257702, 'clip_ratio/high_max': 0.027308928780257702, 'clip_ratio/region_mean': 0.050457064528018236, 'reward_total_mean': 0.8657643795013428, 'reward_meter_mean': 0.9986767768859863, 'reward_meter_std': 0.0004087568959221244, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9632352590560913, 'reward_repeat_penalty_std': 0.04376610368490219, 'reward_total_composite_mean': 0.8657643795013428, 'reward_total_composite_std': 0.039332009851932526, 'epoch': 0.11}

 82%|████████▏ | 2718/3300 [3:59:24<1:02:46,  6.47s/it]INFO 04-12 02:36:15 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2719/3300 [3:59:28<55:48,  5.76s/it]  2026-04-12 02:36:19,719 | INFO | train_grpo_train | metrics_logged mode=train step=2719

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.7636363636363638e-06, 'num_tokens': 6160514.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 4.7315745177911595e-05, 'sampling/sampling_logp_difference/max': 0.0008822012459859252, 'sampling/importance_sampling_ratio/min': 0.9999879002571106, 'sampling/importance_sampling_ratio/mean': 1.0000470876693726, 'sampling/importance_sampling_ratio/max': 1.000882625579834, 'entropy': 0.00040178035851567984, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 82%|████████▏ | 2719/3300 [3:59:28<55:48,  5.76s/it]INFO 04-12 02:36:19 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2720/3300 [3:59:33<53:30,  5.54s/it]2026-04-12 02:36:24,715 | INFO | train_grpo_train | metrics_logged mode=train step=2720

                                                     
{'loss': 0.0004, 'grad_norm': 0.7155827283859253, 'learning_rate': 1.7606060606060606e-06, 'num_tokens': 6162606.0, 'completions/mean_length': 98.5, 'completions/min_length': 98.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.5, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.9994127750396729, 'rewards/meter/std': 2.115286042680964e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994127750396729, 'rewards/total_composite/std': 2.115286042680964e-05, 'reward': 0.9994127750396729, 'reward_std': 2.114395101671107e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008277514949440956, 'sampling/sampling_logp_difference/max': 0.9306962490081787, 'sampling/importance_sampling_ratio/min': 0.3942791223526001, 'sampling/importance_sampling_ratio/mean': 1.0007652044296265, 'sampling/importance_sampling_ratio/max': 1.7030972242355347, 'entropy': 0.03585473005659878, 'clip_ratio/low_mean': 0.0024999999441206455, 'clip_ratio/low_min': 0.0024999999441206455, 'clip_ratio/high_mean': 0.005076272878795862, 'clip_ratio/high_max': 0.005076272878795862, 'clip_ratio/region_mean': 0.007576272822916508, 'reward_total_mean': 0.9994127750396729, 'reward_meter_mean': 0.9994127750396729, 'reward_meter_std': 2.115286042680964e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994127750396729, 'reward_total_composite_std': 2.115286042680964e-05, 'epoch': 0.11}

 82%|████████▏ | 2720/3300 [3:59:33<53:30,  5.54s/it]INFO 04-12 02:36:24 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2721/3300 [3:59:38<50:52,  5.27s/it]2026-04-12 02:36:29,378 | INFO | train_grpo_train | metrics_logged mode=train step=2721

                                                     
{'loss': 0.0456, 'grad_norm': 5.66716194152832, 'learning_rate': 1.7575757575757577e-06, 'num_tokens': 6164482.0, 'completions/mean_length': 70.5, 'completions/min_length': 66.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9914366006851196, 'rewards/meter/std': 0.004417366813868284, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9914366006851196, 'rewards/total_composite/std': 0.004417366813868284, 'reward': 0.9914366006851196, 'reward_std': 0.004417361691594124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054880011826753616, 'sampling/sampling_logp_difference/max': 1.349165916442871, 'sampling/importance_sampling_ratio/min': 0.2594565749168396, 'sampling/importance_sampling_ratio/mean': 0.998794436454773, 'sampling/importance_sampling_ratio/max': 1.5672898292541504, 'entropy': 0.3494097013026476, 'clip_ratio/low_mean': 0.017159562092274427, 'clip_ratio/low_min': 0.017159562092274427, 'clip_ratio/high_mean': 0.02390979148913175, 'clip_ratio/high_max': 0.02390979148913175, 'clip_ratio/region_mean': 0.041069353581406176, 'reward_total_mean': 0.9914366006851196, 'reward_meter_mean': 0.9914366006851196, 'reward_meter_std': 0.004417366813868284, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9914366006851196, 'reward_total_composite_std': 0.004417366813868284, 'epoch': 0.11}

 82%|████████▏ | 2721/3300 [3:59:38<50:52,  5.27s/it]INFO 04-12 02:36:29 [block_pool.py:378] Successfully reset prefix cache

 82%|████████▏ | 2722/3300 [3:59:42<48:54,  5.08s/it]2026-04-12 02:36:34,000 | INFO | train_grpo_train | metrics_logged mode=train step=2722

                                                     
{'loss': 0.0176, 'grad_norm': 11.017292976379395, 'learning_rate': 1.7545454545454545e-06, 'num_tokens': 6166347.0, 'completions/mean_length': 69.125, 'completions/min_length': 68.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9949984550476074, 'rewards/meter/std': 0.011957395821809769, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949984550476074, 'rewards/total_composite/std': 0.011957395821809769, 'reward': 0.9949984550476074, 'reward_std': 0.01195740420371294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05068688467144966, 'sampling/sampling_logp_difference/max': 2.848238945007324, 'sampling/importance_sampling_ratio/min': 0.05794627591967583, 'sampling/importance_sampling_ratio/mean': 0.996598482131958, 'sampling/importance_sampling_ratio/max': 1.5986580848693848, 'entropy': 0.22193360701203346, 'clip_ratio/low_mean': 0.005281690042465925, 'clip_ratio/low_min': 0.005281690042465925, 'clip_ratio/high_mean': 0.027233949513174593, 'clip_ratio/high_max': 0.027233949513174593, 'clip_ratio/region_mean': 0.03251563955564052, 'reward_total_mean': 0.9949984550476074, 'reward_meter_mean': 0.9949984550476074, 'reward_meter_std': 0.011957395821809769, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949984550476074, 'reward_total_composite_std': 0.011957395821809769, 'epoch': 0.11}

 82%|████████▏ | 2722/3300 [3:59:42<48:54,  5.08s/it]INFO 04-12 02:36:34 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2723/3300 [3:59:49<53:22,  5.55s/it]2026-04-12 02:36:40,656 | INFO | train_grpo_train | metrics_logged mode=train step=2723

                                                     
{'loss': 0.0093, 'grad_norm': 2.760967254638672, 'learning_rate': 1.7515151515151516e-06, 'num_tokens': 6169685.0, 'completions/mean_length': 235.25, 'completions/min_length': 225.0, 'completions/max_length': 242.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 235.25, 'completions/min_terminated_length': 225.0, 'completions/max_terminated_length': 242.0, 'rewards/meter/mean': 0.9979515671730042, 'rewards/meter/std': 0.0015659787459298968, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9711538553237915, 'rewards/repeat_penalty/std': 0.039811473339796066, 'rewards/total_composite/mean': 0.969149649143219, 'rewards/total_composite/std': 0.0393538773059845, 'reward': 0.969149649143219, 'reward_std': 0.0393538735806942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05202290415763855, 'sampling/sampling_logp_difference/max': 1.2681152820587158, 'sampling/importance_sampling_ratio/min': 0.2813614308834076, 'sampling/importance_sampling_ratio/mean': 1.0089517831802368, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5150004588067532, 'clip_ratio/low_mean': 0.013798366067931056, 'clip_ratio/low_min': 0.013798366067931056, 'clip_ratio/high_mean': 0.03575678775086999, 'clip_ratio/high_max': 0.03575678775086999, 'clip_ratio/region_mean': 0.049555153818801045, 'reward_total_mean': 0.969149649143219, 'reward_meter_mean': 0.9979515671730042, 'reward_meter_std': 0.0015659787459298968, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9711538553237915, 'reward_repeat_penalty_std': 0.039811473339796066, 'reward_total_composite_mean': 0.969149649143219, 'reward_total_composite_std': 0.0393538773059845, 'epoch': 0.11}

 83%|████████▎ | 2723/3300 [3:59:49<53:22,  5.55s/it]INFO 04-12 02:36:40 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2724/3300 [3:59:54<50:23,  5.25s/it]2026-04-12 02:36:45,202 | INFO | train_grpo_train | metrics_logged mode=train step=2724

                                                     
{'loss': 0.0275, 'grad_norm': 13.60006046295166, 'learning_rate': 1.7484848484848486e-06, 'num_tokens': 6171486.0, 'completions/mean_length': 65.125, 'completions/min_length': 61.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9050167798995972, 'rewards/meter/std': 0.11941422522068024, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9050167798995972, 'rewards/total_composite/std': 0.11941422522068024, 'reward': 0.9050167798995972, 'reward_std': 0.11941422522068024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07930425554513931, 'sampling/sampling_logp_difference/max': 1.9395782947540283, 'sampling/importance_sampling_ratio/min': 0.14376457035541534, 'sampling/importance_sampling_ratio/mean': 0.9989842772483826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36046687327325344, 'clip_ratio/low_mean': 0.014925372786819935, 'clip_ratio/low_min': 0.014925372786819935, 'clip_ratio/high_mean': 0.0387622892158106, 'clip_ratio/high_max': 0.0387622892158106, 'clip_ratio/region_mean': 0.05368766200263053, 'reward_total_mean': 0.9050167798995972, 'reward_meter_mean': 0.9050167798995972, 'reward_meter_std': 0.11941422522068024, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9050167798995972, 'reward_total_composite_std': 0.11941422522068024, 'epoch': 0.11}

 83%|████████▎ | 2724/3300 [3:59:54<50:23,  5.25s/it]INFO 04-12 02:36:45 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2725/3300 [3:59:58<48:35,  5.07s/it]2026-04-12 02:36:49,854 | INFO | train_grpo_train | metrics_logged mode=train step=2725

                                                     
{'loss': -0.0057, 'grad_norm': 5.4172844886779785, 'learning_rate': 1.7454545454545456e-06, 'num_tokens': 6173408.0, 'completions/mean_length': 70.25, 'completions/min_length': 68.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9948872923851013, 'rewards/meter/std': 0.0014726987574249506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9948872923851013, 'rewards/total_composite/std': 0.0014726987574249506, 'reward': 0.9948872923851013, 'reward_std': 0.0014726977096870542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04584960639476776, 'sampling/sampling_logp_difference/max': 1.2650136947631836, 'sampling/importance_sampling_ratio/min': 0.28223544359207153, 'sampling/importance_sampling_ratio/mean': 1.0125356912612915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2636737637221813, 'clip_ratio/low_mean': 0.01608805637806654, 'clip_ratio/low_min': 0.01608805637806654, 'clip_ratio/high_mean': 0.012341346475295722, 'clip_ratio/high_max': 0.012341346475295722, 'clip_ratio/region_mean': 0.028429402853362262, 'reward_total_mean': 0.9948872923851013, 'reward_meter_mean': 0.9948872923851013, 'reward_meter_std': 0.0014726987574249506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9948872923851013, 'reward_total_composite_std': 0.0014726987574249506, 'epoch': 0.11}

 83%|████████▎ | 2725/3300 [3:59:58<48:35,  5.07s/it]INFO 04-12 02:36:50 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2726/3300 [4:00:03<48:26,  5.06s/it]2026-04-12 02:36:54,905 | INFO | train_grpo_train | metrics_logged mode=train step=2726

                                                     
{'loss': -0.004, 'grad_norm': 2.82613468170166, 'learning_rate': 1.7424242424242427e-06, 'num_tokens': 6175647.0, 'completions/mean_length': 103.875, 'completions/min_length': 102.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.875, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9729191064834595, 'rewards/meter/std': 0.010998339392244816, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9729191064834595, 'rewards/total_composite/std': 0.010998339392244816, 'reward': 0.9729191064834595, 'reward_std': 0.010998345911502838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0373404435813427, 'sampling/sampling_logp_difference/max': 1.3049564361572266, 'sampling/importance_sampling_ratio/min': 0.27118435502052307, 'sampling/importance_sampling_ratio/mean': 1.0128881931304932, 'sampling/importance_sampling_ratio/max': 1.7675621509552002, 'entropy': 0.3632195107638836, 'clip_ratio/low_mean': 0.01458781841211021, 'clip_ratio/low_min': 0.01458781841211021, 'clip_ratio/high_mean': 0.009652552893385291, 'clip_ratio/high_max': 0.009652552893385291, 'clip_ratio/region_mean': 0.0242403713054955, 'reward_total_mean': 0.9729191064834595, 'reward_meter_mean': 0.9729191064834595, 'reward_meter_std': 0.010998339392244816, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9729191064834595, 'reward_total_composite_std': 0.010998339392244816, 'epoch': 0.11}

 83%|████████▎ | 2726/3300 [4:00:03<48:26,  5.06s/it]INFO 04-12 02:36:55 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2727/3300 [4:00:08<47:02,  4.93s/it]2026-04-12 02:36:59,508 | INFO | train_grpo_train | metrics_logged mode=train step=2727

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.7393939393939397e-06, 'num_tokens': 6177392.0, 'completions/mean_length': 71.125, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.99944007396698, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99944007396698, 'rewards/total_composite/std': 0.0, 'reward': 0.99944007396698, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.004035182762891054, 'sampling/sampling_logp_difference/max': 0.31750011444091797, 'sampling/importance_sampling_ratio/min': 0.7279666066169739, 'sampling/importance_sampling_ratio/mean': 1.0021454095840454, 'sampling/importance_sampling_ratio/max': 1.0543627738952637, 'entropy': 0.03847737517207861, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.99944007396698, 'reward_meter_mean': 0.99944007396698, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99944007396698, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 83%|████████▎ | 2727/3300 [4:00:08<47:02,  4.93s/it]INFO 04-12 02:36:59 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2728/3300 [4:00:12<44:32,  4.67s/it]2026-04-12 02:37:03,594 | INFO | train_grpo_train | metrics_logged mode=train step=2728

                                                     
{'loss': -0.0004, 'grad_norm': 1.235407829284668, 'learning_rate': 1.7363636363636366e-06, 'num_tokens': 6178952.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992912411689758, 'rewards/meter/std': 1.968258584383875e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992912411689758, 'rewards/total_composite/std': 1.968258584383875e-05, 'reward': 0.9992912411689758, 'reward_std': 1.9688604879775085e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0030222726054489613, 'sampling/sampling_logp_difference/max': 0.4845571517944336, 'sampling/importance_sampling_ratio/min': 0.6159699559211731, 'sampling/importance_sampling_ratio/mean': 0.9995523691177368, 'sampling/importance_sampling_ratio/max': 1.0773866176605225, 'entropy': 0.011132915853522718, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992912411689758, 'reward_meter_mean': 0.9992912411689758, 'reward_meter_std': 1.968258584383875e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992912411689758, 'reward_total_composite_std': 1.968258584383875e-05, 'epoch': 0.11}

 83%|████████▎ | 2728/3300 [4:00:12<44:32,  4.67s/it]INFO 04-12 02:37:03 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2729/3300 [4:00:20<54:06,  5.69s/it]2026-04-12 02:37:11,638 | INFO | train_grpo_train | metrics_logged mode=train step=2729

                                                     
{'loss': 0.0017, 'grad_norm': 1.5231090784072876, 'learning_rate': 1.7333333333333336e-06, 'num_tokens': 6183584.0, 'completions/mean_length': 345.0, 'completions/min_length': 339.0, 'completions/max_length': 359.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 345.0, 'completions/min_terminated_length': 339.0, 'completions/max_terminated_length': 359.0, 'rewards/meter/mean': 0.9989666938781738, 'rewards/meter/std': 0.00017414879403077066, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9558823108673096, 'rewards/repeat_penalty/std': 0.0608881339430809, 'rewards/total_composite/mean': 0.8594046235084534, 'rewards/total_composite/std': 0.05473008379340172, 'reward': 0.8594046235084534, 'reward_std': 0.05473008751869202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05097750574350357, 'sampling/sampling_logp_difference/max': 1.327122688293457, 'sampling/importance_sampling_ratio/min': 0.2652393579483032, 'sampling/importance_sampling_ratio/mean': 1.0122843980789185, 'sampling/importance_sampling_ratio/max': 1.6215828657150269, 'entropy': 0.4432450830936432, 'clip_ratio/low_mean': 0.01417952380143106, 'clip_ratio/low_min': 0.01417952380143106, 'clip_ratio/high_mean': 0.0216883085668087, 'clip_ratio/high_max': 0.0216883085668087, 'clip_ratio/region_mean': 0.03586783236823976, 'reward_total_mean': 0.8594046235084534, 'reward_meter_mean': 0.9989666938781738, 'reward_meter_std': 0.00017414879403077066, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9558823108673096, 'reward_repeat_penalty_std': 0.0608881339430809, 'reward_total_composite_mean': 0.8594046235084534, 'reward_total_composite_std': 0.05473008379340172, 'epoch': 0.11}

 83%|████████▎ | 2729/3300 [4:00:20<54:06,  5.69s/it]INFO 04-12 02:37:11 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2730/3300 [4:00:24<50:13,  5.29s/it]2026-04-12 02:37:15,993 | INFO | train_grpo_train | metrics_logged mode=train step=2730

                                                     
{'loss': -0.0005, 'grad_norm': 14.159218788146973, 'learning_rate': 1.7303030303030304e-06, 'num_tokens': 6185301.0, 'completions/mean_length': 44.625, 'completions/min_length': 43.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9303584098815918, 'rewards/meter/std': 0.03593900799751282, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9303584098815918, 'rewards/total_composite/std': 0.03593900799751282, 'reward': 0.9303584098815918, 'reward_std': 0.035938993096351624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1046677976846695, 'sampling/sampling_logp_difference/max': 6.265340328216553, 'sampling/importance_sampling_ratio/min': 0.0019010662799701095, 'sampling/importance_sampling_ratio/mean': 1.024146556854248, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37994090653955936, 'clip_ratio/low_mean': 0.017045455053448677, 'clip_ratio/low_min': 0.017045455053448677, 'clip_ratio/high_mean': 0.061030346201732755, 'clip_ratio/high_max': 0.061030346201732755, 'clip_ratio/region_mean': 0.07807580125518143, 'reward_total_mean': 0.9303584098815918, 'reward_meter_mean': 0.9303584098815918, 'reward_meter_std': 0.03593900799751282, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9303584098815918, 'reward_total_composite_std': 0.03593900799751282, 'epoch': 0.11}

 83%|████████▎ | 2730/3300 [4:00:24<50:13,  5.29s/it]INFO 04-12 02:37:16 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2731/3300 [4:00:29<47:47,  5.04s/it]2026-04-12 02:37:20,458 | INFO | train_grpo_train | metrics_logged mode=train step=2731

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.7272727272727275e-06, 'num_tokens': 6187213.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005151918157935143, 'sampling/sampling_logp_difference/max': 0.0688324123620987, 'sampling/importance_sampling_ratio/min': 0.9334831237792969, 'sampling/importance_sampling_ratio/mean': 0.9999112486839294, 'sampling/importance_sampling_ratio/max': 1.0230506658554077, 'entropy': 0.003038628783542663, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 83%|████████▎ | 2731/3300 [4:00:29<47:47,  5.04s/it]INFO 04-12 02:37:20 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2732/3300 [4:00:33<45:05,  4.76s/it]2026-04-12 02:37:24,578 | INFO | train_grpo_train | metrics_logged mode=train step=2732

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.7242424242424243e-06, 'num_tokens': 6188637.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0001443149521946907, 'sampling/sampling_logp_difference/max': 0.003842935897409916, 'sampling/importance_sampling_ratio/min': 0.9961644411087036, 'sampling/importance_sampling_ratio/mean': 1.0000722408294678, 'sampling/importance_sampling_ratio/max': 1.0033706426620483, 'entropy': 0.0015087932988535613, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 83%|████████▎ | 2732/3300 [4:00:33<45:05,  4.76s/it]INFO 04-12 02:37:24 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2733/3300 [4:00:37<43:38,  4.62s/it]2026-04-12 02:37:28,858 | INFO | train_grpo_train | metrics_logged mode=train step=2733

                                                     
{'loss': 0.0137, 'grad_norm': 22.582195281982422, 'learning_rate': 1.7212121212121214e-06, 'num_tokens': 6190107.0, 'completions/mean_length': 34.75, 'completions/min_length': 34.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.945763885974884, 'rewards/meter/std': 0.12533576786518097, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.945763885974884, 'rewards/total_composite/std': 0.12533576786518097, 'reward': 0.945763885974884, 'reward_std': 0.12533576786518097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02166467346251011, 'sampling/sampling_logp_difference/max': 1.478475570678711, 'sampling/importance_sampling_ratio/min': 0.2279849797487259, 'sampling/importance_sampling_ratio/mean': 1.0029886960983276, 'sampling/importance_sampling_ratio/max': 1.4781155586242676, 'entropy': 0.12251334544271231, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.0035714285913854837, 'clip_ratio/high_max': 0.0035714285913854837, 'clip_ratio/region_mean': 0.0070436508394777775, 'reward_total_mean': 0.945763885974884, 'reward_meter_mean': 0.945763885974884, 'reward_meter_std': 0.12533576786518097, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.945763885974884, 'reward_total_composite_std': 0.12533576786518097, 'epoch': 0.11}

 83%|████████▎ | 2733/3300 [4:00:37<43:38,  4.62s/it]INFO 04-12 02:37:29 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2734/3300 [4:00:44<50:29,  5.35s/it]2026-04-12 02:37:35,923 | INFO | train_grpo_train | metrics_logged mode=train step=2734

                                                     
{'loss': 0.0021, 'grad_norm': 2.762172222137451, 'learning_rate': 1.7181818181818182e-06, 'num_tokens': 6193858.0, 'completions/mean_length': 268.875, 'completions/min_length': 255.0, 'completions/max_length': 276.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 268.875, 'completions/min_terminated_length': 255.0, 'completions/max_terminated_length': 276.0, 'rewards/meter/mean': 0.9990511536598206, 'rewards/meter/std': 0.00018971240206155926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.990384578704834, 'rewards/repeat_penalty/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.9894447326660156, 'rewards/total_composite/std': 0.027163289487361908, 'reward': 0.9894447326660156, 'reward_std': 0.02716328203678131, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04823000356554985, 'sampling/sampling_logp_difference/max': 2.1149091720581055, 'sampling/importance_sampling_ratio/min': 0.12064424902200699, 'sampling/importance_sampling_ratio/mean': 1.0080488920211792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43885183334350586, 'clip_ratio/low_mean': 0.004182156175374985, 'clip_ratio/low_min': 0.004182156175374985, 'clip_ratio/high_mean': 0.033498246455565095, 'clip_ratio/high_max': 0.033498246455565095, 'clip_ratio/region_mean': 0.03768040263094008, 'reward_total_mean': 0.9894447326660156, 'reward_meter_mean': 0.9990511536598206, 'reward_meter_std': 0.00018971240206155926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.990384578704834, 'reward_repeat_penalty_std': 0.027196412906050682, 'reward_total_composite_mean': 0.9894447326660156, 'reward_total_composite_std': 0.027163289487361908, 'epoch': 0.11}

 83%|████████▎ | 2734/3300 [4:00:44<50:29,  5.35s/it]INFO 04-12 02:37:36 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2735/3300 [4:00:49<48:16,  5.13s/it]2026-04-12 02:37:40,524 | INFO | train_grpo_train | metrics_logged mode=train step=2735

                                                     
{'loss': -0.0023, 'grad_norm': 2.1266746520996094, 'learning_rate': 1.7151515151515152e-06, 'num_tokens': 6195785.0, 'completions/mean_length': 70.875, 'completions/min_length': 70.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.999404788017273, 'rewards/meter/std': 0.00010114238102687523, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999404788017273, 'rewards/total_composite/std': 0.00010114238102687523, 'reward': 0.999404788017273, 'reward_std': 0.00010115459008375183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007956513203680515, 'sampling/sampling_logp_difference/max': 0.6195621490478516, 'sampling/importance_sampling_ratio/min': 0.5381800532341003, 'sampling/importance_sampling_ratio/mean': 1.0012547969818115, 'sampling/importance_sampling_ratio/max': 1.3098119497299194, 'entropy': 0.05470029590651393, 'clip_ratio/low_mean': 0.0053571430034935474, 'clip_ratio/low_min': 0.0053571430034935474, 'clip_ratio/high_mean': 0.0035211266949772835, 'clip_ratio/high_max': 0.0035211266949772835, 'clip_ratio/region_mean': 0.008878269698470831, 'reward_total_mean': 0.999404788017273, 'reward_meter_mean': 0.999404788017273, 'reward_meter_std': 0.00010114238102687523, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999404788017273, 'reward_total_composite_std': 0.00010114238102687523, 'epoch': 0.11}

 83%|████████▎ | 2735/3300 [4:00:49<48:16,  5.13s/it]INFO 04-12 02:37:40 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2736/3300 [4:00:54<48:08,  5.12s/it]2026-04-12 02:37:45,629 | INFO | train_grpo_train | metrics_logged mode=train step=2736

                                                     
{'loss': 0.0073, 'grad_norm': 2.7460949420928955, 'learning_rate': 1.7121212121212123e-06, 'num_tokens': 6198070.0, 'completions/mean_length': 122.625, 'completions/min_length': 121.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.625, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9976873397827148, 'rewards/meter/std': 0.0002810863661579788, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9442275762557983, 'rewards/total_composite/std': 0.07359237223863602, 'reward': 0.9442275762557983, 'reward_std': 0.07359235733747482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02476387470960617, 'sampling/sampling_logp_difference/max': 0.8794323205947876, 'sampling/importance_sampling_ratio/min': 0.4150184392929077, 'sampling/importance_sampling_ratio/mean': 0.9990551471710205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1498414622619748, 'clip_ratio/low_mean': 0.004048783564940095, 'clip_ratio/low_min': 0.004048783564940095, 'clip_ratio/high_mean': 0.023400326492264867, 'clip_ratio/high_max': 0.023400326492264867, 'clip_ratio/region_mean': 0.027449110057204962, 'reward_total_mean': 0.9442275762557983, 'reward_meter_mean': 0.9976873397827148, 'reward_meter_std': 0.0002810863661579788, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9442275762557983, 'reward_total_composite_std': 0.07359237223863602, 'epoch': 0.11}

 83%|████████▎ | 2736/3300 [4:00:54<48:08,  5.12s/it]INFO 04-12 02:37:45 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2737/3300 [4:00:59<47:02,  5.01s/it]2026-04-12 02:37:50,388 | INFO | train_grpo_train | metrics_logged mode=train step=2737

                                                     
{'loss': -0.001, 'grad_norm': 0.5398823618888855, 'learning_rate': 1.7090909090909091e-06, 'num_tokens': 6200301.0, 'completions/mean_length': 97.875, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994151592254639, 'rewards/meter/std': 1.6079073247965425e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994151592254639, 'rewards/total_composite/std': 1.6079073247965425e-05, 'reward': 0.9994151592254639, 'reward_std': 1.6088057236629538e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0048310887068510056, 'sampling/sampling_logp_difference/max': 1.8052539825439453, 'sampling/importance_sampling_ratio/min': 0.16443268954753876, 'sampling/importance_sampling_ratio/mean': 0.9999629259109497, 'sampling/importance_sampling_ratio/max': 1.3984962701797485, 'entropy': 0.015126468031667173, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0012755101779475808, 'clip_ratio/high_max': 0.0012755101779475808, 'clip_ratio/region_mean': 0.0012755101779475808, 'reward_total_mean': 0.9994151592254639, 'reward_meter_mean': 0.9994151592254639, 'reward_meter_std': 1.6079073247965425e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994151592254639, 'reward_total_composite_std': 1.6079073247965425e-05, 'epoch': 0.11}

 83%|████████▎ | 2737/3300 [4:00:59<47:02,  5.01s/it]INFO 04-12 02:37:50 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2738/3300 [4:01:04<46:17,  4.94s/it]2026-04-12 02:37:55,163 | INFO | train_grpo_train | metrics_logged mode=train step=2738

                                                     
{'loss': -0.0017, 'grad_norm': 0.7919180393218994, 'learning_rate': 1.7060606060606062e-06, 'num_tokens': 6202503.0, 'completions/mean_length': 97.25, 'completions/min_length': 96.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.25, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9980356693267822, 'rewards/meter/std': 6.821734132245183e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980356693267822, 'rewards/total_composite/std': 6.821734132245183e-05, 'reward': 0.9980356693267822, 'reward_std': 6.822366412961856e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007202071137726307, 'sampling/sampling_logp_difference/max': 0.41276001930236816, 'sampling/importance_sampling_ratio/min': 0.6618210673332214, 'sampling/importance_sampling_ratio/mean': 1.003057837486267, 'sampling/importance_sampling_ratio/max': 1.4862256050109863, 'entropy': 0.06638350430876017, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006403850042261183, 'clip_ratio/high_max': 0.006403850042261183, 'clip_ratio/region_mean': 0.006403850042261183, 'reward_total_mean': 0.9980356693267822, 'reward_meter_mean': 0.9980356693267822, 'reward_meter_std': 6.821734132245183e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980356693267822, 'reward_total_composite_std': 6.821734132245183e-05, 'epoch': 0.11}

 83%|████████▎ | 2738/3300 [4:01:04<46:17,  4.94s/it]INFO 04-12 02:37:55 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2739/3300 [4:01:08<44:21,  4.74s/it]2026-04-12 02:37:59,447 | INFO | train_grpo_train | metrics_logged mode=train step=2739

                                                     
{'loss': -0.0016, 'grad_norm': 4.850316524505615, 'learning_rate': 1.703030303030303e-06, 'num_tokens': 6204047.0, 'completions/mean_length': 44.0, 'completions/min_length': 42.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9376137256622314, 'rewards/meter/std': 0.004042606335133314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9376137256622314, 'rewards/total_composite/std': 0.004042606335133314, 'reward': 0.9376137256622314, 'reward_std': 0.004042605869472027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0746898204088211, 'sampling/sampling_logp_difference/max': 2.011808395385742, 'sampling/importance_sampling_ratio/min': 0.13374659419059753, 'sampling/importance_sampling_ratio/mean': 0.9946960210800171, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21985023096203804, 'clip_ratio/low_mean': 0.04031385388225317, 'clip_ratio/low_min': 0.04031385388225317, 'clip_ratio/high_mean': 0.016798419412225485, 'clip_ratio/high_max': 0.016798419412225485, 'clip_ratio/region_mean': 0.057112273294478655, 'reward_total_mean': 0.9376137256622314, 'reward_meter_mean': 0.9376137256622314, 'reward_meter_std': 0.004042606335133314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9376137256622314, 'reward_total_composite_std': 0.004042606335133314, 'epoch': 0.11}

 83%|████████▎ | 2739/3300 [4:01:08<44:21,  4.74s/it]INFO 04-12 02:37:59 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2740/3300 [4:01:12<43:30,  4.66s/it]2026-04-12 02:38:03,919 | INFO | train_grpo_train | metrics_logged mode=train step=2740

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.7000000000000002e-06, 'num_tokens': 6205946.0, 'completions/mean_length': 66.375, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0036742938682436943, 'sampling/sampling_logp_difference/max': 0.5825433731079102, 'sampling/importance_sampling_ratio/min': 0.5584760904312134, 'sampling/importance_sampling_ratio/mean': 0.9993410110473633, 'sampling/importance_sampling_ratio/max': 1.19320809841156, 'entropy': 0.018359567504376173, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 83%|████████▎ | 2740/3300 [4:01:12<43:30,  4.66s/it]INFO 04-12 02:38:04 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2741/3300 [4:01:17<44:09,  4.74s/it]2026-04-12 02:38:08,840 | INFO | train_grpo_train | metrics_logged mode=train step=2741

                                                     
{'loss': -0.0202, 'grad_norm': 1.060721755027771, 'learning_rate': 1.6969696969696973e-06, 'num_tokens': 6207995.0, 'completions/mean_length': 96.125, 'completions/min_length': 92.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.125, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9955133199691772, 'rewards/meter/std': 0.004802301991730928, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955133199691772, 'rewards/total_composite/std': 0.004802301991730928, 'reward': 0.9955133199691772, 'reward_std': 0.004802299663424492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006459005642682314, 'sampling/sampling_logp_difference/max': 0.5894908905029297, 'sampling/importance_sampling_ratio/min': 0.5616306066513062, 'sampling/importance_sampling_ratio/mean': 1.0021657943725586, 'sampling/importance_sampling_ratio/max': 1.8030701875686646, 'entropy': 0.07564361020922661, 'clip_ratio/low_mean': 0.005348057369701564, 'clip_ratio/low_min': 0.005348057369701564, 'clip_ratio/high_mean': 0.0012886597542092204, 'clip_ratio/high_max': 0.0012886597542092204, 'clip_ratio/region_mean': 0.006636717123910785, 'reward_total_mean': 0.9955133199691772, 'reward_meter_mean': 0.9955133199691772, 'reward_meter_std': 0.004802301991730928, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955133199691772, 'reward_total_composite_std': 0.004802301991730928, 'epoch': 0.11}

 83%|████████▎ | 2741/3300 [4:01:17<44:09,  4.74s/it]INFO 04-12 02:38:09 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2742/3300 [4:01:22<44:49,  4.82s/it]2026-04-12 02:38:13,846 | INFO | train_grpo_train | metrics_logged mode=train step=2742

                                                     
{'loss': 0.0053, 'grad_norm': 3.1250481605529785, 'learning_rate': 1.6939393939393941e-06, 'num_tokens': 6210017.0, 'completions/mean_length': 100.75, 'completions/min_length': 96.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.75, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9990947842597961, 'rewards/meter/std': 0.0001982693502213806, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990947842597961, 'rewards/total_composite/std': 0.0001982693502213806, 'reward': 0.9990947842597961, 'reward_std': 0.00019826588686555624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043827299028635025, 'sampling/sampling_logp_difference/max': 2.5691757202148438, 'sampling/importance_sampling_ratio/min': 0.07659865915775299, 'sampling/importance_sampling_ratio/mean': 1.0069540739059448, 'sampling/importance_sampling_ratio/max': 1.9947516918182373, 'entropy': 0.329579945653677, 'clip_ratio/low_mean': 0.011077517876401544, 'clip_ratio/low_min': 0.011077517876401544, 'clip_ratio/high_mean': 0.02870650147087872, 'clip_ratio/high_max': 0.02870650147087872, 'clip_ratio/region_mean': 0.039784019347280264, 'reward_total_mean': 0.9990947842597961, 'reward_meter_mean': 0.9990947842597961, 'reward_meter_std': 0.0001982693502213806, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990947842597961, 'reward_total_composite_std': 0.0001982693502213806, 'epoch': 0.11}

 83%|████████▎ | 2742/3300 [4:01:22<44:49,  4.82s/it]INFO 04-12 02:38:14 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2743/3300 [4:01:27<44:21,  4.78s/it]2026-04-12 02:38:18,535 | INFO | train_grpo_train | metrics_logged mode=train step=2743

                                                     
{'loss': 0.0, 'grad_norm': 2.3616504669189453, 'learning_rate': 1.6909090909090912e-06, 'num_tokens': 6211811.0, 'completions/mean_length': 69.25, 'completions/min_length': 67.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9709841012954712, 'rewards/meter/std': 0.02855425886809826, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9709841012954712, 'rewards/total_composite/std': 0.02855425886809826, 'reward': 0.9709841012954712, 'reward_std': 0.028554245829582214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04137599095702171, 'sampling/sampling_logp_difference/max': 1.5528497695922852, 'sampling/importance_sampling_ratio/min': 0.3372916579246521, 'sampling/importance_sampling_ratio/mean': 1.0066807270050049, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33481148816645145, 'clip_ratio/low_mean': 0.014685500762425363, 'clip_ratio/low_min': 0.014685500762425363, 'clip_ratio/high_mean': 0.01622902590315789, 'clip_ratio/high_max': 0.01622902590315789, 'clip_ratio/region_mean': 0.030914526665583253, 'reward_total_mean': 0.9709841012954712, 'reward_meter_mean': 0.9709841012954712, 'reward_meter_std': 0.02855425886809826, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9709841012954712, 'reward_total_composite_std': 0.02855425886809826, 'epoch': 0.11}

 83%|████████▎ | 2743/3300 [4:01:27<44:21,  4.78s/it]INFO 04-12 02:38:18 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2744/3300 [4:01:35<52:07,  5.62s/it]2026-04-12 02:38:26,129 | INFO | train_grpo_train | metrics_logged mode=train step=2744

                                                     
{'loss': -0.0272, 'grad_norm': 3.2804408073425293, 'learning_rate': 1.687878787878788e-06, 'num_tokens': 6215771.0, 'completions/mean_length': 294.0, 'completions/min_length': 272.0, 'completions/max_length': 308.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 294.0, 'completions/min_terminated_length': 272.0, 'completions/max_terminated_length': 308.0, 'rewards/meter/mean': 0.996207058429718, 'rewards/meter/std': 0.0008007285068742931, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.04629101976752281, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9663312435150146, 'rewards/repeat_penalty/std': 0.039662934839725494, 'rewards/total_composite/mean': 0.9384926557540894, 'rewards/total_composite/std': 0.05730018764734268, 'reward': 0.9384926557540894, 'reward_std': 0.057300180196762085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04739288613200188, 'sampling/sampling_logp_difference/max': 1.8937530517578125, 'sampling/importance_sampling_ratio/min': 0.1505059003829956, 'sampling/importance_sampling_ratio/mean': 1.002259373664856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35994909703731537, 'clip_ratio/low_mean': 0.013101526303216815, 'clip_ratio/low_min': 0.013101526303216815, 'clip_ratio/high_mean': 0.026779879350215197, 'clip_ratio/high_max': 0.026779879350215197, 'clip_ratio/region_mean': 0.03988140565343201, 'reward_total_mean': 0.9384926557540894, 'reward_meter_mean': 0.996207058429718, 'reward_meter_std': 0.0008007285068742931, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.04629101976752281, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9663312435150146, 'reward_repeat_penalty_std': 0.039662934839725494, 'reward_total_composite_mean': 0.9384926557540894, 'reward_total_composite_std': 0.05730018764734268, 'epoch': 0.11}

 83%|████████▎ | 2744/3300 [4:01:35<52:07,  5.62s/it]INFO 04-12 02:38:26 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2745/3300 [4:01:45<1:04:38,  6.99s/it]2026-04-12 02:38:36,304 | INFO | train_grpo_train | metrics_logged mode=train step=2745

                                                       
{'loss': -0.2843, 'grad_norm': 0.6794636249542236, 'learning_rate': 1.684848484848485e-06, 'num_tokens': 6220927.0, 'completions/mean_length': 481.5, 'completions/min_length': 458.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 477.14288330078125, 'completions/min_terminated_length': 458.0, 'completions/max_terminated_length': 497.0, 'rewards/meter/mean': 0.9529277086257935, 'rewards/meter/std': 0.12985190749168396, 'rewards/count_adherence/mean': 0.8515625, 'rewards/count_adherence/std': 0.04650149121880531, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8053357601165771, 'rewards/repeat_penalty/std': 0.025981293991208076, 'rewards/total_composite/mean': 0.6496540307998657, 'rewards/total_composite/std': 0.07202339172363281, 'reward': 0.6496540307998657, 'reward_std': 0.07202339917421341, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024653667584061623, 'sampling/sampling_logp_difference/max': 2.698049783706665, 'sampling/importance_sampling_ratio/min': 0.06733670830726624, 'sampling/importance_sampling_ratio/mean': 1.0057053565979004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17778486385941505, 'clip_ratio/low_mean': 0.0029986235313117504, 'clip_ratio/low_min': 0.0029986235313117504, 'clip_ratio/high_mean': 0.012085011578164995, 'clip_ratio/high_max': 0.012085011578164995, 'clip_ratio/region_mean': 0.015083635109476745, 'reward_total_mean': 0.6496540307998657, 'reward_meter_mean': 0.9529277086257935, 'reward_meter_std': 0.12985190749168396, 'reward_count_adherence_mean': 0.8515625, 'reward_count_adherence_std': 0.04650149121880531, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8053357601165771, 'reward_repeat_penalty_std': 0.025981293991208076, 'reward_total_composite_mean': 0.6496540307998657, 'reward_total_composite_std': 0.07202339172363281, 'epoch': 0.11}

 83%|████████▎ | 2745/3300 [4:01:45<1:04:38,  6.99s/it]INFO 04-12 02:38:36 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2746/3300 [4:01:49<57:25,  6.22s/it]  2026-04-12 02:38:40,725 | INFO | train_grpo_train | metrics_logged mode=train step=2746

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.6818181818181819e-06, 'num_tokens': 6222631.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003325922298245132, 'sampling/sampling_logp_difference/max': 0.06264449656009674, 'sampling/importance_sampling_ratio/min': 0.9392772912979126, 'sampling/importance_sampling_ratio/mean': 0.9999988079071045, 'sampling/importance_sampling_ratio/max': 1.0122551918029785, 'entropy': 0.002097451186273247, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 83%|████████▎ | 2746/3300 [4:01:49<57:25,  6.22s/it]INFO 04-12 02:38:40 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2747/3300 [4:01:54<52:50,  5.73s/it]2026-04-12 02:38:45,326 | INFO | train_grpo_train | metrics_logged mode=train step=2747

                                                     
{'loss': 0.0146, 'grad_norm': 3.3234918117523193, 'learning_rate': 1.678787878787879e-06, 'num_tokens': 6224387.0, 'completions/mean_length': 68.5, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.973109781742096, 'rewards/meter/std': 0.031855545938014984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.973109781742096, 'rewards/total_composite/std': 0.031855545938014984, 'reward': 0.973109781742096, 'reward_std': 0.031855542212724686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038049403578042984, 'sampling/sampling_logp_difference/max': 1.5401935577392578, 'sampling/importance_sampling_ratio/min': 0.21433962881565094, 'sampling/importance_sampling_ratio/mean': 1.0069949626922607, 'sampling/importance_sampling_ratio/max': 1.755409598350525, 'entropy': 0.2892153840512037, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.020309405983425677, 'clip_ratio/high_max': 0.020309405983425677, 'clip_ratio/region_mean': 0.02209512027911842, 'reward_total_mean': 0.973109781742096, 'reward_meter_mean': 0.973109781742096, 'reward_meter_std': 0.031855545938014984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.973109781742096, 'reward_total_composite_std': 0.031855545938014984, 'epoch': 0.11}

 83%|████████▎ | 2747/3300 [4:01:54<52:50,  5.73s/it]INFO 04-12 02:38:45 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2748/3300 [4:02:01<57:49,  6.29s/it]2026-04-12 02:38:52,900 | INFO | train_grpo_train | metrics_logged mode=train step=2748

                                                     
{'loss': 0.0236, 'grad_norm': 1.725966215133667, 'learning_rate': 1.675757575757576e-06, 'num_tokens': 6228377.0, 'completions/mean_length': 296.75, 'completions/min_length': 268.0, 'completions/max_length': 309.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 296.75, 'completions/min_terminated_length': 268.0, 'completions/max_terminated_length': 309.0, 'rewards/meter/mean': 0.9991159439086914, 'rewards/meter/std': 0.0003877849958371371, 'rewards/count_adherence/mean': 0.890625, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9171568751335144, 'rewards/repeat_penalty/std': 0.033501796424388885, 'rewards/total_composite/mean': 0.8153371810913086, 'rewards/total_composite/std': 0.031080491840839386, 'reward': 0.8153371810913086, 'reward_std': 0.031080493703484535, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025589879602193832, 'sampling/sampling_logp_difference/max': 3.3428661823272705, 'sampling/importance_sampling_ratio/min': 0.03533553332090378, 'sampling/importance_sampling_ratio/mean': 1.0036824941635132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16522958502173424, 'clip_ratio/low_mean': 0.005000000121071935, 'clip_ratio/low_min': 0.005000000121071935, 'clip_ratio/high_mean': 0.01721590873785317, 'clip_ratio/high_max': 0.01721590873785317, 'clip_ratio/region_mean': 0.022215908858925104, 'reward_total_mean': 0.8153371810913086, 'reward_meter_mean': 0.9991159439086914, 'reward_meter_std': 0.0003877849958371371, 'reward_count_adherence_mean': 0.890625, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9171568751335144, 'reward_repeat_penalty_std': 0.033501796424388885, 'reward_total_composite_mean': 0.8153371810913086, 'reward_total_composite_std': 0.031080491840839386, 'epoch': 0.11}

 83%|████████▎ | 2748/3300 [4:02:01<57:49,  6.29s/it]INFO 04-12 02:38:53 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2749/3300 [4:02:06<54:22,  5.92s/it]2026-04-12 02:38:57,975 | INFO | train_grpo_train | metrics_logged mode=train step=2749

                                                     
{'loss': -0.0163, 'grad_norm': 3.2416765689849854, 'learning_rate': 1.6727272727272728e-06, 'num_tokens': 6230560.0, 'completions/mean_length': 101.875, 'completions/min_length': 99.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.875, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9238500595092773, 'rewards/meter/std': 0.10599032789468765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9238500595092773, 'rewards/total_composite/std': 0.10599032789468765, 'reward': 0.9238500595092773, 'reward_std': 0.10599032789468765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03779112920165062, 'sampling/sampling_logp_difference/max': 1.4791641235351562, 'sampling/importance_sampling_ratio/min': 0.22782805562019348, 'sampling/importance_sampling_ratio/mean': 1.008776307106018, 'sampling/importance_sampling_ratio/max': 1.7313110828399658, 'entropy': 0.36418476328253746, 'clip_ratio/low_mean': 0.011289363959804177, 'clip_ratio/low_min': 0.011289363959804177, 'clip_ratio/high_mean': 0.019543501897715032, 'clip_ratio/high_max': 0.019543501897715032, 'clip_ratio/region_mean': 0.03083286585751921, 'reward_total_mean': 0.9238500595092773, 'reward_meter_mean': 0.9238500595092773, 'reward_meter_std': 0.10599032789468765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9238500595092773, 'reward_total_composite_std': 0.10599032789468765, 'epoch': 0.11}

 83%|████████▎ | 2749/3300 [4:02:06<54:22,  5.92s/it]INFO 04-12 02:38:58 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2750/3300 [4:02:11<51:42,  5.64s/it]2026-04-12 02:39:02,961 | INFO | train_grpo_train | metrics_logged mode=train step=2750

                                                     
{'loss': -0.0042, 'grad_norm': 3.4270999431610107, 'learning_rate': 1.6696969696969698e-06, 'num_tokens': 6232648.0, 'completions/mean_length': 97.0, 'completions/min_length': 96.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.0, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9976779818534851, 'rewards/meter/std': 0.0009901836747303605, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976779818534851, 'rewards/total_composite/std': 0.0009901836747303605, 'reward': 0.9976779818534851, 'reward_std': 0.0009901747107505798, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005543787498027086, 'sampling/sampling_logp_difference/max': 0.7107429504394531, 'sampling/importance_sampling_ratio/min': 0.49127906560897827, 'sampling/importance_sampling_ratio/mean': 1.0021347999572754, 'sampling/importance_sampling_ratio/max': 1.6211403608322144, 'entropy': 0.07637950498610735, 'clip_ratio/low_mean': 0.0013020833721384406, 'clip_ratio/low_min': 0.0013020833721384406, 'clip_ratio/high_mean': 0.008981169667094946, 'clip_ratio/high_max': 0.008981169667094946, 'clip_ratio/region_mean': 0.010283253039233387, 'reward_total_mean': 0.9976779818534851, 'reward_meter_mean': 0.9976779818534851, 'reward_meter_std': 0.0009901836747303605, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976779818534851, 'reward_total_composite_std': 0.0009901836747303605, 'epoch': 0.11}

 83%|████████▎ | 2750/3300 [4:02:11<51:42,  5.64s/it]INFO 04-12 02:39:03 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:36,  3.36s/it]

 23%|██▎       | 3/13 [00:10<00:35,  3.55s/it]

 31%|███       | 4/13 [00:16<00:39,  4.43s/it]

 38%|███▊      | 5/13 [00:21<00:37,  4.70s/it]

 46%|████▌     | 6/13 [00:28<00:37,  5.32s/it]

 54%|█████▍    | 7/13 [00:33<00:31,  5.28s/it]

 62%|██████▏   | 8/13 [00:40<00:29,  5.84s/it]

 69%|██████▉   | 9/13 [00:47<00:25,  6.26s/it]

 77%|███████▋  | 10/13 [00:52<00:17,  5.88s/it]

 85%|████████▍ | 11/13 [00:58<00:11,  5.83s/it]

 92%|█████████▏| 12/13 [01:04<00:05,  5.90s/it]

100%|██████████| 13/13 [01:09<00:00,  5.70s/it]2026-04-12 02:40:19,435 | INFO | train_grpo_train | metrics_logged mode=eval step=2750

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 76.3731, 'eval_samples_per_second': 1.362, 'eval_steps_per_second': 0.17, 'eval_num_tokens': 6232648.0, 'eval_completions/mean_length': 208.81730769230768, 'eval_completions/min_length': 59.30769230769231, 'eval_completions/max_length': 410.2307692307692, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 208.81730769230768, 'eval_completions/min_terminated_length': 59.30769230769231, 'eval_completions/max_terminated_length': 410.2307692307692, 'eval_rewards/meter/mean': 0.8095979323753943, 'eval_rewards/meter/std': 0.2978983329465756, 'eval_rewards/count_adherence/mean': 0.9510142528093778, 'eval_rewards/count_adherence/std': 0.06824518969425789, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.9246861888812139, 'eval_rewards/repeat_penalty/std': 0.09499467995304328, 'eval_rewards/total_composite/mean': 0.7168137293595535, 'eval_rewards/total_composite/std': 0.2932796369378383, 'eval_reward': 0.7168137293595535, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03025670349597931, 'eval_sampling/sampling_logp_difference/max': 1.3938810641948993, 'eval_sampling/importance_sampling_ratio/min': 0.26402970059559894, 'eval_sampling/importance_sampling_ratio/mean': 1.0077975529890795, 'eval_sampling/importance_sampling_ratio/max': 1.4980517717508168, 'eval_entropy': 0.32508696615695953, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7168137293595535, 'eval_reward_meter_mean': 0.8095979323753943, 'eval_reward_meter_std': 0.2978983329465756, 'eval_reward_count_adherence_mean': 0.9510142528093778, 'eval_reward_count_adherence_std': 0.06824518969425789, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.9246861888812139, 'eval_reward_repeat_penalty_std': 0.09499467995304328, 'eval_reward_total_composite_mean': 0.7168137293595535, 'eval_reward_total_composite_std': 0.2932796369378383, 'epoch': 0.11}

 83%|████████▎ | 2750/3300 [4:03:28<51:42,  5.64s/it]

100%|██████████| 13/13 [01:09<00:00,  5.70s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 83%|████████▎ | 2751/3300 [4:03:38<4:34:15, 29.97s/it]2026-04-12 02:40:29,786 | INFO | train_grpo_train | metrics_logged mode=train step=2751

                                                       
{'loss': -0.004, 'grad_norm': 1.9583728313446045, 'learning_rate': 1.6666666666666667e-06, 'num_tokens': 6236726.0, 'completions/mean_length': 270.75, 'completions/min_length': 266.0, 'completions/max_length': 276.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 270.75, 'completions/min_terminated_length': 266.0, 'completions/max_terminated_length': 276.0, 'rewards/meter/mean': 0.9987819194793701, 'rewards/meter/std': 0.0005108492914587259, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987819194793701, 'rewards/total_composite/std': 0.0005108492914587259, 'reward': 0.9987819194793701, 'reward_std': 0.0005108449840918183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05048610270023346, 'sampling/sampling_logp_difference/max': 1.3633785247802734, 'sampling/importance_sampling_ratio/min': 0.25579512119293213, 'sampling/importance_sampling_ratio/mean': 1.0091570615768433, 'sampling/importance_sampling_ratio/max': 1.7512770891189575, 'entropy': 0.44462598487734795, 'clip_ratio/low_mean': 0.010291178710758686, 'clip_ratio/low_min': 0.010291178710758686, 'clip_ratio/high_mean': 0.029523379867896438, 'clip_ratio/high_max': 0.029523379867896438, 'clip_ratio/region_mean': 0.039814558578655124, 'reward_total_mean': 0.9987819194793701, 'reward_meter_mean': 0.9987819194793701, 'reward_meter_std': 0.0005108492914587259, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987819194793701, 'reward_total_composite_std': 0.0005108492914587259, 'epoch': 0.11}

 83%|████████▎ | 2751/3300 [4:03:38<4:34:15, 29.97s/it]INFO 04-12 02:40:30 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2752/3300 [4:03:47<3:34:58, 23.54s/it]2026-04-12 02:40:38,301 | INFO | train_grpo_train | metrics_logged mode=train step=2752

                                                       
{'loss': 0.0122, 'grad_norm': 2.32265567779541, 'learning_rate': 1.6636363636363637e-06, 'num_tokens': 6240441.0, 'completions/mean_length': 237.375, 'completions/min_length': 214.0, 'completions/max_length': 252.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 237.375, 'completions/min_terminated_length': 214.0, 'completions/max_terminated_length': 252.0, 'rewards/meter/mean': 0.9126553535461426, 'rewards/meter/std': 0.10828904807567596, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8564560413360596, 'rewards/repeat_penalty/std': 0.10748697072267532, 'rewards/total_composite/mean': 0.7755623459815979, 'rewards/total_composite/std': 0.09196323156356812, 'reward': 0.7755623459815979, 'reward_std': 0.09196322411298752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04030514135956764, 'sampling/sampling_logp_difference/max': 2.249701738357544, 'sampling/importance_sampling_ratio/min': 0.10543066263198853, 'sampling/importance_sampling_ratio/mean': 1.007071614265442, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41799643263220787, 'clip_ratio/low_mean': 0.01261913578491658, 'clip_ratio/low_min': 0.01261913578491658, 'clip_ratio/high_mean': 0.011767410207539797, 'clip_ratio/high_max': 0.011767410207539797, 'clip_ratio/region_mean': 0.024386545992456377, 'reward_total_mean': 0.7755623459815979, 'reward_meter_mean': 0.9126553535461426, 'reward_meter_std': 0.10828904807567596, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8564560413360596, 'reward_repeat_penalty_std': 0.10748697072267532, 'reward_total_composite_mean': 0.7755623459815979, 'reward_total_composite_std': 0.09196323156356812, 'epoch': 0.11}

 83%|████████▎ | 2752/3300 [4:03:47<3:34:58, 23.54s/it]INFO 04-12 02:40:38 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2753/3300 [4:03:52<2:45:28, 18.15s/it]2026-04-12 02:40:43,875 | INFO | train_grpo_train | metrics_logged mode=train step=2753

                                                       
{'loss': 0.0132, 'grad_norm': 12.199896812438965, 'learning_rate': 1.6606060606060605e-06, 'num_tokens': 6241917.0, 'completions/mean_length': 44.5, 'completions/min_length': 41.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.930524468421936, 'rewards/meter/std': 0.031606100499629974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.930524468421936, 'rewards/total_composite/std': 0.031606100499629974, 'reward': 0.930524468421936, 'reward_std': 0.03160610795021057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07848232239484787, 'sampling/sampling_logp_difference/max': 2.279867649078369, 'sampling/importance_sampling_ratio/min': 0.10229774564504623, 'sampling/importance_sampling_ratio/mean': 0.9975581765174866, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2707332409918308, 'clip_ratio/low_mean': 0.0029761905316263437, 'clip_ratio/low_min': 0.0029761905316263437, 'clip_ratio/high_mean': 0.05645520123653114, 'clip_ratio/high_max': 0.05645520123653114, 'clip_ratio/region_mean': 0.05943139176815748, 'reward_total_mean': 0.930524468421936, 'reward_meter_mean': 0.930524468421936, 'reward_meter_std': 0.031606100499629974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.930524468421936, 'reward_total_composite_std': 0.031606100499629974, 'epoch': 0.11}

 83%|████████▎ | 2753/3300 [4:03:52<2:45:28, 18.15s/it]INFO 04-12 02:40:44 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2754/3300 [4:04:03<2:25:45, 16.02s/it]2026-04-12 02:40:54,921 | INFO | train_grpo_train | metrics_logged mode=train step=2754

                                                       
{'loss': 0.0001, 'grad_norm': 1.868078351020813, 'learning_rate': 1.6575757575757578e-06, 'num_tokens': 6246685.0, 'completions/mean_length': 400.0, 'completions/min_length': 363.0, 'completions/max_length': 450.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 400.0, 'completions/min_terminated_length': 363.0, 'completions/max_terminated_length': 450.0, 'rewards/meter/mean': 0.9454325437545776, 'rewards/meter/std': 0.0641726702451706, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.033064987510442734, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8041722774505615, 'rewards/repeat_penalty/std': 0.09782867133617401, 'rewards/total_composite/mean': 0.667221188545227, 'rewards/total_composite/std': 0.10969909280538559, 'reward': 0.667221188545227, 'reward_std': 0.10969908535480499, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048431482166051865, 'sampling/sampling_logp_difference/max': 10.797205924987793, 'sampling/importance_sampling_ratio/min': 2.045658038696274e-05, 'sampling/importance_sampling_ratio/mean': 1.0145169496536255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45330196619033813, 'clip_ratio/low_mean': 0.017272761091589928, 'clip_ratio/low_min': 0.017272761091589928, 'clip_ratio/high_mean': 0.016477052122354507, 'clip_ratio/high_max': 0.016477052122354507, 'clip_ratio/region_mean': 0.033749813213944435, 'reward_total_mean': 0.667221188545227, 'reward_meter_mean': 0.9454325437545776, 'reward_meter_std': 0.0641726702451706, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.033064987510442734, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8041722774505615, 'reward_repeat_penalty_std': 0.09782867133617401, 'reward_total_composite_mean': 0.667221188545227, 'reward_total_composite_std': 0.10969909280538559, 'epoch': 0.11}

 83%|████████▎ | 2754/3300 [4:04:03<2:25:45, 16.02s/it]INFO 04-12 02:40:55 [block_pool.py:378] Successfully reset prefix cache

 83%|████████▎ | 2755/3300 [4:04:09<1:57:59, 12.99s/it]2026-04-12 02:41:00,857 | INFO | train_grpo_train | metrics_logged mode=train step=2755

                                                       
{'loss': -0.0026, 'grad_norm': 0.17323201894760132, 'learning_rate': 1.6545454545454548e-06, 'num_tokens': 6248505.0, 'completions/mean_length': 71.5, 'completions/min_length': 71.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9994492530822754, 'rewards/meter/std': 2.2336023903335445e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994492530822754, 'rewards/total_composite/std': 2.2336023903335445e-05, 'reward': 0.9994492530822754, 'reward_std': 2.232198130514007e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0045864214189350605, 'sampling/sampling_logp_difference/max': 0.5479953289031982, 'sampling/importance_sampling_ratio/min': 0.5781075954437256, 'sampling/importance_sampling_ratio/mean': 1.0008584260940552, 'sampling/importance_sampling_ratio/max': 1.0643366575241089, 'entropy': 0.031015932094305754, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0016666667070239782, 'clip_ratio/high_max': 0.0016666667070239782, 'clip_ratio/region_mean': 0.0016666667070239782, 'reward_total_mean': 0.9994492530822754, 'reward_meter_mean': 0.9994492530822754, 'reward_meter_std': 2.2336023903335445e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994492530822754, 'reward_total_composite_std': 2.2336023903335445e-05, 'epoch': 0.11}

 83%|████████▎ | 2755/3300 [4:04:09<1:57:59, 12.99s/it]INFO 04-12 02:41:01 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2756/3300 [4:04:16<1:41:18, 11.17s/it]2026-04-12 02:41:07,795 | INFO | train_grpo_train | metrics_logged mode=train step=2756

                                                       
{'loss': -0.0331, 'grad_norm': 2.2315144538879395, 'learning_rate': 1.6515151515151517e-06, 'num_tokens': 6251781.0, 'completions/mean_length': 204.5, 'completions/min_length': 181.0, 'completions/max_length': 213.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 204.5, 'completions/min_terminated_length': 181.0, 'completions/max_terminated_length': 213.0, 'rewards/meter/mean': 0.9686325192451477, 'rewards/meter/std': 0.029917554929852486, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8901515007019043, 'rewards/repeat_penalty/std': 0.10736672580242157, 'rewards/total_composite/mean': 0.8484304547309875, 'rewards/total_composite/std': 0.13986076414585114, 'reward': 0.8484304547309875, 'reward_std': 0.13986073434352875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04765074700117111, 'sampling/sampling_logp_difference/max': 2.9453227519989014, 'sampling/importance_sampling_ratio/min': 0.05258508399128914, 'sampling/importance_sampling_ratio/mean': 1.0120221376419067, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3736105412244797, 'clip_ratio/low_mean': 0.009910162014421076, 'clip_ratio/low_min': 0.009910162014421076, 'clip_ratio/high_mean': 0.02566080493852496, 'clip_ratio/high_max': 0.02566080493852496, 'clip_ratio/region_mean': 0.03557096695294604, 'reward_total_mean': 0.8484304547309875, 'reward_meter_mean': 0.9686325192451477, 'reward_meter_std': 0.029917554929852486, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8901515007019043, 'reward_repeat_penalty_std': 0.10736672580242157, 'reward_total_composite_mean': 0.8484304547309875, 'reward_total_composite_std': 0.13986076414585114, 'epoch': 0.11}

 84%|████████▎ | 2756/3300 [4:04:16<1:41:18, 11.17s/it]INFO 04-12 02:41:08 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2757/3300 [4:04:24<1:33:11, 10.30s/it]2026-04-12 02:41:16,030 | INFO | train_grpo_train | metrics_logged mode=train step=2757

                                                       
{'loss': 0.0142, 'grad_norm': 1.6318613290786743, 'learning_rate': 1.6484848484848487e-06, 'num_tokens': 6256090.0, 'completions/mean_length': 332.625, 'completions/min_length': 322.0, 'completions/max_length': 336.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 332.625, 'completions/min_terminated_length': 322.0, 'completions/max_terminated_length': 336.0, 'rewards/meter/mean': 0.9958105087280273, 'rewards/meter/std': 0.006650847382843494, 'rewards/count_adherence/mean': 0.9090909361839294, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9144736528396606, 'rewards/repeat_penalty/std': 0.027239451184868813, 'rewards/total_composite/mean': 0.8277831077575684, 'rewards/total_composite/std': 0.02233772911131382, 'reward': 0.8277831077575684, 'reward_std': 0.022337697446346283, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02632460929453373, 'sampling/sampling_logp_difference/max': 2.554610252380371, 'sampling/importance_sampling_ratio/min': 0.07772251218557358, 'sampling/importance_sampling_ratio/mean': 1.004244089126587, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19466014206409454, 'clip_ratio/low_mean': 0.006716518080793321, 'clip_ratio/low_min': 0.006716518080793321, 'clip_ratio/high_mean': 0.010339219472371042, 'clip_ratio/high_max': 0.010339219472371042, 'clip_ratio/region_mean': 0.017055737553164363, 'reward_total_mean': 0.8277831077575684, 'reward_meter_mean': 0.9958105087280273, 'reward_meter_std': 0.006650847382843494, 'reward_count_adherence_mean': 0.9090909361839294, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9144736528396606, 'reward_repeat_penalty_std': 0.027239451184868813, 'reward_total_composite_mean': 0.8277831077575684, 'reward_total_composite_std': 0.02233772911131382, 'epoch': 0.11}

 84%|████████▎ | 2757/3300 [4:04:25<1:33:11, 10.30s/it]INFO 04-12 02:41:16 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2758/3300 [4:04:29<1:19:02,  8.75s/it]2026-04-12 02:41:21,101 | INFO | train_grpo_train | metrics_logged mode=train step=2758

                                                       
{'loss': 0.0014, 'grad_norm': 2.0312271118164062, 'learning_rate': 1.6454545454545455e-06, 'num_tokens': 6257914.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973218441009521, 'rewards/meter/std': 5.1264036301290616e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973218441009521, 'rewards/total_composite/std': 5.1264036301290616e-05, 'reward': 0.9973218441009521, 'reward_std': 5.1269918913021684e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008693864569067955, 'sampling/sampling_logp_difference/max': 0.5072799921035767, 'sampling/importance_sampling_ratio/min': 0.6021311283111572, 'sampling/importance_sampling_ratio/mean': 0.9973583221435547, 'sampling/importance_sampling_ratio/max': 1.3268663883209229, 'entropy': 0.04664692562073469, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.010245901066809893, 'clip_ratio/high_max': 0.010245901066809893, 'clip_ratio/region_mean': 0.010245901066809893, 'reward_total_mean': 0.9973218441009521, 'reward_meter_mean': 0.9973218441009521, 'reward_meter_std': 5.1264036301290616e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973218441009521, 'reward_total_composite_std': 5.1264036301290616e-05, 'epoch': 0.11}

 84%|████████▎ | 2758/3300 [4:04:30<1:19:02,  8.75s/it]INFO 04-12 02:41:21 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2759/3300 [4:04:38<1:18:30,  8.71s/it]2026-04-12 02:41:29,711 | INFO | train_grpo_train | metrics_logged mode=train step=2759

                                                       
{'loss': -0.0188, 'grad_norm': 1.848800539970398, 'learning_rate': 1.6424242424242426e-06, 'num_tokens': 6262570.0, 'completions/mean_length': 345.0, 'completions/min_length': 321.0, 'completions/max_length': 353.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 345.0, 'completions/min_terminated_length': 321.0, 'completions/max_terminated_length': 353.0, 'rewards/meter/mean': 0.9986103773117065, 'rewards/meter/std': 0.000737982802093029, 'rewards/count_adherence/mean': 0.9861111044883728, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9926470518112183, 'rewards/repeat_penalty/std': 0.020797256380319595, 'rewards/total_composite/mean': 0.9773909449577332, 'rewards/total_composite/std': 0.04150521755218506, 'reward': 0.9773909449577332, 'reward_std': 0.04150520637631416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04865448549389839, 'sampling/sampling_logp_difference/max': 1.8600101470947266, 'sampling/importance_sampling_ratio/min': 0.15567104518413544, 'sampling/importance_sampling_ratio/mean': 1.0085657835006714, 'sampling/importance_sampling_ratio/max': 1.8393433094024658, 'entropy': 0.4418103024363518, 'clip_ratio/low_mean': 0.006415694952011108, 'clip_ratio/low_min': 0.006415694952011108, 'clip_ratio/high_mean': 0.025451545137912035, 'clip_ratio/high_max': 0.025451545137912035, 'clip_ratio/region_mean': 0.03186724008992314, 'reward_total_mean': 0.9773909449577332, 'reward_meter_mean': 0.9986103773117065, 'reward_meter_std': 0.000737982802093029, 'reward_count_adherence_mean': 0.9861111044883728, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9926470518112183, 'reward_repeat_penalty_std': 0.020797256380319595, 'reward_total_composite_mean': 0.9773909449577332, 'reward_total_composite_std': 0.04150521755218506, 'epoch': 0.11}

 84%|████████▎ | 2759/3300 [4:04:38<1:18:30,  8.71s/it]INFO 04-12 02:41:29 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2760/3300 [4:04:43<1:06:56,  7.44s/it]2026-04-12 02:41:34,190 | INFO | train_grpo_train | metrics_logged mode=train step=2760

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.6393939393939396e-06, 'num_tokens': 6263986.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00032912864116951823, 'sampling/sampling_logp_difference/max': 0.0035952008329331875, 'sampling/importance_sampling_ratio/min': 0.9977269768714905, 'sampling/importance_sampling_ratio/mean': 1.000309944152832, 'sampling/importance_sampling_ratio/max': 1.0036016702651978, 'entropy': 0.0025047636299859732, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 84%|████████▎ | 2760/3300 [4:04:43<1:06:56,  7.44s/it]INFO 04-12 02:41:34 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2761/3300 [4:04:51<1:09:07,  7.70s/it]2026-04-12 02:41:42,480 | INFO | train_grpo_train | metrics_logged mode=train step=2761

                                                       
{'loss': 0.0416, 'grad_norm': 2.0906882286071777, 'learning_rate': 1.6363636363636365e-06, 'num_tokens': 6268093.0, 'completions/mean_length': 319.375, 'completions/min_length': 306.0, 'completions/max_length': 342.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 319.375, 'completions/min_terminated_length': 306.0, 'completions/max_terminated_length': 342.0, 'rewards/meter/mean': 0.9989085793495178, 'rewards/meter/std': 0.00021208541875239462, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9769607782363892, 'rewards/repeat_penalty/std': 0.03188912943005562, 'rewards/total_composite/mean': 0.946518063545227, 'rewards/total_composite/std': 0.07988038659095764, 'reward': 0.946518063545227, 'reward_std': 0.07988038659095764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.050234466791152954, 'sampling/sampling_logp_difference/max': 1.6290464401245117, 'sampling/importance_sampling_ratio/min': 0.19611649215221405, 'sampling/importance_sampling_ratio/mean': 1.0140368938446045, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47739846259355545, 'clip_ratio/low_mean': 0.007310603512451053, 'clip_ratio/low_min': 0.007310603512451053, 'clip_ratio/high_mean': 0.026291712652891874, 'clip_ratio/high_max': 0.026291712652891874, 'clip_ratio/region_mean': 0.03360231616534293, 'reward_total_mean': 0.946518063545227, 'reward_meter_mean': 0.9989085793495178, 'reward_meter_std': 0.00021208541875239462, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9769607782363892, 'reward_repeat_penalty_std': 0.03188912943005562, 'reward_total_composite_mean': 0.946518063545227, 'reward_total_composite_std': 0.07988038659095764, 'epoch': 0.11}

 84%|████████▎ | 2761/3300 [4:04:51<1:09:07,  7.70s/it]INFO 04-12 02:41:42 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2762/3300 [4:04:56<1:01:21,  6.84s/it]2026-04-12 02:41:47,345 | INFO | train_grpo_train | metrics_logged mode=train step=2762

                                                       
{'loss': 0.0269, 'grad_norm': 7.329155921936035, 'learning_rate': 1.6333333333333335e-06, 'num_tokens': 6269931.0, 'completions/mean_length': 67.75, 'completions/min_length': 66.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.977131724357605, 'rewards/meter/std': 0.06190915405750275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.977131724357605, 'rewards/total_composite/std': 0.06190915405750275, 'reward': 0.977131724357605, 'reward_std': 0.06190915033221245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04112435504794121, 'sampling/sampling_logp_difference/max': 1.3632383346557617, 'sampling/importance_sampling_ratio/min': 0.2558309733867645, 'sampling/importance_sampling_ratio/mean': 0.9983816742897034, 'sampling/importance_sampling_ratio/max': 1.5230276584625244, 'entropy': 0.2622176371514797, 'clip_ratio/low_mean': 0.0017857142956927419, 'clip_ratio/low_min': 0.0017857142956927419, 'clip_ratio/high_mean': 0.02216856903396547, 'clip_ratio/high_max': 0.02216856903396547, 'clip_ratio/region_mean': 0.02395428332965821, 'reward_total_mean': 0.977131724357605, 'reward_meter_mean': 0.977131724357605, 'reward_meter_std': 0.06190915405750275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.977131724357605, 'reward_total_composite_std': 0.06190915405750275, 'epoch': 0.11}

 84%|████████▎ | 2762/3300 [4:04:56<1:01:21,  6.84s/it]INFO 04-12 02:41:47 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▎ | 2763/3300 [4:05:01<56:45,  6.34s/it]  2026-04-12 02:41:52,507 | INFO | train_grpo_train | metrics_logged mode=train step=2763

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.6303030303030303e-06, 'num_tokens': 6271724.0, 'completions/mean_length': 66.125, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.002258425345644355, 'sampling/sampling_logp_difference/max': 0.5159242153167725, 'sampling/importance_sampling_ratio/min': 0.9706259965896606, 'sampling/importance_sampling_ratio/mean': 1.0024110078811646, 'sampling/importance_sampling_ratio/max': 1.6751859188079834, 'entropy': 0.014110149233601987, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 84%|████████▎ | 2763/3300 [4:05:01<56:45,  6.34s/it]INFO 04-12 02:41:52 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2764/3300 [4:05:06<52:58,  5.93s/it]2026-04-12 02:41:57,481 | INFO | train_grpo_train | metrics_logged mode=train step=2764

                                                     
{'loss': -0.0057, 'grad_norm': 3.616687774658203, 'learning_rate': 1.6272727272727274e-06, 'num_tokens': 6273656.0, 'completions/mean_length': 79.5, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.5, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9986147880554199, 'rewards/meter/std': 0.0008511117775924504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9986147880554199, 'rewards/total_composite/std': 0.0008511117775924504, 'reward': 0.9986147880554199, 'reward_std': 0.0008511117775924504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029622886329889297, 'sampling/sampling_logp_difference/max': 1.190237045288086, 'sampling/importance_sampling_ratio/min': 0.30414915084838867, 'sampling/importance_sampling_ratio/mean': 1.0039986371994019, 'sampling/importance_sampling_ratio/max': 1.7587852478027344, 'entropy': 0.20842239819467068, 'clip_ratio/low_mean': 0.00474833813495934, 'clip_ratio/low_min': 0.00474833813495934, 'clip_ratio/high_mean': 0.00474683556240052, 'clip_ratio/high_max': 0.00474683556240052, 'clip_ratio/region_mean': 0.00949517369735986, 'reward_total_mean': 0.9986147880554199, 'reward_meter_mean': 0.9986147880554199, 'reward_meter_std': 0.0008511117775924504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9986147880554199, 'reward_total_composite_std': 0.0008511117775924504, 'epoch': 0.11}

 84%|████████▍ | 2764/3300 [4:05:06<52:58,  5.93s/it]INFO 04-12 02:41:57 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2765/3300 [4:05:11<51:57,  5.83s/it]2026-04-12 02:42:03,065 | INFO | train_grpo_train | metrics_logged mode=train step=2765

                                                     
{'loss': -0.0034, 'grad_norm': 6.583384037017822, 'learning_rate': 1.6242424242424242e-06, 'num_tokens': 6275778.0, 'completions/mean_length': 102.25, 'completions/min_length': 99.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.25, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9677326679229736, 'rewards/meter/std': 0.08840977400541306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9677326679229736, 'rewards/total_composite/std': 0.08840977400541306, 'reward': 0.9677326679229736, 'reward_std': 0.08840975910425186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05027589201927185, 'sampling/sampling_logp_difference/max': 1.538498878479004, 'sampling/importance_sampling_ratio/min': 0.21470315754413605, 'sampling/importance_sampling_ratio/mean': 1.0036083459854126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3880934603512287, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.03991973074153066, 'clip_ratio/high_max': 0.03991973074153066, 'clip_ratio/region_mean': 0.04370760964229703, 'reward_total_mean': 0.9677326679229736, 'reward_meter_mean': 0.9677326679229736, 'reward_meter_std': 0.08840977400541306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9677326679229736, 'reward_total_composite_std': 0.08840977400541306, 'epoch': 0.11}

 84%|████████▍ | 2765/3300 [4:05:12<51:57,  5.83s/it]INFO 04-12 02:42:03 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2766/3300 [4:05:17<50:41,  5.70s/it]2026-04-12 02:42:08,455 | INFO | train_grpo_train | metrics_logged mode=train step=2766

                                                     
{'loss': -0.0051, 'grad_norm': 2.6680572032928467, 'learning_rate': 1.6212121212121213e-06, 'num_tokens': 6278065.0, 'completions/mean_length': 116.875, 'completions/min_length': 115.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.875, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9987287521362305, 'rewards/meter/std': 0.0006380022969096899, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987287521362305, 'rewards/total_composite/std': 0.0006380022969096899, 'reward': 0.9987287521362305, 'reward_std': 0.0006380106206052005, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02768939547240734, 'sampling/sampling_logp_difference/max': 0.7849941253662109, 'sampling/importance_sampling_ratio/min': 0.45612236857414246, 'sampling/importance_sampling_ratio/mean': 1.0104761123657227, 'sampling/importance_sampling_ratio/max': 1.6575981378555298, 'entropy': 0.2904528006911278, 'clip_ratio/low_mean': 0.008585114032030106, 'clip_ratio/low_min': 0.008585114032030106, 'clip_ratio/high_mean': 0.013889989699237049, 'clip_ratio/high_max': 0.013889989699237049, 'clip_ratio/region_mean': 0.022475103731267154, 'reward_total_mean': 0.9987287521362305, 'reward_meter_mean': 0.9987287521362305, 'reward_meter_std': 0.0006380022969096899, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987287521362305, 'reward_total_composite_std': 0.0006380022969096899, 'epoch': 0.11}

 84%|████████▍ | 2766/3300 [4:05:17<50:41,  5.70s/it]INFO 04-12 02:42:08 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2767/3300 [4:05:24<53:30,  6.02s/it]2026-04-12 02:42:15,249 | INFO | train_grpo_train | metrics_logged mode=train step=2767

                                                     
{'loss': 0.004, 'grad_norm': 1.0399538278579712, 'learning_rate': 1.618181818181818e-06, 'num_tokens': 6281566.0, 'completions/mean_length': 246.625, 'completions/min_length': 244.0, 'completions/max_length': 248.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 246.625, 'completions/min_terminated_length': 244.0, 'completions/max_terminated_length': 248.0, 'rewards/meter/mean': 0.9990943074226379, 'rewards/meter/std': 0.0001148659794125706, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7403846383094788, 'rewards/repeat_penalty/std': 0.05723259598016739, 'rewards/total_composite/mean': 0.7397112250328064, 'rewards/total_composite/std': 0.05713575705885887, 'reward': 0.7397112250328064, 'reward_std': 0.05713575705885887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014130277559161186, 'sampling/sampling_logp_difference/max': 1.4010510444641113, 'sampling/importance_sampling_ratio/min': 0.24633793532848358, 'sampling/importance_sampling_ratio/mean': 1.0045311450958252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09486087318509817, 'clip_ratio/low_mean': 0.0040465755155310035, 'clip_ratio/low_min': 0.0040465755155310035, 'clip_ratio/high_mean': 0.005081536248326302, 'clip_ratio/high_max': 0.005081536248326302, 'clip_ratio/region_mean': 0.009128111763857305, 'reward_total_mean': 0.7397112250328064, 'reward_meter_mean': 0.9990943074226379, 'reward_meter_std': 0.0001148659794125706, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7403846383094788, 'reward_repeat_penalty_std': 0.05723259598016739, 'reward_total_composite_mean': 0.7397112250328064, 'reward_total_composite_std': 0.05713575705885887, 'epoch': 0.11}

 84%|████████▍ | 2767/3300 [4:05:24<53:30,  6.02s/it]INFO 04-12 02:42:15 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2768/3300 [4:05:29<51:52,  5.85s/it]2026-04-12 02:42:20,699 | INFO | train_grpo_train | metrics_logged mode=train step=2768

                                                     
{'loss': 0.0045, 'grad_norm': 2.360358476638794, 'learning_rate': 1.6151515151515153e-06, 'num_tokens': 6283917.0, 'completions/mean_length': 141.875, 'completions/min_length': 141.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.875, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9991145133972168, 'rewards/meter/std': 0.00046132790157571435, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8920483589172363, 'rewards/total_composite/std': 0.06576977670192719, 'reward': 0.8920483589172363, 'reward_std': 0.06576978415250778, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012811451219022274, 'sampling/sampling_logp_difference/max': 0.977135419845581, 'sampling/importance_sampling_ratio/min': 0.3763877749443054, 'sampling/importance_sampling_ratio/mean': 1.0041320323944092, 'sampling/importance_sampling_ratio/max': 1.6703215837478638, 'entropy': 0.08987067546695471, 'clip_ratio/low_mean': 0.002622719621285796, 'clip_ratio/low_min': 0.002622719621285796, 'clip_ratio/high_mean': 0.004426380852237344, 'clip_ratio/high_max': 0.004426380852237344, 'clip_ratio/region_mean': 0.00704910047352314, 'reward_total_mean': 0.8920483589172363, 'reward_meter_mean': 0.9991145133972168, 'reward_meter_std': 0.00046132790157571435, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8920483589172363, 'reward_total_composite_std': 0.06576977670192719, 'epoch': 0.11}

 84%|████████▍ | 2768/3300 [4:05:29<51:52,  5.85s/it]INFO 04-12 02:42:20 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2769/3300 [4:05:39<1:03:27,  7.17s/it]2026-04-12 02:42:30,943 | INFO | train_grpo_train | metrics_logged mode=train step=2769

                                                       
{'loss': 0.0647, 'grad_norm': 1.8239376544952393, 'learning_rate': 1.6121212121212124e-06, 'num_tokens': 6288893.0, 'completions/mean_length': 501.0, 'completions/min_length': 475.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 497.3333435058594, 'completions/min_terminated_length': 475.0, 'completions/max_terminated_length': 506.0, 'rewards/meter/mean': 0.9878989458084106, 'rewards/meter/std': 0.03154898062348366, 'rewards/count_adherence/mean': 0.7573529481887817, 'rewards/count_adherence/std': 0.020797256380319595, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.97975754737854, 'rewards/repeat_penalty/std': 0.021684719249606133, 'rewards/total_composite/mean': 0.7339591979980469, 'rewards/total_composite/std': 0.05060318857431412, 'reward': 0.7339591979980469, 'reward_std': 0.050603192299604416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05508255213499069, 'sampling/sampling_logp_difference/max': 2.9603583812713623, 'sampling/importance_sampling_ratio/min': 0.05180035158991814, 'sampling/importance_sampling_ratio/mean': 1.0112419128417969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37871410325169563, 'clip_ratio/low_mean': 0.010852963663637638, 'clip_ratio/low_min': 0.010852963663637638, 'clip_ratio/high_mean': 0.014446925837546587, 'clip_ratio/high_max': 0.014446925837546587, 'clip_ratio/region_mean': 0.025299889501184225, 'reward_total_mean': 0.7339591979980469, 'reward_meter_mean': 0.9878989458084106, 'reward_meter_std': 0.03154898062348366, 'reward_count_adherence_mean': 0.7573529481887817, 'reward_count_adherence_std': 0.020797256380319595, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.97975754737854, 'reward_repeat_penalty_std': 0.021684719249606133, 'reward_total_composite_mean': 0.7339591979980469, 'reward_total_composite_std': 0.05060318857431412, 'epoch': 0.11}

 84%|████████▍ | 2769/3300 [4:05:39<1:03:27,  7.17s/it]INFO 04-12 02:42:31 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2770/3300 [4:05:44<56:44,  6.42s/it]  2026-04-12 02:42:35,628 | INFO | train_grpo_train | metrics_logged mode=train step=2770

                                                     
{'loss': -0.0003, 'grad_norm': 0.033189766108989716, 'learning_rate': 1.6090909090909092e-06, 'num_tokens': 6290734.0, 'completions/mean_length': 71.125, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994409084320068, 'rewards/meter/std': 2.465912302795914e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994409084320068, 'rewards/total_composite/std': 2.465912302795914e-06, 'reward': 0.9994409084320068, 'reward_std': 2.4565813419030746e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0038025446701794863, 'sampling/sampling_logp_difference/max': 0.2554647922515869, 'sampling/importance_sampling_ratio/min': 0.7745563983917236, 'sampling/importance_sampling_ratio/mean': 1.0016711950302124, 'sampling/importance_sampling_ratio/max': 1.0757405757904053, 'entropy': 0.03055504336953163, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.0017605633474886417, 'clip_ratio/high_max': 0.0017605633474886417, 'clip_ratio/region_mean': 0.0034966744715347886, 'reward_total_mean': 0.9994409084320068, 'reward_meter_mean': 0.9994409084320068, 'reward_meter_std': 2.465912302795914e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994409084320068, 'reward_total_composite_std': 2.465912302795914e-06, 'epoch': 0.11}

 84%|████████▍ | 2770/3300 [4:05:44<56:44,  6.42s/it]INFO 04-12 02:42:35 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2771/3300 [4:05:49<53:58,  6.12s/it]2026-04-12 02:42:41,047 | INFO | train_grpo_train | metrics_logged mode=train step=2771

                                                     
{'loss': 0.0169, 'grad_norm': 4.650167465209961, 'learning_rate': 1.6060606060606063e-06, 'num_tokens': 6293263.0, 'completions/mean_length': 135.125, 'completions/min_length': 130.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.125, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.9964702725410461, 'rewards/meter/std': 0.003808342618867755, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9964702725410461, 'rewards/total_composite/std': 0.003808342618867755, 'reward': 0.9964702725410461, 'reward_std': 0.003808349836617708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04582271724939346, 'sampling/sampling_logp_difference/max': 1.5504403114318848, 'sampling/importance_sampling_ratio/min': 0.2121545374393463, 'sampling/importance_sampling_ratio/mean': 1.0039819478988647, 'sampling/importance_sampling_ratio/max': 1.650173544883728, 'entropy': 0.4557015933096409, 'clip_ratio/low_mean': 0.009322975529357791, 'clip_ratio/low_min': 0.009322975529357791, 'clip_ratio/high_mean': 0.023043142282404006, 'clip_ratio/high_max': 0.023043142282404006, 'clip_ratio/region_mean': 0.032366117811761796, 'reward_total_mean': 0.9964702725410461, 'reward_meter_mean': 0.9964702725410461, 'reward_meter_std': 0.003808342618867755, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9964702725410461, 'reward_total_composite_std': 0.003808342618867755, 'epoch': 0.11}

 84%|████████▍ | 2771/3300 [4:05:50<53:58,  6.12s/it]INFO 04-12 02:42:41 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2772/3300 [4:05:54<50:06,  5.69s/it]2026-04-12 02:42:45,737 | INFO | train_grpo_train | metrics_logged mode=train step=2772

                                                     
{'loss': 0.0239, 'grad_norm': 2.930152416229248, 'learning_rate': 1.6030303030303033e-06, 'num_tokens': 6295009.0, 'completions/mean_length': 67.25, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9888796210289001, 'rewards/meter/std': 0.006786399520933628, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9888796210289001, 'rewards/total_composite/std': 0.006786399520933628, 'reward': 0.9888796210289001, 'reward_std': 0.006786399520933628, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028971223160624504, 'sampling/sampling_logp_difference/max': 1.6678972244262695, 'sampling/importance_sampling_ratio/min': 0.18864330649375916, 'sampling/importance_sampling_ratio/mean': 1.0060391426086426, 'sampling/importance_sampling_ratio/max': 1.6033055782318115, 'entropy': 0.2543229255825281, 'clip_ratio/low_mean': 0.008928571594879031, 'clip_ratio/low_min': 0.008928571594879031, 'clip_ratio/high_mean': 0.00938489381223917, 'clip_ratio/high_max': 0.00938489381223917, 'clip_ratio/region_mean': 0.0183134654071182, 'reward_total_mean': 0.9888796210289001, 'reward_meter_mean': 0.9888796210289001, 'reward_meter_std': 0.006786399520933628, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9888796210289001, 'reward_total_composite_std': 0.006786399520933628, 'epoch': 0.11}

 84%|████████▍ | 2772/3300 [4:05:54<50:06,  5.69s/it]INFO 04-12 02:42:45 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2773/3300 [4:05:59<46:49,  5.33s/it]2026-04-12 02:42:50,223 | INFO | train_grpo_train | metrics_logged mode=train step=2773

                                                     
{'loss': 0.0032, 'grad_norm': 5.738377094268799, 'learning_rate': 1.6000000000000001e-06, 'num_tokens': 6296752.0, 'completions/mean_length': 60.875, 'completions/min_length': 60.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9969919919967651, 'rewards/meter/std': 0.0007282263832166791, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9969919919967651, 'rewards/total_composite/std': 0.0007282263832166791, 'reward': 0.9969919919967651, 'reward_std': 0.000728218408767134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01134834811091423, 'sampling/sampling_logp_difference/max': 0.8574669361114502, 'sampling/importance_sampling_ratio/min': 0.6171859502792358, 'sampling/importance_sampling_ratio/mean': 1.0058006048202515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.060230674454942346, 'clip_ratio/low_mean': 0.012500000186264515, 'clip_ratio/low_min': 0.012500000186264515, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.014549180399626493, 'reward_total_mean': 0.9969919919967651, 'reward_meter_mean': 0.9969919919967651, 'reward_meter_std': 0.0007282263832166791, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9969919919967651, 'reward_total_composite_std': 0.0007282263832166791, 'epoch': 0.11}

 84%|████████▍ | 2773/3300 [4:05:59<46:49,  5.33s/it]INFO 04-12 02:42:50 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2774/3300 [4:06:05<48:19,  5.51s/it]2026-04-12 02:42:56,160 | INFO | train_grpo_train | metrics_logged mode=train step=2774

                                                     
{'loss': 0.0355, 'grad_norm': 2.424381971359253, 'learning_rate': 1.5969696969696972e-06, 'num_tokens': 6299532.0, 'completions/mean_length': 173.5, 'completions/min_length': 166.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.5, 'completions/min_terminated_length': 166.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.9787227511405945, 'rewards/meter/std': 0.03248432278633118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.08399210125207901, 'rewards/total_composite/mean': 0.9236654043197632, 'rewards/total_composite/std': 0.08021914213895798, 'reward': 0.9236654043197632, 'reward_std': 0.08021914213895798, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04096399247646332, 'sampling/sampling_logp_difference/max': 1.5549349784851074, 'sampling/importance_sampling_ratio/min': 0.21120309829711914, 'sampling/importance_sampling_ratio/mean': 1.0065571069717407, 'sampling/importance_sampling_ratio/max': 1.7961452007293701, 'entropy': 0.36277873255312443, 'clip_ratio/low_mean': 0.013373102992773056, 'clip_ratio/low_min': 0.013373102992773056, 'clip_ratio/high_mean': 0.01708856108598411, 'clip_ratio/high_max': 0.01708856108598411, 'clip_ratio/region_mean': 0.030461664078757167, 'reward_total_mean': 0.9236654043197632, 'reward_meter_mean': 0.9787227511405945, 'reward_meter_std': 0.03248432278633118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.08399210125207901, 'reward_total_composite_mean': 0.9236654043197632, 'reward_total_composite_std': 0.08021914213895798, 'epoch': 0.11}

 84%|████████▍ | 2774/3300 [4:06:05<48:19,  5.51s/it]INFO 04-12 02:42:56 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2775/3300 [4:06:09<46:27,  5.31s/it]2026-04-12 02:43:00,993 | INFO | train_grpo_train | metrics_logged mode=train step=2775

                                                     
{'loss': 0.0058, 'grad_norm': 9.112265586853027, 'learning_rate': 1.593939393939394e-06, 'num_tokens': 6301620.0, 'completions/mean_length': 86.0, 'completions/min_length': 84.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9534580707550049, 'rewards/meter/std': 0.004275289364159107, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.10628911107778549, 'rewards/total_composite/mean': 0.9026405811309814, 'rewards/total_composite/std': 0.1037164255976677, 'reward': 0.9026405811309814, 'reward_std': 0.1037164255976677, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06749268621206284, 'sampling/sampling_logp_difference/max': 2.1043763160705566, 'sampling/importance_sampling_ratio/min': 0.12192168831825256, 'sampling/importance_sampling_ratio/mean': 1.0034648180007935, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2445132229477167, 'clip_ratio/low_mean': 0.013254429679363966, 'clip_ratio/low_min': 0.013254429679363966, 'clip_ratio/high_mean': 0.03781960904598236, 'clip_ratio/high_max': 0.03781960904598236, 'clip_ratio/region_mean': 0.05107403872534633, 'reward_total_mean': 0.9026405811309814, 'reward_meter_mean': 0.9534580707550049, 'reward_meter_std': 0.004275289364159107, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.10628911107778549, 'reward_total_composite_mean': 0.9026405811309814, 'reward_total_composite_std': 0.1037164255976677, 'epoch': 0.11}

 84%|████████▍ | 2775/3300 [4:06:09<46:27,  5.31s/it]INFO 04-12 02:43:01 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2776/3300 [4:06:14<44:30,  5.10s/it]2026-04-12 02:43:05,597 | INFO | train_grpo_train | metrics_logged mode=train step=2776

                                                     
{'loss': 0.0002, 'grad_norm': 2.9178199768066406, 'learning_rate': 1.590909090909091e-06, 'num_tokens': 6303505.0, 'completions/mean_length': 65.625, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9525059461593628, 'rewards/meter/std': 0.0829981118440628, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9525059461593628, 'rewards/total_composite/std': 0.0829981118440628, 'reward': 0.9525059461593628, 'reward_std': 0.082998126745224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028366295620799065, 'sampling/sampling_logp_difference/max': 1.2556953430175781, 'sampling/importance_sampling_ratio/min': 0.2848776876926422, 'sampling/importance_sampling_ratio/mean': 1.003873586654663, 'sampling/importance_sampling_ratio/max': 1.5501948595046997, 'entropy': 0.22676505334675312, 'clip_ratio/low_mean': 0.001923076924867928, 'clip_ratio/low_min': 0.001923076924867928, 'clip_ratio/high_mean': 0.02643295784946531, 'clip_ratio/high_max': 0.02643295784946531, 'clip_ratio/region_mean': 0.02835603477433324, 'reward_total_mean': 0.9525059461593628, 'reward_meter_mean': 0.9525059461593628, 'reward_meter_std': 0.0829981118440628, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9525059461593628, 'reward_total_composite_std': 0.0829981118440628, 'epoch': 0.11}

 84%|████████▍ | 2776/3300 [4:06:14<44:30,  5.10s/it]INFO 04-12 02:43:05 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2777/3300 [4:06:19<43:41,  5.01s/it]2026-04-12 02:43:10,413 | INFO | train_grpo_train | metrics_logged mode=train step=2777

                                                     
{'loss': -0.004, 'grad_norm': 2.4066951274871826, 'learning_rate': 1.5878787878787879e-06, 'num_tokens': 6305374.0, 'completions/mean_length': 79.625, 'completions/min_length': 79.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.625, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9987850189208984, 'rewards/meter/std': 0.0002843133988790214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987850189208984, 'rewards/total_composite/std': 0.0002843133988790214, 'reward': 0.9987850189208984, 'reward_std': 0.00028430239763110876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02783510647714138, 'sampling/sampling_logp_difference/max': 1.6474170684814453, 'sampling/importance_sampling_ratio/min': 0.19254659116268158, 'sampling/importance_sampling_ratio/mean': 1.0044121742248535, 'sampling/importance_sampling_ratio/max': 1.57016921043396, 'entropy': 0.22624119743704796, 'clip_ratio/low_mean': 0.0031250000465661287, 'clip_ratio/low_min': 0.0031250000465661287, 'clip_ratio/high_mean': 0.010977057158015668, 'clip_ratio/high_max': 0.010977057158015668, 'clip_ratio/region_mean': 0.014102057204581797, 'reward_total_mean': 0.9987850189208984, 'reward_meter_mean': 0.9987850189208984, 'reward_meter_std': 0.0002843133988790214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987850189208984, 'reward_total_composite_std': 0.0002843133988790214, 'epoch': 0.11}

 84%|████████▍ | 2777/3300 [4:06:19<43:41,  5.01s/it]INFO 04-12 02:43:10 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2778/3300 [4:06:24<44:26,  5.11s/it]2026-04-12 02:43:15,747 | INFO | train_grpo_train | metrics_logged mode=train step=2778

                                                     
{'loss': 0.0001, 'grad_norm': 2.4368531703948975, 'learning_rate': 1.584848484848485e-06, 'num_tokens': 6307953.0, 'completions/mean_length': 132.375, 'completions/min_length': 132.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.375, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9603220820426941, 'rewards/meter/std': 0.11049286276102066, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9603220820426941, 'rewards/total_composite/std': 0.11049286276102066, 'reward': 0.9603220820426941, 'reward_std': 0.11049285531044006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007737660314887762, 'sampling/sampling_logp_difference/max': 1.6687531471252441, 'sampling/importance_sampling_ratio/min': 0.18848192691802979, 'sampling/importance_sampling_ratio/mean': 1.0007283687591553, 'sampling/importance_sampling_ratio/max': 1.631596565246582, 'entropy': 0.051168760284781456, 'clip_ratio/low_mean': 0.0009469697251915932, 'clip_ratio/low_min': 0.0009469697251915932, 'clip_ratio/high_mean': 0.0084732057293877, 'clip_ratio/high_max': 0.0084732057293877, 'clip_ratio/region_mean': 0.009420175454579294, 'reward_total_mean': 0.9603220820426941, 'reward_meter_mean': 0.9603220820426941, 'reward_meter_std': 0.11049286276102066, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9603220820426941, 'reward_total_composite_std': 0.11049286276102066, 'epoch': 0.11}

 84%|████████▍ | 2778/3300 [4:06:24<44:26,  5.11s/it]INFO 04-12 02:43:15 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2779/3300 [4:06:29<44:38,  5.14s/it]2026-04-12 02:43:20,964 | INFO | train_grpo_train | metrics_logged mode=train step=2779

                                                     
{'loss': 0.0001, 'grad_norm': 2.0250144004821777, 'learning_rate': 1.5818181818181818e-06, 'num_tokens': 6310483.0, 'completions/mean_length': 123.25, 'completions/min_length': 122.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.25, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9977097511291504, 'rewards/meter/std': 0.0001580109674250707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977097511291504, 'rewards/total_composite/std': 0.0001580109674250707, 'reward': 0.9977097511291504, 'reward_std': 0.00015800821711309254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030101941898465157, 'sampling/sampling_logp_difference/max': 1.549009919166565, 'sampling/importance_sampling_ratio/min': 0.21245822310447693, 'sampling/importance_sampling_ratio/mean': 1.0012949705123901, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20793312788009644, 'clip_ratio/low_mean': 0.013203723356127739, 'clip_ratio/low_min': 0.013203723356127739, 'clip_ratio/high_mean': 0.011195790022611618, 'clip_ratio/high_max': 0.011195790022611618, 'clip_ratio/region_mean': 0.024399513378739357, 'reward_total_mean': 0.9977097511291504, 'reward_meter_mean': 0.9977097511291504, 'reward_meter_std': 0.0001580109674250707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977097511291504, 'reward_total_composite_std': 0.0001580109674250707, 'epoch': 0.11}

 84%|████████▍ | 2779/3300 [4:06:29<44:38,  5.14s/it]INFO 04-12 02:43:21 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2780/3300 [4:06:37<51:53,  5.99s/it]2026-04-12 02:43:28,925 | INFO | train_grpo_train | metrics_logged mode=train step=2780

                                                     
{'loss': 0.0047, 'grad_norm': 1.3197673559188843, 'learning_rate': 1.5787878787878788e-06, 'num_tokens': 6314718.0, 'completions/mean_length': 323.375, 'completions/min_length': 317.0, 'completions/max_length': 355.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 323.375, 'completions/min_terminated_length': 317.0, 'completions/max_terminated_length': 355.0, 'rewards/meter/mean': 0.9988690614700317, 'rewards/meter/std': 0.00020919565577059984, 'rewards/count_adherence/mean': 0.9124999642372131, 'rewards/count_adherence/std': 0.0353553481400013, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8053405284881592, 'rewards/repeat_penalty/std': 0.046673085540533066, 'rewards/total_composite/mean': 0.7331855297088623, 'rewards/total_composite/std': 0.03377552330493927, 'reward': 0.7331855297088623, 'reward_std': 0.033775512129068375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0219265203922987, 'sampling/sampling_logp_difference/max': 4.522955417633057, 'sampling/importance_sampling_ratio/min': 0.010856889188289642, 'sampling/importance_sampling_ratio/mean': 1.0070363283157349, 'sampling/importance_sampling_ratio/max': 1.8094199895858765, 'entropy': 0.19129730947315693, 'clip_ratio/low_mean': 0.0015637245669495314, 'clip_ratio/low_min': 0.0015637245669495314, 'clip_ratio/high_mean': 0.010822767158970237, 'clip_ratio/high_max': 0.010822767158970237, 'clip_ratio/region_mean': 0.012386491725919768, 'reward_total_mean': 0.7331855297088623, 'reward_meter_mean': 0.9988690614700317, 'reward_meter_std': 0.00020919565577059984, 'reward_count_adherence_mean': 0.9124999642372131, 'reward_count_adherence_std': 0.0353553481400013, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8053405284881592, 'reward_repeat_penalty_std': 0.046673085540533066, 'reward_total_composite_mean': 0.7331855297088623, 'reward_total_composite_std': 0.03377552330493927, 'epoch': 0.11}

 84%|████████▍ | 2780/3300 [4:06:37<51:53,  5.99s/it]INFO 04-12 02:43:29 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2781/3300 [4:06:42<47:37,  5.51s/it]2026-04-12 02:43:33,307 | INFO | train_grpo_train | metrics_logged mode=train step=2781

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.5757575757575759e-06, 'num_tokens': 6316374.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00022916619491297752, 'sampling/sampling_logp_difference/max': 0.002041937317699194, 'sampling/importance_sampling_ratio/min': 0.9993596076965332, 'sampling/importance_sampling_ratio/mean': 1.0002243518829346, 'sampling/importance_sampling_ratio/max': 1.0020439624786377, 'entropy': 0.0019741212599910796, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 84%|████████▍ | 2781/3300 [4:06:42<47:37,  5.51s/it]INFO 04-12 02:43:33 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2782/3300 [4:06:47<46:22,  5.37s/it]2026-04-12 02:43:38,363 | INFO | train_grpo_train | metrics_logged mode=train step=2782

                                                     
{'loss': -0.0038, 'grad_norm': 3.458383560180664, 'learning_rate': 1.572727272727273e-06, 'num_tokens': 6318445.0, 'completions/mean_length': 101.875, 'completions/min_length': 98.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.875, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.998963475227356, 'rewards/meter/std': 0.0003066870558541268, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998963475227356, 'rewards/total_composite/std': 0.0003066870558541268, 'reward': 0.998963475227356, 'reward_std': 0.00030668076942674816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059917621314525604, 'sampling/sampling_logp_difference/max': 3.0749526023864746, 'sampling/importance_sampling_ratio/min': 0.04619181901216507, 'sampling/importance_sampling_ratio/mean': 1.0039867162704468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3706792779266834, 'clip_ratio/low_mean': 0.01848576357588172, 'clip_ratio/low_min': 0.01848576357588172, 'clip_ratio/high_mean': 0.02071394305676222, 'clip_ratio/high_max': 0.02071394305676222, 'clip_ratio/region_mean': 0.03919970663264394, 'reward_total_mean': 0.998963475227356, 'reward_meter_mean': 0.998963475227356, 'reward_meter_std': 0.0003066870558541268, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998963475227356, 'reward_total_composite_std': 0.0003066870558541268, 'epoch': 0.11}

 84%|████████▍ | 2782/3300 [4:06:47<46:22,  5.37s/it]INFO 04-12 02:43:38 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2783/3300 [4:06:51<43:47,  5.08s/it]2026-04-12 02:43:42,770 | INFO | train_grpo_train | metrics_logged mode=train step=2783

                                                     
{'loss': 0.0798, 'grad_norm': 190.96791076660156, 'learning_rate': 1.56969696969697e-06, 'num_tokens': 6320008.0, 'completions/mean_length': 42.375, 'completions/min_length': 40.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9555168151855469, 'rewards/meter/std': 0.011990510858595371, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9555168151855469, 'rewards/total_composite/std': 0.011990510858595371, 'reward': 0.9555168151855469, 'reward_std': 0.011990513652563095, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08888134360313416, 'sampling/sampling_logp_difference/max': 1.91815185546875, 'sampling/importance_sampling_ratio/min': 0.14687815308570862, 'sampling/importance_sampling_ratio/mean': 0.9958891868591309, 'sampling/importance_sampling_ratio/max': 1.9535908699035645, 'entropy': 0.3651922196149826, 'clip_ratio/low_mean': 0.05653973203152418, 'clip_ratio/low_min': 0.05653973203152418, 'clip_ratio/high_mean': 0.014478484634310007, 'clip_ratio/high_max': 0.014478484634310007, 'clip_ratio/region_mean': 0.07101821666583419, 'reward_total_mean': 0.9555168151855469, 'reward_meter_mean': 0.9555168151855469, 'reward_meter_std': 0.011990510858595371, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9555168151855469, 'reward_total_composite_std': 0.011990510858595371, 'epoch': 0.11}

 84%|████████▍ | 2783/3300 [4:06:51<43:47,  5.08s/it]INFO 04-12 02:43:43 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2784/3300 [4:06:56<42:02,  4.89s/it]2026-04-12 02:43:47,212 | INFO | train_grpo_train | metrics_logged mode=train step=2784

                                                     
{'loss': 0.0038, 'grad_norm': 5.360074043273926, 'learning_rate': 1.566666666666667e-06, 'num_tokens': 6321854.0, 'completions/mean_length': 60.75, 'completions/min_length': 60.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.75, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9968235492706299, 'rewards/meter/std': 0.0013836961006745696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9968235492706299, 'rewards/total_composite/std': 0.0013836961006745696, 'reward': 0.9968235492706299, 'reward_std': 0.001383687136694789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014301088638603687, 'sampling/sampling_logp_difference/max': 1.6227011680603027, 'sampling/importance_sampling_ratio/min': 0.19736485183238983, 'sampling/importance_sampling_ratio/mean': 1.002347707748413, 'sampling/importance_sampling_ratio/max': 1.77481210231781, 'entropy': 0.07807908114045858, 'clip_ratio/low_mean': 0.008333333767950535, 'clip_ratio/low_min': 0.008333333767950535, 'clip_ratio/high_mean': 0.010314207524061203, 'clip_ratio/high_max': 0.010314207524061203, 'clip_ratio/region_mean': 0.018647541292011738, 'reward_total_mean': 0.9968235492706299, 'reward_meter_mean': 0.9968235492706299, 'reward_meter_std': 0.0013836961006745696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9968235492706299, 'reward_total_composite_std': 0.0013836961006745696, 'epoch': 0.11}

 84%|████████▍ | 2784/3300 [4:06:56<42:02,  4.89s/it]INFO 04-12 02:43:47 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2785/3300 [4:07:00<41:57,  4.89s/it]2026-04-12 02:43:52,101 | INFO | train_grpo_train | metrics_logged mode=train step=2785

                                                     
{'loss': 0.0043, 'grad_norm': 2.2449779510498047, 'learning_rate': 1.5636363636363638e-06, 'num_tokens': 6324095.0, 'completions/mean_length': 92.125, 'completions/min_length': 91.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.125, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9976515173912048, 'rewards/meter/std': 0.00014370458666235209, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976515173912048, 'rewards/total_composite/std': 0.00014370458666235209, 'reward': 0.9976515173912048, 'reward_std': 0.00014370400458574295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02354567125439644, 'sampling/sampling_logp_difference/max': 1.614687442779541, 'sampling/importance_sampling_ratio/min': 0.1989528387784958, 'sampling/importance_sampling_ratio/mean': 1.0068784952163696, 'sampling/importance_sampling_ratio/max': 1.9402447938919067, 'entropy': 0.14724664017558098, 'clip_ratio/low_mean': 0.019023986998945475, 'clip_ratio/low_min': 0.019023986998945475, 'clip_ratio/high_mean': 0.0054351037833839655, 'clip_ratio/high_max': 0.0054351037833839655, 'clip_ratio/region_mean': 0.02445909078232944, 'reward_total_mean': 0.9976515173912048, 'reward_meter_mean': 0.9976515173912048, 'reward_meter_std': 0.00014370458666235209, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976515173912048, 'reward_total_composite_std': 0.00014370458666235209, 'epoch': 0.11}

 84%|████████▍ | 2785/3300 [4:07:01<41:57,  4.89s/it]INFO 04-12 02:43:52 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2786/3300 [4:07:05<39:59,  4.67s/it]2026-04-12 02:43:56,252 | INFO | train_grpo_train | metrics_logged mode=train step=2786

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.5606060606060609e-06, 'num_tokens': 6325535.0, 'completions/mean_length': 36.0, 'completions/min_length': 36.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9996045231819153, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9996045231819153, 'rewards/total_composite/std': 0.0, 'reward': 0.9996045231819153, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0015392457135021687, 'sampling/sampling_logp_difference/max': 0.029237329959869385, 'sampling/importance_sampling_ratio/min': 0.9971944093704224, 'sampling/importance_sampling_ratio/mean': 1.0015150308609009, 'sampling/importance_sampling_ratio/max': 1.029668927192688, 'entropy': 0.01709412969648838, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9996045231819153, 'reward_meter_mean': 0.9996045231819153, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9996045231819153, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 84%|████████▍ | 2786/3300 [4:07:05<39:59,  4.67s/it]INFO 04-12 02:43:56 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2787/3300 [4:07:09<38:28,  4.50s/it]2026-04-12 02:44:00,359 | INFO | train_grpo_train | metrics_logged mode=train step=2787

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.5575757575757577e-06, 'num_tokens': 6327007.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 4.562741378322244e-05, 'sampling/sampling_logp_difference/max': 0.0008772250730544329, 'sampling/importance_sampling_ratio/min': 0.9996870756149292, 'sampling/importance_sampling_ratio/mean': 1.00003981590271, 'sampling/importance_sampling_ratio/max': 1.0008776187896729, 'entropy': 0.00042030583426821977, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 84%|████████▍ | 2787/3300 [4:07:09<38:28,  4.50s/it]INFO 04-12 02:44:00 [block_pool.py:378] Successfully reset prefix cache

 84%|████████▍ | 2788/3300 [4:07:14<39:48,  4.66s/it]2026-04-12 02:44:05,406 | INFO | train_grpo_train | metrics_logged mode=train step=2788

                                                     
{'loss': 0.0103, 'grad_norm': 6.924910545349121, 'learning_rate': 1.5545454545454547e-06, 'num_tokens': 6329047.0, 'completions/mean_length': 100.0, 'completions/min_length': 93.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.0, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9535719156265259, 'rewards/meter/std': 0.12491101026535034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9535719156265259, 'rewards/total_composite/std': 0.12491101026535034, 'reward': 0.9535719156265259, 'reward_std': 0.12491098791360855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.041653111577034, 'sampling/sampling_logp_difference/max': 1.1875276565551758, 'sampling/importance_sampling_ratio/min': 0.30497434735298157, 'sampling/importance_sampling_ratio/mean': 0.9999960064888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.349405812099576, 'clip_ratio/low_mean': 0.007281553465873003, 'clip_ratio/low_min': 0.007281553465873003, 'clip_ratio/high_mean': 0.03628689446486533, 'clip_ratio/high_max': 0.03628689446486533, 'clip_ratio/region_mean': 0.04356844793073833, 'reward_total_mean': 0.9535719156265259, 'reward_meter_mean': 0.9535719156265259, 'reward_meter_std': 0.12491101026535034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9535719156265259, 'reward_total_composite_std': 0.12491101026535034, 'epoch': 0.11}

 84%|████████▍ | 2788/3300 [4:07:14<39:48,  4.66s/it]INFO 04-12 02:44:05 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2789/3300 [4:07:18<38:54,  4.57s/it]2026-04-12 02:44:09,756 | INFO | train_grpo_train | metrics_logged mode=train step=2789

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.5515151515151516e-06, 'num_tokens': 6330791.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00020510550530161709, 'sampling/sampling_logp_difference/max': 0.002508021891117096, 'sampling/importance_sampling_ratio/min': 0.9999552369117737, 'sampling/importance_sampling_ratio/mean': 1.0002046823501587, 'sampling/importance_sampling_ratio/max': 1.0025111436843872, 'entropy': 0.0018322104588150978, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 85%|████████▍ | 2789/3300 [4:07:18<38:54,  4.57s/it]INFO 04-12 02:44:09 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2790/3300 [4:07:22<37:48,  4.45s/it]2026-04-12 02:44:13,918 | INFO | train_grpo_train | metrics_logged mode=train step=2790

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.5484848484848486e-06, 'num_tokens': 6332279.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 4.29271676694043e-05, 'sampling/sampling_logp_difference/max': 0.0006322840927168727, 'sampling/importance_sampling_ratio/min': 0.999407947063446, 'sampling/importance_sampling_ratio/mean': 1.000036358833313, 'sampling/importance_sampling_ratio/max': 1.0006325244903564, 'entropy': 0.00035309882514411584, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 85%|████████▍ | 2790/3300 [4:07:22<37:48,  4.45s/it]INFO 04-12 02:44:14 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2791/3300 [4:07:27<38:39,  4.56s/it]2026-04-12 02:44:18,733 | INFO | train_grpo_train | metrics_logged mode=train step=2791

                                                     
{'loss': 0.0171, 'grad_norm': 4.835381984710693, 'learning_rate': 1.5454545454545454e-06, 'num_tokens': 6334035.0, 'completions/mean_length': 66.5, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.98065185546875, 'rewards/meter/std': 0.016748299822211266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.98065185546875, 'rewards/total_composite/std': 0.016748299822211266, 'reward': 0.98065185546875, 'reward_std': 0.016748299822211266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038696978241205215, 'sampling/sampling_logp_difference/max': 1.3247365951538086, 'sampling/importance_sampling_ratio/min': 0.265872985124588, 'sampling/importance_sampling_ratio/mean': 1.0198413133621216, 'sampling/importance_sampling_ratio/max': 1.7263996601104736, 'entropy': 0.35312316194176674, 'clip_ratio/low_mean': 0.010924369795247912, 'clip_ratio/low_min': 0.010924369795247912, 'clip_ratio/high_mean': 0.005740093300119042, 'clip_ratio/high_max': 0.005740093300119042, 'clip_ratio/region_mean': 0.016664463095366955, 'reward_total_mean': 0.98065185546875, 'reward_meter_mean': 0.98065185546875, 'reward_meter_std': 0.016748299822211266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.98065185546875, 'reward_total_composite_std': 0.016748299822211266, 'epoch': 0.11}

 85%|████████▍ | 2791/3300 [4:07:27<38:39,  4.56s/it]INFO 04-12 02:44:18 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2792/3300 [4:07:34<43:47,  5.17s/it]2026-04-12 02:44:25,337 | INFO | train_grpo_train | metrics_logged mode=train step=2792

                                                     
{'loss': 0.007, 'grad_norm': 2.5946123600006104, 'learning_rate': 1.5424242424242425e-06, 'num_tokens': 6337431.0, 'completions/mean_length': 232.5, 'completions/min_length': 231.0, 'completions/max_length': 234.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 232.5, 'completions/min_terminated_length': 231.0, 'completions/max_terminated_length': 234.0, 'rewards/meter/mean': 0.97259521484375, 'rewards/meter/std': 0.05070994421839714, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9230769276618958, 'rewards/repeat_penalty/std': 0.041117113083601, 'rewards/total_composite/mean': 0.8965212106704712, 'rewards/total_composite/std': 0.035956308245658875, 'reward': 0.8965212106704712, 'reward_std': 0.035956308245658875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025431320071220398, 'sampling/sampling_logp_difference/max': 2.821336507797241, 'sampling/importance_sampling_ratio/min': 0.05952633172273636, 'sampling/importance_sampling_ratio/mean': 1.0048270225524902, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19751299265772104, 'clip_ratio/low_mean': 0.008609326789155602, 'clip_ratio/low_min': 0.008609326789155602, 'clip_ratio/high_mean': 0.011832875199615955, 'clip_ratio/high_max': 0.011832875199615955, 'clip_ratio/region_mean': 0.020442201988771558, 'reward_total_mean': 0.8965212106704712, 'reward_meter_mean': 0.97259521484375, 'reward_meter_std': 0.05070994421839714, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9230769276618958, 'reward_repeat_penalty_std': 0.041117113083601, 'reward_total_composite_mean': 0.8965212106704712, 'reward_total_composite_std': 0.035956308245658875, 'epoch': 0.11}

 85%|████████▍ | 2792/3300 [4:07:34<43:47,  5.17s/it]INFO 04-12 02:44:25 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2793/3300 [4:07:39<42:54,  5.08s/it]2026-04-12 02:44:30,195 | INFO | train_grpo_train | metrics_logged mode=train step=2793

                                                     
{'loss': -0.0046, 'grad_norm': 2.338566541671753, 'learning_rate': 1.5393939393939395e-06, 'num_tokens': 6339557.0, 'completions/mean_length': 91.75, 'completions/min_length': 90.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.75, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9973961710929871, 'rewards/meter/std': 0.00040277960943058133, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973961710929871, 'rewards/total_composite/std': 0.00040277960943058133, 'reward': 0.9973961710929871, 'reward_std': 0.00040277454536408186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018009934574365616, 'sampling/sampling_logp_difference/max': 1.1022424697875977, 'sampling/importance_sampling_ratio/min': 0.3321254551410675, 'sampling/importance_sampling_ratio/mean': 0.9954391717910767, 'sampling/importance_sampling_ratio/max': 1.4062292575836182, 'entropy': 0.10903155151754618, 'clip_ratio/low_mean': 0.01233343977946788, 'clip_ratio/low_min': 0.01233343977946788, 'clip_ratio/high_mean': 0.006764258956536651, 'clip_ratio/high_max': 0.006764258956536651, 'clip_ratio/region_mean': 0.01909769873600453, 'reward_total_mean': 0.9973961710929871, 'reward_meter_mean': 0.9973961710929871, 'reward_meter_std': 0.00040277960943058133, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973961710929871, 'reward_total_composite_std': 0.00040277960943058133, 'epoch': 0.11}

 85%|████████▍ | 2793/3300 [4:07:39<42:54,  5.08s/it]INFO 04-12 02:44:30 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2794/3300 [4:07:43<41:53,  4.97s/it]2026-04-12 02:44:34,909 | INFO | train_grpo_train | metrics_logged mode=train step=2794

                                                     
{'loss': -0.0009, 'grad_norm': 4.996799468994141, 'learning_rate': 1.5363636363636364e-06, 'num_tokens': 6341231.0, 'completions/mean_length': 69.25, 'completions/min_length': 67.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9884920120239258, 'rewards/meter/std': 0.01208480168133974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9884920120239258, 'rewards/total_composite/std': 0.01208480168133974, 'reward': 0.9884920120239258, 'reward_std': 0.012084796093404293, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03242041543126106, 'sampling/sampling_logp_difference/max': 1.4227056503295898, 'sampling/importance_sampling_ratio/min': 0.24106091260910034, 'sampling/importance_sampling_ratio/mean': 1.006773591041565, 'sampling/importance_sampling_ratio/max': 1.6237882375717163, 'entropy': 0.29199461452662945, 'clip_ratio/low_mean': 0.007144315168261528, 'clip_ratio/low_min': 0.007144315168261528, 'clip_ratio/high_mean': 0.0126050099497661, 'clip_ratio/high_max': 0.0126050099497661, 'clip_ratio/region_mean': 0.019749325118027627, 'reward_total_mean': 0.9884920120239258, 'reward_meter_mean': 0.9884920120239258, 'reward_meter_std': 0.01208480168133974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9884920120239258, 'reward_total_composite_std': 0.01208480168133974, 'epoch': 0.11}

 85%|████████▍ | 2794/3300 [4:07:43<41:53,  4.97s/it]INFO 04-12 02:44:35 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2795/3300 [4:07:49<42:37,  5.06s/it]2026-04-12 02:44:40,197 | INFO | train_grpo_train | metrics_logged mode=train step=2795

                                                     
{'loss': 0.0081, 'grad_norm': 3.0096166133880615, 'learning_rate': 1.5333333333333334e-06, 'num_tokens': 6343711.0, 'completions/mean_length': 126.0, 'completions/min_length': 123.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.0, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9772056937217712, 'rewards/meter/std': 0.017515793442726135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9772056937217712, 'rewards/total_composite/std': 0.017515793442726135, 'reward': 0.9772056937217712, 'reward_std': 0.017515793442726135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04109743982553482, 'sampling/sampling_logp_difference/max': 1.8138017654418945, 'sampling/importance_sampling_ratio/min': 0.1630331426858902, 'sampling/importance_sampling_ratio/mean': 1.0098986625671387, 'sampling/importance_sampling_ratio/max': 1.8909732103347778, 'entropy': 0.37405042350292206, 'clip_ratio/low_mean': 0.00992763601243496, 'clip_ratio/low_min': 0.00992763601243496, 'clip_ratio/high_mean': 0.01491319842170924, 'clip_ratio/high_max': 0.01491319842170924, 'clip_ratio/region_mean': 0.0248408344341442, 'reward_total_mean': 0.9772056937217712, 'reward_meter_mean': 0.9772056937217712, 'reward_meter_std': 0.017515793442726135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9772056937217712, 'reward_total_composite_std': 0.017515793442726135, 'epoch': 0.11}

 85%|████████▍ | 2795/3300 [4:07:49<42:37,  5.06s/it]INFO 04-12 02:44:40 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2796/3300 [4:07:58<54:31,  6.49s/it]2026-04-12 02:44:50,020 | INFO | train_grpo_train | metrics_logged mode=train step=2796

                                                     
{'loss': -0.029, 'grad_norm': 1.5088597536087036, 'learning_rate': 1.5303030303030302e-06, 'num_tokens': 6349489.0, 'completions/mean_length': 451.25, 'completions/min_length': 425.0, 'completions/max_length': 476.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 451.25, 'completions/min_terminated_length': 425.0, 'completions/max_terminated_length': 476.0, 'rewards/meter/mean': 0.9978024363517761, 'rewards/meter/std': 0.0035957051441073418, 'rewards/count_adherence/mean': 0.7666666507720947, 'rewards/count_adherence/std': 0.035634830594062805, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9777432680130005, 'rewards/repeat_penalty/std': 0.023832013830542564, 'rewards/total_composite/mean': 0.7476283311843872, 'rewards/total_composite/std': 0.03142942115664482, 'reward': 0.7476283311843872, 'reward_std': 0.031429413706064224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05049360170960426, 'sampling/sampling_logp_difference/max': 5.123012065887451, 'sampling/importance_sampling_ratio/min': 0.005958050023764372, 'sampling/importance_sampling_ratio/mean': 1.0113346576690674, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47964709997177124, 'clip_ratio/low_mean': 0.015304939355701208, 'clip_ratio/low_min': 0.015304939355701208, 'clip_ratio/high_mean': 0.019141072407364845, 'clip_ratio/high_max': 0.019141072407364845, 'clip_ratio/region_mean': 0.03444601176306605, 'reward_total_mean': 0.7476283311843872, 'reward_meter_mean': 0.9978024363517761, 'reward_meter_std': 0.0035957051441073418, 'reward_count_adherence_mean': 0.7666666507720947, 'reward_count_adherence_std': 0.035634830594062805, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9777432680130005, 'reward_repeat_penalty_std': 0.023832013830542564, 'reward_total_composite_mean': 0.7476283311843872, 'reward_total_composite_std': 0.03142942115664482, 'epoch': 0.11}

 85%|████████▍ | 2796/3300 [4:07:59<54:31,  6.49s/it]INFO 04-12 02:44:50 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2797/3300 [4:08:03<49:38,  5.92s/it]2026-04-12 02:44:54,622 | INFO | train_grpo_train | metrics_logged mode=train step=2797

                                                     
{'loss': -0.0044, 'grad_norm': 2.2734148502349854, 'learning_rate': 1.5272727272727275e-06, 'num_tokens': 6351421.0, 'completions/mean_length': 66.5, 'completions/min_length': 63.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9991590976715088, 'rewards/meter/std': 0.0001742392487358302, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991590976715088, 'rewards/total_composite/std': 0.0001742392487358302, 'reward': 0.9991590976715088, 'reward_std': 0.00017422223754692823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.028453955426812172, 'sampling/sampling_logp_difference/max': 1.6620092391967773, 'sampling/importance_sampling_ratio/min': 0.1897573173046112, 'sampling/importance_sampling_ratio/mean': 1.0024892091751099, 'sampling/importance_sampling_ratio/max': 1.4288572072982788, 'entropy': 0.21964257210493088, 'clip_ratio/low_mean': 0.009413161547854543, 'clip_ratio/low_min': 0.009413161547854543, 'clip_ratio/high_mean': 0.013506086892448366, 'clip_ratio/high_max': 0.013506086892448366, 'clip_ratio/region_mean': 0.02291924844030291, 'reward_total_mean': 0.9991590976715088, 'reward_meter_mean': 0.9991590976715088, 'reward_meter_std': 0.0001742392487358302, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991590976715088, 'reward_total_composite_std': 0.0001742392487358302, 'epoch': 0.11}

 85%|████████▍ | 2797/3300 [4:08:03<49:38,  5.92s/it]INFO 04-12 02:44:54 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2798/3300 [4:08:08<48:04,  5.75s/it]2026-04-12 02:44:59,940 | INFO | train_grpo_train | metrics_logged mode=train step=2798

                                                     
{'loss': -0.0027, 'grad_norm': 1.5468372106552124, 'learning_rate': 1.5242424242424245e-06, 'num_tokens': 6353888.0, 'completions/mean_length': 131.375, 'completions/min_length': 128.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.375, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9992819428443909, 'rewards/meter/std': 0.00023744572536088526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992819428443909, 'rewards/total_composite/std': 0.00023744572536088526, 'reward': 0.9992819428443909, 'reward_std': 0.00023744485224597156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01363647636026144, 'sampling/sampling_logp_difference/max': 1.2669620513916016, 'sampling/importance_sampling_ratio/min': 0.28168606758117676, 'sampling/importance_sampling_ratio/mean': 1.000964641571045, 'sampling/importance_sampling_ratio/max': 1.7840969562530518, 'entropy': 0.06868221471086144, 'clip_ratio/low_mean': 0.009618075448088348, 'clip_ratio/low_min': 0.009618075448088348, 'clip_ratio/high_mean': 0.0047420773771591485, 'clip_ratio/high_max': 0.0047420773771591485, 'clip_ratio/region_mean': 0.014360152825247496, 'reward_total_mean': 0.9992819428443909, 'reward_meter_mean': 0.9992819428443909, 'reward_meter_std': 0.00023744572536088526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992819428443909, 'reward_total_composite_std': 0.00023744572536088526, 'epoch': 0.11}

 85%|████████▍ | 2798/3300 [4:08:08<48:04,  5.75s/it]INFO 04-12 02:45:00 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2799/3300 [4:08:19<59:16,  7.10s/it]2026-04-12 02:45:10,197 | INFO | train_grpo_train | metrics_logged mode=train step=2799

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.5212121212121214e-06, 'num_tokens': 6355832.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9562926292419434, 'rewards/meter/std': 0.0762549340724945, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.025877464562654495, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9028632640838623, 'rewards/repeat_penalty/std': 0.0416770875453949, 'rewards/total_composite/mean': 0.5374298095703125, 'rewards/total_composite/std': 0.2192150205373764, 'reward': 0.5374298095703125, 'reward_std': 0.2192150205373764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.5374298095703125, 'reward_meter_mean': 0.9562926292419434, 'reward_meter_std': 0.0762549340724945, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.025877464562654495, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9028632640838623, 'reward_repeat_penalty_std': 0.0416770875453949, 'reward_total_composite_mean': 0.5374298095703125, 'reward_total_composite_std': 0.2192150205373764, 'epoch': 0.11}

 85%|████████▍ | 2799/3300 [4:08:19<59:16,  7.10s/it]INFO 04-12 02:45:10 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2800/3300 [4:08:26<59:33,  7.15s/it]2026-04-12 02:45:17,458 | INFO | train_grpo_train | metrics_logged mode=train step=2800

                                                     
{'loss': 0.0229, 'grad_norm': 2.547086715698242, 'learning_rate': 1.5181818181818184e-06, 'num_tokens': 6359592.0, 'completions/mean_length': 271.0, 'completions/min_length': 257.0, 'completions/max_length': 292.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 271.0, 'completions/min_terminated_length': 257.0, 'completions/max_terminated_length': 292.0, 'rewards/meter/mean': 0.9248241186141968, 'rewards/meter/std': 0.17904320359230042, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8916666507720947, 'rewards/repeat_penalty/std': 0.08683133870363235, 'rewards/total_composite/mean': 0.8224738240242004, 'rewards/total_composite/std': 0.1752014011144638, 'reward': 0.8224738240242004, 'reward_std': 0.1752013862133026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04360240325331688, 'sampling/sampling_logp_difference/max': 1.5698328018188477, 'sampling/importance_sampling_ratio/min': 0.2080799639225006, 'sampling/importance_sampling_ratio/mean': 1.0097815990447998, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4257059842348099, 'clip_ratio/low_mean': 0.00767904706299305, 'clip_ratio/low_min': 0.00767904706299305, 'clip_ratio/high_mean': 0.027406520443037152, 'clip_ratio/high_max': 0.027406520443037152, 'clip_ratio/region_mean': 0.0350855675060302, 'reward_total_mean': 0.8224738240242004, 'reward_meter_mean': 0.9248241186141968, 'reward_meter_std': 0.17904320359230042, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8916666507720947, 'reward_repeat_penalty_std': 0.08683133870363235, 'reward_total_composite_mean': 0.8224738240242004, 'reward_total_composite_std': 0.1752014011144638, 'epoch': 0.11}

 85%|████████▍ | 2800/3300 [4:08:26<59:33,  7.15s/it]INFO 04-12 02:45:17 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:39,  3.55s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.48s/it]

 31%|███       | 4/13 [00:16<00:39,  4.40s/it]

 38%|███▊      | 5/13 [00:22<00:39,  4.94s/it]

 46%|████▌     | 6/13 [00:29<00:39,  5.66s/it]

 54%|█████▍    | 7/13 [00:34<00:33,  5.50s/it]

 62%|██████▏   | 8/13 [00:41<00:29,  5.98s/it]

 69%|██████▉   | 9/13 [00:48<00:25,  6.39s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.96s/it]

 85%|████████▍ | 11/13 [00:59<00:11,  5.95s/it]

 92%|█████████▏| 12/13 [01:05<00:05,  5.98s/it]

100%|██████████| 13/13 [01:11<00:00,  5.78s/it]2026-04-12 02:46:35,700 | INFO | train_grpo_train | metrics_logged mode=eval step=2800

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 78.1168, 'eval_samples_per_second': 1.331, 'eval_steps_per_second': 0.166, 'eval_num_tokens': 6359592.0, 'eval_completions/mean_length': 210.54807692307693, 'eval_completions/min_length': 60.69230769230769, 'eval_completions/max_length': 418.84615384615387, 'eval_completions/clipped_ratio': 0.028846153846153848, 'eval_completions/mean_terminated_length': 200.53846271221454, 'eval_completions/min_terminated_length': 60.69230769230769, 'eval_completions/max_terminated_length': 381.15384615384613, 'eval_rewards/meter/mean': 0.7971470860334543, 'eval_rewards/meter/std': 0.31754180788993835, 'eval_rewards/count_adherence/mean': 0.9559012926541842, 'eval_rewards/count_adherence/std': 0.06501825956197885, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.9524615177741418, 'eval_rewards/repeat_penalty/std': 0.07457881277570358, 'eval_rewards/total_composite/mean': 0.717106791642996, 'eval_rewards/total_composite/std': 0.329087950862371, 'eval_reward': 0.717106791642996, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03292630620014209, 'eval_sampling/sampling_logp_difference/max': 1.1617195422832782, 'eval_sampling/importance_sampling_ratio/min': 0.32078430171196276, 'eval_sampling/importance_sampling_ratio/mean': 1.0094741124373217, 'eval_sampling/importance_sampling_ratio/max': 1.5328054244701679, 'eval_entropy': 0.38292053456489855, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.717106791642996, 'eval_reward_meter_mean': 0.7971470860334543, 'eval_reward_meter_std': 0.31754180788993835, 'eval_reward_count_adherence_mean': 0.9559012926541842, 'eval_reward_count_adherence_std': 0.06501825956197885, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.9524615177741418, 'eval_reward_repeat_penalty_std': 0.07457881277570358, 'eval_reward_total_composite_mean': 0.717106791642996, 'eval_reward_total_composite_std': 0.329087950862371, 'epoch': 0.11}

 85%|████████▍ | 2800/3300 [4:09:44<59:33,  7.15s/it]

100%|██████████| 13/13 [01:11<00:00,  5.78s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 85%|████████▍ | 2801/3300 [4:09:52<4:16:35, 30.85s/it]2026-04-12 02:46:43,625 | INFO | train_grpo_train | metrics_logged mode=train step=2801

                                                       
{'loss': -0.0039, 'grad_norm': 2.5363826751708984, 'learning_rate': 1.5151515151515152e-06, 'num_tokens': 6362056.0, 'completions/mean_length': 123.0, 'completions/min_length': 121.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.0, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9976023435592651, 'rewards/meter/std': 0.0004385724605526775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9619686603546143, 'rewards/total_composite/std': 0.06589511781930923, 'reward': 0.9619686603546143, 'reward_std': 0.06589511036872864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022641444578766823, 'sampling/sampling_logp_difference/max': 1.1188087463378906, 'sampling/importance_sampling_ratio/min': 0.3919609487056732, 'sampling/importance_sampling_ratio/mean': 1.0046809911727905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17738944850862026, 'clip_ratio/low_mean': 0.004123763646930456, 'clip_ratio/low_min': 0.004123763646930456, 'clip_ratio/high_mean': 0.016170406714081764, 'clip_ratio/high_max': 0.016170406714081764, 'clip_ratio/region_mean': 0.02029417036101222, 'reward_total_mean': 0.9619686603546143, 'reward_meter_mean': 0.9976023435592651, 'reward_meter_std': 0.0004385724605526775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9619686603546143, 'reward_total_composite_std': 0.06589511781930923, 'epoch': 0.11}

 85%|████████▍ | 2801/3300 [4:09:52<4:16:35, 30.85s/it]INFO 04-12 02:46:43 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2802/3300 [4:09:58<3:13:02, 23.26s/it]2026-04-12 02:46:49,160 | INFO | train_grpo_train | metrics_logged mode=train step=2802

                                                       
{'loss': -0.0145, 'grad_norm': 3.6220641136169434, 'learning_rate': 1.5121212121212123e-06, 'num_tokens': 6364387.0, 'completions/mean_length': 135.375, 'completions/min_length': 129.0, 'completions/max_length': 147.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.375, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 147.0, 'rewards/meter/mean': 0.8960581421852112, 'rewards/meter/std': 0.2624220848083496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8960581421852112, 'rewards/total_composite/std': 0.2624220848083496, 'reward': 0.8960581421852112, 'reward_std': 0.2624220848083496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049931738525629044, 'sampling/sampling_logp_difference/max': 1.0952783823013306, 'sampling/importance_sampling_ratio/min': 0.35183805227279663, 'sampling/importance_sampling_ratio/mean': 1.0152724981307983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4941607192158699, 'clip_ratio/low_mean': 0.0019379844889044762, 'clip_ratio/low_min': 0.0019379844889044762, 'clip_ratio/high_mean': 0.03208270261529833, 'clip_ratio/high_max': 0.03208270261529833, 'clip_ratio/region_mean': 0.03402068710420281, 'reward_total_mean': 0.8960581421852112, 'reward_meter_mean': 0.8960581421852112, 'reward_meter_std': 0.2624220848083496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8960581421852112, 'reward_total_composite_std': 0.2624220848083496, 'epoch': 0.11}

 85%|████████▍ | 2802/3300 [4:09:58<3:13:02, 23.26s/it]INFO 04-12 02:46:49 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2803/3300 [4:10:02<2:26:18, 17.66s/it]2026-04-12 02:46:53,763 | INFO | train_grpo_train | metrics_logged mode=train step=2803

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.5090909090909091e-06, 'num_tokens': 6366291.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00012427744513843209, 'sampling/sampling_logp_difference/max': 0.00598335824906826, 'sampling/importance_sampling_ratio/min': 0.9940345287322998, 'sampling/importance_sampling_ratio/mean': 1.0000933408737183, 'sampling/importance_sampling_ratio/max': 1.0050559043884277, 'entropy': 0.0011047264706576243, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 85%|████████▍ | 2803/3300 [4:10:02<2:26:18, 17.66s/it]INFO 04-12 02:46:53 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▍ | 2804/3300 [4:10:09<1:58:37, 14.35s/it]2026-04-12 02:47:00,389 | INFO | train_grpo_train | metrics_logged mode=train step=2804

                                                       
{'loss': 0.0106, 'grad_norm': 2.9486567974090576, 'learning_rate': 1.5060606060606062e-06, 'num_tokens': 6369524.0, 'completions/mean_length': 235.125, 'completions/min_length': 228.0, 'completions/max_length': 241.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 235.125, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 241.0, 'rewards/meter/mean': 0.9466480016708374, 'rewards/meter/std': 0.14826621115207672, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9886363744735718, 'rewards/repeat_penalty/std': 0.03214120864868164, 'rewards/total_composite/mean': 0.9352947473526001, 'rewards/total_composite/std': 0.14715024828910828, 'reward': 0.9352947473526001, 'reward_std': 0.14715023338794708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04945862293243408, 'sampling/sampling_logp_difference/max': 1.3091225624084473, 'sampling/importance_sampling_ratio/min': 0.27005690336227417, 'sampling/importance_sampling_ratio/mean': 1.0136529207229614, 'sampling/importance_sampling_ratio/max': 1.872164011001587, 'entropy': 0.46530257537961006, 'clip_ratio/low_mean': 0.010395145742222667, 'clip_ratio/low_min': 0.010395145742222667, 'clip_ratio/high_mean': 0.026202646549791098, 'clip_ratio/high_max': 0.026202646549791098, 'clip_ratio/region_mean': 0.036597792292013764, 'reward_total_mean': 0.9352947473526001, 'reward_meter_mean': 0.9466480016708374, 'reward_meter_std': 0.14826621115207672, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9886363744735718, 'reward_repeat_penalty_std': 0.03214120864868164, 'reward_total_composite_mean': 0.9352947473526001, 'reward_total_composite_std': 0.14715024828910828, 'epoch': 0.11}

 85%|████████▍ | 2804/3300 [4:10:09<1:58:37, 14.35s/it]INFO 04-12 02:47:00 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2805/3300 [4:10:14<1:35:35, 11.59s/it]2026-04-12 02:47:05,522 | INFO | train_grpo_train | metrics_logged mode=train step=2805

                                                       
{'loss': -0.0011, 'grad_norm': 4.279446125030518, 'learning_rate': 1.5030303030303032e-06, 'num_tokens': 6371724.0, 'completions/mean_length': 100.0, 'completions/min_length': 97.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9975219964981079, 'rewards/meter/std': 0.0041835736483335495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9728400707244873, 'rewards/total_composite/std': 0.07395301014184952, 'reward': 0.9728400707244873, 'reward_std': 0.07395300269126892, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03659245744347572, 'sampling/sampling_logp_difference/max': 1.4790048599243164, 'sampling/importance_sampling_ratio/min': 0.2278643399477005, 'sampling/importance_sampling_ratio/mean': 1.008768081665039, 'sampling/importance_sampling_ratio/max': 1.6153602600097656, 'entropy': 0.3125297725200653, 'clip_ratio/low_mean': 0.0063775512389838696, 'clip_ratio/low_min': 0.0063775512389838696, 'clip_ratio/high_mean': 0.029976313933730125, 'clip_ratio/high_max': 0.029976313933730125, 'clip_ratio/region_mean': 0.036353865172713995, 'reward_total_mean': 0.9728400707244873, 'reward_meter_mean': 0.9975219964981079, 'reward_meter_std': 0.0041835736483335495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9728400707244873, 'reward_total_composite_std': 0.07395301014184952, 'epoch': 0.11}

 85%|████████▌ | 2805/3300 [4:10:14<1:35:35, 11.59s/it]INFO 04-12 02:47:05 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2806/3300 [4:10:18<1:17:27,  9.41s/it]2026-04-12 02:47:09,845 | INFO | train_grpo_train | metrics_logged mode=train step=2806

                                                       
{'loss': -0.0041, 'grad_norm': 0.8794472813606262, 'learning_rate': 1.5e-06, 'num_tokens': 6373380.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7413170337677002, 'rewards/meter/std': 0.13101767003536224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7413170337677002, 'rewards/total_composite/std': 0.13101767003536224, 'reward': 0.7413170337677002, 'reward_std': 0.13101768493652344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0035149240866303444, 'sampling/sampling_logp_difference/max': 1.4165477752685547, 'sampling/importance_sampling_ratio/min': 0.24254991114139557, 'sampling/importance_sampling_ratio/mean': 0.9984825849533081, 'sampling/importance_sampling_ratio/max': 1.0023080110549927, 'entropy': 0.0018152309057768434, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7413170337677002, 'reward_meter_mean': 0.7413170337677002, 'reward_meter_std': 0.13101767003536224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7413170337677002, 'reward_total_composite_std': 0.13101767003536224, 'epoch': 0.11}

 85%|████████▌ | 2806/3300 [4:10:18<1:17:27,  9.41s/it]INFO 04-12 02:47:10 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2807/3300 [4:10:23<1:06:54,  8.14s/it]2026-04-12 02:47:15,057 | INFO | train_grpo_train | metrics_logged mode=train step=2807

                                                       
{'loss': -0.0079, 'grad_norm': 1.6524964570999146, 'learning_rate': 1.496969696969697e-06, 'num_tokens': 6375595.0, 'completions/mean_length': 118.875, 'completions/min_length': 116.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.875, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9989545345306396, 'rewards/meter/std': 0.00029157428070902824, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989545345306396, 'rewards/total_composite/std': 0.00029157428070902824, 'reward': 0.9989545345306396, 'reward_std': 0.00029156444361433387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03430765122175217, 'sampling/sampling_logp_difference/max': 1.1520414352416992, 'sampling/importance_sampling_ratio/min': 0.31599104404449463, 'sampling/importance_sampling_ratio/mean': 1.0066179037094116, 'sampling/importance_sampling_ratio/max': 1.6717981100082397, 'entropy': 0.30837439745664597, 'clip_ratio/low_mean': 0.004273504368029535, 'clip_ratio/low_min': 0.004273504368029535, 'clip_ratio/high_mean': 0.01668829272966832, 'clip_ratio/high_max': 0.01668829272966832, 'clip_ratio/region_mean': 0.020961797097697854, 'reward_total_mean': 0.9989545345306396, 'reward_meter_mean': 0.9989545345306396, 'reward_meter_std': 0.00029157428070902824, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989545345306396, 'reward_total_composite_std': 0.00029157428070902824, 'epoch': 0.11}

 85%|████████▌ | 2807/3300 [4:10:24<1:06:54,  8.14s/it]INFO 04-12 02:47:15 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2808/3300 [4:10:28<58:52,  7.18s/it]  2026-04-12 02:47:19,980 | INFO | train_grpo_train | metrics_logged mode=train step=2808

                                                     
{'loss': -0.0002, 'grad_norm': 0.13850678503513336, 'learning_rate': 1.493939393939394e-06, 'num_tokens': 6377373.0, 'completions/mean_length': 66.25, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981439709663391, 'rewards/meter/std': 1.6234846043516882e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981439709663391, 'rewards/total_composite/std': 1.6234846043516882e-05, 'reward': 0.9981439709663391, 'reward_std': 1.6221796613535844e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006040393374860287, 'sampling/sampling_logp_difference/max': 0.7629547119140625, 'sampling/importance_sampling_ratio/min': 0.46628665924072266, 'sampling/importance_sampling_ratio/mean': 0.9976470470428467, 'sampling/importance_sampling_ratio/max': 1.0641453266143799, 'entropy': 0.022154160542413592, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.00562528264708817, 'clip_ratio/high_max': 0.00562528264708817, 'clip_ratio/region_mean': 0.007519222097471356, 'reward_total_mean': 0.9981439709663391, 'reward_meter_mean': 0.9981439709663391, 'reward_meter_std': 1.6234846043516882e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981439709663391, 'reward_total_composite_std': 1.6234846043516882e-05, 'epoch': 0.11}

 85%|████████▌ | 2808/3300 [4:10:28<58:52,  7.18s/it]INFO 04-12 02:47:20 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2809/3300 [4:10:33<52:22,  6.40s/it]2026-04-12 02:47:24,559 | INFO | train_grpo_train | metrics_logged mode=train step=2809

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.490909090909091e-06, 'num_tokens': 6379350.0, 'completions/mean_length': 66.125, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0026704377960413694, 'sampling/sampling_logp_difference/max': 0.33742189407348633, 'sampling/importance_sampling_ratio/min': 0.7136077284812927, 'sampling/importance_sampling_ratio/mean': 0.9995959401130676, 'sampling/importance_sampling_ratio/max': 1.0592564344406128, 'entropy': 0.015775780542753637, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 85%|████████▌ | 2809/3300 [4:10:33<52:22,  6.40s/it]INFO 04-12 02:47:24 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2810/3300 [4:10:37<47:10,  5.78s/it]2026-04-12 02:47:28,879 | INFO | train_grpo_train | metrics_logged mode=train step=2810

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.4878787878787878e-06, 'num_tokens': 6380838.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00016406136273872107, 'sampling/sampling_logp_difference/max': 0.0028465576469898224, 'sampling/importance_sampling_ratio/min': 0.9998530745506287, 'sampling/importance_sampling_ratio/mean': 1.0001628398895264, 'sampling/importance_sampling_ratio/max': 1.0028506517410278, 'entropy': 0.0013478934415616095, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 85%|████████▌ | 2810/3300 [4:10:37<47:10,  5.78s/it]INFO 04-12 02:47:29 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2811/3300 [4:10:42<44:31,  5.46s/it]2026-04-12 02:47:33,613 | INFO | train_grpo_train | metrics_logged mode=train step=2811

                                                     
{'loss': 0.0043, 'grad_norm': 4.882931232452393, 'learning_rate': 1.484848484848485e-06, 'num_tokens': 6382557.0, 'completions/mean_length': 65.875, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9871849417686462, 'rewards/meter/std': 0.011940443888306618, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9871849417686462, 'rewards/total_composite/std': 0.011940443888306618, 'reward': 0.9871849417686462, 'reward_std': 0.011940454132854939, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029961178079247475, 'sampling/sampling_logp_difference/max': 1.0541696548461914, 'sampling/importance_sampling_ratio/min': 0.3484816551208496, 'sampling/importance_sampling_ratio/mean': 1.0042191743850708, 'sampling/importance_sampling_ratio/max': 1.5272610187530518, 'entropy': 0.25444527715444565, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.013288452988490462, 'clip_ratio/high_max': 0.013288452988490462, 'clip_ratio/region_mean': 0.015182392438873649, 'reward_total_mean': 0.9871849417686462, 'reward_meter_mean': 0.9871849417686462, 'reward_meter_std': 0.011940443888306618, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9871849417686462, 'reward_total_composite_std': 0.011940443888306618, 'epoch': 0.11}

 85%|████████▌ | 2811/3300 [4:10:42<44:31,  5.46s/it]INFO 04-12 02:47:33 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2812/3300 [4:10:46<41:30,  5.10s/it]2026-04-12 02:47:37,877 | INFO | train_grpo_train | metrics_logged mode=train step=2812

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.481818181818182e-06, 'num_tokens': 6383901.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00016283367585856467, 'sampling/sampling_logp_difference/max': 0.0031713712960481644, 'sampling/importance_sampling_ratio/min': 0.9999763369560242, 'sampling/importance_sampling_ratio/mean': 1.0001628398895264, 'sampling/importance_sampling_ratio/max': 1.0031764507293701, 'entropy': 0.0012169194815214723, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 85%|████████▌ | 2812/3300 [4:10:46<41:30,  5.10s/it]INFO 04-12 02:47:38 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2813/3300 [4:10:55<51:24,  6.33s/it]2026-04-12 02:47:47,082 | INFO | train_grpo_train | metrics_logged mode=train step=2813

                                                     
{'loss': -0.0216, 'grad_norm': 1.4151586294174194, 'learning_rate': 1.478787878787879e-06, 'num_tokens': 6388632.0, 'completions/mean_length': 402.375, 'completions/min_length': 390.0, 'completions/max_length': 438.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 402.375, 'completions/min_terminated_length': 390.0, 'completions/max_terminated_length': 438.0, 'rewards/meter/mean': 0.9989125728607178, 'rewards/meter/std': 0.0006985744112171233, 'rewards/count_adherence/mean': 0.7788461446762085, 'rewards/count_adherence/std': 0.027196412906050682, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9802631139755249, 'rewards/repeat_penalty/std': 0.027239417657256126, 'rewards/total_composite/mean': 0.7628299593925476, 'rewards/total_composite/std': 0.0387784019112587, 'reward': 0.7628299593925476, 'reward_std': 0.038778405636548996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04369629919528961, 'sampling/sampling_logp_difference/max': 2.0083513259887695, 'sampling/importance_sampling_ratio/min': 0.1342097669839859, 'sampling/importance_sampling_ratio/mean': 1.010063886642456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43715671822428703, 'clip_ratio/low_mean': 0.008746044244617224, 'clip_ratio/low_min': 0.008746044244617224, 'clip_ratio/high_mean': 0.02115830732509494, 'clip_ratio/high_max': 0.02115830732509494, 'clip_ratio/region_mean': 0.029904351569712162, 'reward_total_mean': 0.7628299593925476, 'reward_meter_mean': 0.9989125728607178, 'reward_meter_std': 0.0006985744112171233, 'reward_count_adherence_mean': 0.7788461446762085, 'reward_count_adherence_std': 0.027196412906050682, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9802631139755249, 'reward_repeat_penalty_std': 0.027239417657256126, 'reward_total_composite_mean': 0.7628299593925476, 'reward_total_composite_std': 0.0387784019112587, 'epoch': 0.11}

 85%|████████▌ | 2813/3300 [4:10:56<51:24,  6.33s/it]INFO 04-12 02:47:47 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2814/3300 [4:11:01<48:54,  6.04s/it]2026-04-12 02:47:52,425 | INFO | train_grpo_train | metrics_logged mode=train step=2814

                                                     
{'loss': -0.0018, 'grad_norm': 1.2337918281555176, 'learning_rate': 1.475757575757576e-06, 'num_tokens': 6391131.0, 'completions/mean_length': 127.375, 'completions/min_length': 125.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.375, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9961709976196289, 'rewards/meter/std': 0.0034113116562366486, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9428157210350037, 'rewards/total_composite/std': 0.07388003170490265, 'reward': 0.9428157210350037, 'reward_std': 0.07388003170490265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012026426382362843, 'sampling/sampling_logp_difference/max': 1.767538070678711, 'sampling/importance_sampling_ratio/min': 0.1707528531551361, 'sampling/importance_sampling_ratio/mean': 1.0020824670791626, 'sampling/importance_sampling_ratio/max': 1.6902966499328613, 'entropy': 0.11378153134137392, 'clip_ratio/low_mean': 0.0029606895986944437, 'clip_ratio/low_min': 0.0029606895986944437, 'clip_ratio/high_mean': 0.008882812689989805, 'clip_ratio/high_max': 0.008882812689989805, 'clip_ratio/region_mean': 0.011843502288684249, 'reward_total_mean': 0.9428157210350037, 'reward_meter_mean': 0.9961709976196289, 'reward_meter_std': 0.0034113116562366486, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9428157210350037, 'reward_total_composite_std': 0.07388003170490265, 'epoch': 0.11}

 85%|████████▌ | 2814/3300 [4:11:01<48:54,  6.04s/it]INFO 04-12 02:47:52 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2815/3300 [4:11:05<44:30,  5.51s/it]2026-04-12 02:47:56,686 | INFO | train_grpo_train | metrics_logged mode=train step=2815

                                                     
{'loss': 0.013, 'grad_norm': 1.3737421035766602, 'learning_rate': 1.4727272727272728e-06, 'num_tokens': 6392648.0, 'completions/mean_length': 40.625, 'completions/min_length': 39.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.99871426820755, 'rewards/meter/std': 0.0005067276651971042, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99871426820755, 'rewards/total_composite/std': 0.0005067276651971042, 'reward': 0.99871426820755, 'reward_std': 0.0005067184683866799, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01035021897405386, 'sampling/sampling_logp_difference/max': 0.8471136093139648, 'sampling/importance_sampling_ratio/min': 0.4286504089832306, 'sampling/importance_sampling_ratio/mean': 0.9997342824935913, 'sampling/importance_sampling_ratio/max': 1.4657316207885742, 'entropy': 0.0865395087748766, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006330128293484449, 'clip_ratio/high_max': 0.006330128293484449, 'clip_ratio/region_mean': 0.006330128293484449, 'reward_total_mean': 0.99871426820755, 'reward_meter_mean': 0.99871426820755, 'reward_meter_std': 0.0005067276651971042, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99871426820755, 'reward_total_composite_std': 0.0005067276651971042, 'epoch': 0.11}

 85%|████████▌ | 2815/3300 [4:11:05<44:30,  5.51s/it]INFO 04-12 02:47:56 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2816/3300 [4:11:10<42:58,  5.33s/it]2026-04-12 02:48:01,597 | INFO | train_grpo_train | metrics_logged mode=train step=2816

                                                     
{'loss': 0.0064, 'grad_norm': 3.449428081512451, 'learning_rate': 1.4696969696969698e-06, 'num_tokens': 6394769.0, 'completions/mean_length': 102.125, 'completions/min_length': 101.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.125, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9987610578536987, 'rewards/meter/std': 0.0006801035488024354, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987610578536987, 'rewards/total_composite/std': 0.0006801035488024354, 'reward': 0.9987610578536987, 'reward_std': 0.0006801035488024354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03967723622918129, 'sampling/sampling_logp_difference/max': 1.2159053087234497, 'sampling/importance_sampling_ratio/min': 0.29644152522087097, 'sampling/importance_sampling_ratio/mean': 1.0010402202606201, 'sampling/importance_sampling_ratio/max': 1.9102437496185303, 'entropy': 0.33477963879704475, 'clip_ratio/low_mean': 0.01341109094209969, 'clip_ratio/low_min': 0.01341109094209969, 'clip_ratio/high_mean': 0.022014267276972532, 'clip_ratio/high_max': 0.022014267276972532, 'clip_ratio/region_mean': 0.03542535821907222, 'reward_total_mean': 0.9987610578536987, 'reward_meter_mean': 0.9987610578536987, 'reward_meter_std': 0.0006801035488024354, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987610578536987, 'reward_total_composite_std': 0.0006801035488024354, 'epoch': 0.11}

 85%|████████▌ | 2816/3300 [4:11:10<42:58,  5.33s/it]INFO 04-12 02:48:01 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2817/3300 [4:11:14<40:53,  5.08s/it]2026-04-12 02:48:06,098 | INFO | train_grpo_train | metrics_logged mode=train step=2817

                                                     
{'loss': 0.0051, 'grad_norm': 4.832633972167969, 'learning_rate': 1.4666666666666669e-06, 'num_tokens': 6396562.0, 'completions/mean_length': 67.125, 'completions/min_length': 66.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9991535544395447, 'rewards/meter/std': 0.00035637259134091437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991535544395447, 'rewards/total_composite/std': 0.00035637259134091437, 'reward': 0.9991535544395447, 'reward_std': 0.00035637334804050624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03295252099633217, 'sampling/sampling_logp_difference/max': 5.969149589538574, 'sampling/importance_sampling_ratio/min': 0.0025564145762473345, 'sampling/importance_sampling_ratio/mean': 0.9987229108810425, 'sampling/importance_sampling_ratio/max': 1.4108344316482544, 'entropy': 0.15694944839924574, 'clip_ratio/low_mean': 0.0037313431967049837, 'clip_ratio/low_min': 0.0037313431967049837, 'clip_ratio/high_mean': 0.01860184350516647, 'clip_ratio/high_max': 0.01860184350516647, 'clip_ratio/region_mean': 0.022333186701871455, 'reward_total_mean': 0.9991535544395447, 'reward_meter_mean': 0.9991535544395447, 'reward_meter_std': 0.00035637259134091437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991535544395447, 'reward_total_composite_std': 0.00035637259134091437, 'epoch': 0.11}

 85%|████████▌ | 2817/3300 [4:11:15<40:53,  5.08s/it]INFO 04-12 02:48:06 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2818/3300 [4:11:21<43:19,  5.39s/it]2026-04-12 02:48:12,224 | INFO | train_grpo_train | metrics_logged mode=train step=2818

                                                     
{'loss': -0.002, 'grad_norm': 1.7950763702392578, 'learning_rate': 1.4636363636363637e-06, 'num_tokens': 6399800.0, 'completions/mean_length': 198.75, 'completions/min_length': 196.0, 'completions/max_length': 200.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 198.75, 'completions/min_terminated_length': 196.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.9993983507156372, 'rewards/meter/std': 7.34872737666592e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9090909361839294, 'rewards/repeat_penalty/std': 0.048592954874038696, 'rewards/total_composite/mean': 0.9085431098937988, 'rewards/total_composite/std': 0.04854762554168701, 'reward': 0.9085431098937988, 'reward_std': 0.04854761064052582, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014119832776486874, 'sampling/sampling_logp_difference/max': 1.0774822235107422, 'sampling/importance_sampling_ratio/min': 0.3404516279697418, 'sampling/importance_sampling_ratio/mean': 1.003812551498413, 'sampling/importance_sampling_ratio/max': 1.756600022315979, 'entropy': 0.12045034673064947, 'clip_ratio/low_mean': 0.0075952449114993215, 'clip_ratio/low_min': 0.0075952449114993215, 'clip_ratio/high_mean': 0.007499999832361937, 'clip_ratio/high_max': 0.007499999832361937, 'clip_ratio/region_mean': 0.015095244743861258, 'reward_total_mean': 0.9085431098937988, 'reward_meter_mean': 0.9993983507156372, 'reward_meter_std': 7.34872737666592e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9090909361839294, 'reward_repeat_penalty_std': 0.048592954874038696, 'reward_total_composite_mean': 0.9085431098937988, 'reward_total_composite_std': 0.04854762554168701, 'epoch': 0.11}

 85%|████████▌ | 2818/3300 [4:11:21<43:19,  5.39s/it]INFO 04-12 02:48:12 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2819/3300 [4:11:26<42:05,  5.25s/it]2026-04-12 02:48:17,147 | INFO | train_grpo_train | metrics_logged mode=train step=2819

                                                     
{'loss': -0.0006, 'grad_norm': 4.404444217681885, 'learning_rate': 1.4606060606060608e-06, 'num_tokens': 6402006.0, 'completions/mean_length': 106.75, 'completions/min_length': 106.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.75, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9992326498031616, 'rewards/meter/std': 0.0003244501131121069, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992326498031616, 'rewards/total_composite/std': 0.0003244501131121069, 'reward': 0.9992326498031616, 'reward_std': 0.00032444868702441454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01700535975396633, 'sampling/sampling_logp_difference/max': 1.6258955001831055, 'sampling/importance_sampling_ratio/min': 0.1967354267835617, 'sampling/importance_sampling_ratio/mean': 1.0044901371002197, 'sampling/importance_sampling_ratio/max': 1.5127817392349243, 'entropy': 0.12008912675082684, 'clip_ratio/low_mean': 0.003537735901772976, 'clip_ratio/low_min': 0.003537735901772976, 'clip_ratio/high_mean': 0.007020366610959172, 'clip_ratio/high_max': 0.007020366610959172, 'clip_ratio/region_mean': 0.010558102512732148, 'reward_total_mean': 0.9992326498031616, 'reward_meter_mean': 0.9992326498031616, 'reward_meter_std': 0.0003244501131121069, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992326498031616, 'reward_total_composite_std': 0.0003244501131121069, 'epoch': 0.11}

 85%|████████▌ | 2819/3300 [4:11:26<42:05,  5.25s/it]INFO 04-12 02:48:17 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2820/3300 [4:11:30<40:01,  5.00s/it]2026-04-12 02:48:21,565 | INFO | train_grpo_train | metrics_logged mode=train step=2820

                                                     
{'loss': -0.0417, 'grad_norm': 4.645050525665283, 'learning_rate': 1.4575757575757576e-06, 'num_tokens': 6403787.0, 'completions/mean_length': 69.625, 'completions/min_length': 62.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9988994598388672, 'rewards/meter/std': 0.001365774660371244, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988994598388672, 'rewards/total_composite/std': 0.001365774660371244, 'reward': 0.9988994598388672, 'reward_std': 0.0013657818781211972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022290274500846863, 'sampling/sampling_logp_difference/max': 1.9662590026855469, 'sampling/importance_sampling_ratio/min': 0.1399795413017273, 'sampling/importance_sampling_ratio/mean': 0.9964150786399841, 'sampling/importance_sampling_ratio/max': 1.4191588163375854, 'entropy': 0.10764285689219832, 'clip_ratio/low_mean': 0.008064515888690948, 'clip_ratio/low_min': 0.008064515888690948, 'clip_ratio/high_mean': 0.00890342053025961, 'clip_ratio/high_max': 0.00890342053025961, 'clip_ratio/region_mean': 0.016967936418950558, 'reward_total_mean': 0.9988994598388672, 'reward_meter_mean': 0.9988994598388672, 'reward_meter_std': 0.001365774660371244, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988994598388672, 'reward_total_composite_std': 0.001365774660371244, 'epoch': 0.11}

 85%|████████▌ | 2820/3300 [4:11:30<40:01,  5.00s/it]INFO 04-12 02:48:21 [block_pool.py:378] Successfully reset prefix cache

 85%|████████▌ | 2821/3300 [4:11:34<38:46,  4.86s/it]2026-04-12 02:48:26,085 | INFO | train_grpo_train | metrics_logged mode=train step=2821

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.4545454545454546e-06, 'num_tokens': 6405779.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00021103888866491616, 'sampling/sampling_logp_difference/max': 0.002838927786797285, 'sampling/importance_sampling_ratio/min': 0.9997273683547974, 'sampling/importance_sampling_ratio/mean': 1.0002089738845825, 'sampling/importance_sampling_ratio/max': 1.0028430223464966, 'entropy': 0.0020125040755374357, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 85%|████████▌ | 2821/3300 [4:11:35<38:46,  4.86s/it]INFO 04-12 02:48:26 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2822/3300 [4:11:40<40:24,  5.07s/it]2026-04-12 02:48:31,654 | INFO | train_grpo_train | metrics_logged mode=train step=2822

                                                     
{'loss': -0.002, 'grad_norm': 1.043582558631897, 'learning_rate': 1.4515151515151515e-06, 'num_tokens': 6408280.0, 'completions/mean_length': 158.625, 'completions/min_length': 157.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.625, 'completions/min_terminated_length': 157.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.980668306350708, 'rewards/meter/std': 0.04880243167281151, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.9252284169197083, 'rewards/total_composite/std': 0.0608980655670166, 'reward': 0.9252284169197083, 'reward_std': 0.0608980655670166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013364696875214577, 'sampling/sampling_logp_difference/max': 1.1392550468444824, 'sampling/importance_sampling_ratio/min': 0.32005739212036133, 'sampling/importance_sampling_ratio/mean': 1.0040552616119385, 'sampling/importance_sampling_ratio/max': 1.4671419858932495, 'entropy': 0.14032841473817825, 'clip_ratio/low_mean': 0.003149755357299, 'clip_ratio/low_min': 0.003149755357299, 'clip_ratio/high_mean': 0.003930817823857069, 'clip_ratio/high_max': 0.003930817823857069, 'clip_ratio/region_mean': 0.007080573181156069, 'reward_total_mean': 0.9252284169197083, 'reward_meter_mean': 0.980668306350708, 'reward_meter_std': 0.04880243167281151, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.9252284169197083, 'reward_total_composite_std': 0.0608980655670166, 'epoch': 0.11}

 86%|████████▌ | 2822/3300 [4:11:40<40:24,  5.07s/it]INFO 04-12 02:48:31 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2823/3300 [4:11:44<38:30,  4.84s/it]2026-04-12 02:48:35,969 | INFO | train_grpo_train | metrics_logged mode=train step=2823

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.4484848484848485e-06, 'num_tokens': 6409984.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002062379935523495, 'sampling/sampling_logp_difference/max': 0.0020993247162550688, 'sampling/importance_sampling_ratio/min': 0.9997925162315369, 'sampling/importance_sampling_ratio/mean': 1.000205397605896, 'sampling/importance_sampling_ratio/max': 1.0021015405654907, 'entropy': 0.001565092708915472, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2823/3300 [4:11:44<38:30,  4.84s/it]INFO 04-12 02:48:36 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2824/3300 [4:11:49<37:50,  4.77s/it]2026-04-12 02:48:40,574 | INFO | train_grpo_train | metrics_logged mode=train step=2824

                                                     
{'loss': 0.0014, 'grad_norm': 0.7708361148834229, 'learning_rate': 1.4454545454545453e-06, 'num_tokens': 6411779.0, 'completions/mean_length': 71.375, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994142055511475, 'rewards/meter/std': 5.688391684088856e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994142055511475, 'rewards/total_composite/std': 5.688391684088856e-05, 'reward': 0.9994142055511475, 'reward_std': 5.688605597242713e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009274118579924107, 'sampling/sampling_logp_difference/max': 0.677030086517334, 'sampling/importance_sampling_ratio/min': 0.5081238150596619, 'sampling/importance_sampling_ratio/mean': 1.0028364658355713, 'sampling/importance_sampling_ratio/max': 1.4751698970794678, 'entropy': 0.06790398666635156, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.0017361111240461469, 'clip_ratio/high_max': 0.0017361111240461469, 'clip_ratio/region_mean': 0.0034722222480922937, 'reward_total_mean': 0.9994142055511475, 'reward_meter_mean': 0.9994142055511475, 'reward_meter_std': 5.688391684088856e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994142055511475, 'reward_total_composite_std': 5.688391684088856e-05, 'epoch': 0.11}

 86%|████████▌ | 2824/3300 [4:11:49<37:50,  4.77s/it]INFO 04-12 02:48:40 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2825/3300 [4:11:54<38:38,  4.88s/it]2026-04-12 02:48:45,703 | INFO | train_grpo_train | metrics_logged mode=train step=2825

                                                     
{'loss': -0.0007, 'grad_norm': 0.8596204519271851, 'learning_rate': 1.4424242424242426e-06, 'num_tokens': 6413966.0, 'completions/mean_length': 106.375, 'completions/min_length': 104.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.375, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9992821216583252, 'rewards/meter/std': 0.0001443348592147231, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992821216583252, 'rewards/total_composite/std': 0.0001443348592147231, 'reward': 0.9992821216583252, 'reward_std': 0.000144336445373483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018452905118465424, 'sampling/sampling_logp_difference/max': 1.3011884689331055, 'sampling/importance_sampling_ratio/min': 0.2722080945968628, 'sampling/importance_sampling_ratio/mean': 1.001145839691162, 'sampling/importance_sampling_ratio/max': 1.6910252571105957, 'entropy': 0.12096875347197056, 'clip_ratio/low_mean': 0.0035493926843628287, 'clip_ratio/low_min': 0.0035493926843628287, 'clip_ratio/high_mean': 0.009345794096589088, 'clip_ratio/high_max': 0.009345794096589088, 'clip_ratio/region_mean': 0.012895186780951917, 'reward_total_mean': 0.9992821216583252, 'reward_meter_mean': 0.9992821216583252, 'reward_meter_std': 0.0001443348592147231, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992821216583252, 'reward_total_composite_std': 0.0001443348592147231, 'epoch': 0.11}

 86%|████████▌ | 2825/3300 [4:11:54<38:38,  4.88s/it]INFO 04-12 02:48:45 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2826/3300 [4:11:59<37:46,  4.78s/it]2026-04-12 02:48:50,262 | INFO | train_grpo_train | metrics_logged mode=train step=2826

                                                     
{'loss': -0.0004, 'grad_norm': 0.9645191431045532, 'learning_rate': 1.4393939393939396e-06, 'num_tokens': 6415713.0, 'completions/mean_length': 66.375, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981287717819214, 'rewards/meter/std': 3.6517179978545755e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981287717819214, 'rewards/total_composite/std': 3.6517179978545755e-05, 'reward': 0.9981287717819214, 'reward_std': 3.651110455393791e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007696997374296188, 'sampling/sampling_logp_difference/max': 0.9785857200622559, 'sampling/importance_sampling_ratio/min': 0.37584227323532104, 'sampling/importance_sampling_ratio/mean': 0.9988413453102112, 'sampling/importance_sampling_ratio/max': 1.138381838798523, 'entropy': 0.03601772431284189, 'clip_ratio/low_mean': 0.0037313431967049837, 'clip_ratio/low_min': 0.0037313431967049837, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9981287717819214, 'reward_meter_mean': 0.9981287717819214, 'reward_meter_std': 3.6517179978545755e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981287717819214, 'reward_total_composite_std': 3.6517179978545755e-05, 'epoch': 0.11}

 86%|████████▌ | 2826/3300 [4:11:59<37:46,  4.78s/it]INFO 04-12 02:48:50 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2827/3300 [4:12:05<41:14,  5.23s/it]2026-04-12 02:48:56,542 | INFO | train_grpo_train | metrics_logged mode=train step=2827

                                                     
{'loss': -0.0018, 'grad_norm': 2.085141897201538, 'learning_rate': 1.4363636363636365e-06, 'num_tokens': 6418795.0, 'completions/mean_length': 197.25, 'completions/min_length': 192.0, 'completions/max_length': 201.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.25, 'completions/min_terminated_length': 192.0, 'completions/max_terminated_length': 201.0, 'rewards/meter/mean': 0.9994193315505981, 'rewards/meter/std': 5.9720292483689263e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9545454978942871, 'rewards/repeat_penalty/std': 0.0485929399728775, 'rewards/total_composite/mean': 0.9539901614189148, 'rewards/total_composite/std': 0.04854191467165947, 'reward': 0.9539901614189148, 'reward_std': 0.04854191467165947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015652773901820183, 'sampling/sampling_logp_difference/max': 1.3481910228729248, 'sampling/importance_sampling_ratio/min': 0.2597096562385559, 'sampling/importance_sampling_ratio/mean': 1.0017954111099243, 'sampling/importance_sampling_ratio/max': 1.5125662088394165, 'entropy': 0.117509332485497, 'clip_ratio/low_mean': 0.005142431997228414, 'clip_ratio/low_min': 0.005142431997228414, 'clip_ratio/high_mean': 0.008849852601997554, 'clip_ratio/high_max': 0.008849852601997554, 'clip_ratio/region_mean': 0.013992284599225968, 'reward_total_mean': 0.9539901614189148, 'reward_meter_mean': 0.9994193315505981, 'reward_meter_std': 5.9720292483689263e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9545454978942871, 'reward_repeat_penalty_std': 0.0485929399728775, 'reward_total_composite_mean': 0.9539901614189148, 'reward_total_composite_std': 0.04854191467165947, 'epoch': 0.11}

 86%|████████▌ | 2827/3300 [4:12:05<41:14,  5.23s/it]INFO 04-12 02:48:56 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2828/3300 [4:12:09<39:26,  5.01s/it]2026-04-12 02:49:01,041 | INFO | train_grpo_train | metrics_logged mode=train step=2828

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.4333333333333335e-06, 'num_tokens': 6420507.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00012123679334763438, 'sampling/sampling_logp_difference/max': 0.008353465236723423, 'sampling/importance_sampling_ratio/min': 0.9991452097892761, 'sampling/importance_sampling_ratio/mean': 1.0001118183135986, 'sampling/importance_sampling_ratio/max': 1.0083885192871094, 'entropy': 0.0012116766738472506, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2828/3300 [4:12:10<39:26,  5.01s/it]INFO 04-12 02:49:01 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2829/3300 [4:12:14<39:16,  5.00s/it]2026-04-12 02:49:06,021 | INFO | train_grpo_train | metrics_logged mode=train step=2829

                                                     
{'loss': 0.0003, 'grad_norm': 1.4268546104431152, 'learning_rate': 1.4303030303030306e-06, 'num_tokens': 6422684.0, 'completions/mean_length': 93.125, 'completions/min_length': 93.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.125, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9977059364318848, 'rewards/meter/std': 6.68626744300127e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977059364318848, 'rewards/total_composite/std': 6.68626744300127e-05, 'reward': 0.9977059364318848, 'reward_std': 6.687968561891466e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0164635069668293, 'sampling/sampling_logp_difference/max': 1.0106797218322754, 'sampling/importance_sampling_ratio/min': 0.3639715313911438, 'sampling/importance_sampling_ratio/mean': 1.0083869695663452, 'sampling/importance_sampling_ratio/max': 1.625991702079773, 'entropy': 0.10609597992151976, 'clip_ratio/low_mean': 0.014784946106374264, 'clip_ratio/low_min': 0.014784946106374264, 'clip_ratio/high_mean': 0.0013297871919348836, 'clip_ratio/high_max': 0.0013297871919348836, 'clip_ratio/region_mean': 0.016114733298309147, 'reward_total_mean': 0.9977059364318848, 'reward_meter_mean': 0.9977059364318848, 'reward_meter_std': 6.68626744300127e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977059364318848, 'reward_total_composite_std': 6.68626744300127e-05, 'epoch': 0.11}

 86%|████████▌ | 2829/3300 [4:12:15<39:16,  5.00s/it]INFO 04-12 02:49:06 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2830/3300 [4:12:20<40:55,  5.22s/it]2026-04-12 02:49:11,769 | INFO | train_grpo_train | metrics_logged mode=train step=2830

                                                     
{'loss': -0.0018, 'grad_norm': 2.387721538543701, 'learning_rate': 1.4272727272727274e-06, 'num_tokens': 6425415.0, 'completions/mean_length': 158.375, 'completions/min_length': 157.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.375, 'completions/min_terminated_length': 157.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9977245330810547, 'rewards/meter/std': 0.0005517599638551474, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.9561665058135986, 'rewards/total_composite/std': 0.05763240531086922, 'reward': 0.9561665058135986, 'reward_std': 0.05763239413499832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01837637834250927, 'sampling/sampling_logp_difference/max': 1.528066635131836, 'sampling/importance_sampling_ratio/min': 0.21695472300052643, 'sampling/importance_sampling_ratio/mean': 1.0021953582763672, 'sampling/importance_sampling_ratio/max': 1.5251771211624146, 'entropy': 0.16988196410238743, 'clip_ratio/low_mean': 0.004751874483190477, 'clip_ratio/low_min': 0.004751874483190477, 'clip_ratio/high_mean': 0.010205571772530675, 'clip_ratio/high_max': 0.010205571772530675, 'clip_ratio/region_mean': 0.014957446255721152, 'reward_total_mean': 0.9561665058135986, 'reward_meter_mean': 0.9977245330810547, 'reward_meter_std': 0.0005517599638551474, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.9561665058135986, 'reward_total_composite_std': 0.05763240531086922, 'epoch': 0.11}

 86%|████████▌ | 2830/3300 [4:12:20<40:55,  5.22s/it]INFO 04-12 02:49:12 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2831/3300 [4:12:25<40:38,  5.20s/it]2026-04-12 02:49:16,932 | INFO | train_grpo_train | metrics_logged mode=train step=2831

                                                     
{'loss': -0.0056, 'grad_norm': 3.8205645084381104, 'learning_rate': 1.4242424242424244e-06, 'num_tokens': 6427666.0, 'completions/mean_length': 106.375, 'completions/min_length': 104.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.375, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.8774728775024414, 'rewards/meter/std': 0.3444782793521881, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8774728775024414, 'rewards/total_composite/std': 0.3444782793521881, 'reward': 0.8774728775024414, 'reward_std': 0.3444782495498657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01733158528804779, 'sampling/sampling_logp_difference/max': 0.9781675338745117, 'sampling/importance_sampling_ratio/min': 0.37599948048591614, 'sampling/importance_sampling_ratio/mean': 1.0037299394607544, 'sampling/importance_sampling_ratio/max': 1.4661322832107544, 'entropy': 0.1615914311259985, 'clip_ratio/low_mean': 0.0072115384973585606, 'clip_ratio/low_min': 0.0072115384973585606, 'clip_ratio/high_mean': 0.008177978219464421, 'clip_ratio/high_max': 0.008177978219464421, 'clip_ratio/region_mean': 0.015389516716822982, 'reward_total_mean': 0.8774728775024414, 'reward_meter_mean': 0.8774728775024414, 'reward_meter_std': 0.3444782793521881, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8774728775024414, 'reward_total_composite_std': 0.3444782793521881, 'epoch': 0.11}

 86%|████████▌ | 2831/3300 [4:12:25<40:38,  5.20s/it]INFO 04-12 02:49:17 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2832/3300 [4:12:30<39:13,  5.03s/it]2026-04-12 02:49:21,534 | INFO | train_grpo_train | metrics_logged mode=train step=2832

                                                     
{'loss': 0.0047, 'grad_norm': 0.9516367316246033, 'learning_rate': 1.4212121212121213e-06, 'num_tokens': 6429513.0, 'completions/mean_length': 79.875, 'completions/min_length': 79.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7240880727767944, 'rewards/meter/std': 0.014326409436762333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7240880727767944, 'rewards/total_composite/std': 0.014326409436762333, 'reward': 0.7240880727767944, 'reward_std': 0.014326424337923527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002547384472563863, 'sampling/sampling_logp_difference/max': 1.0646743774414062, 'sampling/importance_sampling_ratio/min': 0.3448401391506195, 'sampling/importance_sampling_ratio/mean': 0.9994467496871948, 'sampling/importance_sampling_ratio/max': 1.1789082288742065, 'entropy': 0.007047486782539636, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0015822785208001733, 'clip_ratio/high_max': 0.0015822785208001733, 'clip_ratio/region_mean': 0.0015822785208001733, 'reward_total_mean': 0.7240880727767944, 'reward_meter_mean': 0.7240880727767944, 'reward_meter_std': 0.014326409436762333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7240880727767944, 'reward_total_composite_std': 0.014326409436762333, 'epoch': 0.11}

 86%|████████▌ | 2832/3300 [4:12:30<39:13,  5.03s/it]INFO 04-12 02:49:21 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2833/3300 [4:12:34<37:12,  4.78s/it]2026-04-12 02:49:25,740 | INFO | train_grpo_train | metrics_logged mode=train step=2833

                                                     
{'loss': 0.0162, 'grad_norm': 4.6605939865112305, 'learning_rate': 1.4181818181818183e-06, 'num_tokens': 6430923.0, 'completions/mean_length': 35.25, 'completions/min_length': 34.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9975078105926514, 'rewards/meter/std': 0.0002991736400872469, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975078105926514, 'rewards/total_composite/std': 0.0002991736400872469, 'reward': 0.9975078105926514, 'reward_std': 0.00029916909988969564, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023440107703208923, 'sampling/sampling_logp_difference/max': 1.062718391418457, 'sampling/importance_sampling_ratio/min': 0.34551531076431274, 'sampling/importance_sampling_ratio/mean': 0.9931513667106628, 'sampling/importance_sampling_ratio/max': 1.2360522747039795, 'entropy': 0.12462522741407156, 'clip_ratio/low_mean': 0.003289473708719015, 'clip_ratio/low_min': 0.003289473708719015, 'clip_ratio/high_mean': 0.010135134682059288, 'clip_ratio/high_max': 0.010135134682059288, 'clip_ratio/region_mean': 0.013424608390778303, 'reward_total_mean': 0.9975078105926514, 'reward_meter_mean': 0.9975078105926514, 'reward_meter_std': 0.0002991736400872469, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975078105926514, 'reward_total_composite_std': 0.0002991736400872469, 'epoch': 0.11}

 86%|████████▌ | 2833/3300 [4:12:34<37:12,  4.78s/it]INFO 04-12 02:49:25 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2834/3300 [4:12:40<39:02,  5.03s/it]2026-04-12 02:49:31,338 | INFO | train_grpo_train | metrics_logged mode=train step=2834

                                                     
{'loss': 0.0036, 'grad_norm': 1.4452556371688843, 'learning_rate': 1.4151515151515151e-06, 'num_tokens': 6433732.0, 'completions/mean_length': 158.125, 'completions/min_length': 155.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.125, 'completions/min_terminated_length': 155.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.9938750267028809, 'rewards/meter/std': 0.01016020867973566, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.9522948265075684, 'rewards/total_composite/std': 0.05487677827477455, 'reward': 0.9522948265075684, 'reward_std': 0.054876767098903656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021103786304593086, 'sampling/sampling_logp_difference/max': 1.3641910552978516, 'sampling/importance_sampling_ratio/min': 0.2555873394012451, 'sampling/importance_sampling_ratio/mean': 1.001399040222168, 'sampling/importance_sampling_ratio/max': 1.7436052560806274, 'entropy': 0.19695778377354145, 'clip_ratio/low_mean': 0.0023785202065482736, 'clip_ratio/low_min': 0.0023785202065482736, 'clip_ratio/high_mean': 0.010262453462928534, 'clip_ratio/high_max': 0.010262453462928534, 'clip_ratio/region_mean': 0.012640973669476807, 'reward_total_mean': 0.9522948265075684, 'reward_meter_mean': 0.9938750267028809, 'reward_meter_std': 0.01016020867973566, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.9522948265075684, 'reward_total_composite_std': 0.05487677827477455, 'epoch': 0.11}

 86%|████████▌ | 2834/3300 [4:12:40<39:02,  5.03s/it]INFO 04-12 02:49:31 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2835/3300 [4:12:44<37:35,  4.85s/it]2026-04-12 02:49:35,774 | INFO | train_grpo_train | metrics_logged mode=train step=2835

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.4121212121212122e-06, 'num_tokens': 6435500.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00019080200581811368, 'sampling/sampling_logp_difference/max': 0.009624381549656391, 'sampling/importance_sampling_ratio/min': 0.9914776682853699, 'sampling/importance_sampling_ratio/mean': 1.0001351833343506, 'sampling/importance_sampling_ratio/max': 1.0096708536148071, 'entropy': 0.0018191997078247368, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2835/3300 [4:12:44<37:35,  4.85s/it]INFO 04-12 02:49:36 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2836/3300 [4:12:49<37:31,  4.85s/it]2026-04-12 02:49:40,631 | INFO | train_grpo_train | metrics_logged mode=train step=2836

                                                     
{'loss': -0.0092, 'grad_norm': 3.5270473957061768, 'learning_rate': 1.409090909090909e-06, 'num_tokens': 6437557.0, 'completions/mean_length': 97.125, 'completions/min_length': 94.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.125, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9715797901153564, 'rewards/meter/std': 0.03168648108839989, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9715797901153564, 'rewards/total_composite/std': 0.03168648108839989, 'reward': 0.9715797901153564, 'reward_std': 0.03168648108839989, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047103382647037506, 'sampling/sampling_logp_difference/max': 1.1195144653320312, 'sampling/importance_sampling_ratio/min': 0.3264382481575012, 'sampling/importance_sampling_ratio/mean': 1.0144469738006592, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4101872891187668, 'clip_ratio/low_mean': 0.006538120680488646, 'clip_ratio/low_min': 0.006538120680488646, 'clip_ratio/high_mean': 0.014033293817192316, 'clip_ratio/high_max': 0.014033293817192316, 'clip_ratio/region_mean': 0.020571414497680962, 'reward_total_mean': 0.9715797901153564, 'reward_meter_mean': 0.9715797901153564, 'reward_meter_std': 0.03168648108839989, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9715797901153564, 'reward_total_composite_std': 0.03168648108839989, 'epoch': 0.11}

 86%|████████▌ | 2836/3300 [4:12:49<37:31,  4.85s/it]INFO 04-12 02:49:40 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2837/3300 [4:12:55<39:05,  5.07s/it]2026-04-12 02:49:46,193 | INFO | train_grpo_train | metrics_logged mode=train step=2837

                                                     
{'loss': 0.0116, 'grad_norm': 2.079005002975464, 'learning_rate': 1.406060606060606e-06, 'num_tokens': 6440330.0, 'completions/mean_length': 155.625, 'completions/min_length': 151.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 155.625, 'completions/min_terminated_length': 151.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.9973591566085815, 'rewards/meter/std': 0.001539723016321659, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9696353077888489, 'rewards/total_composite/std': 0.05092417448759079, 'reward': 0.9696353077888489, 'reward_std': 0.05092417448759079, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0309575367718935, 'sampling/sampling_logp_difference/max': 2.1312127113342285, 'sampling/importance_sampling_ratio/min': 0.1186932623386383, 'sampling/importance_sampling_ratio/mean': 1.0057966709136963, 'sampling/importance_sampling_ratio/max': 1.7507938146591187, 'entropy': 0.23529152572155, 'clip_ratio/low_mean': 0.0031847134232521057, 'clip_ratio/low_min': 0.0031847134232521057, 'clip_ratio/high_mean': 0.02511692512780428, 'clip_ratio/high_max': 0.02511692512780428, 'clip_ratio/region_mean': 0.028301638551056385, 'reward_total_mean': 0.9696353077888489, 'reward_meter_mean': 0.9973591566085815, 'reward_meter_std': 0.001539723016321659, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9696353077888489, 'reward_total_composite_std': 0.05092417448759079, 'epoch': 0.11}

 86%|████████▌ | 2837/3300 [4:12:55<39:05,  5.07s/it]INFO 04-12 02:49:46 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2838/3300 [4:13:00<39:48,  5.17s/it]2026-04-12 02:49:51,610 | INFO | train_grpo_train | metrics_logged mode=train step=2838

                                                     
{'loss': -0.0031, 'grad_norm': 2.711418628692627, 'learning_rate': 1.403030303030303e-06, 'num_tokens': 6442782.0, 'completions/mean_length': 134.5, 'completions/min_length': 132.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.5, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.999076247215271, 'rewards/meter/std': 0.00021775417553726584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999076247215271, 'rewards/total_composite/std': 0.00021775417553726584, 'reward': 0.999076247215271, 'reward_std': 0.00021775579079985619, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044978611171245575, 'sampling/sampling_logp_difference/max': 1.342717170715332, 'sampling/importance_sampling_ratio/min': 0.26113519072532654, 'sampling/importance_sampling_ratio/mean': 1.0005239248275757, 'sampling/importance_sampling_ratio/max': 1.565982460975647, 'entropy': 0.37597112730145454, 'clip_ratio/low_mean': 0.023296055383980274, 'clip_ratio/low_min': 0.023296055383980274, 'clip_ratio/high_mean': 0.012895984342321754, 'clip_ratio/high_max': 0.012895984342321754, 'clip_ratio/region_mean': 0.03619203972630203, 'reward_total_mean': 0.999076247215271, 'reward_meter_mean': 0.999076247215271, 'reward_meter_std': 0.00021775417553726584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999076247215271, 'reward_total_composite_std': 0.00021775417553726584, 'epoch': 0.11}

 86%|████████▌ | 2838/3300 [4:13:00<39:48,  5.17s/it]INFO 04-12 02:49:51 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2839/3300 [4:13:04<38:09,  4.97s/it]2026-04-12 02:49:56,101 | INFO | train_grpo_train | metrics_logged mode=train step=2839

                                                     
{'loss': 0.0133, 'grad_norm': 4.530300140380859, 'learning_rate': 1.4000000000000001e-06, 'num_tokens': 6444599.0, 'completions/mean_length': 66.125, 'completions/min_length': 65.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9489889144897461, 'rewards/meter/std': 0.003239382989704609, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9253357648849487, 'rewards/total_composite/std': 0.06812429428100586, 'reward': 0.9253357648849487, 'reward_std': 0.06812427937984467, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039003822952508926, 'sampling/sampling_logp_difference/max': 1.4666242599487305, 'sampling/importance_sampling_ratio/min': 0.2307029664516449, 'sampling/importance_sampling_ratio/mean': 1.0046266317367554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19998472929000854, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.0361340157687664, 'clip_ratio/high_max': 0.0361340157687664, 'clip_ratio/region_mean': 0.04370977357029915, 'reward_total_mean': 0.9253357648849487, 'reward_meter_mean': 0.9489889144897461, 'reward_meter_std': 0.003239382989704609, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9253357648849487, 'reward_total_composite_std': 0.06812429428100586, 'epoch': 0.11}

 86%|████████▌ | 2839/3300 [4:13:05<38:09,  4.97s/it]INFO 04-12 02:49:56 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2840/3300 [4:13:09<37:57,  4.95s/it]2026-04-12 02:50:01,019 | INFO | train_grpo_train | metrics_logged mode=train step=2840

                                                     
{'loss': 0.001, 'grad_norm': 0.9888406991958618, 'learning_rate': 1.3969696969696972e-06, 'num_tokens': 6446736.0, 'completions/mean_length': 98.125, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.125, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9993867874145508, 'rewards/meter/std': 5.2988518291385844e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993867874145508, 'rewards/total_composite/std': 5.2988518291385844e-05, 'reward': 0.9993867874145508, 'reward_std': 5.3000439947936684e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009926079772412777, 'sampling/sampling_logp_difference/max': 1.117924690246582, 'sampling/importance_sampling_ratio/min': 0.326957643032074, 'sampling/importance_sampling_ratio/mean': 1.0002104043960571, 'sampling/importance_sampling_ratio/max': 1.6428247690200806, 'entropy': 0.04481238219887018, 'clip_ratio/low_mean': 0.0038265305338427424, 'clip_ratio/low_min': 0.0038265305338427424, 'clip_ratio/high_mean': 0.005076272878795862, 'clip_ratio/high_max': 0.005076272878795862, 'clip_ratio/region_mean': 0.008902803412638605, 'reward_total_mean': 0.9993867874145508, 'reward_meter_mean': 0.9993867874145508, 'reward_meter_std': 5.2988518291385844e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993867874145508, 'reward_total_composite_std': 5.2988518291385844e-05, 'epoch': 0.11}

 86%|████████▌ | 2840/3300 [4:13:10<37:57,  4.95s/it]INFO 04-12 02:50:01 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2841/3300 [4:13:14<36:28,  4.77s/it]2026-04-12 02:50:05,354 | INFO | train_grpo_train | metrics_logged mode=train step=2841

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.3939393939393942e-06, 'num_tokens': 6448576.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00023380780476145446, 'sampling/sampling_logp_difference/max': 0.0026471014134585857, 'sampling/importance_sampling_ratio/min': 0.9991210103034973, 'sampling/importance_sampling_ratio/mean': 1.000229001045227, 'sampling/importance_sampling_ratio/max': 1.0026506185531616, 'entropy': 0.001806725820642896, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2841/3300 [4:13:14<36:28,  4.77s/it]INFO 04-12 02:50:05 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2842/3300 [4:13:18<35:26,  4.64s/it]2026-04-12 02:50:09,713 | INFO | train_grpo_train | metrics_logged mode=train step=2842

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.390909090909091e-06, 'num_tokens': 6450400.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00023029858130030334, 'sampling/sampling_logp_difference/max': 0.003564245533198118, 'sampling/importance_sampling_ratio/min': 0.9999062418937683, 'sampling/importance_sampling_ratio/mean': 1.0002295970916748, 'sampling/importance_sampling_ratio/max': 1.0035706758499146, 'entropy': 0.002088540146360174, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2842/3300 [4:13:18<35:26,  4.64s/it]INFO 04-12 02:50:09 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2843/3300 [4:13:23<34:49,  4.57s/it]2026-04-12 02:50:14,118 | INFO | train_grpo_train | metrics_logged mode=train step=2843

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.3878787878787881e-06, 'num_tokens': 6452232.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002246771182399243, 'sampling/sampling_logp_difference/max': 0.0022279510740190744, 'sampling/importance_sampling_ratio/min': 0.999897837638855, 'sampling/importance_sampling_ratio/mean': 1.0002238750457764, 'sampling/importance_sampling_ratio/max': 1.0022305250167847, 'entropy': 0.0019821640598820522, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2843/3300 [4:13:23<34:49,  4.57s/it]INFO 04-12 02:50:14 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2844/3300 [4:13:27<34:08,  4.49s/it]2026-04-12 02:50:18,429 | INFO | train_grpo_train | metrics_logged mode=train step=2844

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.384848484848485e-06, 'num_tokens': 6453936.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00025937738246284425, 'sampling/sampling_logp_difference/max': 0.003465580753982067, 'sampling/importance_sampling_ratio/min': 0.9998733997344971, 'sampling/importance_sampling_ratio/mean': 1.0002586841583252, 'sampling/importance_sampling_ratio/max': 1.0034716129302979, 'entropy': 0.0018740687082754448, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2844/3300 [4:13:27<34:08,  4.49s/it]INFO 04-12 02:50:18 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2845/3300 [4:13:31<33:34,  4.43s/it]2026-04-12 02:50:22,697 | INFO | train_grpo_train | metrics_logged mode=train step=2845

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.381818181818182e-06, 'num_tokens': 6455448.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0005941917770542204, 'sampling/sampling_logp_difference/max': 0.01501617580652237, 'sampling/importance_sampling_ratio/min': 0.9850959777832031, 'sampling/importance_sampling_ratio/mean': 1.000425934791565, 'sampling/importance_sampling_ratio/max': 1.008884072303772, 'entropy': 0.004738863033708185, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▌ | 2845/3300 [4:13:31<33:34,  4.43s/it]INFO 04-12 02:50:22 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▌ | 2846/3300 [4:13:37<36:40,  4.85s/it]2026-04-12 02:50:28,522 | INFO | train_grpo_train | metrics_logged mode=train step=2846

                                                     
{'loss': 0.0054, 'grad_norm': 2.306335926055908, 'learning_rate': 1.3787878787878788e-06, 'num_tokens': 6458571.0, 'completions/mean_length': 178.375, 'completions/min_length': 178.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.375, 'completions/min_terminated_length': 178.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.9989786148071289, 'rewards/meter/std': 0.00010886832023970783, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.9434767961502075, 'rewards/total_composite/std': 0.05927610024809837, 'reward': 0.9434767961502075, 'reward_std': 0.05927610024809837, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01770571805536747, 'sampling/sampling_logp_difference/max': 1.0042657852172852, 'sampling/importance_sampling_ratio/min': 0.36631348729133606, 'sampling/importance_sampling_ratio/mean': 1.0038070678710938, 'sampling/importance_sampling_ratio/max': 1.79298734664917, 'entropy': 0.18307929299771786, 'clip_ratio/low_mean': 0.007716872845776379, 'clip_ratio/low_min': 0.007716872845776379, 'clip_ratio/high_mean': 0.008395581040531397, 'clip_ratio/high_max': 0.008395581040531397, 'clip_ratio/region_mean': 0.016112453886307776, 'reward_total_mean': 0.9434767961502075, 'reward_meter_mean': 0.9989786148071289, 'reward_meter_std': 0.00010886832023970783, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.9434767961502075, 'reward_total_composite_std': 0.05927610024809837, 'epoch': 0.11}

 86%|████████▌ | 2846/3300 [4:13:37<36:40,  4.85s/it]INFO 04-12 02:50:28 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▋ | 2847/3300 [4:13:42<36:06,  4.78s/it]2026-04-12 02:50:33,154 | INFO | train_grpo_train | metrics_logged mode=train step=2847

                                                     
{'loss': -0.0027, 'grad_norm': 2.3205626010894775, 'learning_rate': 1.3757575757575759e-06, 'num_tokens': 6460630.0, 'completions/mean_length': 80.375, 'completions/min_length': 79.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9975742101669312, 'rewards/meter/std': 0.002095119096338749, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975742101669312, 'rewards/total_composite/std': 0.002095119096338749, 'reward': 0.9975742101669312, 'reward_std': 0.0020951295737177134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034449756145477295, 'sampling/sampling_logp_difference/max': 1.1230812072753906, 'sampling/importance_sampling_ratio/min': 0.3252760171890259, 'sampling/importance_sampling_ratio/mean': 1.0142433643341064, 'sampling/importance_sampling_ratio/max': 1.7717058658599854, 'entropy': 0.2891568746417761, 'clip_ratio/low_mean': 0.01566602219827473, 'clip_ratio/low_min': 0.01566602219827473, 'clip_ratio/high_mean': 0.016998398234136403, 'clip_ratio/high_max': 0.016998398234136403, 'clip_ratio/region_mean': 0.032664420432411134, 'reward_total_mean': 0.9975742101669312, 'reward_meter_mean': 0.9975742101669312, 'reward_meter_std': 0.002095119096338749, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975742101669312, 'reward_total_composite_std': 0.002095119096338749, 'epoch': 0.11}

 86%|████████▋ | 2847/3300 [4:13:42<36:06,  4.78s/it]INFO 04-12 02:50:33 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▋ | 2848/3300 [4:13:46<34:18,  4.55s/it]2026-04-12 02:50:37,200 | INFO | train_grpo_train | metrics_logged mode=train step=2848

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.3727272727272727e-06, 'num_tokens': 6462110.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0006461237207986414, 'sampling/sampling_logp_difference/max': 0.029664982110261917, 'sampling/importance_sampling_ratio/min': 0.9707706570625305, 'sampling/importance_sampling_ratio/mean': 1.0002989768981934, 'sampling/importance_sampling_ratio/max': 1.017333745956421, 'entropy': 0.005831975577166304, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▋ | 2848/3300 [4:13:46<34:18,  4.55s/it]INFO 04-12 02:50:37 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▋ | 2849/3300 [4:13:51<36:16,  4.83s/it]2026-04-12 02:50:42,635 | INFO | train_grpo_train | metrics_logged mode=train step=2849

                                                     
{'loss': -0.0004, 'grad_norm': 2.436666965484619, 'learning_rate': 1.3696969696969697e-06, 'num_tokens': 6464768.0, 'completions/mean_length': 142.25, 'completions/min_length': 140.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.25, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9991288781166077, 'rewards/meter/std': 0.0001261577708646655, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8928571939468384, 'rewards/repeat_penalty/std': 0.10101524740457535, 'rewards/total_composite/mean': 0.8920784592628479, 'rewards/total_composite/std': 0.10092534869909286, 'reward': 0.8920784592628479, 'reward_std': 0.10092535614967346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019777992740273476, 'sampling/sampling_logp_difference/max': 1.4602470397949219, 'sampling/importance_sampling_ratio/min': 0.2321789264678955, 'sampling/importance_sampling_ratio/mean': 1.0004011392593384, 'sampling/importance_sampling_ratio/max': 1.7561588287353516, 'entropy': 0.15198146365582943, 'clip_ratio/low_mean': 0.005313260713592172, 'clip_ratio/low_min': 0.005313260713592172, 'clip_ratio/high_mean': 0.0078856002073735, 'clip_ratio/high_max': 0.0078856002073735, 'clip_ratio/region_mean': 0.013198860920965672, 'reward_total_mean': 0.8920784592628479, 'reward_meter_mean': 0.9991288781166077, 'reward_meter_std': 0.0001261577708646655, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8928571939468384, 'reward_repeat_penalty_std': 0.10101524740457535, 'reward_total_composite_mean': 0.8920784592628479, 'reward_total_composite_std': 0.10092534869909286, 'epoch': 0.11}

 86%|████████▋ | 2849/3300 [4:13:51<36:16,  4.83s/it]INFO 04-12 02:50:42 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▋ | 2850/3300 [4:13:55<34:50,  4.65s/it]2026-04-12 02:50:46,855 | INFO | train_grpo_train | metrics_logged mode=train step=2850

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.3666666666666668e-06, 'num_tokens': 6466328.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002724671212490648, 'sampling/sampling_logp_difference/max': 0.0029080850072205067, 'sampling/importance_sampling_ratio/min': 0.999262273311615, 'sampling/importance_sampling_ratio/mean': 1.0002659559249878, 'sampling/importance_sampling_ratio/max': 1.0029124021530151, 'entropy': 0.002160783580620773, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 86%|████████▋ | 2850/3300 [4:13:55<34:50,  4.65s/it]INFO 04-12 02:50:47 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:38,  3.54s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.42s/it]

 31%|███       | 4/13 [00:16<00:39,  4.38s/it]

 38%|███▊      | 5/13 [00:21<00:37,  4.64s/it]

 46%|████▌     | 6/13 [00:28<00:38,  5.45s/it]

 54%|█████▍    | 7/13 [00:33<00:32,  5.39s/it]

 62%|██████▏   | 8/13 [00:40<00:28,  5.80s/it]

 69%|██████▉   | 9/13 [00:47<00:25,  6.27s/it]

 77%|███████▋  | 10/13 [00:52<00:17,  5.91s/it]

 85%|████████▍ | 11/13 [00:58<00:11,  5.79s/it]

 92%|█████████▏| 12/13 [01:04<00:05,  5.87s/it]

100%|██████████| 13/13 [01:09<00:00,  5.73s/it]2026-04-12 02:52:04,188 | INFO | train_grpo_train | metrics_logged mode=eval step=2850

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 77.18, 'eval_samples_per_second': 1.347, 'eval_steps_per_second': 0.168, 'eval_num_tokens': 6466328.0, 'eval_completions/mean_length': 211.08653846153845, 'eval_completions/min_length': 59.92307692307692, 'eval_completions/max_length': 411.38461538461536, 'eval_completions/clipped_ratio': 0.028846153846153848, 'eval_completions/mean_terminated_length': 201.30357360839844, 'eval_completions/min_terminated_length': 59.92307692307692, 'eval_completions/max_terminated_length': 386.0769230769231, 'eval_rewards/meter/mean': 0.7710461295568026, 'eval_rewards/meter/std': 0.3627813183344327, 'eval_rewards/count_adherence/mean': 0.9523772459763747, 'eval_rewards/count_adherence/std': 0.06613255492769755, 'eval_rewards/arabic_clean/mean': 0.9615384615384616, 'eval_rewards/arabic_clean/std': 0.10878565678229699, 'eval_rewards/repeat_penalty/mean': 0.9531235878284161, 'eval_rewards/repeat_penalty/std': 0.07651243057961647, 'eval_rewards/total_composite/mean': 0.6904892875598028, 'eval_rewards/total_composite/std': 0.3804844663693355, 'eval_reward': 0.6904892875598028, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03192363708065106, 'eval_sampling/sampling_logp_difference/max': 1.119479619539701, 'eval_sampling/importance_sampling_ratio/min': 0.33661178098275113, 'eval_sampling/importance_sampling_ratio/mean': 1.009327219082759, 'eval_sampling/importance_sampling_ratio/max': 1.5361342888612013, 'eval_entropy': 0.3772245920621432, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.6904892875598028, 'eval_reward_meter_mean': 0.7710461295568026, 'eval_reward_meter_std': 0.3627813183344327, 'eval_reward_count_adherence_mean': 0.9523772459763747, 'eval_reward_count_adherence_std': 0.06613255492769755, 'eval_reward_arabic_clean_mean': 0.9615384615384616, 'eval_reward_arabic_clean_std': 0.10878565678229699, 'eval_reward_repeat_penalty_mean': 0.9531235878284161, 'eval_reward_repeat_penalty_std': 0.07651243057961647, 'eval_reward_total_composite_mean': 0.6904892875598028, 'eval_reward_total_composite_std': 0.3804844663693355, 'epoch': 0.11}

 86%|████████▋ | 2850/3300 [4:15:13<34:50,  4.65s/it]

100%|██████████| 13/13 [01:10<00:00,  5.73s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 86%|████████▋ | 2851/3300 [4:15:23<3:41:04, 29.54s/it]2026-04-12 02:52:14,502 | INFO | train_grpo_train | metrics_logged mode=train step=2851

                                                       
{'loss': 0.0037, 'grad_norm': 1.7204653024673462, 'learning_rate': 1.3636363636363636e-06, 'num_tokens': 6469974.0, 'completions/mean_length': 248.75, 'completions/min_length': 248.0, 'completions/max_length': 250.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 248.75, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 250.0, 'rewards/meter/mean': 0.9985492825508118, 'rewards/meter/std': 0.000639773381408304, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9230769276618958, 'rewards/repeat_penalty/std': 0.07121692597866058, 'rewards/total_composite/mean': 0.9217178821563721, 'rewards/total_composite/std': 0.0708138644695282, 'reward': 0.9217178821563721, 'reward_std': 0.070813849568367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024417459964752197, 'sampling/sampling_logp_difference/max': 1.4727954864501953, 'sampling/importance_sampling_ratio/min': 0.2292836457490921, 'sampling/importance_sampling_ratio/mean': 1.0041165351867676, 'sampling/importance_sampling_ratio/max': 1.587019920349121, 'entropy': 0.24152424931526184, 'clip_ratio/low_mean': 0.004026185255497694, 'clip_ratio/low_min': 0.004026185255497694, 'clip_ratio/high_mean': 0.010042233159765601, 'clip_ratio/high_max': 0.010042233159765601, 'clip_ratio/region_mean': 0.014068418415263295, 'reward_total_mean': 0.9217178821563721, 'reward_meter_mean': 0.9985492825508118, 'reward_meter_std': 0.000639773381408304, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9230769276618958, 'reward_repeat_penalty_std': 0.07121692597866058, 'reward_total_composite_mean': 0.9217178821563721, 'reward_total_composite_std': 0.0708138644695282, 'epoch': 0.11}

 86%|████████▋ | 2851/3300 [4:15:23<3:41:04, 29.54s/it]INFO 04-12 02:52:14 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▋ | 2852/3300 [4:15:28<2:45:00, 22.10s/it]2026-04-12 02:52:19,238 | INFO | train_grpo_train | metrics_logged mode=train step=2852

                                                       
{'loss': -0.0252, 'grad_norm': 8.037017822265625, 'learning_rate': 1.3606060606060607e-06, 'num_tokens': 6471754.0, 'completions/mean_length': 65.5, 'completions/min_length': 60.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.951325535774231, 'rewards/meter/std': 0.1125522255897522, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.951325535774231, 'rewards/total_composite/std': 0.1125522255897522, 'reward': 0.951325535774231, 'reward_std': 0.112552210688591, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042092662304639816, 'sampling/sampling_logp_difference/max': 1.1297359466552734, 'sampling/importance_sampling_ratio/min': 0.3231185972690582, 'sampling/importance_sampling_ratio/mean': 1.0115268230438232, 'sampling/importance_sampling_ratio/max': 1.7486579418182373, 'entropy': 0.38093627616763115, 'clip_ratio/low_mean': 0.008333333767950535, 'clip_ratio/low_min': 0.008333333767950535, 'clip_ratio/high_mean': 0.014985310845077038, 'clip_ratio/high_max': 0.014985310845077038, 'clip_ratio/region_mean': 0.023318644613027573, 'reward_total_mean': 0.951325535774231, 'reward_meter_mean': 0.951325535774231, 'reward_meter_std': 0.1125522255897522, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.951325535774231, 'reward_total_composite_std': 0.1125522255897522, 'epoch': 0.11}

 86%|████████▋ | 2852/3300 [4:15:28<2:45:00, 22.10s/it]INFO 04-12 02:52:19 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▋ | 2853/3300 [4:15:32<2:05:54, 16.90s/it]2026-04-12 02:52:24,006 | INFO | train_grpo_train | metrics_logged mode=train step=2853

                                                       
{'loss': 0.0008, 'grad_norm': 2.253185510635376, 'learning_rate': 1.357575757575758e-06, 'num_tokens': 6473577.0, 'completions/mean_length': 79.875, 'completions/min_length': 79.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9988980293273926, 'rewards/meter/std': 0.0002794552710838616, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988980293273926, 'rewards/total_composite/std': 0.0002794552710838616, 'reward': 0.9988980293273926, 'reward_std': 0.0002794554748106748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027408134192228317, 'sampling/sampling_logp_difference/max': 0.9639253616333008, 'sampling/importance_sampling_ratio/min': 0.38139283657073975, 'sampling/importance_sampling_ratio/mean': 1.0052258968353271, 'sampling/importance_sampling_ratio/max': 1.6166248321533203, 'entropy': 0.2539651710540056, 'clip_ratio/low_mean': 0.004707278567366302, 'clip_ratio/low_min': 0.004707278567366302, 'clip_ratio/high_mean': 0.0204123689327389, 'clip_ratio/high_max': 0.0204123689327389, 'clip_ratio/region_mean': 0.025119647500105202, 'reward_total_mean': 0.9988980293273926, 'reward_meter_mean': 0.9988980293273926, 'reward_meter_std': 0.0002794552710838616, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988980293273926, 'reward_total_composite_std': 0.0002794552710838616, 'epoch': 0.11}

 86%|████████▋ | 2853/3300 [4:15:33<2:05:54, 16.90s/it]INFO 04-12 02:52:24 [block_pool.py:378] Successfully reset prefix cache

 86%|████████▋ | 2854/3300 [4:15:38<1:39:39, 13.41s/it]2026-04-12 02:52:29,265 | INFO | train_grpo_train | metrics_logged mode=train step=2854

                                                       
{'loss': -0.0055, 'grad_norm': 4.1874613761901855, 'learning_rate': 1.3545454545454547e-06, 'num_tokens': 6475980.0, 'completions/mean_length': 126.375, 'completions/min_length': 124.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.375, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9470775127410889, 'rewards/meter/std': 0.12438784539699554, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.929327130317688, 'rewards/total_composite/std': 0.12683995068073273, 'reward': 0.929327130317688, 'reward_std': 0.12683993577957153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0373864509165287, 'sampling/sampling_logp_difference/max': 1.086287498474121, 'sampling/importance_sampling_ratio/min': 0.3374670445919037, 'sampling/importance_sampling_ratio/mean': 1.0064387321472168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34479158371686935, 'clip_ratio/low_mean': 0.004032257944345474, 'clip_ratio/low_min': 0.004032257944345474, 'clip_ratio/high_mean': 0.018657967331819236, 'clip_ratio/high_max': 0.018657967331819236, 'clip_ratio/region_mean': 0.02269022527616471, 'reward_total_mean': 0.929327130317688, 'reward_meter_mean': 0.9470775127410889, 'reward_meter_std': 0.12438784539699554, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.929327130317688, 'reward_total_composite_std': 0.12683995068073273, 'epoch': 0.11}

 86%|████████▋ | 2854/3300 [4:15:38<1:39:39, 13.41s/it]INFO 04-12 02:52:29 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2855/3300 [4:15:42<1:19:30, 10.72s/it]2026-04-12 02:52:33,706 | INFO | train_grpo_train | metrics_logged mode=train step=2855

                                                       
{'loss': 0.0057, 'grad_norm': 5.475809097290039, 'learning_rate': 1.3515151515151518e-06, 'num_tokens': 6477692.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9972702860832214, 'rewards/meter/std': 0.00018694026221055537, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972702860832214, 'rewards/total_composite/std': 0.00018694026221055537, 'reward': 0.9972702860832214, 'reward_std': 0.0001869339175755158, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01253997441381216, 'sampling/sampling_logp_difference/max': 0.8330123424530029, 'sampling/importance_sampling_ratio/min': 0.4347377419471741, 'sampling/importance_sampling_ratio/mean': 1.0026929378509521, 'sampling/importance_sampling_ratio/max': 1.7445498704910278, 'entropy': 0.06320370081812143, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.006147540640085936, 'clip_ratio/high_max': 0.006147540640085936, 'clip_ratio/region_mean': 0.008196720853447914, 'reward_total_mean': 0.9972702860832214, 'reward_meter_mean': 0.9972702860832214, 'reward_meter_std': 0.00018694026221055537, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972702860832214, 'reward_total_composite_std': 0.00018694026221055537, 'epoch': 0.11}

 87%|████████▋ | 2855/3300 [4:15:42<1:19:30, 10.72s/it]INFO 04-12 02:52:33 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2856/3300 [4:15:47<1:06:19,  8.96s/it]2026-04-12 02:52:38,569 | INFO | train_grpo_train | metrics_logged mode=train step=2856

                                                       
{'loss': -0.0073, 'grad_norm': 3.800494909286499, 'learning_rate': 1.3484848484848486e-06, 'num_tokens': 6479747.0, 'completions/mean_length': 100.875, 'completions/min_length': 98.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.875, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9850212931632996, 'rewards/meter/std': 0.02900601737201214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9850212931632996, 'rewards/total_composite/std': 0.02900601737201214, 'reward': 0.9850212931632996, 'reward_std': 0.02900601364672184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031728874891996384, 'sampling/sampling_logp_difference/max': 1.3605375289916992, 'sampling/importance_sampling_ratio/min': 0.25652286410331726, 'sampling/importance_sampling_ratio/mean': 1.0072658061981201, 'sampling/importance_sampling_ratio/max': 1.8992688655853271, 'entropy': 0.2935164403170347, 'clip_ratio/low_mean': 0.0012755101779475808, 'clip_ratio/low_min': 0.0012755101779475808, 'clip_ratio/high_mean': 0.02479799627326429, 'clip_ratio/high_max': 0.02479799627326429, 'clip_ratio/region_mean': 0.02607350645121187, 'reward_total_mean': 0.9850212931632996, 'reward_meter_mean': 0.9850212931632996, 'reward_meter_std': 0.02900601737201214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9850212931632996, 'reward_total_composite_std': 0.02900601737201214, 'epoch': 0.11}

 87%|████████▋ | 2856/3300 [4:15:47<1:06:19,  8.96s/it]INFO 04-12 02:52:38 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2857/3300 [4:15:51<56:05,  7.60s/it]  2026-04-12 02:52:42,974 | INFO | train_grpo_train | metrics_logged mode=train step=2857

                                                     
{'loss': 0.0001, 'grad_norm': 0.22853942215442657, 'learning_rate': 1.3454545454545457e-06, 'num_tokens': 6481515.0, 'completions/mean_length': 66.0, 'completions/min_length': 66.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9981458783149719, 'rewards/meter/std': 1.3051687346887775e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981458783149719, 'rewards/total_composite/std': 1.3051687346887775e-05, 'reward': 0.9981458783149719, 'reward_std': 1.3051648238615599e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005550025030970573, 'sampling/sampling_logp_difference/max': 0.3850884437561035, 'sampling/importance_sampling_ratio/min': 0.8009796142578125, 'sampling/importance_sampling_ratio/mean': 1.002258539199829, 'sampling/importance_sampling_ratio/max': 1.469744324684143, 'entropy': 0.053996546659618616, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.009469697251915932, 'clip_ratio/high_max': 0.009469697251915932, 'clip_ratio/region_mean': 0.009469697251915932, 'reward_total_mean': 0.9981458783149719, 'reward_meter_mean': 0.9981458783149719, 'reward_meter_std': 1.3051687346887775e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981458783149719, 'reward_total_composite_std': 1.3051687346887775e-05, 'epoch': 0.11}

 87%|████████▋ | 2857/3300 [4:15:51<56:05,  7.60s/it]INFO 04-12 02:52:43 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2858/3300 [4:15:59<55:14,  7.50s/it]2026-04-12 02:52:50,251 | INFO | train_grpo_train | metrics_logged mode=train step=2858

                                                     
{'loss': 0.0307, 'grad_norm': 2.1501612663269043, 'learning_rate': 1.3424242424242425e-06, 'num_tokens': 6485258.0, 'completions/mean_length': 279.875, 'completions/min_length': 261.0, 'completions/max_length': 298.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 279.875, 'completions/min_terminated_length': 261.0, 'completions/max_terminated_length': 298.0, 'rewards/meter/mean': 0.9832162857055664, 'rewards/meter/std': 0.0437212772667408, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8965686559677124, 'rewards/repeat_penalty/std': 0.05159881338477135, 'rewards/total_composite/mean': 0.8241852521896362, 'rewards/total_composite/std': 0.05549175664782524, 'reward': 0.8241852521896362, 'reward_std': 0.05549176037311554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03513329103589058, 'sampling/sampling_logp_difference/max': 2.449089527130127, 'sampling/importance_sampling_ratio/min': 0.08637219667434692, 'sampling/importance_sampling_ratio/mean': 1.0033777952194214, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2756977826356888, 'clip_ratio/low_mean': 0.014873409294523299, 'clip_ratio/low_min': 0.014873409294523299, 'clip_ratio/high_mean': 0.0062042842619121075, 'clip_ratio/high_max': 0.0062042842619121075, 'clip_ratio/region_mean': 0.021077693556435406, 'reward_total_mean': 0.8241852521896362, 'reward_meter_mean': 0.9832162857055664, 'reward_meter_std': 0.0437212772667408, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8965686559677124, 'reward_repeat_penalty_std': 0.05159881338477135, 'reward_total_composite_mean': 0.8241852521896362, 'reward_total_composite_std': 0.05549175664782524, 'epoch': 0.11}

 87%|████████▋ | 2858/3300 [4:15:59<55:14,  7.50s/it]INFO 04-12 02:52:50 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2859/3300 [4:16:03<48:26,  6.59s/it]2026-04-12 02:52:54,715 | INFO | train_grpo_train | metrics_logged mode=train step=2859

                                                     
{'loss': -0.0101, 'grad_norm': 3.014261245727539, 'learning_rate': 1.3393939393939395e-06, 'num_tokens': 6487078.0, 'completions/mean_length': 66.5, 'completions/min_length': 65.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9922887086868286, 'rewards/meter/std': 0.0029313687700778246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9922887086868286, 'rewards/total_composite/std': 0.0029313687700778246, 'reward': 0.9922887086868286, 'reward_std': 0.0029313701670616865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031141582876443863, 'sampling/sampling_logp_difference/max': 1.0435974597930908, 'sampling/importance_sampling_ratio/min': 0.3521854281425476, 'sampling/importance_sampling_ratio/mean': 1.0089040994644165, 'sampling/importance_sampling_ratio/max': 1.5588260889053345, 'entropy': 0.23347646743059158, 'clip_ratio/low_mean': 0.009471436496824026, 'clip_ratio/low_min': 0.009471436496824026, 'clip_ratio/high_mean': 0.01487299520522356, 'clip_ratio/high_max': 0.01487299520522356, 'clip_ratio/region_mean': 0.024344431702047586, 'reward_total_mean': 0.9922887086868286, 'reward_meter_mean': 0.9922887086868286, 'reward_meter_std': 0.0029313687700778246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9922887086868286, 'reward_total_composite_std': 0.0029313687700778246, 'epoch': 0.11}

 87%|████████▋ | 2859/3300 [4:16:03<48:26,  6.59s/it]INFO 04-12 02:52:54 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2860/3300 [4:16:08<43:33,  5.94s/it]2026-04-12 02:52:59,138 | INFO | train_grpo_train | metrics_logged mode=train step=2860

                                                     
{'loss': -0.0013, 'grad_norm': 0.3336952030658722, 'learning_rate': 1.3363636363636364e-06, 'num_tokens': 6488807.0, 'completions/mean_length': 66.125, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981107711791992, 'rewards/meter/std': 5.631321982946247e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981107711791992, 'rewards/total_composite/std': 5.631321982946247e-05, 'reward': 0.9981107711791992, 'reward_std': 5.631066233036108e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007993371225893497, 'sampling/sampling_logp_difference/max': 0.5051336288452148, 'sampling/importance_sampling_ratio/min': 0.6034249067306519, 'sampling/importance_sampling_ratio/mean': 1.004895567893982, 'sampling/importance_sampling_ratio/max': 1.383058786392212, 'entropy': 0.05899450369179249, 'clip_ratio/low_mean': 0.011363636702299118, 'clip_ratio/low_min': 0.011363636702299118, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.011363636702299118, 'reward_total_mean': 0.9981107711791992, 'reward_meter_mean': 0.9981107711791992, 'reward_meter_std': 5.631321982946247e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981107711791992, 'reward_total_composite_std': 5.631321982946247e-05, 'epoch': 0.11}

 87%|████████▋ | 2860/3300 [4:16:08<43:33,  5.94s/it]INFO 04-12 02:52:59 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2861/3300 [4:16:11<39:06,  5.35s/it]2026-04-12 02:53:03,097 | INFO | train_grpo_train | metrics_logged mode=train step=2861

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.3333333333333334e-06, 'num_tokens': 6490287.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 3.3981952583417296e-05, 'sampling/sampling_logp_difference/max': 0.0006292310426943004, 'sampling/importance_sampling_ratio/min': 0.9998278617858887, 'sampling/importance_sampling_ratio/mean': 1.0000324249267578, 'sampling/importance_sampling_ratio/max': 1.0006294250488281, 'entropy': 0.0002626157729537226, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.11}

 87%|████████▋ | 2861/3300 [4:16:12<39:06,  5.35s/it]INFO 04-12 02:53:03 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2862/3300 [4:16:18<40:49,  5.59s/it]2026-04-12 02:53:09,263 | INFO | train_grpo_train | metrics_logged mode=train step=2862

                                                     
{'loss': -0.0358, 'grad_norm': 3.7277348041534424, 'learning_rate': 1.3303030303030305e-06, 'num_tokens': 6493447.0, 'completions/mean_length': 195.0, 'completions/min_length': 173.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 195.0, 'completions/min_terminated_length': 173.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.9989534616470337, 'rewards/meter/std': 0.00048265265650115907, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9573421478271484, 'rewards/total_composite/std': 0.07722419500350952, 'reward': 0.9573421478271484, 'reward_std': 0.07722420245409012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05510193854570389, 'sampling/sampling_logp_difference/max': 1.9347803592681885, 'sampling/importance_sampling_ratio/min': 0.14445599913597107, 'sampling/importance_sampling_ratio/mean': 1.0075865983963013, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4370429702103138, 'clip_ratio/low_mean': 0.01037695212289691, 'clip_ratio/low_min': 0.01037695212289691, 'clip_ratio/high_mean': 0.040064468048512936, 'clip_ratio/high_max': 0.040064468048512936, 'clip_ratio/region_mean': 0.050441420171409845, 'reward_total_mean': 0.9573421478271484, 'reward_meter_mean': 0.9989534616470337, 'reward_meter_std': 0.00048265265650115907, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9573421478271484, 'reward_total_composite_std': 0.07722419500350952, 'epoch': 0.11}

 87%|████████▋ | 2862/3300 [4:16:18<40:49,  5.59s/it]INFO 04-12 02:53:09 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2863/3300 [4:16:22<38:11,  5.24s/it]2026-04-12 02:53:13,702 | INFO | train_grpo_train | metrics_logged mode=train step=2863

                                                     
{'loss': 0.0, 'grad_norm': 0.6136480569839478, 'learning_rate': 1.3272727272727273e-06, 'num_tokens': 6495258.0, 'completions/mean_length': 71.375, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994188547134399, 'rewards/meter/std': 3.3046217140508816e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994188547134399, 'rewards/total_composite/std': 3.3046217140508816e-05, 'reward': 0.9994188547134399, 'reward_std': 3.304608617327176e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01137758232653141, 'sampling/sampling_logp_difference/max': 1.0333290100097656, 'sampling/importance_sampling_ratio/min': 0.3558204770088196, 'sampling/importance_sampling_ratio/mean': 1.0004888772964478, 'sampling/importance_sampling_ratio/max': 1.2199125289916992, 'entropy': 0.07642534840852022, 'clip_ratio/low_mean': 0.005208333372138441, 'clip_ratio/low_min': 0.005208333372138441, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.005208333372138441, 'reward_total_mean': 0.9994188547134399, 'reward_meter_mean': 0.9994188547134399, 'reward_meter_std': 3.3046217140508816e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994188547134399, 'reward_total_composite_std': 3.3046217140508816e-05, 'epoch': 0.11}

 87%|████████▋ | 2863/3300 [4:16:22<38:11,  5.24s/it]INFO 04-12 02:53:13 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2864/3300 [4:16:28<38:53,  5.35s/it]2026-04-12 02:53:19,304 | INFO | train_grpo_train | metrics_logged mode=train step=2864

                                                     
{'loss': 0.0017, 'grad_norm': 2.3443167209625244, 'learning_rate': 1.3242424242424243e-06, 'num_tokens': 6498081.0, 'completions/mean_length': 164.875, 'completions/min_length': 162.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 164.875, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9993668794631958, 'rewards/meter/std': 0.0001238392578670755, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9861111044883728, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.9854899644851685, 'rewards/total_composite/std': 0.03934282064437866, 'reward': 0.9854899644851685, 'reward_std': 0.039342816919088364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024905560538172722, 'sampling/sampling_logp_difference/max': 1.7190942764282227, 'sampling/importance_sampling_ratio/min': 0.17922841012477875, 'sampling/importance_sampling_ratio/mean': 1.000923991203308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15149082150310278, 'clip_ratio/low_mean': 0.0030303029343485832, 'clip_ratio/low_min': 0.0030303029343485832, 'clip_ratio/high_mean': 0.013674213085323572, 'clip_ratio/high_max': 0.013674213085323572, 'clip_ratio/region_mean': 0.016704516019672155, 'reward_total_mean': 0.9854899644851685, 'reward_meter_mean': 0.9993668794631958, 'reward_meter_std': 0.0001238392578670755, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9861111044883728, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.9854899644851685, 'reward_total_composite_std': 0.03934282064437866, 'epoch': 0.12}

 87%|████████▋ | 2864/3300 [4:16:28<38:53,  5.35s/it]INFO 04-12 02:53:19 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2865/3300 [4:16:37<47:36,  6.57s/it]2026-04-12 02:53:28,700 | INFO | train_grpo_train | metrics_logged mode=train step=2865

                                                     
{'loss': -0.0115, 'grad_norm': 1.4304178953170776, 'learning_rate': 1.3212121212121212e-06, 'num_tokens': 6503361.0, 'completions/mean_length': 442.0, 'completions/min_length': 430.0, 'completions/max_length': 470.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 442.0, 'completions/min_terminated_length': 430.0, 'completions/max_terminated_length': 470.0, 'rewards/meter/mean': 0.9988847970962524, 'rewards/meter/std': 0.0004295332182664424, 'rewards/count_adherence/mean': 0.8035714030265808, 'rewards/count_adherence/std': 0.03306501731276512, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9593685269355774, 'rewards/repeat_penalty/std': 0.030346577987074852, 'rewards/total_composite/mean': 0.7700097560882568, 'rewards/total_composite/std': 0.03843219205737114, 'reward': 0.7700097560882568, 'reward_std': 0.03843218460679054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05211654305458069, 'sampling/sampling_logp_difference/max': 1.5292320251464844, 'sampling/importance_sampling_ratio/min': 0.21670202910900116, 'sampling/importance_sampling_ratio/mean': 1.010416030883789, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4746365137398243, 'clip_ratio/low_mean': 0.008053920813836157, 'clip_ratio/low_min': 0.008053920813836157, 'clip_ratio/high_mean': 0.019567076116800308, 'clip_ratio/high_max': 0.019567076116800308, 'clip_ratio/region_mean': 0.027620996930636466, 'reward_total_mean': 0.7700097560882568, 'reward_meter_mean': 0.9988847970962524, 'reward_meter_std': 0.0004295332182664424, 'reward_count_adherence_mean': 0.8035714030265808, 'reward_count_adherence_std': 0.03306501731276512, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9593685269355774, 'reward_repeat_penalty_std': 0.030346577987074852, 'reward_total_composite_mean': 0.7700097560882568, 'reward_total_composite_std': 0.03843219205737114, 'epoch': 0.12}

 87%|████████▋ | 2865/3300 [4:16:37<47:36,  6.57s/it]INFO 04-12 02:53:28 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2866/3300 [4:16:41<41:52,  5.79s/it]2026-04-12 02:53:32,679 | INFO | train_grpo_train | metrics_logged mode=train step=2866

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.3181818181818182e-06, 'num_tokens': 6504873.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0006467251223511994, 'sampling/sampling_logp_difference/max': 0.01518770307302475, 'sampling/importance_sampling_ratio/min': 0.9985290169715881, 'sampling/importance_sampling_ratio/mean': 1.0006215572357178, 'sampling/importance_sampling_ratio/max': 1.015303611755371, 'entropy': 0.0056428477400913835, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 87%|████████▋ | 2866/3300 [4:16:41<41:52,  5.79s/it]INFO 04-12 02:53:32 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2867/3300 [4:16:46<38:50,  5.38s/it]2026-04-12 02:53:37,106 | INFO | train_grpo_train | metrics_logged mode=train step=2867

                                                     
{'loss': 0.005, 'grad_norm': 4.305418968200684, 'learning_rate': 1.315151515151515e-06, 'num_tokens': 6506642.0, 'completions/mean_length': 64.125, 'completions/min_length': 64.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9991264343261719, 'rewards/meter/std': 0.0007721302099525928, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991264343261719, 'rewards/total_composite/std': 0.0007721302099525928, 'reward': 0.9991264343261719, 'reward_std': 0.0007721302681602538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00290035386569798, 'sampling/sampling_logp_difference/max': 1.102513313293457, 'sampling/importance_sampling_ratio/min': 0.33203551173210144, 'sampling/importance_sampling_ratio/mean': 0.9988252520561218, 'sampling/importance_sampling_ratio/max': 1.0192989110946655, 'entropy': 0.0063372578588314354, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9991264343261719, 'reward_meter_mean': 0.9991264343261719, 'reward_meter_std': 0.0007721302099525928, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991264343261719, 'reward_total_composite_std': 0.0007721302099525928, 'epoch': 0.12}

 87%|████████▋ | 2867/3300 [4:16:46<38:50,  5.38s/it]INFO 04-12 02:53:37 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2868/3300 [4:16:50<37:05,  5.15s/it]2026-04-12 02:53:41,726 | INFO | train_grpo_train | metrics_logged mode=train step=2868

                                                     
{'loss': -0.0079, 'grad_norm': 2.3818113803863525, 'learning_rate': 1.3121212121212123e-06, 'num_tokens': 6508612.0, 'completions/mean_length': 80.25, 'completions/min_length': 77.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.25, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9989343285560608, 'rewards/meter/std': 0.00022084804368205369, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989343285560608, 'rewards/total_composite/std': 0.00022084804368205369, 'reward': 0.9989343285560608, 'reward_std': 0.00022085075033828616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023430120199918747, 'sampling/sampling_logp_difference/max': 1.1872577667236328, 'sampling/importance_sampling_ratio/min': 0.3050566613674164, 'sampling/importance_sampling_ratio/mean': 1.0084359645843506, 'sampling/importance_sampling_ratio/max': 1.6601855754852295, 'entropy': 0.21316692978143692, 'clip_ratio/low_mean': 0.003246753243729472, 'clip_ratio/low_min': 0.003246753243729472, 'clip_ratio/high_mean': 0.006192129687406123, 'clip_ratio/high_max': 0.006192129687406123, 'clip_ratio/region_mean': 0.009438882931135595, 'reward_total_mean': 0.9989343285560608, 'reward_meter_mean': 0.9989343285560608, 'reward_meter_std': 0.00022084804368205369, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989343285560608, 'reward_total_composite_std': 0.00022084804368205369, 'epoch': 0.12}

 87%|████████▋ | 2868/3300 [4:16:50<37:05,  5.15s/it]INFO 04-12 02:53:41 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2869/3300 [4:16:55<36:21,  5.06s/it]2026-04-12 02:53:46,574 | INFO | train_grpo_train | metrics_logged mode=train step=2869

                                                     
{'loss': -0.0098, 'grad_norm': 2.163902997970581, 'learning_rate': 1.3090909090909093e-06, 'num_tokens': 6510609.0, 'completions/mean_length': 96.625, 'completions/min_length': 94.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.625, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9977865219116211, 'rewards/meter/std': 0.0006117028533481061, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977865219116211, 'rewards/total_composite/std': 0.0006117028533481061, 'reward': 0.9977865219116211, 'reward_std': 0.0006116984295658767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016453739255666733, 'sampling/sampling_logp_difference/max': 1.150710105895996, 'sampling/importance_sampling_ratio/min': 0.3164120018482208, 'sampling/importance_sampling_ratio/mean': 1.0011489391326904, 'sampling/importance_sampling_ratio/max': 1.840904712677002, 'entropy': 0.14486887026578188, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006416999618522823, 'clip_ratio/high_max': 0.006416999618522823, 'clip_ratio/region_mean': 0.006416999618522823, 'reward_total_mean': 0.9977865219116211, 'reward_meter_mean': 0.9977865219116211, 'reward_meter_std': 0.0006117028533481061, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977865219116211, 'reward_total_composite_std': 0.0006117028533481061, 'epoch': 0.12}

 87%|████████▋ | 2869/3300 [4:16:55<36:21,  5.06s/it]INFO 04-12 02:53:46 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2870/3300 [4:16:59<34:41,  4.84s/it]2026-04-12 02:53:50,902 | INFO | train_grpo_train | metrics_logged mode=train step=2870

                                                     
{'loss': -0.0, 'grad_norm': 1.1164172887802124, 'learning_rate': 1.3060606060606062e-06, 'num_tokens': 6512337.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973324537277222, 'rewards/meter/std': 2.97028473141836e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973324537277222, 'rewards/total_composite/std': 2.97028473141836e-05, 'reward': 0.9973324537277222, 'reward_std': 2.9693699616473168e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010174022056162357, 'sampling/sampling_logp_difference/max': 0.505859375, 'sampling/importance_sampling_ratio/min': 0.6083394289016724, 'sampling/importance_sampling_ratio/mean': 1.003736138343811, 'sampling/importance_sampling_ratio/max': 1.6584101915359497, 'entropy': 0.05785668967291713, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.0020491802133619785, 'reward_total_mean': 0.9973324537277222, 'reward_meter_mean': 0.9973324537277222, 'reward_meter_std': 2.97028473141836e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973324537277222, 'reward_total_composite_std': 2.97028473141836e-05, 'epoch': 0.12}

 87%|████████▋ | 2870/3300 [4:16:59<34:41,  4.84s/it]INFO 04-12 02:53:51 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2871/3300 [4:17:04<34:12,  4.78s/it]2026-04-12 02:53:55,551 | INFO | train_grpo_train | metrics_logged mode=train step=2871

                                                     
{'loss': 0.0021, 'grad_norm': 1.2812321186065674, 'learning_rate': 1.3030303030303032e-06, 'num_tokens': 6514415.0, 'completions/mean_length': 79.75, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.75, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9988850355148315, 'rewards/meter/std': 0.00015237655316013843, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988850355148315, 'rewards/total_composite/std': 0.00015237655316013843, 'reward': 0.9988850355148315, 'reward_std': 0.00015238845662679523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02608993463218212, 'sampling/sampling_logp_difference/max': 0.9019899368286133, 'sampling/importance_sampling_ratio/min': 0.4057614207267761, 'sampling/importance_sampling_ratio/mean': 1.0047295093536377, 'sampling/importance_sampling_ratio/max': 1.7389991283416748, 'entropy': 0.22218100167810917, 'clip_ratio/low_mean': 0.012681030901148915, 'clip_ratio/low_min': 0.012681030901148915, 'clip_ratio/high_mean': 0.01095727866049856, 'clip_ratio/high_max': 0.01095727866049856, 'clip_ratio/region_mean': 0.023638309561647475, 'reward_total_mean': 0.9988850355148315, 'reward_meter_mean': 0.9988850355148315, 'reward_meter_std': 0.00015237655316013843, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988850355148315, 'reward_total_composite_std': 0.00015237655316013843, 'epoch': 0.12}

 87%|████████▋ | 2871/3300 [4:17:04<34:12,  4.78s/it]INFO 04-12 02:53:55 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2872/3300 [4:17:11<38:05,  5.34s/it]2026-04-12 02:54:02,185 | INFO | train_grpo_train | metrics_logged mode=train step=2872

                                                     
{'loss': 0.0065, 'grad_norm': 1.5020183324813843, 'learning_rate': 1.3e-06, 'num_tokens': 6517818.0, 'completions/mean_length': 249.375, 'completions/min_length': 248.0, 'completions/max_length': 251.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 249.375, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 251.0, 'rewards/meter/mean': 0.9985396862030029, 'rewards/meter/std': 0.0008591708028689027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9519230723381042, 'rewards/repeat_penalty/std': 0.05723259598016739, 'rewards/total_composite/mean': 0.9505188465118408, 'rewards/total_composite/std': 0.05688954144716263, 'reward': 0.9505188465118408, 'reward_std': 0.05688953027129173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025731699541211128, 'sampling/sampling_logp_difference/max': 1.5636186599731445, 'sampling/importance_sampling_ratio/min': 0.20937703549861908, 'sampling/importance_sampling_ratio/mean': 1.0057145357131958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.27019675448536873, 'clip_ratio/low_mean': 0.003998056286945939, 'clip_ratio/low_min': 0.003998056286945939, 'clip_ratio/high_mean': 0.0125664914958179, 'clip_ratio/high_max': 0.0125664914958179, 'clip_ratio/region_mean': 0.01656454778276384, 'reward_total_mean': 0.9505188465118408, 'reward_meter_mean': 0.9985396862030029, 'reward_meter_std': 0.0008591708028689027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9519230723381042, 'reward_repeat_penalty_std': 0.05723259598016739, 'reward_total_composite_mean': 0.9505188465118408, 'reward_total_composite_std': 0.05688954144716263, 'epoch': 0.12}

 87%|████████▋ | 2872/3300 [4:17:11<38:05,  5.34s/it]INFO 04-12 02:54:02 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2873/3300 [4:17:15<36:00,  5.06s/it]2026-04-12 02:54:06,596 | INFO | train_grpo_train | metrics_logged mode=train step=2873

                                                     
{'loss': -0.0005, 'grad_norm': 0.2703946828842163, 'learning_rate': 1.296969696969697e-06, 'num_tokens': 6519782.0, 'completions/mean_length': 66.5, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981412291526794, 'rewards/meter/std': 1.5779543900862336e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981412291526794, 'rewards/total_composite/std': 1.5779543900862336e-05, 'reward': 0.9981412291526794, 'reward_std': 1.5772715414641425e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009107197634875774, 'sampling/sampling_logp_difference/max': 0.5398116111755371, 'sampling/importance_sampling_ratio/min': 0.5828580856323242, 'sampling/importance_sampling_ratio/mean': 1.0045442581176758, 'sampling/importance_sampling_ratio/max': 1.4784173965454102, 'entropy': 0.057810590136796236, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.9981412291526794, 'reward_meter_mean': 0.9981412291526794, 'reward_meter_std': 1.5779543900862336e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981412291526794, 'reward_total_composite_std': 1.5779543900862336e-05, 'epoch': 0.12}

 87%|████████▋ | 2873/3300 [4:17:15<36:00,  5.06s/it]INFO 04-12 02:54:06 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2874/3300 [4:17:19<33:54,  4.78s/it]2026-04-12 02:54:10,712 | INFO | train_grpo_train | metrics_logged mode=train step=2874

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2939393939393941e-06, 'num_tokens': 6521271.0, 'completions/mean_length': 36.125, 'completions/min_length': 36.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9996045231819153, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9996045231819153, 'rewards/total_composite/std': 0.0, 'reward': 0.9996045231819153, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0043750000186264515, 'sampling/sampling_logp_difference/max': 0.496179461479187, 'sampling/importance_sampling_ratio/min': 0.6088523864746094, 'sampling/importance_sampling_ratio/mean': 0.9987858533859253, 'sampling/importance_sampling_ratio/max': 1.0343594551086426, 'entropy': 0.02826439938507974, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9996045231819153, 'reward_meter_mean': 0.9996045231819153, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9996045231819153, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 87%|████████▋ | 2874/3300 [4:17:19<33:54,  4.78s/it]INFO 04-12 02:54:10 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2875/3300 [4:17:24<33:11,  4.69s/it]2026-04-12 02:54:15,180 | INFO | train_grpo_train | metrics_logged mode=train step=2875

                                                     
{'loss': 0.0072, 'grad_norm': 5.8196916580200195, 'learning_rate': 1.290909090909091e-06, 'num_tokens': 6523051.0, 'completions/mean_length': 66.5, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9894192218780518, 'rewards/meter/std': 0.010708401910960674, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9894192218780518, 'rewards/total_composite/std': 0.010708401910960674, 'reward': 0.9894192218780518, 'reward_std': 0.01070839911699295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02634219266474247, 'sampling/sampling_logp_difference/max': 2.222754955291748, 'sampling/importance_sampling_ratio/min': 0.10831031203269958, 'sampling/importance_sampling_ratio/mean': 1.0059653520584106, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.17508152220398188, 'clip_ratio/low_mean': 0.0018115942366421223, 'clip_ratio/low_min': 0.0018115942366421223, 'clip_ratio/high_mean': 0.007520091836340725, 'clip_ratio/high_max': 0.007520091836340725, 'clip_ratio/region_mean': 0.009331686072982848, 'reward_total_mean': 0.9894192218780518, 'reward_meter_mean': 0.9894192218780518, 'reward_meter_std': 0.010708401910960674, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9894192218780518, 'reward_total_composite_std': 0.010708401910960674, 'epoch': 0.12}

 87%|████████▋ | 2875/3300 [4:17:24<33:11,  4.69s/it]INFO 04-12 02:54:15 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2876/3300 [4:17:28<32:17,  4.57s/it]2026-04-12 02:54:19,491 | INFO | train_grpo_train | metrics_logged mode=train step=2876

                                                     
{'loss': -0.0001, 'grad_norm': 0.7274681925773621, 'learning_rate': 1.287878787878788e-06, 'num_tokens': 6525011.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973265528678894, 'rewards/meter/std': 2.102993312291801e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973265528678894, 'rewards/total_composite/std': 2.102993312291801e-05, 'reward': 0.9973265528678894, 'reward_std': 2.101841710100416e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010243063792586327, 'sampling/sampling_logp_difference/max': 0.8772482872009277, 'sampling/importance_sampling_ratio/min': 0.41592586040496826, 'sampling/importance_sampling_ratio/mean': 0.9998522400856018, 'sampling/importance_sampling_ratio/max': 1.5096032619476318, 'entropy': 0.05711106630042195, 'clip_ratio/low_mean': 0.004098360426723957, 'clip_ratio/low_min': 0.004098360426723957, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.006147540640085936, 'reward_total_mean': 0.9973265528678894, 'reward_meter_mean': 0.9973265528678894, 'reward_meter_std': 2.102993312291801e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973265528678894, 'reward_total_composite_std': 2.102993312291801e-05, 'epoch': 0.12}

 87%|████████▋ | 2876/3300 [4:17:28<32:17,  4.57s/it]INFO 04-12 02:54:19 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2877/3300 [4:17:33<33:31,  4.76s/it]2026-04-12 02:54:24,679 | INFO | train_grpo_train | metrics_logged mode=train step=2877

                                                     
{'loss': -0.0056, 'grad_norm': 2.641002655029297, 'learning_rate': 1.2848484848484848e-06, 'num_tokens': 6527326.0, 'completions/mean_length': 118.375, 'completions/min_length': 115.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.375, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9990575313568115, 'rewards/meter/std': 0.00029957492370158434, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990575313568115, 'rewards/total_composite/std': 0.00029957492370158434, 'reward': 0.9990575313568115, 'reward_std': 0.00029957492370158434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0404638797044754, 'sampling/sampling_logp_difference/max': 1.9206485748291016, 'sampling/importance_sampling_ratio/min': 0.14651189744472504, 'sampling/importance_sampling_ratio/mean': 1.0039846897125244, 'sampling/importance_sampling_ratio/max': 1.9006513357162476, 'entropy': 0.35712629184126854, 'clip_ratio/low_mean': 0.015123688150197268, 'clip_ratio/low_min': 0.015123688150197268, 'clip_ratio/high_mean': 0.023920452571474016, 'clip_ratio/high_max': 0.023920452571474016, 'clip_ratio/region_mean': 0.03904414072167128, 'reward_total_mean': 0.9990575313568115, 'reward_meter_mean': 0.9990575313568115, 'reward_meter_std': 0.00029957492370158434, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990575313568115, 'reward_total_composite_std': 0.00029957492370158434, 'epoch': 0.12}

 87%|████████▋ | 2877/3300 [4:17:33<33:31,  4.76s/it]INFO 04-12 02:54:24 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2878/3300 [4:17:41<39:12,  5.57s/it]2026-04-12 02:54:32,157 | INFO | train_grpo_train | metrics_logged mode=train step=2878

                                                     
{'loss': 0.0099, 'grad_norm': 2.2012033462524414, 'learning_rate': 1.2818181818181819e-06, 'num_tokens': 6531545.0, 'completions/mean_length': 302.375, 'completions/min_length': 294.0, 'completions/max_length': 310.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 302.375, 'completions/min_terminated_length': 294.0, 'completions/max_terminated_length': 310.0, 'rewards/meter/mean': 0.9952223896980286, 'rewards/meter/std': 0.0032623247243463993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9802631139755249, 'rewards/repeat_penalty/std': 0.027239417657256126, 'rewards/total_composite/mean': 0.975556492805481, 'rewards/total_composite/std': 0.02634783275425434, 'reward': 0.975556492805481, 'reward_std': 0.026347849518060684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05190505087375641, 'sampling/sampling_logp_difference/max': 1.4298324584960938, 'sampling/importance_sampling_ratio/min': 0.23934902250766754, 'sampling/importance_sampling_ratio/mean': 1.0102460384368896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47275901213288307, 'clip_ratio/low_mean': 0.012780054472386837, 'clip_ratio/low_min': 0.012780054472386837, 'clip_ratio/high_mean': 0.030017988989129663, 'clip_ratio/high_max': 0.030017988989129663, 'clip_ratio/region_mean': 0.0427980434615165, 'reward_total_mean': 0.975556492805481, 'reward_meter_mean': 0.9952223896980286, 'reward_meter_std': 0.0032623247243463993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9802631139755249, 'reward_repeat_penalty_std': 0.027239417657256126, 'reward_total_composite_mean': 0.975556492805481, 'reward_total_composite_std': 0.02634783275425434, 'epoch': 0.12}

 87%|████████▋ | 2878/3300 [4:17:41<39:12,  5.57s/it]INFO 04-12 02:54:32 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2879/3300 [4:17:45<37:40,  5.37s/it]2026-04-12 02:54:37,052 | INFO | train_grpo_train | metrics_logged mode=train step=2879

                                                     
{'loss': -0.0009, 'grad_norm': 0.22587938606739044, 'learning_rate': 1.2787878787878787e-06, 'num_tokens': 6533695.0, 'completions/mean_length': 97.75, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.75, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994171857833862, 'rewards/meter/std': 3.359419133630581e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994171857833862, 'rewards/total_composite/std': 3.359419133630581e-05, 'reward': 0.9994171857833862, 'reward_std': 3.360588743817061e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009699815884232521, 'sampling/sampling_logp_difference/max': 2.2170257568359375, 'sampling/importance_sampling_ratio/min': 0.569092333316803, 'sampling/importance_sampling_ratio/mean': 1.0023701190948486, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04796775570139289, 'clip_ratio/low_mean': 0.003839680110104382, 'clip_ratio/low_min': 0.003839680110104382, 'clip_ratio/high_mean': 0.0025510203558951616, 'clip_ratio/high_max': 0.0025510203558951616, 'clip_ratio/region_mean': 0.006390700465999544, 'reward_total_mean': 0.9994171857833862, 'reward_meter_mean': 0.9994171857833862, 'reward_meter_std': 3.359419133630581e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994171857833862, 'reward_total_composite_std': 3.359419133630581e-05, 'epoch': 0.12}

 87%|████████▋ | 2879/3300 [4:17:46<37:40,  5.37s/it]INFO 04-12 02:54:37 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2880/3300 [4:17:51<37:40,  5.38s/it]2026-04-12 02:54:42,463 | INFO | train_grpo_train | metrics_logged mode=train step=2880

                                                     
{'loss': 0.0276, 'grad_norm': 3.3986284732818604, 'learning_rate': 1.2757575757575758e-06, 'num_tokens': 6536129.0, 'completions/mean_length': 131.25, 'completions/min_length': 127.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.25, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9821168184280396, 'rewards/meter/std': 0.010360779240727425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9821168184280396, 'rewards/total_composite/std': 0.010360779240727425, 'reward': 0.9821168184280396, 'reward_std': 0.010360777378082275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.045867305248975754, 'sampling/sampling_logp_difference/max': 1.2023324966430664, 'sampling/importance_sampling_ratio/min': 0.3004924952983856, 'sampling/importance_sampling_ratio/mean': 1.0092159509658813, 'sampling/importance_sampling_ratio/max': 1.8226743936538696, 'entropy': 0.46811263263225555, 'clip_ratio/low_mean': 0.014072955353185534, 'clip_ratio/low_min': 0.014072955353185534, 'clip_ratio/high_mean': 0.01734496164135635, 'clip_ratio/high_max': 0.01734496164135635, 'clip_ratio/region_mean': 0.03141791699454188, 'reward_total_mean': 0.9821168184280396, 'reward_meter_mean': 0.9821168184280396, 'reward_meter_std': 0.010360779240727425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9821168184280396, 'reward_total_composite_std': 0.010360779240727425, 'epoch': 0.12}

 87%|████████▋ | 2880/3300 [4:17:51<37:40,  5.38s/it]INFO 04-12 02:54:42 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2881/3300 [4:17:56<36:45,  5.26s/it]2026-04-12 02:54:47,451 | INFO | train_grpo_train | metrics_logged mode=train step=2881

                                                     
{'loss': -0.0002, 'grad_norm': 0.9868201613426208, 'learning_rate': 1.2727272727272728e-06, 'num_tokens': 6538228.0, 'completions/mean_length': 97.375, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.375, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9979463815689087, 'rewards/meter/std': 6.646273686783388e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979463815689087, 'rewards/total_composite/std': 6.646273686783388e-05, 'reward': 0.9979463815689087, 'reward_std': 6.644289533142e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013669819571077824, 'sampling/sampling_logp_difference/max': 0.7892913818359375, 'sampling/importance_sampling_ratio/min': 0.4541665017604828, 'sampling/importance_sampling_ratio/mean': 1.0037821531295776, 'sampling/importance_sampling_ratio/max': 1.4697163105010986, 'entropy': 0.1293118530884385, 'clip_ratio/low_mean': 0.005141489440575242, 'clip_ratio/low_min': 0.005141489440575242, 'clip_ratio/high_mean': 0.0038528296863660216, 'clip_ratio/high_max': 0.0038528296863660216, 'clip_ratio/region_mean': 0.008994319126941264, 'reward_total_mean': 0.9979463815689087, 'reward_meter_mean': 0.9979463815689087, 'reward_meter_std': 6.646273686783388e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979463815689087, 'reward_total_composite_std': 6.646273686783388e-05, 'epoch': 0.12}

 87%|████████▋ | 2881/3300 [4:17:56<36:45,  5.26s/it]INFO 04-12 02:54:47 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2882/3300 [4:18:00<34:59,  5.02s/it]2026-04-12 02:54:51,908 | INFO | train_grpo_train | metrics_logged mode=train step=2882

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2696969696969698e-06, 'num_tokens': 6540020.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00015525527123827487, 'sampling/sampling_logp_difference/max': 0.005098958499729633, 'sampling/importance_sampling_ratio/min': 0.9974856972694397, 'sampling/importance_sampling_ratio/mean': 1.0001249313354492, 'sampling/importance_sampling_ratio/max': 1.0051120519638062, 'entropy': 0.0014212291716830805, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 87%|████████▋ | 2882/3300 [4:18:00<34:59,  5.02s/it]INFO 04-12 02:54:52 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2883/3300 [4:18:06<35:51,  5.16s/it]2026-04-12 02:54:57,384 | INFO | train_grpo_train | metrics_logged mode=train step=2883

                                                     
{'loss': -0.0014, 'grad_norm': 1.7158304452896118, 'learning_rate': 1.2666666666666669e-06, 'num_tokens': 6542928.0, 'completions/mean_length': 142.5, 'completions/min_length': 141.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.5, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.998982846736908, 'rewards/meter/std': 0.00020013147150166333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998982846736908, 'rewards/total_composite/std': 0.00020013147150166333, 'reward': 0.998982846736908, 'reward_std': 0.00020013845642097294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022738341242074966, 'sampling/sampling_logp_difference/max': 0.9792003631591797, 'sampling/importance_sampling_ratio/min': 0.3756113350391388, 'sampling/importance_sampling_ratio/mean': 1.0040578842163086, 'sampling/importance_sampling_ratio/max': 1.6984710693359375, 'entropy': 0.2192006167024374, 'clip_ratio/low_mean': 0.004407651489600539, 'clip_ratio/low_min': 0.004407651489600539, 'clip_ratio/high_mean': 0.013993188505992293, 'clip_ratio/high_max': 0.013993188505992293, 'clip_ratio/region_mean': 0.018400839995592833, 'reward_total_mean': 0.998982846736908, 'reward_meter_mean': 0.998982846736908, 'reward_meter_std': 0.00020013147150166333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998982846736908, 'reward_total_composite_std': 0.00020013147150166333, 'epoch': 0.12}

 87%|████████▋ | 2883/3300 [4:18:06<35:51,  5.16s/it]INFO 04-12 02:54:57 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2884/3300 [4:18:10<34:25,  4.97s/it]2026-04-12 02:55:01,898 | INFO | train_grpo_train | metrics_logged mode=train step=2884

                                                     
{'loss': 0.0006, 'grad_norm': 2.683151960372925, 'learning_rate': 1.2636363636363637e-06, 'num_tokens': 6544901.0, 'completions/mean_length': 67.625, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9992280006408691, 'rewards/meter/std': 0.00013651978224515915, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992280006408691, 'rewards/total_composite/std': 0.00013651978224515915, 'reward': 0.9992280006408691, 'reward_std': 0.00013652884808834642, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019809167832136154, 'sampling/sampling_logp_difference/max': 1.097256064414978, 'sampling/importance_sampling_ratio/min': 0.3337857127189636, 'sampling/importance_sampling_ratio/mean': 0.999349057674408, 'sampling/importance_sampling_ratio/max': 1.5717170238494873, 'entropy': 0.14564206078648567, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.014788191299885511, 'clip_ratio/high_max': 0.014788191299885511, 'clip_ratio/region_mean': 0.014788191299885511, 'reward_total_mean': 0.9992280006408691, 'reward_meter_mean': 0.9992280006408691, 'reward_meter_std': 0.00013651978224515915, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992280006408691, 'reward_total_composite_std': 0.00013651978224515915, 'epoch': 0.12}

 87%|████████▋ | 2884/3300 [4:18:10<34:25,  4.97s/it]INFO 04-12 02:55:02 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2885/3300 [4:18:15<33:34,  4.85s/it]2026-04-12 02:55:06,492 | INFO | train_grpo_train | metrics_logged mode=train step=2885

                                                     
{'loss': 0.0008, 'grad_norm': 0.26706913113594055, 'learning_rate': 1.2606060606060608e-06, 'num_tokens': 6546912.0, 'completions/mean_length': 71.375, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994363784790039, 'rewards/meter/std': 2.7232801585341804e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994363784790039, 'rewards/total_composite/std': 2.7232801585341804e-05, 'reward': 0.9994363784790039, 'reward_std': 2.7221845812164247e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010549667291343212, 'sampling/sampling_logp_difference/max': 0.676356315612793, 'sampling/importance_sampling_ratio/min': 0.5084663033485413, 'sampling/importance_sampling_ratio/mean': 1.002943754196167, 'sampling/importance_sampling_ratio/max': 1.2115890979766846, 'entropy': 0.08360666781663895, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.0034966744715347886, 'clip_ratio/high_max': 0.0034966744715347886, 'clip_ratio/region_mean': 0.0052327855955809355, 'reward_total_mean': 0.9994363784790039, 'reward_meter_mean': 0.9994363784790039, 'reward_meter_std': 2.7232801585341804e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994363784790039, 'reward_total_composite_std': 2.7232801585341804e-05, 'epoch': 0.12}

 87%|████████▋ | 2885/3300 [4:18:15<33:34,  4.85s/it]INFO 04-12 02:55:06 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2886/3300 [4:18:19<32:58,  4.78s/it]2026-04-12 02:55:11,098 | INFO | train_grpo_train | metrics_logged mode=train step=2886

                                                     
{'loss': 0.0053, 'grad_norm': 2.8855912685394287, 'learning_rate': 1.2575757575757578e-06, 'num_tokens': 6548792.0, 'completions/mean_length': 67.0, 'completions/min_length': 66.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9954104423522949, 'rewards/meter/std': 0.007760278414934874, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9954104423522949, 'rewards/total_composite/std': 0.007760278414934874, 'reward': 0.9954104423522949, 'reward_std': 0.007760272361338139, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009755668230354786, 'sampling/sampling_logp_difference/max': 0.9411745071411133, 'sampling/importance_sampling_ratio/min': 0.39016932249069214, 'sampling/importance_sampling_ratio/mean': 1.0019506216049194, 'sampling/importance_sampling_ratio/max': 1.3913602828979492, 'entropy': 0.09443793492391706, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.005542142200283706, 'reward_total_mean': 0.9954104423522949, 'reward_meter_mean': 0.9954104423522949, 'reward_meter_std': 0.007760278414934874, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9954104423522949, 'reward_total_composite_std': 0.007760278414934874, 'epoch': 0.12}

 87%|████████▋ | 2886/3300 [4:18:20<32:58,  4.78s/it]INFO 04-12 02:55:11 [block_pool.py:378] Successfully reset prefix cache

 87%|████████▋ | 2887/3300 [4:18:24<32:28,  4.72s/it]2026-04-12 02:55:15,669 | INFO | train_grpo_train | metrics_logged mode=train step=2887

                                                     
{'loss': -0.0051, 'grad_norm': 3.4663033485412598, 'learning_rate': 1.2545454545454546e-06, 'num_tokens': 6550591.0, 'completions/mean_length': 65.875, 'completions/min_length': 65.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9924707412719727, 'rewards/meter/std': 0.0012687068665400147, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924707412719727, 'rewards/total_composite/std': 0.0012687068665400147, 'reward': 0.9924707412719727, 'reward_std': 0.001268712687306106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02419329807162285, 'sampling/sampling_logp_difference/max': 0.6981816291809082, 'sampling/importance_sampling_ratio/min': 0.4974890947341919, 'sampling/importance_sampling_ratio/mean': 1.0042961835861206, 'sampling/importance_sampling_ratio/max': 1.3884730339050293, 'entropy': 0.17512462474405766, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.007520923274569213, 'clip_ratio/high_max': 0.007520923274569213, 'clip_ratio/region_mean': 0.0094148627249524, 'reward_total_mean': 0.9924707412719727, 'reward_meter_mean': 0.9924707412719727, 'reward_meter_std': 0.0012687068665400147, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924707412719727, 'reward_total_composite_std': 0.0012687068665400147, 'epoch': 0.12}

 87%|████████▋ | 2887/3300 [4:18:24<32:28,  4.72s/it]INFO 04-12 02:55:15 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2888/3300 [4:18:29<33:00,  4.81s/it]2026-04-12 02:55:20,683 | INFO | train_grpo_train | metrics_logged mode=train step=2888

                                                     
{'loss': -0.001, 'grad_norm': 0.5309938192367554, 'learning_rate': 1.2515151515151517e-06, 'num_tokens': 6552707.0, 'completions/mean_length': 97.5, 'completions/min_length': 96.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.5, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9980380535125732, 'rewards/meter/std': 4.4257390982238576e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980380535125732, 'rewards/total_composite/std': 4.4257390982238576e-05, 'reward': 0.9980380535125732, 'reward_std': 4.425693259690888e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015023077838122845, 'sampling/sampling_logp_difference/max': 0.9142270088195801, 'sampling/importance_sampling_ratio/min': 0.4008263647556305, 'sampling/importance_sampling_ratio/mean': 1.001752257347107, 'sampling/importance_sampling_ratio/max': 1.4897899627685547, 'entropy': 0.1272029634565115, 'clip_ratio/low_mean': 0.003839680110104382, 'clip_ratio/low_min': 0.003839680110104382, 'clip_ratio/high_mean': 0.011493272730149329, 'clip_ratio/high_max': 0.011493272730149329, 'clip_ratio/region_mean': 0.01533295284025371, 'reward_total_mean': 0.9980380535125732, 'reward_meter_mean': 0.9980380535125732, 'reward_meter_std': 4.4257390982238576e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980380535125732, 'reward_total_composite_std': 4.4257390982238576e-05, 'epoch': 0.12}

 88%|████████▊ | 2888/3300 [4:18:29<33:00,  4.81s/it]INFO 04-12 02:55:20 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2889/3300 [4:18:34<32:49,  4.79s/it]2026-04-12 02:55:25,448 | INFO | train_grpo_train | metrics_logged mode=train step=2889

                                                     
{'loss': -0.0002, 'grad_norm': 1.1036797761917114, 'learning_rate': 1.2484848484848485e-06, 'num_tokens': 6554583.0, 'completions/mean_length': 79.5, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.5, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.998902440071106, 'rewards/meter/std': 0.00011711295519489795, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998902440071106, 'rewards/total_composite/std': 0.00011711295519489795, 'reward': 0.998902440071106, 'reward_std': 0.00011710206308634952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02463226206600666, 'sampling/sampling_logp_difference/max': 1.763596534729004, 'sampling/importance_sampling_ratio/min': 0.17142720520496368, 'sampling/importance_sampling_ratio/mean': 1.008751392364502, 'sampling/importance_sampling_ratio/max': 1.4718900918960571, 'entropy': 0.21664478071033955, 'clip_ratio/low_mean': 0.004707766929641366, 'clip_ratio/low_min': 0.004707766929641366, 'clip_ratio/high_mean': 0.007871835259720683, 'clip_ratio/high_max': 0.007871835259720683, 'clip_ratio/region_mean': 0.012579602189362049, 'reward_total_mean': 0.998902440071106, 'reward_meter_mean': 0.998902440071106, 'reward_meter_std': 0.00011711295519489795, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998902440071106, 'reward_total_composite_std': 0.00011711295519489795, 'epoch': 0.12}

 88%|████████▊ | 2889/3300 [4:18:34<32:49,  4.79s/it]INFO 04-12 02:55:25 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2890/3300 [4:18:39<32:47,  4.80s/it]2026-04-12 02:55:30,256 | INFO | train_grpo_train | metrics_logged mode=train step=2890

                                                     
{'loss': -0.0123, 'grad_norm': 2.1610708236694336, 'learning_rate': 1.2454545454545456e-06, 'num_tokens': 6556527.0, 'completions/mean_length': 92.0, 'completions/min_length': 87.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.0, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9974758625030518, 'rewards/meter/std': 0.0004072287702001631, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9974758625030518, 'rewards/total_composite/std': 0.0004072287702001631, 'reward': 0.9974758625030518, 'reward_std': 0.0004072203009855002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01821918971836567, 'sampling/sampling_logp_difference/max': 1.333749771118164, 'sampling/importance_sampling_ratio/min': 0.26348739862442017, 'sampling/importance_sampling_ratio/mean': 1.0036619901657104, 'sampling/importance_sampling_ratio/max': 1.664230227470398, 'entropy': 0.14203907828778028, 'clip_ratio/low_mean': 0.005469039548188448, 'clip_ratio/low_min': 0.005469039548188448, 'clip_ratio/high_mean': 0.013470079400576651, 'clip_ratio/high_max': 0.013470079400576651, 'clip_ratio/region_mean': 0.0189391189487651, 'reward_total_mean': 0.9974758625030518, 'reward_meter_mean': 0.9974758625030518, 'reward_meter_std': 0.0004072287702001631, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9974758625030518, 'reward_total_composite_std': 0.0004072287702001631, 'epoch': 0.12}

 88%|████████▊ | 2890/3300 [4:18:39<32:47,  4.80s/it]INFO 04-12 02:55:30 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2891/3300 [4:18:44<33:00,  4.84s/it]2026-04-12 02:55:35,206 | INFO | train_grpo_train | metrics_logged mode=train step=2891

                                                     
{'loss': 0.0102, 'grad_norm': 3.022761106491089, 'learning_rate': 1.2424242424242424e-06, 'num_tokens': 6558597.0, 'completions/mean_length': 97.75, 'completions/min_length': 96.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.75, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9901492595672607, 'rewards/meter/std': 0.005893507041037083, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9901492595672607, 'rewards/total_composite/std': 0.005893507041037083, 'reward': 0.9901492595672607, 'reward_std': 0.005893507041037083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032126907259225845, 'sampling/sampling_logp_difference/max': 1.2647018432617188, 'sampling/importance_sampling_ratio/min': 0.2823234498500824, 'sampling/importance_sampling_ratio/mean': 1.0117164850234985, 'sampling/importance_sampling_ratio/max': 1.6783695220947266, 'entropy': 0.2828730084002018, 'clip_ratio/low_mean': 0.008739139186218381, 'clip_ratio/low_min': 0.008739139186218381, 'clip_ratio/high_mean': 0.01412348123267293, 'clip_ratio/high_max': 0.01412348123267293, 'clip_ratio/region_mean': 0.02286262041889131, 'reward_total_mean': 0.9901492595672607, 'reward_meter_mean': 0.9901492595672607, 'reward_meter_std': 0.005893507041037083, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9901492595672607, 'reward_total_composite_std': 0.005893507041037083, 'epoch': 0.12}

 88%|████████▊ | 2891/3300 [4:18:44<33:00,  4.84s/it]INFO 04-12 02:55:35 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2892/3300 [4:18:48<31:19,  4.61s/it]2026-04-12 02:55:39,252 | INFO | train_grpo_train | metrics_logged mode=train step=2892

                                                     
{'loss': 0.0466, 'grad_norm': 15.570990562438965, 'learning_rate': 1.2393939393939394e-06, 'num_tokens': 6559937.0, 'completions/mean_length': 23.5, 'completions/min_length': 21.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.895869255065918, 'rewards/meter/std': 0.033730946481227875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.895869255065918, 'rewards/total_composite/std': 0.033730946481227875, 'reward': 0.895869255065918, 'reward_std': 0.03373093158006668, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0897468775510788, 'sampling/sampling_logp_difference/max': 1.4235849380493164, 'sampling/importance_sampling_ratio/min': 0.24084904789924622, 'sampling/importance_sampling_ratio/mean': 0.9834393858909607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37175317015498877, 'clip_ratio/low_mean': 0.020869565196335316, 'clip_ratio/low_min': 0.020869565196335316, 'clip_ratio/high_mean': 0.06117365509271622, 'clip_ratio/high_max': 0.06117365509271622, 'clip_ratio/region_mean': 0.08204322028905153, 'reward_total_mean': 0.895869255065918, 'reward_meter_mean': 0.895869255065918, 'reward_meter_std': 0.033730946481227875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.895869255065918, 'reward_total_composite_std': 0.033730946481227875, 'epoch': 0.12}

 88%|████████▊ | 2892/3300 [4:18:48<31:19,  4.61s/it]INFO 04-12 02:55:39 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2893/3300 [4:18:52<30:40,  4.52s/it]2026-04-12 02:55:43,586 | INFO | train_grpo_train | metrics_logged mode=train step=2893

                                                     
{'loss': -0.0319, 'grad_norm': 11.483199119567871, 'learning_rate': 1.2363636363636365e-06, 'num_tokens': 6561687.0, 'completions/mean_length': 59.75, 'completions/min_length': 56.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9967870712280273, 'rewards/meter/std': 0.0010188753949478269, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9967870712280273, 'rewards/total_composite/std': 0.0010188753949478269, 'reward': 0.9967870712280273, 'reward_std': 0.0010188753949478269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01920030452311039, 'sampling/sampling_logp_difference/max': 1.7169630527496338, 'sampling/importance_sampling_ratio/min': 0.17961078882217407, 'sampling/importance_sampling_ratio/mean': 1.0000895261764526, 'sampling/importance_sampling_ratio/max': 1.8504060506820679, 'entropy': 0.07222167262807488, 'clip_ratio/low_mean': 0.006696428870782256, 'clip_ratio/low_min': 0.006696428870782256, 'clip_ratio/high_mean': 0.010245901066809893, 'clip_ratio/high_max': 0.010245901066809893, 'clip_ratio/region_mean': 0.01694232993759215, 'reward_total_mean': 0.9967870712280273, 'reward_meter_mean': 0.9967870712280273, 'reward_meter_std': 0.0010188753949478269, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9967870712280273, 'reward_total_composite_std': 0.0010188753949478269, 'epoch': 0.12}

 88%|████████▊ | 2893/3300 [4:18:52<30:40,  4.52s/it]INFO 04-12 02:55:43 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2894/3300 [4:18:58<32:47,  4.85s/it]2026-04-12 02:55:49,181 | INFO | train_grpo_train | metrics_logged mode=train step=2894

                                                     
{'loss': -0.0994, 'grad_norm': 1.820000171661377, 'learning_rate': 1.2333333333333335e-06, 'num_tokens': 6564523.0, 'completions/mean_length': 154.5, 'completions/min_length': 108.0, 'completions/max_length': 166.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.5, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.9889054298400879, 'rewards/meter/std': 0.003125808434560895, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9117738604545593, 'rewards/total_composite/std': 0.13682788610458374, 'reward': 0.9117738604545593, 'reward_std': 0.13682788610458374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04758564755320549, 'sampling/sampling_logp_difference/max': 1.404418706893921, 'sampling/importance_sampling_ratio/min': 0.24550971388816833, 'sampling/importance_sampling_ratio/mean': 1.008368730545044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5260906405746937, 'clip_ratio/low_mean': 0.007427186123095453, 'clip_ratio/low_min': 0.007427186123095453, 'clip_ratio/high_mean': 0.025422006146982312, 'clip_ratio/high_max': 0.025422006146982312, 'clip_ratio/region_mean': 0.032849192270077765, 'reward_total_mean': 0.9117738604545593, 'reward_meter_mean': 0.9889054298400879, 'reward_meter_std': 0.003125808434560895, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.1414213478565216, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9117738604545593, 'reward_total_composite_std': 0.13682788610458374, 'epoch': 0.12}

 88%|████████▊ | 2894/3300 [4:18:58<32:47,  4.85s/it]INFO 04-12 02:55:49 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2895/3300 [4:19:03<33:14,  4.92s/it]2026-04-12 02:55:54,290 | INFO | train_grpo_train | metrics_logged mode=train step=2895

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2303030303030304e-06, 'num_tokens': 6567027.0, 'completions/mean_length': 132.0, 'completions/min_length': 132.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.0, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.6684597730636597, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7777777910232544, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.5199131369590759, 'rewards/total_composite/std': 0.0, 'reward': 0.5199131369590759, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0006711529567837715, 'sampling/sampling_logp_difference/max': 0.02449747547507286, 'sampling/importance_sampling_ratio/min': 0.9979704022407532, 'sampling/importance_sampling_ratio/mean': 1.0006672143936157, 'sampling/importance_sampling_ratio/max': 1.0248000621795654, 'entropy': 0.005916567693930119, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.5199131369590759, 'reward_meter_mean': 0.6684597730636597, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7777777910232544, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.5199131369590759, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 88%|████████▊ | 2895/3300 [4:19:03<33:14,  4.92s/it]INFO 04-12 02:55:54 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2896/3300 [4:19:07<31:47,  4.72s/it]2026-04-12 02:55:58,539 | INFO | train_grpo_train | metrics_logged mode=train step=2896

                                                     
{'loss': 0.0006, 'grad_norm': 3.799595355987549, 'learning_rate': 1.2272727272727274e-06, 'num_tokens': 6568899.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9972846508026123, 'rewards/meter/std': 0.00015232501027639955, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972846508026123, 'rewards/total_composite/std': 0.00015232501027639955, 'reward': 0.9972846508026123, 'reward_std': 0.00015233403246384114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005107899196445942, 'sampling/sampling_logp_difference/max': 0.5471892356872559, 'sampling/importance_sampling_ratio/min': 0.5785737633705139, 'sampling/importance_sampling_ratio/mean': 1.0012083053588867, 'sampling/importance_sampling_ratio/max': 1.18209969997406, 'entropy': 0.035036823246628046, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.004098360426723957, 'clip_ratio/high_max': 0.004098360426723957, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.9972846508026123, 'reward_meter_mean': 0.9972846508026123, 'reward_meter_std': 0.00015232501027639955, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972846508026123, 'reward_total_composite_std': 0.00015232501027639955, 'epoch': 0.12}

 88%|████████▊ | 2896/3300 [4:19:07<31:47,  4.72s/it]INFO 04-12 02:55:58 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2897/3300 [4:19:12<32:23,  4.82s/it]2026-04-12 02:56:03,595 | INFO | train_grpo_train | metrics_logged mode=train step=2897

                                                     
{'loss': 0.0028, 'grad_norm': 1.706589937210083, 'learning_rate': 1.2242424242424242e-06, 'num_tokens': 6571375.0, 'completions/mean_length': 128.5, 'completions/min_length': 127.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.5, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9972264766693115, 'rewards/meter/std': 0.002113205147907138, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.9259439706802368, 'rewards/total_composite/std': 0.07545064389705658, 'reward': 0.9259439706802368, 'reward_std': 0.07545064389705658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016268158331513405, 'sampling/sampling_logp_difference/max': 0.9007883071899414, 'sampling/importance_sampling_ratio/min': 0.4062493145465851, 'sampling/importance_sampling_ratio/mean': 1.0051052570343018, 'sampling/importance_sampling_ratio/max': 1.4906891584396362, 'entropy': 0.17375356331467628, 'clip_ratio/low_mean': 0.005829093977808952, 'clip_ratio/low_min': 0.005829093977808952, 'clip_ratio/high_mean': 0.008736189920455217, 'clip_ratio/high_max': 0.008736189920455217, 'clip_ratio/region_mean': 0.01456528389826417, 'reward_total_mean': 0.9259439706802368, 'reward_meter_mean': 0.9972264766693115, 'reward_meter_std': 0.002113205147907138, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.9259439706802368, 'reward_total_composite_std': 0.07545064389705658, 'epoch': 0.12}

 88%|████████▊ | 2897/3300 [4:19:12<32:23,  4.82s/it]INFO 04-12 02:56:03 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2898/3300 [4:19:16<31:26,  4.69s/it]2026-04-12 02:56:07,990 | INFO | train_grpo_train | metrics_logged mode=train step=2898

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2212121212121213e-06, 'num_tokens': 6573071.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002584453613962978, 'sampling/sampling_logp_difference/max': 0.004060388542711735, 'sampling/importance_sampling_ratio/min': 0.9999027252197266, 'sampling/importance_sampling_ratio/mean': 1.0002572536468506, 'sampling/importance_sampling_ratio/max': 1.0040686130523682, 'entropy': 0.0020958341483492404, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 88%|████████▊ | 2898/3300 [4:19:16<31:26,  4.69s/it]INFO 04-12 02:56:08 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2899/3300 [4:19:21<30:49,  4.61s/it]2026-04-12 02:56:12,410 | INFO | train_grpo_train | metrics_logged mode=train step=2899

                                                     
{'loss': -0.0001, 'grad_norm': 0.039646849036216736, 'learning_rate': 1.2181818181818183e-06, 'num_tokens': 6574871.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973360300064087, 'rewards/meter/std': 8.113272997434251e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973360300064087, 'rewards/total_composite/std': 8.113272997434251e-06, 'reward': 0.9973360300064087, 'reward_std': 8.10424353403505e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004976366646587849, 'sampling/sampling_logp_difference/max': 0.4815742075443268, 'sampling/importance_sampling_ratio/min': 0.617810070514679, 'sampling/importance_sampling_ratio/mean': 1.0008740425109863, 'sampling/importance_sampling_ratio/max': 1.1416552066802979, 'entropy': 0.034302944084629416, 'clip_ratio/low_mean': 0.004098360426723957, 'clip_ratio/low_min': 0.004098360426723957, 'clip_ratio/high_mean': 0.004098360426723957, 'clip_ratio/high_max': 0.004098360426723957, 'clip_ratio/region_mean': 0.008196720853447914, 'reward_total_mean': 0.9973360300064087, 'reward_meter_mean': 0.9973360300064087, 'reward_meter_std': 8.113272997434251e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973360300064087, 'reward_total_composite_std': 8.113272997434251e-06, 'epoch': 0.12}

 88%|████████▊ | 2899/3300 [4:19:21<30:49,  4.61s/it]INFO 04-12 02:56:12 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2900/3300 [4:19:25<30:18,  4.55s/it]2026-04-12 02:56:16,806 | INFO | train_grpo_train | metrics_logged mode=train step=2900

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2151515151515154e-06, 'num_tokens': 6576751.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002209599915659055, 'sampling/sampling_logp_difference/max': 0.009038996882736683, 'sampling/importance_sampling_ratio/min': 0.9910017251968384, 'sampling/importance_sampling_ratio/mean': 1.000099539756775, 'sampling/importance_sampling_ratio/max': 1.006137490272522, 'entropy': 0.0027679620689013973, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 88%|████████▊ | 2900/3300 [4:19:25<30:18,  4.55s/it]INFO 04-12 02:56:17 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:38,  3.53s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.46s/it]

 31%|███       | 4/13 [00:16<00:39,  4.42s/it]

 38%|███▊      | 5/13 [00:22<00:38,  4.85s/it]

 46%|████▌     | 6/13 [00:29<00:38,  5.56s/it]

 54%|█████▍    | 7/13 [00:34<00:32,  5.43s/it]

 62%|██████▏   | 8/13 [00:41<00:30,  6.00s/it]

 69%|██████▉   | 9/13 [00:48<00:25,  6.40s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.92s/it]

 85%|████████▍ | 11/13 [00:59<00:11,  5.90s/it]

 92%|█████████▏| 12/13 [01:05<00:05,  5.98s/it]

100%|██████████| 13/13 [01:11<00:00,  5.85s/it]2026-04-12 02:57:35,316 | INFO | train_grpo_train | metrics_logged mode=eval step=2900

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 78.4008, 'eval_samples_per_second': 1.327, 'eval_steps_per_second': 0.166, 'eval_num_tokens': 6576751.0, 'eval_completions/mean_length': 211.68269230769232, 'eval_completions/min_length': 61.84615384615385, 'eval_completions/max_length': 422.84615384615387, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 205.16483600323016, 'eval_completions/min_terminated_length': 61.84615384615385, 'eval_completions/max_terminated_length': 402.46153846153845, 'eval_rewards/meter/mean': 0.7883604077192453, 'eval_rewards/meter/std': 0.34034269847548926, 'eval_rewards/count_adherence/mean': 0.9613341138913081, 'eval_rewards/count_adherence/std': 0.06466435583738181, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.955476293197045, 'eval_rewards/repeat_penalty/std': 0.07249205201291122, 'eval_rewards/total_composite/mean': 0.7234287353662344, 'eval_rewards/total_composite/std': 0.343703310077007, 'eval_reward': 0.7234287353662344, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03452126896725251, 'eval_sampling/sampling_logp_difference/max': 1.1897428219134991, 'eval_sampling/importance_sampling_ratio/min': 0.3125262191662422, 'eval_sampling/importance_sampling_ratio/mean': 1.0095937985640306, 'eval_sampling/importance_sampling_ratio/max': 1.529663553604713, 'eval_entropy': 0.4129388790864211, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7234287353662344, 'eval_reward_meter_mean': 0.7883604077192453, 'eval_reward_meter_std': 0.34034269847548926, 'eval_reward_count_adherence_mean': 0.9613341138913081, 'eval_reward_count_adherence_std': 0.06466435583738181, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.955476293197045, 'eval_reward_repeat_penalty_std': 0.07249205201291122, 'eval_reward_total_composite_mean': 0.7234287353662344, 'eval_reward_total_composite_std': 0.343703310077007, 'epoch': 0.12}

 88%|████████▊ | 2900/3300 [4:20:44<30:18,  4.55s/it]

100%|██████████| 13/13 [01:11<00:00,  5.85s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 88%|████████▊ | 2901/3300 [4:20:51<3:12:52, 29.00s/it]2026-04-12 02:57:42,905 | INFO | train_grpo_train | metrics_logged mode=train step=2901

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2121212121212122e-06, 'num_tokens': 6578415.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0006013626698404551, 'sampling/sampling_logp_difference/max': 0.007742481306195259, 'sampling/importance_sampling_ratio/min': 0.9922874569892883, 'sampling/importance_sampling_ratio/mean': 1.0005377531051636, 'sampling/importance_sampling_ratio/max': 1.0070579051971436, 'entropy': 0.005171062133740634, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 88%|████████▊ | 2901/3300 [4:20:51<3:12:52, 29.00s/it]INFO 04-12 02:57:43 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2902/3300 [4:21:02<2:35:11, 23.40s/it]2026-04-12 02:57:53,192 | INFO | train_grpo_train | metrics_logged mode=train step=2902

                                                       
{'loss': 0.2202, 'grad_norm': 13.265780448913574, 'learning_rate': 1.2090909090909092e-06, 'num_tokens': 6582881.0, 'completions/mean_length': 501.25, 'completions/min_length': 485.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 497.66668701171875, 'completions/min_terminated_length': 485.0, 'completions/max_terminated_length': 512.0, 'rewards/meter/mean': 0.9986035823822021, 'rewards/meter/std': 0.0011026532156392932, 'rewards/count_adherence/mean': 0.796875, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9701460003852844, 'rewards/repeat_penalty/std': 0.01849004067480564, 'rewards/total_composite/mean': 0.7718735337257385, 'rewards/total_composite/std': 0.04241907224059105, 'reward': 0.7718735337257385, 'reward_std': 0.042419061064720154, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04629309102892876, 'sampling/sampling_logp_difference/max': 2.552377700805664, 'sampling/importance_sampling_ratio/min': 0.07789622992277145, 'sampling/importance_sampling_ratio/mean': 1.0130130052566528, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3398243226110935, 'clip_ratio/low_mean': 0.010757095878943801, 'clip_ratio/low_min': 0.010757095878943801, 'clip_ratio/high_mean': 0.010358819272369146, 'clip_ratio/high_max': 0.010358819272369146, 'clip_ratio/region_mean': 0.021115915151312947, 'reward_total_mean': 0.7718735337257385, 'reward_meter_mean': 0.9986035823822021, 'reward_meter_std': 0.0011026532156392932, 'reward_count_adherence_mean': 0.796875, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9701460003852844, 'reward_repeat_penalty_std': 0.01849004067480564, 'reward_total_composite_mean': 0.7718735337257385, 'reward_total_composite_std': 0.04241907224059105, 'epoch': 0.12}

 88%|████████▊ | 2902/3300 [4:21:02<2:35:11, 23.40s/it]INFO 04-12 02:57:53 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2903/3300 [4:21:10<2:04:51, 18.87s/it]2026-04-12 02:58:01,499 | INFO | train_grpo_train | metrics_logged mode=train step=2903

                                                       
{'loss': -0.0437, 'grad_norm': 2.858194351196289, 'learning_rate': 1.206060606060606e-06, 'num_tokens': 6587110.0, 'completions/mean_length': 325.625, 'completions/min_length': 304.0, 'completions/max_length': 349.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 325.625, 'completions/min_terminated_length': 304.0, 'completions/max_terminated_length': 349.0, 'rewards/meter/mean': 0.9878900051116943, 'rewards/meter/std': 0.025390086695551872, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.0534522607922554, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9380475282669067, 'rewards/total_composite/std': 0.04961733520030975, 'reward': 0.9380475282669067, 'reward_std': 0.049617357552051544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06375819444656372, 'sampling/sampling_logp_difference/max': 2.556863784790039, 'sampling/importance_sampling_ratio/min': 0.0775475725531578, 'sampling/importance_sampling_ratio/mean': 1.017285943031311, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.633208267390728, 'clip_ratio/low_mean': 0.025543570518493652, 'clip_ratio/low_min': 0.025543570518493652, 'clip_ratio/high_mean': 0.01695435866713524, 'clip_ratio/high_max': 0.01695435866713524, 'clip_ratio/region_mean': 0.04249792918562889, 'reward_total_mean': 0.9380475282669067, 'reward_meter_mean': 0.9878900051116943, 'reward_meter_std': 0.025390086695551872, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.0534522607922554, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9380475282669067, 'reward_total_composite_std': 0.04961733520030975, 'epoch': 0.12}

 88%|████████▊ | 2903/3300 [4:21:10<2:04:51, 18.87s/it]INFO 04-12 02:58:01 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2904/3300 [4:21:14<1:36:10, 14.57s/it]2026-04-12 02:58:06,042 | INFO | train_grpo_train | metrics_logged mode=train step=2904

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2030303030303031e-06, 'num_tokens': 6588910.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00021447156905196607, 'sampling/sampling_logp_difference/max': 0.01532711274921894, 'sampling/importance_sampling_ratio/min': 0.9847897291183472, 'sampling/importance_sampling_ratio/mean': 1.0000479221343994, 'sampling/importance_sampling_ratio/max': 1.0044045448303223, 'entropy': 0.0017551528289914131, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 88%|████████▊ | 2904/3300 [4:21:15<1:36:10, 14.57s/it]INFO 04-12 02:58:06 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2905/3300 [4:21:19<1:16:42, 11.65s/it]2026-04-12 02:58:10,877 | INFO | train_grpo_train | metrics_logged mode=train step=2905

                                                       
{'loss': 0.0003, 'grad_norm': 0.42203962802886963, 'learning_rate': 1.2000000000000002e-06, 'num_tokens': 6591138.0, 'completions/mean_length': 97.5, 'completions/min_length': 96.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.5, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9980453252792358, 'rewards/meter/std': 4.257191903889179e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980453252792358, 'rewards/total_composite/std': 4.257191903889179e-05, 'reward': 0.9980453252792358, 'reward_std': 4.25745893153362e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014815653674304485, 'sampling/sampling_logp_difference/max': 1.1947669982910156, 'sampling/importance_sampling_ratio/min': 0.30277448892593384, 'sampling/importance_sampling_ratio/mean': 1.002492904663086, 'sampling/importance_sampling_ratio/max': 1.5006216764450073, 'entropy': 0.11437637638300657, 'clip_ratio/low_mean': 0.005181760177947581, 'clip_ratio/low_min': 0.005181760177947581, 'clip_ratio/high_mean': 0.008968020090833306, 'clip_ratio/high_max': 0.008968020090833306, 'clip_ratio/region_mean': 0.014149780268780887, 'reward_total_mean': 0.9980453252792358, 'reward_meter_mean': 0.9980453252792358, 'reward_meter_std': 4.257191903889179e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980453252792358, 'reward_total_composite_std': 4.257191903889179e-05, 'epoch': 0.12}

 88%|████████▊ | 2905/3300 [4:21:19<1:16:42, 11.65s/it]INFO 04-12 02:58:11 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2906/3300 [4:21:24<1:02:06,  9.46s/it]2026-04-12 02:58:15,217 | INFO | train_grpo_train | metrics_logged mode=train step=2906

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.196969696969697e-06, 'num_tokens': 6592674.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00027867371682077646, 'sampling/sampling_logp_difference/max': 0.003768636379390955, 'sampling/importance_sampling_ratio/min': 0.9966413378715515, 'sampling/importance_sampling_ratio/mean': 1.0002539157867432, 'sampling/importance_sampling_ratio/max': 1.003775715827942, 'entropy': 0.002388683380559087, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 88%|████████▊ | 2906/3300 [4:21:24<1:02:06,  9.46s/it]INFO 04-12 02:58:15 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2907/3300 [4:21:28<51:47,  7.91s/it]  2026-04-12 02:58:19,506 | INFO | train_grpo_train | metrics_logged mode=train step=2907

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.193939393939394e-06, 'num_tokens': 6594410.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00022611531312577426, 'sampling/sampling_logp_difference/max': 0.0029838387854397297, 'sampling/importance_sampling_ratio/min': 0.9999564290046692, 'sampling/importance_sampling_ratio/mean': 1.0002257823944092, 'sampling/importance_sampling_ratio/max': 1.002988338470459, 'entropy': 0.001633810141356662, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 88%|████████▊ | 2907/3300 [4:21:28<51:47,  7.91s/it]INFO 04-12 02:58:19 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2908/3300 [4:21:32<44:13,  6.77s/it]2026-04-12 02:58:23,618 | INFO | train_grpo_train | metrics_logged mode=train step=2908

                                                     
{'loss': -0.0005, 'grad_norm': 0.9043018221855164, 'learning_rate': 1.190909090909091e-06, 'num_tokens': 6595818.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9995806217193604, 'rewards/meter/std': 1.9758595954044722e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9995806217193604, 'rewards/total_composite/std': 1.9758595954044722e-05, 'reward': 0.9995806217193604, 'reward_std': 1.9746988982660696e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005839377176016569, 'sampling/sampling_logp_difference/max': 0.7329778671264648, 'sampling/importance_sampling_ratio/min': 0.4804760813713074, 'sampling/importance_sampling_ratio/mean': 1.0006860494613647, 'sampling/importance_sampling_ratio/max': 1.241074562072754, 'entropy': 0.03689346136525273, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9995806217193604, 'reward_meter_mean': 0.9995806217193604, 'reward_meter_std': 1.9758595954044722e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9995806217193604, 'reward_total_composite_std': 1.9758595954044722e-05, 'epoch': 0.12}

 88%|████████▊ | 2908/3300 [4:21:32<44:13,  6.77s/it]INFO 04-12 02:58:23 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2909/3300 [4:21:40<45:49,  7.03s/it]2026-04-12 02:58:31,258 | INFO | train_grpo_train | metrics_logged mode=train step=2909

                                                     
{'loss': -0.0212, 'grad_norm': 1.9418765306472778, 'learning_rate': 1.187878787878788e-06, 'num_tokens': 6600002.0, 'completions/mean_length': 306.0, 'completions/min_length': 284.0, 'completions/max_length': 319.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 306.0, 'completions/min_terminated_length': 284.0, 'completions/max_terminated_length': 319.0, 'rewards/meter/mean': 0.973264217376709, 'rewards/meter/std': 0.044348716735839844, 'rewards/count_adherence/mean': 0.887499988079071, 'rewards/count_adherence/std': 0.035355325788259506, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.962775707244873, 'rewards/repeat_penalty/std': 0.044049203395843506, 'rewards/total_composite/mean': 0.7390196323394775, 'rewards/total_composite/std': 0.3020605444908142, 'reward': 0.7390196323394775, 'reward_std': 0.3020605444908142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048917196691036224, 'sampling/sampling_logp_difference/max': 1.7572402954101562, 'sampling/importance_sampling_ratio/min': 0.17252032458782196, 'sampling/importance_sampling_ratio/mean': 1.0099446773529053, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5155597105622292, 'clip_ratio/low_mean': 0.0035211266949772835, 'clip_ratio/low_min': 0.0035211266949772835, 'clip_ratio/high_mean': 0.032285032561048865, 'clip_ratio/high_max': 0.032285032561048865, 'clip_ratio/region_mean': 0.03580615925602615, 'reward_total_mean': 0.7390196323394775, 'reward_meter_mean': 0.973264217376709, 'reward_meter_std': 0.044348716735839844, 'reward_count_adherence_mean': 0.887499988079071, 'reward_count_adherence_std': 0.035355325788259506, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.962775707244873, 'reward_repeat_penalty_std': 0.044049203395843506, 'reward_total_composite_mean': 0.7390196323394775, 'reward_total_composite_std': 0.3020605444908142, 'epoch': 0.12}

 88%|████████▊ | 2909/3300 [4:21:40<45:49,  7.03s/it]INFO 04-12 02:58:31 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2910/3300 [4:21:44<39:52,  6.13s/it]2026-04-12 02:58:35,304 | INFO | train_grpo_train | metrics_logged mode=train step=2910

                                                     
{'loss': 0.0001, 'grad_norm': 0.2488185316324234, 'learning_rate': 1.184848484848485e-06, 'num_tokens': 6601362.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9995900392532349, 'rewards/meter/std': 2.9798916330037173e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9995900392532349, 'rewards/total_composite/std': 2.9798916330037173e-06, 'reward': 0.9995900392532349, 'reward_std': 2.9798916330037173e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0049346331506967545, 'sampling/sampling_logp_difference/max': 0.7781157493591309, 'sampling/importance_sampling_ratio/min': 0.4592705965042114, 'sampling/importance_sampling_ratio/mean': 0.9999272227287292, 'sampling/importance_sampling_ratio/max': 1.1393342018127441, 'entropy': 0.02543548052199185, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9995900392532349, 'reward_meter_mean': 0.9995900392532349, 'reward_meter_std': 2.9798916330037173e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9995900392532349, 'reward_total_composite_std': 2.9798916330037173e-06, 'epoch': 0.12}

 88%|████████▊ | 2910/3300 [4:21:44<39:52,  6.13s/it]INFO 04-12 02:58:35 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2911/3300 [4:21:48<36:15,  5.59s/it]2026-04-12 02:58:39,626 | INFO | train_grpo_train | metrics_logged mode=train step=2911

                                                     
{'loss': -0.0001, 'grad_norm': 0.16482257843017578, 'learning_rate': 1.181818181818182e-06, 'num_tokens': 6603186.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981510639190674, 'rewards/meter/std': 1.0189157364948187e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981510639190674, 'rewards/total_composite/std': 1.0189157364948187e-05, 'reward': 0.9981510639190674, 'reward_std': 1.0199873031524476e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005824689287692308, 'sampling/sampling_logp_difference/max': 0.6331427097320557, 'sampling/importance_sampling_ratio/min': 0.530920684337616, 'sampling/importance_sampling_ratio/mean': 0.9997050762176514, 'sampling/importance_sampling_ratio/max': 1.1907520294189453, 'entropy': 0.03822743846103549, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981510639190674, 'reward_meter_mean': 0.9981510639190674, 'reward_meter_std': 1.0189157364948187e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981510639190674, 'reward_total_composite_std': 1.0189157364948187e-05, 'epoch': 0.12}

 88%|████████▊ | 2911/3300 [4:21:48<36:15,  5.59s/it]INFO 04-12 02:58:39 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2912/3300 [4:21:52<33:46,  5.22s/it]2026-04-12 02:58:43,990 | INFO | train_grpo_train | metrics_logged mode=train step=2912

                                                     
{'loss': -0.0024, 'grad_norm': 1.688647747039795, 'learning_rate': 1.1787878787878788e-06, 'num_tokens': 6605005.0, 'completions/mean_length': 67.375, 'completions/min_length': 66.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.99924635887146, 'rewards/meter/std': 0.0001372320402879268, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99924635887146, 'rewards/total_composite/std': 0.0001372320402879268, 'reward': 0.99924635887146, 'reward_std': 0.0001372358383378014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01559353992342949, 'sampling/sampling_logp_difference/max': 1.1167762279510498, 'sampling/importance_sampling_ratio/min': 0.32733336091041565, 'sampling/importance_sampling_ratio/mean': 1.0067250728607178, 'sampling/importance_sampling_ratio/max': 1.6964365243911743, 'entropy': 0.12552092969417572, 'clip_ratio/low_mean': 0.0074360816506668925, 'clip_ratio/low_min': 0.0074360816506668925, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.009301753249019384, 'reward_total_mean': 0.99924635887146, 'reward_meter_mean': 0.99924635887146, 'reward_meter_std': 0.0001372320402879268, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99924635887146, 'reward_total_composite_std': 0.0001372320402879268, 'epoch': 0.12}

 88%|████████▊ | 2912/3300 [4:21:52<33:46,  5.22s/it]INFO 04-12 02:58:44 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2913/3300 [4:21:57<32:15,  5.00s/it]2026-04-12 02:58:48,475 | INFO | train_grpo_train | metrics_logged mode=train step=2913

                                                     
{'loss': -0.0167, 'grad_norm': 5.435182094573975, 'learning_rate': 1.1757575757575759e-06, 'num_tokens': 6606638.0, 'completions/mean_length': 69.125, 'completions/min_length': 66.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9914849996566772, 'rewards/meter/std': 0.013738686218857765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9914849996566772, 'rewards/total_composite/std': 0.013738686218857765, 'reward': 0.9914849996566772, 'reward_std': 0.013738670386373997, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027403326705098152, 'sampling/sampling_logp_difference/max': 1.0196411609649658, 'sampling/importance_sampling_ratio/min': 0.3607243597507477, 'sampling/importance_sampling_ratio/mean': 1.0020140409469604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23639377392828465, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.024883363395929337, 'clip_ratio/high_max': 0.024883363395929337, 'clip_ratio/region_mean': 0.024883363395929337, 'reward_total_mean': 0.9914849996566772, 'reward_meter_mean': 0.9914849996566772, 'reward_meter_std': 0.013738686218857765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9914849996566772, 'reward_total_composite_std': 0.013738686218857765, 'epoch': 0.12}

 88%|████████▊ | 2913/3300 [4:21:57<32:15,  5.00s/it]INFO 04-12 02:58:48 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2914/3300 [4:22:04<36:45,  5.71s/it]2026-04-12 02:58:55,853 | INFO | train_grpo_train | metrics_logged mode=train step=2914

                                                     
{'loss': 0.004, 'grad_norm': 2.8395214080810547, 'learning_rate': 1.172727272727273e-06, 'num_tokens': 6610549.0, 'completions/mean_length': 302.875, 'completions/min_length': 291.0, 'completions/max_length': 312.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 302.875, 'completions/min_terminated_length': 291.0, 'completions/max_terminated_length': 312.0, 'rewards/meter/mean': 0.9837737083435059, 'rewards/meter/std': 0.04140230268239975, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9926470518112183, 'rewards/repeat_penalty/std': 0.020797256380319595, 'rewards/total_composite/mean': 0.8662627935409546, 'rewards/total_composite/std': 0.3506163954734802, 'reward': 0.8662627935409546, 'reward_std': 0.3506163954734802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059933777898550034, 'sampling/sampling_logp_difference/max': 1.6421480178833008, 'sampling/importance_sampling_ratio/min': 0.1935638189315796, 'sampling/importance_sampling_ratio/mean': 1.010536551475525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6028463244438171, 'clip_ratio/low_mean': 0.004125412553548813, 'clip_ratio/low_min': 0.004125412553548813, 'clip_ratio/high_mean': 0.052266900427639484, 'clip_ratio/high_max': 0.052266900427639484, 'clip_ratio/region_mean': 0.0563923129811883, 'reward_total_mean': 0.8662627935409546, 'reward_meter_mean': 0.9837737083435059, 'reward_meter_std': 0.04140230268239975, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9926470518112183, 'reward_repeat_penalty_std': 0.020797256380319595, 'reward_total_composite_mean': 0.8662627935409546, 'reward_total_composite_std': 0.3506163954734802, 'epoch': 0.12}

 88%|████████▊ | 2914/3300 [4:22:04<36:45,  5.71s/it]INFO 04-12 02:58:56 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2915/3300 [4:22:11<39:17,  6.12s/it]2026-04-12 02:59:02,931 | INFO | train_grpo_train | metrics_logged mode=train step=2915

                                                     
{'loss': 0.0239, 'grad_norm': 2.1781363487243652, 'learning_rate': 1.1696969696969697e-06, 'num_tokens': 6614339.0, 'completions/mean_length': 264.75, 'completions/min_length': 256.0, 'completions/max_length': 286.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 264.75, 'completions/min_terminated_length': 256.0, 'completions/max_terminated_length': 286.0, 'rewards/meter/mean': 0.9970918893814087, 'rewards/meter/std': 0.005339814815670252, 'rewards/count_adherence/mean': 0.984375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9676470756530762, 'rewards/repeat_penalty/std': 0.03468189761042595, 'rewards/total_composite/mean': 0.9504234790802002, 'rewards/total_composite/std': 0.06568735092878342, 'reward': 0.9504234790802002, 'reward_std': 0.06568736582994461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033314015716314316, 'sampling/sampling_logp_difference/max': 1.4124369621276855, 'sampling/importance_sampling_ratio/min': 0.24354904890060425, 'sampling/importance_sampling_ratio/mean': 1.0066348314285278, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3080044835805893, 'clip_ratio/low_mean': 0.017094367765821517, 'clip_ratio/low_min': 0.017094367765821517, 'clip_ratio/high_mean': 0.012379350140690804, 'clip_ratio/high_max': 0.012379350140690804, 'clip_ratio/region_mean': 0.02947371790651232, 'reward_total_mean': 0.9504234790802002, 'reward_meter_mean': 0.9970918893814087, 'reward_meter_std': 0.005339814815670252, 'reward_count_adherence_mean': 0.984375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9676470756530762, 'reward_repeat_penalty_std': 0.03468189761042595, 'reward_total_composite_mean': 0.9504234790802002, 'reward_total_composite_std': 0.06568735092878342, 'epoch': 0.12}

 88%|████████▊ | 2915/3300 [4:22:11<39:17,  6.12s/it]INFO 04-12 02:59:03 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2916/3300 [4:22:17<39:11,  6.12s/it]2026-04-12 02:59:09,063 | INFO | train_grpo_train | metrics_logged mode=train step=2916

                                                     
{'loss': -0.0014, 'grad_norm': 1.5669199228286743, 'learning_rate': 1.1666666666666668e-06, 'num_tokens': 6617355.0, 'completions/mean_length': 198.0, 'completions/min_length': 196.0, 'completions/max_length': 201.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 198.0, 'completions/min_terminated_length': 196.0, 'completions/max_terminated_length': 201.0, 'rewards/meter/mean': 0.9989873170852661, 'rewards/meter/std': 0.000616980018094182, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989873170852661, 'rewards/total_composite/std': 0.000616980018094182, 'reward': 0.9989873170852661, 'reward_std': 0.0006169742555357516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03752731904387474, 'sampling/sampling_logp_difference/max': 1.0977716445922852, 'sampling/importance_sampling_ratio/min': 0.33361366391181946, 'sampling/importance_sampling_ratio/mean': 1.0102366209030151, 'sampling/importance_sampling_ratio/max': 1.8285475969314575, 'entropy': 0.3740917034447193, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.020882656681351364, 'clip_ratio/high_max': 0.020882656681351364, 'clip_ratio/region_mean': 0.02277659613173455, 'reward_total_mean': 0.9989873170852661, 'reward_meter_mean': 0.9989873170852661, 'reward_meter_std': 0.000616980018094182, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989873170852661, 'reward_total_composite_std': 0.000616980018094182, 'epoch': 0.12}

 88%|████████▊ | 2916/3300 [4:22:18<39:11,  6.12s/it]INFO 04-12 02:59:09 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2917/3300 [4:22:22<36:33,  5.73s/it]2026-04-12 02:59:13,857 | INFO | train_grpo_train | metrics_logged mode=train step=2917

                                                     
{'loss': -0.0003, 'grad_norm': 2.548100233078003, 'learning_rate': 1.1636363636363638e-06, 'num_tokens': 6619535.0, 'completions/mean_length': 92.5, 'completions/min_length': 91.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.5, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9972745180130005, 'rewards/meter/std': 0.00045341101940721273, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972745180130005, 'rewards/total_composite/std': 0.00045341101940721273, 'reward': 0.9972745180130005, 'reward_std': 0.00045341774239204824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016107600182294846, 'sampling/sampling_logp_difference/max': 1.0648202896118164, 'sampling/importance_sampling_ratio/min': 0.3447898030281067, 'sampling/importance_sampling_ratio/mean': 1.0014476776123047, 'sampling/importance_sampling_ratio/max': 1.420615792274475, 'entropy': 0.12023854069411755, 'clip_ratio/low_mean': 0.008138206438161433, 'clip_ratio/low_min': 0.008138206438161433, 'clip_ratio/high_mean': 0.006749970489181578, 'clip_ratio/high_max': 0.006749970489181578, 'clip_ratio/region_mean': 0.014888176927343011, 'reward_total_mean': 0.9972745180130005, 'reward_meter_mean': 0.9972745180130005, 'reward_meter_std': 0.00045341101940721273, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972745180130005, 'reward_total_composite_std': 0.00045341101940721273, 'epoch': 0.12}

 88%|████████▊ | 2917/3300 [4:22:22<36:33,  5.73s/it]INFO 04-12 02:59:14 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2918/3300 [4:22:27<35:10,  5.53s/it]2026-04-12 02:59:18,912 | INFO | train_grpo_train | metrics_logged mode=train step=2918

                                                     
{'loss': 0.0044, 'grad_norm': 3.6516404151916504, 'learning_rate': 1.1606060606060607e-06, 'num_tokens': 6621518.0, 'completions/mean_length': 103.875, 'completions/min_length': 101.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.875, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9887614250183105, 'rewards/meter/std': 0.016851743683218956, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9887614250183105, 'rewards/total_composite/std': 0.016851743683218956, 'reward': 0.9887614250183105, 'reward_std': 0.01685173623263836, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038346562534570694, 'sampling/sampling_logp_difference/max': 1.4495925903320312, 'sampling/importance_sampling_ratio/min': 0.2346658855676651, 'sampling/importance_sampling_ratio/mean': 1.0085445642471313, 'sampling/importance_sampling_ratio/max': 1.9795823097229004, 'entropy': 0.2945725992321968, 'clip_ratio/low_mean': 0.0024271844886243343, 'clip_ratio/low_min': 0.0024271844886243343, 'clip_ratio/high_mean': 0.019217372057028115, 'clip_ratio/high_max': 0.019217372057028115, 'clip_ratio/region_mean': 0.02164455654565245, 'reward_total_mean': 0.9887614250183105, 'reward_meter_mean': 0.9887614250183105, 'reward_meter_std': 0.016851743683218956, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9887614250183105, 'reward_total_composite_std': 0.016851743683218956, 'epoch': 0.12}

 88%|████████▊ | 2918/3300 [4:22:27<35:10,  5.53s/it]INFO 04-12 02:59:19 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2919/3300 [4:22:32<34:08,  5.38s/it]2026-04-12 02:59:23,953 | INFO | train_grpo_train | metrics_logged mode=train step=2919

                                                     
{'loss': -0.0035, 'grad_norm': 1.5092811584472656, 'learning_rate': 1.1575757575757577e-06, 'num_tokens': 6623884.0, 'completions/mean_length': 106.75, 'completions/min_length': 106.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.75, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.999142587184906, 'rewards/meter/std': 0.0002281271299580112, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999142587184906, 'rewards/total_composite/std': 0.0002281271299580112, 'reward': 0.999142587184906, 'reward_std': 0.00022812940005678684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016602281481027603, 'sampling/sampling_logp_difference/max': 1.0120248794555664, 'sampling/importance_sampling_ratio/min': 0.3634822368621826, 'sampling/importance_sampling_ratio/mean': 1.0026243925094604, 'sampling/importance_sampling_ratio/max': 1.5648819208145142, 'entropy': 0.14118101634085178, 'clip_ratio/low_mean': 0.004705960163846612, 'clip_ratio/low_min': 0.004705960163846612, 'clip_ratio/high_mean': 0.003504672786220908, 'clip_ratio/high_max': 0.003504672786220908, 'clip_ratio/region_mean': 0.00821063295006752, 'reward_total_mean': 0.999142587184906, 'reward_meter_mean': 0.999142587184906, 'reward_meter_std': 0.0002281271299580112, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999142587184906, 'reward_total_composite_std': 0.0002281271299580112, 'epoch': 0.12}

 88%|████████▊ | 2919/3300 [4:22:32<34:08,  5.38s/it]INFO 04-12 02:59:24 [block_pool.py:378] Successfully reset prefix cache

 88%|████████▊ | 2920/3300 [4:22:37<33:27,  5.28s/it]2026-04-12 02:59:29,010 | INFO | train_grpo_train | metrics_logged mode=train step=2920

                                                     
{'loss': 0.002, 'grad_norm': 0.6831678152084351, 'learning_rate': 1.1545454545454545e-06, 'num_tokens': 6626176.0, 'completions/mean_length': 106.5, 'completions/min_length': 106.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.5, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9992160201072693, 'rewards/meter/std': 5.9427493397379294e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992160201072693, 'rewards/total_composite/std': 5.9427493397379294e-05, 'reward': 0.9992160201072693, 'reward_std': 5.943301584920846e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013828075490891933, 'sampling/sampling_logp_difference/max': 0.6171674728393555, 'sampling/importance_sampling_ratio/min': 0.5394703149795532, 'sampling/importance_sampling_ratio/mean': 1.0034005641937256, 'sampling/importance_sampling_ratio/max': 1.5806970596313477, 'entropy': 0.13282244093716145, 'clip_ratio/low_mean': 0.002358490601181984, 'clip_ratio/low_min': 0.002358490601181984, 'clip_ratio/high_mean': 0.00352671486325562, 'clip_ratio/high_max': 0.00352671486325562, 'clip_ratio/region_mean': 0.005885205464437604, 'reward_total_mean': 0.9992160201072693, 'reward_meter_mean': 0.9992160201072693, 'reward_meter_std': 5.9427493397379294e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992160201072693, 'reward_total_composite_std': 5.9427493397379294e-05, 'epoch': 0.12}

 88%|████████▊ | 2920/3300 [4:22:38<33:27,  5.28s/it]INFO 04-12 02:59:29 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2921/3300 [4:22:42<31:40,  5.01s/it]2026-04-12 02:59:33,396 | INFO | train_grpo_train | metrics_logged mode=train step=2921

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.1515151515151516e-06, 'num_tokens': 6627968.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0001586712896823883, 'sampling/sampling_logp_difference/max': 0.006689060479402542, 'sampling/importance_sampling_ratio/min': 0.9933332204818726, 'sampling/importance_sampling_ratio/mean': 1.0000739097595215, 'sampling/importance_sampling_ratio/max': 1.0063046216964722, 'entropy': 0.0019735290406970307, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▊ | 2921/3300 [4:22:42<31:40,  5.01s/it]INFO 04-12 02:59:33 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2922/3300 [4:22:46<30:16,  4.81s/it]2026-04-12 02:59:37,714 | INFO | train_grpo_train | metrics_logged mode=train step=2922

                                                     
{'loss': 0.0033, 'grad_norm': 0.311759889125824, 'learning_rate': 1.1484848484848486e-06, 'num_tokens': 6629489.0, 'completions/mean_length': 34.125, 'completions/min_length': 34.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9968487024307251, 'rewards/meter/std': 0.0025374931283295155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9968487024307251, 'rewards/total_composite/std': 0.0025374931283295155, 'reward': 0.9968487024307251, 'reward_std': 0.0025374931283295155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006284201052039862, 'sampling/sampling_logp_difference/max': 0.8642997741699219, 'sampling/importance_sampling_ratio/min': 0.4213464856147766, 'sampling/importance_sampling_ratio/mean': 1.0004445314407349, 'sampling/importance_sampling_ratio/max': 1.059367060661316, 'entropy': 0.03445210144855082, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0035714285913854837, 'reward_total_mean': 0.9968487024307251, 'reward_meter_mean': 0.9968487024307251, 'reward_meter_std': 0.0025374931283295155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9968487024307251, 'reward_total_composite_std': 0.0025374931283295155, 'epoch': 0.12}

 89%|████████▊ | 2922/3300 [4:22:46<30:16,  4.81s/it]INFO 04-12 02:59:37 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2923/3300 [4:22:52<32:10,  5.12s/it]2026-04-12 02:59:43,570 | INFO | train_grpo_train | metrics_logged mode=train step=2923

                                                     
{'loss': 0.0051, 'grad_norm': 4.833311557769775, 'learning_rate': 1.1454545454545457e-06, 'num_tokens': 6632353.0, 'completions/mean_length': 168.0, 'completions/min_length': 164.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.0, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.9983302354812622, 'rewards/meter/std': 0.0015840368578210473, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9983302354812622, 'rewards/total_composite/std': 0.0015840368578210473, 'reward': 0.9983302354812622, 'reward_std': 0.0015840278938412666, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04776981472969055, 'sampling/sampling_logp_difference/max': 1.0252695083618164, 'sampling/importance_sampling_ratio/min': 0.3586997985839844, 'sampling/importance_sampling_ratio/mean': 1.0118147134780884, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4091906175017357, 'clip_ratio/low_mean': 0.00808823574334383, 'clip_ratio/low_min': 0.00808823574334383, 'clip_ratio/high_mean': 0.03803631942719221, 'clip_ratio/high_max': 0.03803631942719221, 'clip_ratio/region_mean': 0.04612455517053604, 'reward_total_mean': 0.9983302354812622, 'reward_meter_mean': 0.9983302354812622, 'reward_meter_std': 0.0015840368578210473, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9983302354812622, 'reward_total_composite_std': 0.0015840368578210473, 'epoch': 0.12}

 89%|████████▊ | 2923/3300 [4:22:52<32:10,  5.12s/it]INFO 04-12 02:59:43 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2924/3300 [4:22:56<30:57,  4.94s/it]2026-04-12 02:59:48,092 | INFO | train_grpo_train | metrics_logged mode=train step=2924

                                                     
{'loss': -0.0, 'grad_norm': 0.0032307966612279415, 'learning_rate': 1.1424242424242425e-06, 'num_tokens': 6634097.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973385334014893, 'rewards/meter/std': 1.0115243185282452e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973385334014893, 'rewards/total_composite/std': 1.0115243185282452e-06, 'reward': 0.9973385334014893, 'reward_std': 1.0115243185282452e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0019246861338615417, 'sampling/sampling_logp_difference/max': 0.2358248233795166, 'sampling/importance_sampling_ratio/min': 0.9778724312782288, 'sampling/importance_sampling_ratio/mean': 1.0016769170761108, 'sampling/importance_sampling_ratio/max': 1.2659525871276855, 'entropy': 0.0171608105301857, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0020491802133619785, 'reward_total_mean': 0.9973385334014893, 'reward_meter_mean': 0.9973385334014893, 'reward_meter_std': 1.0115243185282452e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973385334014893, 'reward_total_composite_std': 1.0115243185282452e-06, 'epoch': 0.12}

 89%|████████▊ | 2924/3300 [4:22:57<30:57,  4.94s/it]INFO 04-12 02:59:48 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2925/3300 [4:23:02<31:55,  5.11s/it]2026-04-12 02:59:53,590 | INFO | train_grpo_train | metrics_logged mode=train step=2925

                                                     
{'loss': 0.0039, 'grad_norm': 1.0003553628921509, 'learning_rate': 1.1393939393939395e-06, 'num_tokens': 6636616.0, 'completions/mean_length': 141.875, 'completions/min_length': 140.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.875, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.999082088470459, 'rewards/meter/std': 0.00011351890861988068, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999082088470459, 'rewards/total_composite/std': 0.00011351890861988068, 'reward': 0.999082088470459, 'reward_std': 0.00011351749708410352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02242162451148033, 'sampling/sampling_logp_difference/max': 1.4046411514282227, 'sampling/importance_sampling_ratio/min': 0.24545511603355408, 'sampling/importance_sampling_ratio/mean': 1.0028555393218994, 'sampling/importance_sampling_ratio/max': 1.6239076852798462, 'entropy': 0.1869067270308733, 'clip_ratio/low_mean': 0.007904067635536194, 'clip_ratio/low_min': 0.007904067635536194, 'clip_ratio/high_mean': 0.00530050863744691, 'clip_ratio/high_max': 0.00530050863744691, 'clip_ratio/region_mean': 0.013204576272983104, 'reward_total_mean': 0.999082088470459, 'reward_meter_mean': 0.999082088470459, 'reward_meter_std': 0.00011351890861988068, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999082088470459, 'reward_total_composite_std': 0.00011351890861988068, 'epoch': 0.12}

 89%|████████▊ | 2925/3300 [4:23:02<31:55,  5.11s/it]INFO 04-12 02:59:53 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2926/3300 [4:23:08<34:03,  5.46s/it]2026-04-12 02:59:59,904 | INFO | train_grpo_train | metrics_logged mode=train step=2926

                                                     
{'loss': -0.0091, 'grad_norm': 3.0424153804779053, 'learning_rate': 1.1363636363636364e-06, 'num_tokens': 6639784.0, 'completions/mean_length': 208.0, 'completions/min_length': 201.0, 'completions/max_length': 220.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 208.0, 'completions/min_terminated_length': 201.0, 'completions/max_terminated_length': 220.0, 'rewards/meter/mean': 0.8537713289260864, 'rewards/meter/std': 0.2598452866077423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9204545617103577, 'rewards/repeat_penalty/std': 0.0758657231926918, 'rewards/total_composite/mean': 0.7895516157150269, 'rewards/total_composite/std': 0.25156062841415405, 'reward': 0.7895516157150269, 'reward_std': 0.25156062841415405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.046991463750600815, 'sampling/sampling_logp_difference/max': 1.488713264465332, 'sampling/importance_sampling_ratio/min': 0.22566285729408264, 'sampling/importance_sampling_ratio/mean': 1.012603998184204, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47782545536756516, 'clip_ratio/low_mean': 0.00918415142223239, 'clip_ratio/low_min': 0.00918415142223239, 'clip_ratio/high_mean': 0.029786661034449935, 'clip_ratio/high_max': 0.029786661034449935, 'clip_ratio/region_mean': 0.038970812456682324, 'reward_total_mean': 0.7895516157150269, 'reward_meter_mean': 0.8537713289260864, 'reward_meter_std': 0.2598452866077423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9204545617103577, 'reward_repeat_penalty_std': 0.0758657231926918, 'reward_total_composite_mean': 0.7895516157150269, 'reward_total_composite_std': 0.25156062841415405, 'epoch': 0.12}

 89%|████████▊ | 2926/3300 [4:23:08<34:03,  5.46s/it]INFO 04-12 03:00:00 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2927/3300 [4:23:12<31:28,  5.06s/it]2026-04-12 03:00:04,007 | INFO | train_grpo_train | metrics_logged mode=train step=2927

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.1333333333333334e-06, 'num_tokens': 6641264.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 4.449375410331413e-05, 'sampling/sampling_logp_difference/max': 0.00045358005445450544, 'sampling/importance_sampling_ratio/min': 0.999942421913147, 'sampling/importance_sampling_ratio/mean': 1.0000441074371338, 'sampling/importance_sampling_ratio/max': 1.0004537105560303, 'entropy': 0.0003692922255140729, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▊ | 2927/3300 [4:23:13<31:28,  5.06s/it]INFO 04-12 03:00:04 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▊ | 2928/3300 [4:23:19<34:01,  5.49s/it]2026-04-12 03:00:10,488 | INFO | train_grpo_train | metrics_logged mode=train step=2928

                                                     
{'loss': 0.003, 'grad_norm': 1.2671390771865845, 'learning_rate': 1.1303030303030305e-06, 'num_tokens': 6644813.0, 'completions/mean_length': 236.625, 'completions/min_length': 233.0, 'completions/max_length': 239.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 236.625, 'completions/min_terminated_length': 233.0, 'completions/max_terminated_length': 239.0, 'rewards/meter/mean': 0.9991750121116638, 'rewards/meter/std': 0.00013307879271451384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991750121116638, 'rewards/total_composite/std': 0.00013307879271451384, 'reward': 0.9991750121116638, 'reward_std': 0.00013307768676895648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035225577652454376, 'sampling/sampling_logp_difference/max': 1.075662612915039, 'sampling/importance_sampling_ratio/min': 0.3410716950893402, 'sampling/importance_sampling_ratio/mean': 1.0087391138076782, 'sampling/importance_sampling_ratio/max': 1.8521274328231812, 'entropy': 0.3537592738866806, 'clip_ratio/low_mean': 0.012113503995351493, 'clip_ratio/low_min': 0.012113503995351493, 'clip_ratio/high_mean': 0.007456094026565552, 'clip_ratio/high_max': 0.007456094026565552, 'clip_ratio/region_mean': 0.019569598021917045, 'reward_total_mean': 0.9991750121116638, 'reward_meter_mean': 0.9991750121116638, 'reward_meter_std': 0.00013307879271451384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991750121116638, 'reward_total_composite_std': 0.00013307879271451384, 'epoch': 0.12}

 89%|████████▊ | 2928/3300 [4:23:19<34:01,  5.49s/it]INFO 04-12 03:00:10 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2929/3300 [4:23:23<31:58,  5.17s/it]2026-04-12 03:00:14,926 | INFO | train_grpo_train | metrics_logged mode=train step=2929

                                                     
{'loss': -0.0, 'grad_norm': 0.040891580283641815, 'learning_rate': 1.1272727272727275e-06, 'num_tokens': 6646628.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981529116630554, 'rewards/meter/std': 2.428059815429151e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981529116630554, 'rewards/total_composite/std': 2.428059815429151e-06, 'reward': 0.9981529116630554, 'reward_std': 2.428059815429151e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0030419672839343548, 'sampling/sampling_logp_difference/max': 0.3047952651977539, 'sampling/importance_sampling_ratio/min': 0.7372743487358093, 'sampling/importance_sampling_ratio/mean': 1.0010944604873657, 'sampling/importance_sampling_ratio/max': 1.0871691703796387, 'entropy': 0.029220073018223047, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018939394503831863, 'reward_total_mean': 0.9981529116630554, 'reward_meter_mean': 0.9981529116630554, 'reward_meter_std': 2.428059815429151e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981529116630554, 'reward_total_composite_std': 2.428059815429151e-06, 'epoch': 0.12}

 89%|████████▉ | 2929/3300 [4:23:23<31:58,  5.17s/it]INFO 04-12 03:00:15 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2930/3300 [4:23:28<30:33,  4.96s/it]2026-04-12 03:00:19,373 | INFO | train_grpo_train | metrics_logged mode=train step=2930

                                                     
{'loss': -0.0005, 'grad_norm': 0.1988767832517624, 'learning_rate': 1.1242424242424243e-06, 'num_tokens': 6648339.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.998143196105957, 'rewards/meter/std': 1.6269057596218772e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998143196105957, 'rewards/total_composite/std': 1.6269057596218772e-05, 'reward': 0.998143196105957, 'reward_std': 1.6278554539894685e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007081957999616861, 'sampling/sampling_logp_difference/max': 0.8161485195159912, 'sampling/importance_sampling_ratio/min': 0.44213125109672546, 'sampling/importance_sampling_ratio/mean': 0.9996000528335571, 'sampling/importance_sampling_ratio/max': 1.115285038948059, 'entropy': 0.03760868404060602, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.998143196105957, 'reward_meter_mean': 0.998143196105957, 'reward_meter_std': 1.6269057596218772e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998143196105957, 'reward_total_composite_std': 1.6269057596218772e-05, 'epoch': 0.12}

 89%|████████▉ | 2930/3300 [4:23:28<30:33,  4.96s/it]INFO 04-12 03:00:19 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2931/3300 [4:23:32<29:46,  4.84s/it]2026-04-12 03:00:23,947 | INFO | train_grpo_train | metrics_logged mode=train step=2931

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.1212121212121214e-06, 'num_tokens': 6650299.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0029783351346850395, 'sampling/sampling_logp_difference/max': 1.665567398071289, 'sampling/importance_sampling_ratio/min': 0.18908332288265228, 'sampling/importance_sampling_ratio/mean': 0.9990981221199036, 'sampling/importance_sampling_ratio/max': 1.0083446502685547, 'entropy': 0.002932642324594781, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▉ | 2931/3300 [4:23:32<29:46,  4.84s/it]INFO 04-12 03:00:24 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2932/3300 [4:23:36<28:11,  4.60s/it]2026-04-12 03:00:27,968 | INFO | train_grpo_train | metrics_logged mode=train step=2932

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.1181818181818182e-06, 'num_tokens': 6651915.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 7.779466250212863e-05, 'sampling/sampling_logp_difference/max': 0.0008399296784773469, 'sampling/importance_sampling_ratio/min': 0.9997352361679077, 'sampling/importance_sampling_ratio/mean': 1.0000754594802856, 'sampling/importance_sampling_ratio/max': 1.000840187072754, 'entropy': 0.0006702992213831749, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▉ | 2932/3300 [4:23:36<28:11,  4.60s/it]INFO 04-12 03:00:28 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2933/3300 [4:23:41<27:43,  4.53s/it]2026-04-12 03:00:32,358 | INFO | train_grpo_train | metrics_logged mode=train step=2933

                                                     
{'loss': -0.0002, 'grad_norm': 0.009405174292623997, 'learning_rate': 1.1151515151515153e-06, 'num_tokens': 6653707.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981515407562256, 'rewards/meter/std': 1.501989686403249e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981515407562256, 'rewards/total_composite/std': 1.501989686403249e-06, 'reward': 0.9981515407562256, 'reward_std': 1.5057863720358e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0026725942734628916, 'sampling/sampling_logp_difference/max': 0.09437625110149384, 'sampling/importance_sampling_ratio/min': 0.9517659544944763, 'sampling/importance_sampling_ratio/mean': 1.002063512802124, 'sampling/importance_sampling_ratio/max': 1.0989731550216675, 'entropy': 0.028828308917582035, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.9981515407562256, 'reward_meter_mean': 0.9981515407562256, 'reward_meter_std': 1.501989686403249e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981515407562256, 'reward_total_composite_std': 1.501989686403249e-06, 'epoch': 0.12}

 89%|████████▉ | 2933/3300 [4:23:41<27:43,  4.53s/it]INFO 04-12 03:00:32 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2934/3300 [4:23:45<27:54,  4.58s/it]2026-04-12 03:00:37,029 | INFO | train_grpo_train | metrics_logged mode=train step=2934

                                                     
{'loss': -0.0064, 'grad_norm': 2.1487016677856445, 'learning_rate': 1.112121212121212e-06, 'num_tokens': 6655568.0, 'completions/mean_length': 78.625, 'completions/min_length': 77.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9986746311187744, 'rewards/meter/std': 0.0008188642095774412, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9986746311187744, 'rewards/total_composite/std': 0.0008188642095774412, 'reward': 0.9986746311187744, 'reward_std': 0.0008188657811842859, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020405031740665436, 'sampling/sampling_logp_difference/max': 0.8371596336364746, 'sampling/importance_sampling_ratio/min': 0.43293848633766174, 'sampling/importance_sampling_ratio/mean': 1.002895712852478, 'sampling/importance_sampling_ratio/max': 1.461742877960205, 'entropy': 0.18547690846025944, 'clip_ratio/low_mean': 0.001623376621864736, 'clip_ratio/low_min': 0.001623376621864736, 'clip_ratio/high_mean': 0.01744664926081896, 'clip_ratio/high_max': 0.01744664926081896, 'clip_ratio/region_mean': 0.019070025882683694, 'reward_total_mean': 0.9986746311187744, 'reward_meter_mean': 0.9986746311187744, 'reward_meter_std': 0.0008188642095774412, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9986746311187744, 'reward_total_composite_std': 0.0008188642095774412, 'epoch': 0.12}

 89%|████████▉ | 2934/3300 [4:23:46<27:54,  4.58s/it]INFO 04-12 03:00:37 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2935/3300 [4:23:50<28:20,  4.66s/it]2026-04-12 03:00:41,885 | INFO | train_grpo_train | metrics_logged mode=train step=2935

                                                     
{'loss': 0.0009, 'grad_norm': 0.6610411405563354, 'learning_rate': 1.1090909090909093e-06, 'num_tokens': 6657678.0, 'completions/mean_length': 92.75, 'completions/min_length': 92.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.75, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9977371692657471, 'rewards/meter/std': 7.424020441249013e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977371692657471, 'rewards/total_composite/std': 7.424020441249013e-05, 'reward': 0.9977371692657471, 'reward_std': 7.424017530865967e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011522673070430756, 'sampling/sampling_logp_difference/max': 0.7292871475219727, 'sampling/importance_sampling_ratio/min': 0.48225268721580505, 'sampling/importance_sampling_ratio/mean': 1.0004804134368896, 'sampling/importance_sampling_ratio/max': 1.3887077569961548, 'entropy': 0.09037415590137243, 'clip_ratio/low_mean': 0.005390953738242388, 'clip_ratio/low_min': 0.005390953738242388, 'clip_ratio/high_mean': 0.008094357093796134, 'clip_ratio/high_max': 0.008094357093796134, 'clip_ratio/region_mean': 0.013485310832038522, 'reward_total_mean': 0.9977371692657471, 'reward_meter_mean': 0.9977371692657471, 'reward_meter_std': 7.424020441249013e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977371692657471, 'reward_total_composite_std': 7.424020441249013e-05, 'epoch': 0.12}

 89%|████████▉ | 2935/3300 [4:23:50<28:20,  4.66s/it]INFO 04-12 03:00:42 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2936/3300 [4:24:00<38:13,  6.30s/it]2026-04-12 03:00:52,012 | INFO | train_grpo_train | metrics_logged mode=train step=2936

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.1060606060606062e-06, 'num_tokens': 6659326.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9992560148239136, 'rewards/meter/std': 9.200150816468522e-05, 'rewards/count_adherence/mean': 0.6499999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9709615111351013, 'rewards/repeat_penalty/std': 0.027279434725642204, 'rewards/total_composite/mean': 0.6306551694869995, 'rewards/total_composite/std': 0.01770690269768238, 'reward': 0.6306551694869995, 'reward_std': 0.017706912010908127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.6306551694869995, 'reward_meter_mean': 0.9992560148239136, 'reward_meter_std': 9.200150816468522e-05, 'reward_count_adherence_mean': 0.6499999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9709615111351013, 'reward_repeat_penalty_std': 0.027279434725642204, 'reward_total_composite_mean': 0.6306551694869995, 'reward_total_composite_std': 0.01770690269768238, 'epoch': 0.12}

 89%|████████▉ | 2936/3300 [4:24:01<38:13,  6.30s/it]INFO 04-12 03:00:52 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2937/3300 [4:24:05<35:03,  5.79s/it]2026-04-12 03:00:56,631 | INFO | train_grpo_train | metrics_logged mode=train step=2937

                                                     
{'loss': 0.0054, 'grad_norm': 1.2551523447036743, 'learning_rate': 1.1030303030303032e-06, 'num_tokens': 6661254.0, 'completions/mean_length': 79.0, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.0, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9990774989128113, 'rewards/meter/std': 0.00010208567255176604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990774989128113, 'rewards/total_composite/std': 0.00010208567255176604, 'reward': 0.9990774989128113, 'reward_std': 0.00010208544699708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01773812435567379, 'sampling/sampling_logp_difference/max': 0.474831223487854, 'sampling/importance_sampling_ratio/min': 0.6219900250434875, 'sampling/importance_sampling_ratio/mean': 1.0067074298858643, 'sampling/importance_sampling_ratio/max': 1.4563788175582886, 'entropy': 0.16411693580448627, 'clip_ratio/low_mean': 0.007854048046283424, 'clip_ratio/low_min': 0.007854048046283424, 'clip_ratio/high_mean': 0.006369685288518667, 'clip_ratio/high_max': 0.006369685288518667, 'clip_ratio/region_mean': 0.014223733334802091, 'reward_total_mean': 0.9990774989128113, 'reward_meter_mean': 0.9990774989128113, 'reward_meter_std': 0.00010208567255176604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990774989128113, 'reward_total_composite_std': 0.00010208567255176604, 'epoch': 0.12}

 89%|████████▉ | 2937/3300 [4:24:05<35:03,  5.79s/it]INFO 04-12 03:00:56 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2938/3300 [4:24:09<31:46,  5.27s/it]2026-04-12 03:01:00,664 | INFO | train_grpo_train | metrics_logged mode=train step=2938

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.1e-06, 'num_tokens': 6662614.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 9.101704199565575e-05, 'sampling/sampling_logp_difference/max': 0.003561503253877163, 'sampling/importance_sampling_ratio/min': 0.9964448809623718, 'sampling/importance_sampling_ratio/mean': 1.00003981590271, 'sampling/importance_sampling_ratio/max': 1.001272439956665, 'entropy': 0.0009229831339325756, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▉ | 2938/3300 [4:24:09<31:46,  5.27s/it]INFO 04-12 03:01:00 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2939/3300 [4:24:13<29:32,  4.91s/it]2026-04-12 03:01:04,743 | INFO | train_grpo_train | metrics_logged mode=train step=2939

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.096969696969697e-06, 'num_tokens': 6664078.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9995916485786438, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9995916485786438, 'rewards/total_composite/std': 0.0, 'reward': 0.9995916485786438, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.002187325619161129, 'sampling/sampling_logp_difference/max': 0.01813378371298313, 'sampling/importance_sampling_ratio/min': 0.9820296168327332, 'sampling/importance_sampling_ratio/mean': 1.001636266708374, 'sampling/importance_sampling_ratio/max': 1.0143886804580688, 'entropy': 0.023612842429429293, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9995916485786438, 'reward_meter_mean': 0.9995916485786438, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9995916485786438, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▉ | 2939/3300 [4:24:13<29:32,  4.91s/it]INFO 04-12 03:01:04 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2940/3300 [4:24:21<34:34,  5.76s/it]2026-04-12 03:01:12,497 | INFO | train_grpo_train | metrics_logged mode=train step=2940

                                                     
{'loss': 0.0184, 'grad_norm': 1.6798667907714844, 'learning_rate': 1.093939393939394e-06, 'num_tokens': 6668028.0, 'completions/mean_length': 290.75, 'completions/min_length': 284.0, 'completions/max_length': 304.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 290.75, 'completions/min_terminated_length': 284.0, 'completions/max_terminated_length': 304.0, 'rewards/meter/mean': 0.9990341663360596, 'rewards/meter/std': 9.884726023301482e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9598958492279053, 'rewards/repeat_penalty/std': 0.04713716357946396, 'rewards/total_composite/mean': 0.9589698314666748, 'rewards/total_composite/std': 0.04711604490876198, 'reward': 0.9589698314666748, 'reward_std': 0.04711604863405228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026417069137096405, 'sampling/sampling_logp_difference/max': 1.65692138671875, 'sampling/importance_sampling_ratio/min': 0.19072523713111877, 'sampling/importance_sampling_ratio/mean': 1.0048670768737793, 'sampling/importance_sampling_ratio/max': 1.720551609992981, 'entropy': 0.299512580037117, 'clip_ratio/low_mean': 0.007625087164342403, 'clip_ratio/low_min': 0.007625087164342403, 'clip_ratio/high_mean': 0.009173946105875075, 'clip_ratio/high_max': 0.009173946105875075, 'clip_ratio/region_mean': 0.01679903327021748, 'reward_total_mean': 0.9589698314666748, 'reward_meter_mean': 0.9990341663360596, 'reward_meter_std': 9.884726023301482e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9598958492279053, 'reward_repeat_penalty_std': 0.04713716357946396, 'reward_total_composite_mean': 0.9589698314666748, 'reward_total_composite_std': 0.04711604490876198, 'epoch': 0.12}

 89%|████████▉ | 2940/3300 [4:24:21<34:34,  5.76s/it]INFO 04-12 03:01:12 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2941/3300 [4:24:27<34:28,  5.76s/it]2026-04-12 03:01:18,260 | INFO | train_grpo_train | metrics_logged mode=train step=2941

                                                     
{'loss': 0.0035, 'grad_norm': 1.4318662881851196, 'learning_rate': 1.090909090909091e-06, 'num_tokens': 6670786.0, 'completions/mean_length': 165.75, 'completions/min_length': 164.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.75, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9992879629135132, 'rewards/meter/std': 0.0003045983612537384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9715272188186646, 'rewards/total_composite/std': 0.05134394392371178, 'reward': 0.9715272188186646, 'reward_std': 0.05134394019842148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015097087249159813, 'sampling/sampling_logp_difference/max': 1.1712437868118286, 'sampling/importance_sampling_ratio/min': 0.3099811375141144, 'sampling/importance_sampling_ratio/mean': 1.0033372640609741, 'sampling/importance_sampling_ratio/max': 1.513466238975525, 'entropy': 0.11971556767821312, 'clip_ratio/low_mean': 0.003765060333535075, 'clip_ratio/low_min': 0.003765060333535075, 'clip_ratio/high_mean': 0.009821378625929356, 'clip_ratio/high_max': 0.009821378625929356, 'clip_ratio/region_mean': 0.01358643895946443, 'reward_total_mean': 0.9715272188186646, 'reward_meter_mean': 0.9992879629135132, 'reward_meter_std': 0.0003045983612537384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9715272188186646, 'reward_total_composite_std': 0.05134394392371178, 'epoch': 0.12}

 89%|████████▉ | 2941/3300 [4:24:27<34:28,  5.76s/it]INFO 04-12 03:01:18 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2942/3300 [4:24:34<36:53,  6.18s/it]2026-04-12 03:01:25,424 | INFO | train_grpo_train | metrics_logged mode=train step=2942

                                                     
{'loss': -0.0003, 'grad_norm': 2.797619581222534, 'learning_rate': 1.087878787878788e-06, 'num_tokens': 6674569.0, 'completions/mean_length': 278.875, 'completions/min_length': 275.0, 'completions/max_length': 282.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 278.875, 'completions/min_terminated_length': 275.0, 'completions/max_terminated_length': 282.0, 'rewards/meter/mean': 0.9974960684776306, 'rewards/meter/std': 0.000377756921807304, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9338235855102539, 'rewards/repeat_penalty/std': 0.020797256380319595, 'rewards/total_composite/mean': 0.8150506019592285, 'rewards/total_composite/std': 0.018201656639575958, 'reward': 0.8150506019592285, 'reward_std': 0.018201671540737152, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03948317840695381, 'sampling/sampling_logp_difference/max': 1.47027587890625, 'sampling/importance_sampling_ratio/min': 0.22986207902431488, 'sampling/importance_sampling_ratio/mean': 1.0059531927108765, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31575205735862255, 'clip_ratio/low_mean': 0.0027173913549631834, 'clip_ratio/low_min': 0.0027173913549631834, 'clip_ratio/high_mean': 0.026906014885753393, 'clip_ratio/high_max': 0.026906014885753393, 'clip_ratio/region_mean': 0.029623406240716577, 'reward_total_mean': 0.8150506019592285, 'reward_meter_mean': 0.9974960684776306, 'reward_meter_std': 0.000377756921807304, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9338235855102539, 'reward_repeat_penalty_std': 0.020797256380319595, 'reward_total_composite_mean': 0.8150506019592285, 'reward_total_composite_std': 0.018201656639575958, 'epoch': 0.12}

 89%|████████▉ | 2942/3300 [4:24:34<36:53,  6.18s/it]INFO 04-12 03:01:25 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2943/3300 [4:24:41<38:42,  6.51s/it]2026-04-12 03:01:32,695 | INFO | train_grpo_train | metrics_logged mode=train step=2943

                                                     
{'loss': 0.0233, 'grad_norm': 3.1147727966308594, 'learning_rate': 1.084848484848485e-06, 'num_tokens': 6678237.0, 'completions/mean_length': 271.5, 'completions/min_length': 258.0, 'completions/max_length': 284.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 271.5, 'completions/min_terminated_length': 258.0, 'completions/max_terminated_length': 284.0, 'rewards/meter/mean': 0.9941989779472351, 'rewards/meter/std': 0.0028337803669273853, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9489378929138184, 'rewards/repeat_penalty/std': 0.02066386677324772, 'rewards/total_composite/mean': 0.8254706859588623, 'rewards/total_composite/std': 0.016226978972554207, 'reward': 0.8254706859588623, 'reward_std': 0.0162269975990057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.049873415380716324, 'sampling/sampling_logp_difference/max': 3.8973119258880615, 'sampling/importance_sampling_ratio/min': 0.02029639668762684, 'sampling/importance_sampling_ratio/mean': 1.0050019025802612, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37596092373132706, 'clip_ratio/low_mean': 0.02794346760492772, 'clip_ratio/low_min': 0.02794346760492772, 'clip_ratio/high_mean': 0.006298449821770191, 'clip_ratio/high_max': 0.006298449821770191, 'clip_ratio/region_mean': 0.03424191742669791, 'reward_total_mean': 0.8254706859588623, 'reward_meter_mean': 0.9941989779472351, 'reward_meter_std': 0.0028337803669273853, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9489378929138184, 'reward_repeat_penalty_std': 0.02066386677324772, 'reward_total_composite_mean': 0.8254706859588623, 'reward_total_composite_std': 0.016226978972554207, 'epoch': 0.12}

 89%|████████▉ | 2943/3300 [4:24:41<38:42,  6.51s/it]INFO 04-12 03:01:32 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2944/3300 [4:24:46<35:05,  5.91s/it]2026-04-12 03:01:37,215 | INFO | train_grpo_train | metrics_logged mode=train step=2944

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.081818181818182e-06, 'num_tokens': 6680078.0, 'completions/mean_length': 67.125, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9981522560119629, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981522560119629, 'rewards/total_composite/std': 0.0, 'reward': 0.9981522560119629, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0032630774658173323, 'sampling/sampling_logp_difference/max': 0.3072841167449951, 'sampling/importance_sampling_ratio/min': 0.946239173412323, 'sampling/importance_sampling_ratio/mean': 1.0029489994049072, 'sampling/importance_sampling_ratio/max': 1.359727144241333, 'entropy': 0.027955329976975918, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9981522560119629, 'reward_meter_mean': 0.9981522560119629, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981522560119629, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▉ | 2944/3300 [4:24:46<35:05,  5.91s/it]INFO 04-12 03:01:37 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2945/3300 [4:24:50<33:03,  5.59s/it]2026-04-12 03:01:42,044 | INFO | train_grpo_train | metrics_logged mode=train step=2945

                                                     
{'loss': -0.0019, 'grad_norm': 1.8361525535583496, 'learning_rate': 1.078787878787879e-06, 'num_tokens': 6681590.0, 'completions/mean_length': 68.0, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9993088841438293, 'rewards/meter/std': 0.0001494528987677768, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993088841438293, 'rewards/total_composite/std': 0.0001494528987677768, 'reward': 0.9993088841438293, 'reward_std': 0.0001494480820838362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01226960588246584, 'sampling/sampling_logp_difference/max': 0.6763467788696289, 'sampling/importance_sampling_ratio/min': 0.5084711313247681, 'sampling/importance_sampling_ratio/mean': 1.0019557476043701, 'sampling/importance_sampling_ratio/max': 1.2217551469802856, 'entropy': 0.08658721391111612, 'clip_ratio/low_mean': 0.01105684821959585, 'clip_ratio/low_min': 0.01105684821959585, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.012895083520561457, 'reward_total_mean': 0.9993088841438293, 'reward_meter_mean': 0.9993088841438293, 'reward_meter_std': 0.0001494528987677768, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993088841438293, 'reward_total_composite_std': 0.0001494528987677768, 'epoch': 0.12}

 89%|████████▉ | 2945/3300 [4:24:51<33:03,  5.59s/it]INFO 04-12 03:01:42 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2946/3300 [4:24:55<31:00,  5.26s/it]2026-04-12 03:01:46,528 | INFO | train_grpo_train | metrics_logged mode=train step=2946

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.0757575757575758e-06, 'num_tokens': 6683270.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003203663509339094, 'sampling/sampling_logp_difference/max': 0.004554374143481255, 'sampling/importance_sampling_ratio/min': 0.9995915293693542, 'sampling/importance_sampling_ratio/mean': 1.0003174543380737, 'sampling/importance_sampling_ratio/max': 1.0045647621154785, 'entropy': 0.0023213086824398488, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 89%|████████▉ | 2946/3300 [4:24:55<31:00,  5.26s/it]INFO 04-12 03:01:46 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2947/3300 [4:25:00<31:03,  5.28s/it]2026-04-12 03:01:51,860 | INFO | train_grpo_train | metrics_logged mode=train step=2947

                                                     
{'loss': 0.0031, 'grad_norm': 3.6917295455932617, 'learning_rate': 1.0727272727272728e-06, 'num_tokens': 6685877.0, 'completions/mean_length': 118.875, 'completions/min_length': 115.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.875, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9956615567207336, 'rewards/meter/std': 0.0021291025914251804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.9246125221252441, 'rewards/total_composite/std': 0.07701455056667328, 'reward': 0.9246125221252441, 'reward_std': 0.07701455056667328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03363678976893425, 'sampling/sampling_logp_difference/max': 1.5754187107086182, 'sampling/importance_sampling_ratio/min': 0.20692089200019836, 'sampling/importance_sampling_ratio/mean': 1.0035653114318848, 'sampling/importance_sampling_ratio/max': 1.86605966091156, 'entropy': 0.20067131519317627, 'clip_ratio/low_mean': 0.02329266769811511, 'clip_ratio/low_min': 0.02329266769811511, 'clip_ratio/high_mean': 0.015964585822075605, 'clip_ratio/high_max': 0.015964585822075605, 'clip_ratio/region_mean': 0.039257253520190716, 'reward_total_mean': 0.9246125221252441, 'reward_meter_mean': 0.9956615567207336, 'reward_meter_std': 0.0021291025914251804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.9246125221252441, 'reward_total_composite_std': 0.07701455056667328, 'epoch': 0.12}

 89%|████████▉ | 2947/3300 [4:25:00<31:03,  5.28s/it]INFO 04-12 03:01:52 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2948/3300 [4:25:05<30:14,  5.16s/it]2026-04-12 03:01:56,728 | INFO | train_grpo_train | metrics_logged mode=train step=2948

                                                     
{'loss': 0.0056, 'grad_norm': 3.9713573455810547, 'learning_rate': 1.0696969696969696e-06, 'num_tokens': 6687719.0, 'completions/mean_length': 69.25, 'completions/min_length': 66.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9902843236923218, 'rewards/meter/std': 0.01087514590471983, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9902843236923218, 'rewards/total_composite/std': 0.01087514590471983, 'reward': 0.9902843236923218, 'reward_std': 0.010875147767364979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040240850299596786, 'sampling/sampling_logp_difference/max': 1.782811164855957, 'sampling/importance_sampling_ratio/min': 0.16816474497318268, 'sampling/importance_sampling_ratio/mean': 0.9977670311927795, 'sampling/importance_sampling_ratio/max': 1.6766802072525024, 'entropy': 0.31914436258375645, 'clip_ratio/low_mean': 0.007194617064669728, 'clip_ratio/low_min': 0.007194617064669728, 'clip_ratio/high_mean': 0.01954207243397832, 'clip_ratio/high_max': 0.01954207243397832, 'clip_ratio/region_mean': 0.026736689498648047, 'reward_total_mean': 0.9902843236923218, 'reward_meter_mean': 0.9902843236923218, 'reward_meter_std': 0.01087514590471983, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9902843236923218, 'reward_total_composite_std': 0.01087514590471983, 'epoch': 0.12}

 89%|████████▉ | 2948/3300 [4:25:05<30:14,  5.16s/it]INFO 04-12 03:01:56 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2949/3300 [4:25:11<30:42,  5.25s/it]2026-04-12 03:02:02,193 | INFO | train_grpo_train | metrics_logged mode=train step=2949

                                                     
{'loss': -0.001, 'grad_norm': 2.246262311935425, 'learning_rate': 1.066666666666667e-06, 'num_tokens': 6690261.0, 'completions/mean_length': 128.75, 'completions/min_length': 128.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.75, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9919617176055908, 'rewards/meter/std': 0.01668241061270237, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9919617176055908, 'rewards/total_composite/std': 0.01668241061270237, 'reward': 0.9919617176055908, 'reward_std': 0.01668240688741207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017122527584433556, 'sampling/sampling_logp_difference/max': 1.4149751663208008, 'sampling/importance_sampling_ratio/min': 0.24293164908885956, 'sampling/importance_sampling_ratio/mean': 0.999458909034729, 'sampling/importance_sampling_ratio/max': 1.6150113344192505, 'entropy': 0.15123706310987473, 'clip_ratio/low_mean': 0.0019379844889044762, 'clip_ratio/low_min': 0.0019379844889044762, 'clip_ratio/high_mean': 0.01359617244452238, 'clip_ratio/high_max': 0.01359617244452238, 'clip_ratio/region_mean': 0.015534156933426857, 'reward_total_mean': 0.9919617176055908, 'reward_meter_mean': 0.9919617176055908, 'reward_meter_std': 0.01668241061270237, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9919617176055908, 'reward_total_composite_std': 0.01668241061270237, 'epoch': 0.12}

 89%|████████▉ | 2949/3300 [4:25:11<30:42,  5.25s/it]INFO 04-12 03:02:02 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2950/3300 [4:25:16<30:27,  5.22s/it]2026-04-12 03:02:07,356 | INFO | train_grpo_train | metrics_logged mode=train step=2950

                                                     
{'loss': -0.0107, 'grad_norm': 4.775022029876709, 'learning_rate': 1.0636363636363637e-06, 'num_tokens': 6692334.0, 'completions/mean_length': 90.125, 'completions/min_length': 88.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.125, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9957476258277893, 'rewards/meter/std': 0.0016542106168344617, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.8999999761581421, 'rewards/repeat_penalty/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.8962882161140442, 'rewards/total_composite/std': 0.10756382346153259, 'reward': 0.8962882161140442, 'reward_std': 0.107563816010952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03922773897647858, 'sampling/sampling_logp_difference/max': 3.326974630355835, 'sampling/importance_sampling_ratio/min': 0.03590155765414238, 'sampling/importance_sampling_ratio/mean': 0.9992499947547913, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16245986334979534, 'clip_ratio/low_mean': 0.009927221923135221, 'clip_ratio/low_min': 0.009927221923135221, 'clip_ratio/high_mean': 0.009542036801576614, 'clip_ratio/high_max': 0.009542036801576614, 'clip_ratio/region_mean': 0.019469258724711835, 'reward_total_mean': 0.8962882161140442, 'reward_meter_mean': 0.9957476258277893, 'reward_meter_std': 0.0016542106168344617, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.8999999761581421, 'reward_repeat_penalty_std': 0.10690449178218842, 'reward_total_composite_mean': 0.8962882161140442, 'reward_total_composite_std': 0.10756382346153259, 'epoch': 0.12}

 89%|████████▉ | 2950/3300 [4:25:16<30:27,  5.22s/it]INFO 04-12 03:02:07 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:38,  3.48s/it]

 23%|██▎       | 3/13 [00:10<00:35,  3.50s/it]

 31%|███       | 4/13 [00:16<00:38,  4.29s/it]

 38%|███▊      | 5/13 [00:21<00:36,  4.58s/it]

 46%|████▌     | 6/13 [00:27<00:36,  5.25s/it]

 54%|█████▍    | 7/13 [00:32<00:31,  5.22s/it]

 62%|██████▏   | 8/13 [00:39<00:28,  5.66s/it]

 69%|██████▉   | 9/13 [00:46<00:24,  6.18s/it]

 77%|███████▋  | 10/13 [00:51<00:17,  5.71s/it]

 85%|████████▍ | 11/13 [00:57<00:11,  5.75s/it]

 92%|█████████▏| 12/13 [01:03<00:05,  5.88s/it]

100%|██████████| 13/13 [01:08<00:00,  5.70s/it]2026-04-12 03:03:22,934 | INFO | train_grpo_train | metrics_logged mode=eval step=2950

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 75.4625, 'eval_samples_per_second': 1.378, 'eval_steps_per_second': 0.172, 'eval_num_tokens': 6692334.0, 'eval_completions/mean_length': 208.25, 'eval_completions/min_length': 60.15384615384615, 'eval_completions/max_length': 403.61538461538464, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 205.135989849384, 'eval_completions/min_terminated_length': 60.15384615384615, 'eval_completions/max_terminated_length': 395.0769230769231, 'eval_rewards/meter/mean': 0.7935946950545678, 'eval_rewards/meter/std': 0.32434708754030556, 'eval_rewards/count_adherence/mean': 0.9516645761636587, 'eval_rewards/count_adherence/std': 0.06901963714223641, 'eval_rewards/arabic_clean/mean': 0.9711538461538461, 'eval_rewards/arabic_clean/std': 0.06280488005051246, 'eval_rewards/repeat_penalty/mean': 0.9551343642748319, 'eval_rewards/repeat_penalty/std': 0.08046436768311721, 'eval_rewards/total_composite/mean': 0.7131056900207813, 'eval_rewards/total_composite/std': 0.3318366717833739, 'eval_reward': 0.7131056900207813, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.032634207692283854, 'eval_sampling/sampling_logp_difference/max': 1.3049519062042236, 'eval_sampling/importance_sampling_ratio/min': 0.27521977172448087, 'eval_sampling/importance_sampling_ratio/mean': 1.0077741604584913, 'eval_sampling/importance_sampling_ratio/max': 1.446980045391963, 'eval_entropy': 0.38887230020302993, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7131056900207813, 'eval_reward_meter_mean': 0.7935946950545678, 'eval_reward_meter_std': 0.32434708754030556, 'eval_reward_count_adherence_mean': 0.9516645761636587, 'eval_reward_count_adherence_std': 0.06901963714223641, 'eval_reward_arabic_clean_mean': 0.9711538461538461, 'eval_reward_arabic_clean_std': 0.06280488005051246, 'eval_reward_repeat_penalty_mean': 0.9551343642748319, 'eval_reward_repeat_penalty_std': 0.08046436768311721, 'eval_reward_total_composite_mean': 0.7131056900207813, 'eval_reward_total_composite_std': 0.3318366717833739, 'epoch': 0.12}

 89%|████████▉ | 2950/3300 [4:26:31<30:27,  5.22s/it]

100%|██████████| 13/13 [01:08<00:00,  5.70s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 89%|████████▉ | 2951/3300 [4:26:38<2:45:22, 28.43s/it]2026-04-12 03:03:29,949 | INFO | train_grpo_train | metrics_logged mode=train step=2951

                                                       
{'loss': -0.0003, 'grad_norm': 7.092353820800781, 'learning_rate': 1.0606060606060608e-06, 'num_tokens': 6693891.0, 'completions/mean_length': 33.625, 'completions/min_length': 32.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9904308915138245, 'rewards/meter/std': 0.004542448092252016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9904308915138245, 'rewards/total_composite/std': 0.004542448092252016, 'reward': 0.9904308915138245, 'reward_std': 0.004542447626590729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021938247606158257, 'sampling/sampling_logp_difference/max': 1.2846660614013672, 'sampling/importance_sampling_ratio/min': 0.27674299478530884, 'sampling/importance_sampling_ratio/mean': 1.0013483762741089, 'sampling/importance_sampling_ratio/max': 1.398740530014038, 'entropy': 0.1007662764750421, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0037878789007663727, 'reward_total_mean': 0.9904308915138245, 'reward_meter_mean': 0.9904308915138245, 'reward_meter_std': 0.004542448092252016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9904308915138245, 'reward_total_composite_std': 0.004542448092252016, 'epoch': 0.12}

 89%|████████▉ | 2951/3300 [4:26:38<2:45:22, 28.43s/it]INFO 04-12 03:03:30 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2952/3300 [4:26:43<2:03:30, 21.30s/it]2026-04-12 03:03:34,591 | INFO | train_grpo_train | metrics_logged mode=train step=2952

                                                       
{'loss': 0.0018, 'grad_norm': 0.5699337124824524, 'learning_rate': 1.0575757575757576e-06, 'num_tokens': 6695487.0, 'completions/mean_length': 36.5, 'completions/min_length': 36.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9995974898338318, 'rewards/meter/std': 1.9893313947250135e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9995974898338318, 'rewards/total_composite/std': 1.9893313947250135e-05, 'reward': 0.9995974898338318, 'reward_std': 1.9893312128260732e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009155918844044209, 'sampling/sampling_logp_difference/max': 0.8658480644226074, 'sampling/importance_sampling_ratio/min': 0.42069464921951294, 'sampling/importance_sampling_ratio/mean': 0.9967590570449829, 'sampling/importance_sampling_ratio/max': 1.0766030550003052, 'entropy': 0.03357855952344835, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.013795045204460621, 'clip_ratio/high_max': 0.013795045204460621, 'clip_ratio/region_mean': 0.013795045204460621, 'reward_total_mean': 0.9995974898338318, 'reward_meter_mean': 0.9995974898338318, 'reward_meter_std': 1.9893313947250135e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9995974898338318, 'reward_total_composite_std': 1.9893313947250135e-05, 'epoch': 0.12}

 89%|████████▉ | 2952/3300 [4:26:43<2:03:30, 21.30s/it]INFO 04-12 03:03:34 [block_pool.py:378] Successfully reset prefix cache

 89%|████████▉ | 2953/3300 [4:26:48<1:35:23, 16.49s/it]2026-04-12 03:03:39,871 | INFO | train_grpo_train | metrics_logged mode=train step=2953

                                                       
{'loss': -0.0001, 'grad_norm': 1.022800087928772, 'learning_rate': 1.0545454545454547e-06, 'num_tokens': 6697761.0, 'completions/mean_length': 107.25, 'completions/min_length': 106.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9992892742156982, 'rewards/meter/std': 9.131882688961923e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992892742156982, 'rewards/total_composite/std': 9.131882688961923e-05, 'reward': 0.9992892742156982, 'reward_std': 9.13253243197687e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016012487933039665, 'sampling/sampling_logp_difference/max': 0.9576692581176758, 'sampling/importance_sampling_ratio/min': 0.3837863802909851, 'sampling/importance_sampling_ratio/mean': 1.0017812252044678, 'sampling/importance_sampling_ratio/max': 1.6504535675048828, 'entropy': 0.13113553449511528, 'clip_ratio/low_mean': 0.002336448524147272, 'clip_ratio/low_min': 0.002336448524147272, 'clip_ratio/high_mean': 0.00930431904271245, 'clip_ratio/high_max': 0.00930431904271245, 'clip_ratio/region_mean': 0.011640767566859722, 'reward_total_mean': 0.9992892742156982, 'reward_meter_mean': 0.9992892742156982, 'reward_meter_std': 9.131882688961923e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992892742156982, 'reward_total_composite_std': 9.131882688961923e-05, 'epoch': 0.12}

 89%|████████▉ | 2953/3300 [4:26:48<1:35:23, 16.49s/it]INFO 04-12 03:03:40 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2954/3300 [4:26:53<1:14:34, 12.93s/it]2026-04-12 03:03:44,504 | INFO | train_grpo_train | metrics_logged mode=train step=2954

                                                       
{'loss': -0.0034, 'grad_norm': 10.395458221435547, 'learning_rate': 1.0515151515151515e-06, 'num_tokens': 6699386.0, 'completions/mean_length': 42.125, 'completions/min_length': 40.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9281213283538818, 'rewards/meter/std': 0.031521886587142944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9281213283538818, 'rewards/total_composite/std': 0.031521886587142944, 'reward': 0.9281213283538818, 'reward_std': 0.03152187913656235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08637839555740356, 'sampling/sampling_logp_difference/max': 2.8028721809387207, 'sampling/importance_sampling_ratio/min': 0.060635656118392944, 'sampling/importance_sampling_ratio/mean': 0.9940267205238342, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30831943079829216, 'clip_ratio/low_mean': 0.024695121683180332, 'clip_ratio/low_min': 0.024695121683180332, 'clip_ratio/high_mean': 0.04968656366690993, 'clip_ratio/high_max': 0.04968656366690993, 'clip_ratio/region_mean': 0.07438168535009027, 'reward_total_mean': 0.9281213283538818, 'reward_meter_mean': 0.9281213283538818, 'reward_meter_std': 0.031521886587142944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9281213283538818, 'reward_total_composite_std': 0.031521886587142944, 'epoch': 0.12}

 90%|████████▉ | 2954/3300 [4:26:53<1:14:34, 12.93s/it]INFO 04-12 03:03:44 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2955/3300 [4:26:59<1:02:49, 10.93s/it]2026-04-12 03:03:50,750 | INFO | train_grpo_train | metrics_logged mode=train step=2955

                                                       
{'loss': 0.0177, 'grad_norm': 3.04614520072937, 'learning_rate': 1.0484848484848485e-06, 'num_tokens': 6702346.0, 'completions/mean_length': 174.0, 'completions/min_length': 166.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 174.0, 'completions/min_terminated_length': 166.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.9902915358543396, 'rewards/meter/std': 0.007541172672063112, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9902915358543396, 'rewards/total_composite/std': 0.007541172672063112, 'reward': 0.9902915358543396, 'reward_std': 0.007541188970208168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04366350546479225, 'sampling/sampling_logp_difference/max': 1.370412826538086, 'sampling/importance_sampling_ratio/min': 0.25400206446647644, 'sampling/importance_sampling_ratio/mean': 1.0090171098709106, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3894418552517891, 'clip_ratio/low_mean': 0.00557116128038615, 'clip_ratio/low_min': 0.00557116128038615, 'clip_ratio/high_mean': 0.022244701394811273, 'clip_ratio/high_max': 0.022244701394811273, 'clip_ratio/region_mean': 0.027815862675197423, 'reward_total_mean': 0.9902915358543396, 'reward_meter_mean': 0.9902915358543396, 'reward_meter_std': 0.007541172672063112, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9902915358543396, 'reward_total_composite_std': 0.007541172672063112, 'epoch': 0.12}

 90%|████████▉ | 2955/3300 [4:26:59<1:02:49, 10.93s/it]INFO 04-12 03:03:50 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2956/3300 [4:27:04<52:07,  9.09s/it]  2026-04-12 03:03:55,568 | INFO | train_grpo_train | metrics_logged mode=train step=2956

                                                     
{'loss': 0.0014, 'grad_norm': 2.19303035736084, 'learning_rate': 1.0454545454545456e-06, 'num_tokens': 6704123.0, 'completions/mean_length': 68.125, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9994041919708252, 'rewards/meter/std': 0.0001539249497000128, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994041919708252, 'rewards/total_composite/std': 0.0001539249497000128, 'reward': 0.9994041919708252, 'reward_std': 0.0001539262884762138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012835102155804634, 'sampling/sampling_logp_difference/max': 1.0354909896850586, 'sampling/importance_sampling_ratio/min': 0.35505202412605286, 'sampling/importance_sampling_ratio/mean': 1.0002238750457764, 'sampling/importance_sampling_ratio/max': 1.366014838218689, 'entropy': 0.0865520965307951, 'clip_ratio/low_mean': 0.007299659075215459, 'clip_ratio/low_min': 0.007299659075215459, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.009137894376181066, 'reward_total_mean': 0.9994041919708252, 'reward_meter_mean': 0.9994041919708252, 'reward_meter_std': 0.0001539249497000128, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994041919708252, 'reward_total_composite_std': 0.0001539249497000128, 'epoch': 0.12}

 90%|████████▉ | 2956/3300 [4:27:04<52:07,  9.09s/it]INFO 04-12 03:03:55 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2957/3300 [4:27:09<44:30,  7.79s/it]2026-04-12 03:04:00,303 | INFO | train_grpo_train | metrics_logged mode=train step=2957

                                                     
{'loss': 0.0006, 'grad_norm': 0.0930638462305069, 'learning_rate': 1.0424242424242426e-06, 'num_tokens': 6705889.0, 'completions/mean_length': 66.75, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.998155951499939, 'rewards/meter/std': 9.95435857475968e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998155951499939, 'rewards/total_composite/std': 9.95435857475968e-06, 'reward': 0.998155951499939, 'reward_std': 9.94737092696596e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008290099911391735, 'sampling/sampling_logp_difference/max': 0.8198513984680176, 'sampling/importance_sampling_ratio/min': 0.44049713015556335, 'sampling/importance_sampling_ratio/mean': 0.9991642832756042, 'sampling/importance_sampling_ratio/max': 1.253778100013733, 'entropy': 0.04649371886625886, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0056535504991188645, 'clip_ratio/high_max': 0.0056535504991188645, 'clip_ratio/region_mean': 0.0056535504991188645, 'reward_total_mean': 0.998155951499939, 'reward_meter_mean': 0.998155951499939, 'reward_meter_std': 9.95435857475968e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998155951499939, 'reward_total_composite_std': 9.95435857475968e-06, 'epoch': 0.12}

 90%|████████▉ | 2957/3300 [4:27:09<44:30,  7.79s/it]INFO 04-12 03:04:00 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2958/3300 [4:27:17<45:05,  7.91s/it]2026-04-12 03:04:08,509 | INFO | train_grpo_train | metrics_logged mode=train step=2958

                                                     
{'loss': -0.0383, 'grad_norm': 2.60581636428833, 'learning_rate': 1.0393939393939394e-06, 'num_tokens': 6710000.0, 'completions/mean_length': 319.875, 'completions/min_length': 293.0, 'completions/max_length': 348.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 319.875, 'completions/min_terminated_length': 293.0, 'completions/max_terminated_length': 348.0, 'rewards/meter/mean': 0.9985268712043762, 'rewards/meter/std': 0.0011142482981085777, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.0534522607922554, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9794891476631165, 'rewards/repeat_penalty/std': 0.028372056782245636, 'rewards/total_composite/mean': 0.9288632869720459, 'rewards/total_composite/std': 0.05354485660791397, 'reward': 0.9288632869720459, 'reward_std': 0.05354484170675278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0562576949596405, 'sampling/sampling_logp_difference/max': 1.9502592086791992, 'sampling/importance_sampling_ratio/min': 0.14223718643188477, 'sampling/importance_sampling_ratio/mean': 1.0072740316390991, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47900502756237984, 'clip_ratio/low_mean': 0.016178023535758257, 'clip_ratio/low_min': 0.016178023535758257, 'clip_ratio/high_mean': 0.023225491400808096, 'clip_ratio/high_max': 0.023225491400808096, 'clip_ratio/region_mean': 0.03940351493656635, 'reward_total_mean': 0.9288632869720459, 'reward_meter_mean': 0.9985268712043762, 'reward_meter_std': 0.0011142482981085777, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.0534522607922554, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9794891476631165, 'reward_repeat_penalty_std': 0.028372056782245636, 'reward_total_composite_mean': 0.9288632869720459, 'reward_total_composite_std': 0.05354485660791397, 'epoch': 0.12}

 90%|████████▉ | 2958/3300 [4:27:17<45:05,  7.91s/it]INFO 04-12 03:04:08 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2959/3300 [4:27:25<45:49,  8.06s/it]2026-04-12 03:04:16,907 | INFO | train_grpo_train | metrics_logged mode=train step=2959

                                                     
{'loss': -0.0149, 'grad_norm': 1.92487370967865, 'learning_rate': 1.0363636363636365e-06, 'num_tokens': 6714601.0, 'completions/mean_length': 358.125, 'completions/min_length': 348.0, 'completions/max_length': 382.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 358.125, 'completions/min_terminated_length': 348.0, 'completions/max_terminated_length': 382.0, 'rewards/meter/mean': 0.9991495013237, 'rewards/meter/std': 0.0001215613738168031, 'rewards/count_adherence/mean': 0.8295454978942871, 'rewards/count_adherence/std': 0.03214123100042343, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.985702633857727, 'rewards/repeat_penalty/std': 0.02648802287876606, 'rewards/total_composite/mean': 0.8171521425247192, 'rewards/total_composite/std': 0.04246782884001732, 'reward': 0.8171521425247192, 'reward_std': 0.042467836290597916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04391275718808174, 'sampling/sampling_logp_difference/max': 1.476973533630371, 'sampling/importance_sampling_ratio/min': 0.228327676653862, 'sampling/importance_sampling_ratio/mean': 1.0104491710662842, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40936222299933434, 'clip_ratio/low_mean': 0.004202505107969046, 'clip_ratio/low_min': 0.004202505107969046, 'clip_ratio/high_mean': 0.022940220427699387, 'clip_ratio/high_max': 0.022940220427699387, 'clip_ratio/region_mean': 0.027142725535668433, 'reward_total_mean': 0.8171521425247192, 'reward_meter_mean': 0.9991495013237, 'reward_meter_std': 0.0001215613738168031, 'reward_count_adherence_mean': 0.8295454978942871, 'reward_count_adherence_std': 0.03214123100042343, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.985702633857727, 'reward_repeat_penalty_std': 0.02648802287876606, 'reward_total_composite_mean': 0.8171521425247192, 'reward_total_composite_std': 0.04246782884001732, 'epoch': 0.12}

 90%|████████▉ | 2959/3300 [4:27:25<45:49,  8.06s/it]INFO 04-12 03:04:17 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2960/3300 [4:27:33<45:50,  8.09s/it]2026-04-12 03:04:25,063 | INFO | train_grpo_train | metrics_logged mode=train step=2960

                                                     
{'loss': -0.0038, 'grad_norm': 2.4538395404815674, 'learning_rate': 1.0333333333333333e-06, 'num_tokens': 6719098.0, 'completions/mean_length': 355.125, 'completions/min_length': 340.0, 'completions/max_length': 360.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 355.125, 'completions/min_terminated_length': 340.0, 'completions/max_terminated_length': 360.0, 'rewards/meter/mean': 0.9989354610443115, 'rewards/meter/std': 0.0001577387156430632, 'rewards/count_adherence/mean': 0.8977272510528564, 'rewards/count_adherence/std': 0.03214123100042343, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9736841917037964, 'rewards/repeat_penalty/std': 0.02813275158405304, 'rewards/total_composite/mean': 0.8728748559951782, 'rewards/total_composite/std': 0.03264994919300079, 'reward': 0.8728748559951782, 'reward_std': 0.032649945467710495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022427460178732872, 'sampling/sampling_logp_difference/max': 1.2122278213500977, 'sampling/importance_sampling_ratio/min': 0.29753369092941284, 'sampling/importance_sampling_ratio/mean': 1.0067881345748901, 'sampling/importance_sampling_ratio/max': 1.6595386266708374, 'entropy': 0.3090662658214569, 'clip_ratio/low_mean': 0.009166883770376444, 'clip_ratio/low_min': 0.009166883770376444, 'clip_ratio/high_mean': 0.004899043240584433, 'clip_ratio/high_max': 0.004899043240584433, 'clip_ratio/region_mean': 0.014065927010960877, 'reward_total_mean': 0.8728748559951782, 'reward_meter_mean': 0.9989354610443115, 'reward_meter_std': 0.0001577387156430632, 'reward_count_adherence_mean': 0.8977272510528564, 'reward_count_adherence_std': 0.03214123100042343, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9736841917037964, 'reward_repeat_penalty_std': 0.02813275158405304, 'reward_total_composite_mean': 0.8728748559951782, 'reward_total_composite_std': 0.03264994919300079, 'epoch': 0.12}

 90%|████████▉ | 2960/3300 [4:27:34<45:50,  8.09s/it]INFO 04-12 03:04:25 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2961/3300 [4:27:38<40:19,  7.14s/it]2026-04-12 03:04:29,975 | INFO | train_grpo_train | metrics_logged mode=train step=2961

                                                     
{'loss': -0.0069, 'grad_norm': 9.514266967773438, 'learning_rate': 1.0303030303030304e-06, 'num_tokens': 6721217.0, 'completions/mean_length': 85.875, 'completions/min_length': 83.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.875, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9182481169700623, 'rewards/meter/std': 0.11503338068723679, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9182481169700623, 'rewards/total_composite/std': 0.11503338068723679, 'reward': 0.9182481169700623, 'reward_std': 0.11503338813781738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06353027373552322, 'sampling/sampling_logp_difference/max': 2.5414462089538574, 'sampling/importance_sampling_ratio/min': 0.07875242084264755, 'sampling/importance_sampling_ratio/mean': 0.9983007907867432, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31395105831325054, 'clip_ratio/low_mean': 0.013554217293858528, 'clip_ratio/low_min': 0.013554217293858528, 'clip_ratio/high_mean': 0.0521149430423975, 'clip_ratio/high_max': 0.0521149430423975, 'clip_ratio/region_mean': 0.06566916033625603, 'reward_total_mean': 0.9182481169700623, 'reward_meter_mean': 0.9182481169700623, 'reward_meter_std': 0.11503338068723679, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9182481169700623, 'reward_total_composite_std': 0.11503338068723679, 'epoch': 0.12}

 90%|████████▉ | 2961/3300 [4:27:38<40:19,  7.14s/it]INFO 04-12 03:04:30 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2962/3300 [4:27:43<35:46,  6.35s/it]2026-04-12 03:04:34,492 | INFO | train_grpo_train | metrics_logged mode=train step=2962

                                                     
{'loss': -0.0033, 'grad_norm': 4.379561424255371, 'learning_rate': 1.0272727272727274e-06, 'num_tokens': 6722966.0, 'completions/mean_length': 63.625, 'completions/min_length': 63.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9992475509643555, 'rewards/meter/std': 0.00020958359527867287, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992475509643555, 'rewards/total_composite/std': 0.00020958359527867287, 'reward': 0.9992475509643555, 'reward_std': 0.00020958769891876727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00583175802603364, 'sampling/sampling_logp_difference/max': 1.3357278108596802, 'sampling/importance_sampling_ratio/min': 0.26296672224998474, 'sampling/importance_sampling_ratio/mean': 1.0014468431472778, 'sampling/importance_sampling_ratio/max': 1.7847542762756348, 'entropy': 0.009632107801735401, 'clip_ratio/low_mean': 0.0019841270986944437, 'clip_ratio/low_min': 0.0019841270986944437, 'clip_ratio/high_mean': 0.001953125, 'clip_ratio/high_max': 0.001953125, 'clip_ratio/region_mean': 0.003937252098694444, 'reward_total_mean': 0.9992475509643555, 'reward_meter_mean': 0.9992475509643555, 'reward_meter_std': 0.00020958359527867287, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992475509643555, 'reward_total_composite_std': 0.00020958359527867287, 'epoch': 0.12}

 90%|████████▉ | 2962/3300 [4:27:43<35:46,  6.35s/it]INFO 04-12 03:04:34 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2963/3300 [4:27:49<34:39,  6.17s/it]2026-04-12 03:04:40,248 | INFO | train_grpo_train | metrics_logged mode=train step=2963

                                                     
{'loss': 0.0074, 'grad_norm': 2.2698678970336914, 'learning_rate': 1.0242424242424242e-06, 'num_tokens': 6725671.0, 'completions/mean_length': 157.125, 'completions/min_length': 152.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.125, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9989545345306396, 'rewards/meter/std': 0.0008401040686294436, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9811090230941772, 'rewards/total_composite/std': 0.05032341554760933, 'reward': 0.9811090230941772, 'reward_std': 0.050323426723480225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03372993692755699, 'sampling/sampling_logp_difference/max': 1.6377344131469727, 'sampling/importance_sampling_ratio/min': 0.19442002475261688, 'sampling/importance_sampling_ratio/mean': 1.0084627866744995, 'sampling/importance_sampling_ratio/max': 1.5869454145431519, 'entropy': 0.3180831354111433, 'clip_ratio/low_mean': 0.004687500186264515, 'clip_ratio/low_min': 0.004687500186264515, 'clip_ratio/high_mean': 0.023915290948934853, 'clip_ratio/high_max': 0.023915290948934853, 'clip_ratio/region_mean': 0.028602791135199368, 'reward_total_mean': 0.9811090230941772, 'reward_meter_mean': 0.9989545345306396, 'reward_meter_std': 0.0008401040686294436, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9811090230941772, 'reward_total_composite_std': 0.05032341554760933, 'epoch': 0.12}

 90%|████████▉ | 2963/3300 [4:27:49<34:39,  6.17s/it]INFO 04-12 03:04:40 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2964/3300 [4:27:53<31:56,  5.70s/it]2026-04-12 03:04:44,863 | INFO | train_grpo_train | metrics_logged mode=train step=2964

                                                     
{'loss': -0.0008, 'grad_norm': 1.2236340045928955, 'learning_rate': 1.0212121212121213e-06, 'num_tokens': 6727399.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993772506713867, 'rewards/meter/std': 6.269344157772139e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993772506713867, 'rewards/total_composite/std': 6.269344157772139e-05, 'reward': 0.9993772506713867, 'reward_std': 6.270247104112059e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002508246572688222, 'sampling/sampling_logp_difference/max': 0.5242347717285156, 'sampling/importance_sampling_ratio/min': 0.5920082330703735, 'sampling/importance_sampling_ratio/mean': 0.9991680383682251, 'sampling/importance_sampling_ratio/max': 1.1811754703521729, 'entropy': 0.00610040313040372, 'clip_ratio/low_mean': 0.001953125, 'clip_ratio/low_min': 0.001953125, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.001953125, 'reward_total_mean': 0.9993772506713867, 'reward_meter_mean': 0.9993772506713867, 'reward_meter_std': 6.269344157772139e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993772506713867, 'reward_total_composite_std': 6.269344157772139e-05, 'epoch': 0.12}

 90%|████████▉ | 2964/3300 [4:27:53<31:56,  5.70s/it]INFO 04-12 03:04:45 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2965/3300 [4:27:58<30:00,  5.37s/it]2026-04-12 03:04:49,459 | INFO | train_grpo_train | metrics_logged mode=train step=2965

                                                     
{'loss': -0.0025, 'grad_norm': 0.8599671721458435, 'learning_rate': 1.0181818181818183e-06, 'num_tokens': 6729286.0, 'completions/mean_length': 67.875, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9994280338287354, 'rewards/meter/std': 7.756530249025673e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994280338287354, 'rewards/total_composite/std': 7.756530249025673e-05, 'reward': 0.9994280338287354, 'reward_std': 7.756453851470724e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010112082585692406, 'sampling/sampling_logp_difference/max': 0.8493509292602539, 'sampling/importance_sampling_ratio/min': 0.4276924431324005, 'sampling/importance_sampling_ratio/mean': 0.9992300271987915, 'sampling/importance_sampling_ratio/max': 1.371290683746338, 'entropy': 0.0738078411668539, 'clip_ratio/low_mean': 0.005542142200283706, 'clip_ratio/low_min': 0.005542142200283706, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.007380377501249313, 'reward_total_mean': 0.9994280338287354, 'reward_meter_mean': 0.9994280338287354, 'reward_meter_std': 7.756530249025673e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994280338287354, 'reward_total_composite_std': 7.756530249025673e-05, 'epoch': 0.12}

 90%|████████▉ | 2965/3300 [4:27:58<30:00,  5.37s/it]INFO 04-12 03:04:49 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2966/3300 [4:28:05<33:31,  6.02s/it]2026-04-12 03:04:57,004 | INFO | train_grpo_train | metrics_logged mode=train step=2966

                                                     
{'loss': 0.0035, 'grad_norm': 1.9423378705978394, 'learning_rate': 1.0151515151515152e-06, 'num_tokens': 6733705.0, 'completions/mean_length': 296.375, 'completions/min_length': 291.0, 'completions/max_length': 301.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 296.375, 'completions/min_terminated_length': 291.0, 'completions/max_terminated_length': 301.0, 'rewards/meter/mean': 0.9927277565002441, 'rewards/meter/std': 0.018433313816785812, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9264706373214722, 'rewards/repeat_penalty/std': 0.04159451648592949, 'rewards/total_composite/mean': 0.9196349382400513, 'rewards/total_composite/std': 0.042633701115846634, 'reward': 0.9196349382400513, 'reward_std': 0.04263370484113693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029081236571073532, 'sampling/sampling_logp_difference/max': 1.037057876586914, 'sampling/importance_sampling_ratio/min': 0.3544961214065552, 'sampling/importance_sampling_ratio/mean': 1.0065155029296875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3008219301700592, 'clip_ratio/low_mean': 0.008422989631071687, 'clip_ratio/low_min': 0.008422989631071687, 'clip_ratio/high_mean': 0.014326842967420816, 'clip_ratio/high_max': 0.014326842967420816, 'clip_ratio/region_mean': 0.022749832598492503, 'reward_total_mean': 0.9196349382400513, 'reward_meter_mean': 0.9927277565002441, 'reward_meter_std': 0.018433313816785812, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9264706373214722, 'reward_repeat_penalty_std': 0.04159451648592949, 'reward_total_composite_mean': 0.9196349382400513, 'reward_total_composite_std': 0.042633701115846634, 'epoch': 0.12}

 90%|████████▉ | 2966/3300 [4:28:06<33:31,  6.02s/it]INFO 04-12 03:04:57 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2967/3300 [4:28:10<31:48,  5.73s/it]2026-04-12 03:05:02,049 | INFO | train_grpo_train | metrics_logged mode=train step=2967

                                                     
{'loss': 0.0063, 'grad_norm': 2.049762487411499, 'learning_rate': 1.0121212121212122e-06, 'num_tokens': 6735749.0, 'completions/mean_length': 92.5, 'completions/min_length': 91.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.5, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9977041482925415, 'rewards/meter/std': 0.00018006080063059926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977041482925415, 'rewards/total_composite/std': 0.00018006080063059926, 'reward': 0.9977041482925415, 'reward_std': 0.0001800469763111323, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01588408276438713, 'sampling/sampling_logp_difference/max': 1.0536683797836304, 'sampling/importance_sampling_ratio/min': 0.34865638613700867, 'sampling/importance_sampling_ratio/mean': 1.0007275342941284, 'sampling/importance_sampling_ratio/max': 1.9789665937423706, 'entropy': 0.11041967011988163, 'clip_ratio/low_mean': 0.008021619636565447, 'clip_ratio/low_min': 0.008021619636565447, 'clip_ratio/high_mean': 0.008167747058905661, 'clip_ratio/high_max': 0.008167747058905661, 'clip_ratio/region_mean': 0.016189366695471108, 'reward_total_mean': 0.9977041482925415, 'reward_meter_mean': 0.9977041482925415, 'reward_meter_std': 0.00018006080063059926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977041482925415, 'reward_total_composite_std': 0.00018006080063059926, 'epoch': 0.12}

 90%|████████▉ | 2967/3300 [4:28:11<31:48,  5.73s/it]INFO 04-12 03:05:02 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2968/3300 [4:28:19<35:36,  6.44s/it]2026-04-12 03:05:10,129 | INFO | train_grpo_train | metrics_logged mode=train step=2968

                                                     
{'loss': 0.0007, 'grad_norm': 1.8433130979537964, 'learning_rate': 1.0090909090909092e-06, 'num_tokens': 6740073.0, 'completions/mean_length': 337.5, 'completions/min_length': 332.0, 'completions/max_length': 347.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 337.5, 'completions/min_terminated_length': 332.0, 'completions/max_terminated_length': 347.0, 'rewards/meter/mean': 0.9978909492492676, 'rewards/meter/std': 0.0014596291584894061, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9868420958518982, 'rewards/repeat_penalty/std': 0.024363677948713303, 'rewards/total_composite/mean': 0.9847573637962341, 'rewards/total_composite/std': 0.024200566112995148, 'reward': 0.9847573637962341, 'reward_std': 0.024200553074479103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051781829446554184, 'sampling/sampling_logp_difference/max': 2.833930492401123, 'sampling/importance_sampling_ratio/min': 0.05878135934472084, 'sampling/importance_sampling_ratio/mean': 1.0109304189682007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4550737701356411, 'clip_ratio/low_mean': 0.00747274118475616, 'clip_ratio/low_min': 0.00747274118475616, 'clip_ratio/high_mean': 0.0354147027246654, 'clip_ratio/high_max': 0.0354147027246654, 'clip_ratio/region_mean': 0.04288744390942156, 'reward_total_mean': 0.9847573637962341, 'reward_meter_mean': 0.9978909492492676, 'reward_meter_std': 0.0014596291584894061, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9868420958518982, 'reward_repeat_penalty_std': 0.024363677948713303, 'reward_total_composite_mean': 0.9847573637962341, 'reward_total_composite_std': 0.024200566112995148, 'epoch': 0.12}

 90%|████████▉ | 2968/3300 [4:28:19<35:36,  6.44s/it]INFO 04-12 03:05:10 [block_pool.py:378] Successfully reset prefix cache

 90%|████████▉ | 2969/3300 [4:28:23<31:44,  5.75s/it]2026-04-12 03:05:14,289 | INFO | train_grpo_train | metrics_logged mode=train step=2969

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.006060606060606e-06, 'num_tokens': 6741609.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0006444337195716798, 'sampling/sampling_logp_difference/max': 0.0378272607922554, 'sampling/importance_sampling_ratio/min': 0.9628793001174927, 'sampling/importance_sampling_ratio/mean': 1.000121831893921, 'sampling/importance_sampling_ratio/max': 1.0287812948226929, 'entropy': 0.005584679980529472, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 90%|████████▉ | 2969/3300 [4:28:23<31:44,  5.75s/it]INFO 04-12 03:05:14 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2970/3300 [4:28:28<31:13,  5.68s/it]2026-04-12 03:05:19,786 | INFO | train_grpo_train | metrics_logged mode=train step=2970

                                                     
{'loss': 0.0005, 'grad_norm': 0.3558986186981201, 'learning_rate': 1.0030303030303031e-06, 'num_tokens': 6744177.0, 'completions/mean_length': 132.0, 'completions/min_length': 131.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.0, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9994352459907532, 'rewards/meter/std': 1.4545888916472904e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994352459907532, 'rewards/total_composite/std': 1.4545888916472904e-05, 'reward': 0.9994352459907532, 'reward_std': 1.4537939932779409e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010489360429346561, 'sampling/sampling_logp_difference/max': 0.6865215301513672, 'sampling/importance_sampling_ratio/min': 0.5033237934112549, 'sampling/importance_sampling_ratio/mean': 1.0008372068405151, 'sampling/importance_sampling_ratio/max': 1.4119945764541626, 'entropy': 0.09148383047431707, 'clip_ratio/low_mean': 0.005667578196153045, 'clip_ratio/low_min': 0.005667578196153045, 'clip_ratio/high_mean': 0.006643245578743517, 'clip_ratio/high_max': 0.006643245578743517, 'clip_ratio/region_mean': 0.012310823774896562, 'reward_total_mean': 0.9994352459907532, 'reward_meter_mean': 0.9994352459907532, 'reward_meter_std': 1.4545888916472904e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994352459907532, 'reward_total_composite_std': 1.4545888916472904e-05, 'epoch': 0.12}

 90%|█████████ | 2970/3300 [4:28:28<31:13,  5.68s/it]INFO 04-12 03:05:20 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2971/3300 [4:28:33<30:07,  5.49s/it]2026-04-12 03:05:24,856 | INFO | train_grpo_train | metrics_logged mode=train step=2971

                                                     
{'loss': 0.001, 'grad_norm': 0.8809167742729187, 'learning_rate': 1.0000000000000002e-06, 'num_tokens': 6746402.0, 'completions/mean_length': 98.125, 'completions/min_length': 98.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.125, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9994017481803894, 'rewards/meter/std': 5.140481152920984e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994017481803894, 'rewards/total_composite/std': 5.140481152920984e-05, 'reward': 0.9994017481803894, 'reward_std': 5.139390123076737e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011422020383179188, 'sampling/sampling_logp_difference/max': 1.0734617710113525, 'sampling/importance_sampling_ratio/min': 0.5932945609092712, 'sampling/importance_sampling_ratio/mean': 1.0018893480300903, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.051710725761950016, 'clip_ratio/low_mean': 0.006377551006153226, 'clip_ratio/low_min': 0.006377551006153226, 'clip_ratio/high_mean': 0.0012755101779475808, 'clip_ratio/high_max': 0.0012755101779475808, 'clip_ratio/region_mean': 0.007653061184100807, 'reward_total_mean': 0.9994017481803894, 'reward_meter_mean': 0.9994017481803894, 'reward_meter_std': 5.140481152920984e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994017481803894, 'reward_total_composite_std': 5.140481152920984e-05, 'epoch': 0.12}

 90%|█████████ | 2971/3300 [4:28:33<30:07,  5.49s/it]INFO 04-12 03:05:25 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2972/3300 [4:28:37<27:53,  5.10s/it]2026-04-12 03:05:29,044 | INFO | train_grpo_train | metrics_logged mode=train step=2972

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.96969696969697e-07, 'num_tokens': 6747826.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0012666110415011644, 'sampling/sampling_logp_difference/max': 0.08357194811105728, 'sampling/importance_sampling_ratio/min': 0.9198248982429504, 'sampling/importance_sampling_ratio/mean': 1.000320315361023, 'sampling/importance_sampling_ratio/max': 1.056694507598877, 'entropy': 0.010805643978528678, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 90%|█████████ | 2972/3300 [4:28:38<27:53,  5.10s/it]INFO 04-12 03:05:29 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2973/3300 [4:28:44<30:55,  5.67s/it]2026-04-12 03:05:36,126 | INFO | train_grpo_train | metrics_logged mode=train step=2973

                                                     
{'loss': 0.0073, 'grad_norm': 1.3921781778335571, 'learning_rate': 9.93939393939394e-07, 'num_tokens': 6751625.0, 'completions/mean_length': 248.875, 'completions/min_length': 246.0, 'completions/max_length': 253.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 248.875, 'completions/min_terminated_length': 246.0, 'completions/max_terminated_length': 253.0, 'rewards/meter/mean': 0.9990034699440002, 'rewards/meter/std': 0.00021533701510634273, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9711538553237915, 'rewards/repeat_penalty/std': 0.039811473339796066, 'rewards/total_composite/mean': 0.9701830148696899, 'rewards/total_composite/std': 0.03968827798962593, 'reward': 0.9701830148696899, 'reward_std': 0.039688270539045334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020545274019241333, 'sampling/sampling_logp_difference/max': 1.2074041366577148, 'sampling/importance_sampling_ratio/min': 0.29897233843803406, 'sampling/importance_sampling_ratio/mean': 1.0035536289215088, 'sampling/importance_sampling_ratio/max': 1.81640625, 'entropy': 0.2675408758223057, 'clip_ratio/low_mean': 0.002494166372343898, 'clip_ratio/low_min': 0.002494166372343898, 'clip_ratio/high_mean': 0.012083206791430712, 'clip_ratio/high_max': 0.012083206791430712, 'clip_ratio/region_mean': 0.01457737316377461, 'reward_total_mean': 0.9701830148696899, 'reward_meter_mean': 0.9990034699440002, 'reward_meter_std': 0.00021533701510634273, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9711538553237915, 'reward_repeat_penalty_std': 0.039811473339796066, 'reward_total_composite_mean': 0.9701830148696899, 'reward_total_composite_std': 0.03968827798962593, 'epoch': 0.12}

 90%|█████████ | 2973/3300 [4:28:45<30:55,  5.67s/it]INFO 04-12 03:05:36 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2974/3300 [4:28:52<34:21,  6.33s/it]2026-04-12 03:05:43,899 | INFO | train_grpo_train | metrics_logged mode=train step=2974

                                                     
{'loss': -0.0, 'grad_norm': 2.7557997703552246, 'learning_rate': 9.90909090909091e-07, 'num_tokens': 6755543.0, 'completions/mean_length': 292.75, 'completions/min_length': 279.0, 'completions/max_length': 309.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 292.75, 'completions/min_terminated_length': 279.0, 'completions/max_terminated_length': 309.0, 'rewards/meter/mean': 0.9976239204406738, 'rewards/meter/std': 0.0013418762246146798, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9558823108673096, 'rewards/repeat_penalty/std': 0.06852733343839645, 'rewards/total_composite/mean': 0.8287225961685181, 'rewards/total_composite/std': 0.3412354588508606, 'reward': 0.8287225961685181, 'reward_std': 0.3412354588508606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05460209771990776, 'sampling/sampling_logp_difference/max': 2.425222635269165, 'sampling/importance_sampling_ratio/min': 0.08845842629671097, 'sampling/importance_sampling_ratio/mean': 1.0052495002746582, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4930482190102339, 'clip_ratio/low_mean': 0.007359597599133849, 'clip_ratio/low_min': 0.007359597599133849, 'clip_ratio/high_mean': 0.03571874415501952, 'clip_ratio/high_max': 0.03571874415501952, 'clip_ratio/region_mean': 0.04307834175415337, 'reward_total_mean': 0.8287225961685181, 'reward_meter_mean': 0.9976239204406738, 'reward_meter_std': 0.0013418762246146798, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9558823108673096, 'reward_repeat_penalty_std': 0.06852733343839645, 'reward_total_composite_mean': 0.8287225961685181, 'reward_total_composite_std': 0.3412354588508606, 'epoch': 0.12}

 90%|█████████ | 2974/3300 [4:28:52<34:21,  6.33s/it]INFO 04-12 03:05:44 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2975/3300 [4:28:57<32:05,  5.92s/it]2026-04-12 03:05:48,926 | INFO | train_grpo_train | metrics_logged mode=train step=2975

                                                     
{'loss': -0.0015, 'grad_norm': 1.2692075967788696, 'learning_rate': 9.87878787878788e-07, 'num_tokens': 6757327.0, 'completions/mean_length': 68.0, 'completions/min_length': 68.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9994783401489258, 'rewards/meter/std': 9.867444896372035e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994783401489258, 'rewards/total_composite/std': 9.867444896372035e-05, 'reward': 0.9994783401489258, 'reward_std': 9.867054177448153e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009787042625248432, 'sampling/sampling_logp_difference/max': 0.6712450981140137, 'sampling/importance_sampling_ratio/min': 0.5110718607902527, 'sampling/importance_sampling_ratio/mean': 0.9991775751113892, 'sampling/importance_sampling_ratio/max': 1.392788290977478, 'entropy': 0.07629173761233687, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.009191176504828036, 'clip_ratio/high_max': 0.009191176504828036, 'clip_ratio/region_mean': 0.011029411805793643, 'reward_total_mean': 0.9994783401489258, 'reward_meter_mean': 0.9994783401489258, 'reward_meter_std': 9.867444896372035e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994783401489258, 'reward_total_composite_std': 9.867444896372035e-05, 'epoch': 0.12}

 90%|█████████ | 2975/3300 [4:28:57<32:05,  5.92s/it]INFO 04-12 03:05:49 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2976/3300 [4:29:03<31:09,  5.77s/it]2026-04-12 03:05:54,322 | INFO | train_grpo_train | metrics_logged mode=train step=2976

                                                     
{'loss': -0.0012, 'grad_norm': 1.1907378435134888, 'learning_rate': 9.84848484848485e-07, 'num_tokens': 6759949.0, 'completions/mean_length': 131.75, 'completions/min_length': 130.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.75, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.9993938207626343, 'rewards/meter/std': 7.404623465845361e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993938207626343, 'rewards/total_composite/std': 7.404623465845361e-05, 'reward': 0.9993938207626343, 'reward_std': 7.40409450372681e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008480089716613293, 'sampling/sampling_logp_difference/max': 0.5749435424804688, 'sampling/importance_sampling_ratio/min': 0.5627366304397583, 'sampling/importance_sampling_ratio/mean': 1.0010247230529785, 'sampling/importance_sampling_ratio/max': 1.4190884828567505, 'entropy': 0.07698143133893609, 'clip_ratio/low_mean': 0.000961538462433964, 'clip_ratio/low_min': 0.000961538462433964, 'clip_ratio/high_mean': 0.006628788076341152, 'clip_ratio/high_max': 0.006628788076341152, 'clip_ratio/region_mean': 0.007590326538775116, 'reward_total_mean': 0.9993938207626343, 'reward_meter_mean': 0.9993938207626343, 'reward_meter_std': 7.404623465845361e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993938207626343, 'reward_total_composite_std': 7.404623465845361e-05, 'epoch': 0.12}

 90%|█████████ | 2976/3300 [4:29:03<31:09,  5.77s/it]INFO 04-12 03:05:54 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2977/3300 [4:29:08<30:00,  5.57s/it]2026-04-12 03:05:59,418 | INFO | train_grpo_train | metrics_logged mode=train step=2977

                                                     
{'loss': 0.0003, 'grad_norm': 0.025195496156811714, 'learning_rate': 9.818181818181818e-07, 'num_tokens': 6762117.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994038343429565, 'rewards/meter/std': 3.7851127672183793e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994038343429565, 'rewards/total_composite/std': 3.7851127672183793e-06, 'reward': 0.9994038343429565, 'reward_std': 3.7985312246746616e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0029753749258816242, 'sampling/sampling_logp_difference/max': 0.2579946517944336, 'sampling/importance_sampling_ratio/min': 0.8075622916221619, 'sampling/importance_sampling_ratio/mean': 1.000174880027771, 'sampling/importance_sampling_ratio/max': 1.2943319082260132, 'entropy': 0.0308038501534611, 'clip_ratio/low_mean': 0.0025510203558951616, 'clip_ratio/low_min': 0.0025510203558951616, 'clip_ratio/high_mean': 0.0012755101779475808, 'clip_ratio/high_max': 0.0012755101779475808, 'clip_ratio/region_mean': 0.0038265305338427424, 'reward_total_mean': 0.9994038343429565, 'reward_meter_mean': 0.9994038343429565, 'reward_meter_std': 3.7851127672183793e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994038343429565, 'reward_total_composite_std': 3.7851127672183793e-06, 'epoch': 0.12}

 90%|█████████ | 2977/3300 [4:29:08<30:00,  5.57s/it]INFO 04-12 03:05:59 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2978/3300 [4:29:14<30:16,  5.64s/it]2026-04-12 03:06:05,213 | INFO | train_grpo_train | metrics_logged mode=train step=2978

                                                     
{'loss': -0.0016, 'grad_norm': 3.375030994415283, 'learning_rate': 9.787878787878788e-07, 'num_tokens': 6764878.0, 'completions/mean_length': 167.125, 'completions/min_length': 162.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 167.125, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.9967660903930664, 'rewards/meter/std': 0.00537865748628974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9690313339233398, 'rewards/total_composite/std': 0.05059126019477844, 'reward': 0.9690313339233398, 'reward_std': 0.050591256469488144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03752928227186203, 'sampling/sampling_logp_difference/max': 1.3720521926879883, 'sampling/importance_sampling_ratio/min': 0.2535859942436218, 'sampling/importance_sampling_ratio/mean': 1.0075362920761108, 'sampling/importance_sampling_ratio/max': 1.8077061176300049, 'entropy': 0.30791075527668, 'clip_ratio/low_mean': 0.003756533144041896, 'clip_ratio/low_min': 0.003756533144041896, 'clip_ratio/high_mean': 0.02096639317460358, 'clip_ratio/high_max': 0.02096639317460358, 'clip_ratio/region_mean': 0.024722926318645477, 'reward_total_mean': 0.9690313339233398, 'reward_meter_mean': 0.9967660903930664, 'reward_meter_std': 0.00537865748628974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9690313339233398, 'reward_total_composite_std': 0.05059126019477844, 'epoch': 0.12}

 90%|█████████ | 2978/3300 [4:29:14<30:16,  5.64s/it]INFO 04-12 03:06:05 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2979/3300 [4:29:18<28:42,  5.37s/it]2026-04-12 03:06:09,941 | INFO | train_grpo_train | metrics_logged mode=train step=2979

                                                     
{'loss': 0.015, 'grad_norm': 6.203634738922119, 'learning_rate': 9.757575757575759e-07, 'num_tokens': 6766789.0, 'completions/mean_length': 70.875, 'completions/min_length': 70.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9866373538970947, 'rewards/meter/std': 0.01733180694282055, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9866373538970947, 'rewards/total_composite/std': 0.01733180694282055, 'reward': 0.9866373538970947, 'reward_std': 0.017331810668110847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025828473269939423, 'sampling/sampling_logp_difference/max': 0.6619188785552979, 'sampling/importance_sampling_ratio/min': 0.5158604979515076, 'sampling/importance_sampling_ratio/mean': 1.0079153776168823, 'sampling/importance_sampling_ratio/max': 1.5757840871810913, 'entropy': 0.27640925347805023, 'clip_ratio/low_mean': 0.008778364514000714, 'clip_ratio/low_min': 0.008778364514000714, 'clip_ratio/high_mean': 0.02296627010218799, 'clip_ratio/high_max': 0.02296627010218799, 'clip_ratio/region_mean': 0.031744634616188705, 'reward_total_mean': 0.9866373538970947, 'reward_meter_mean': 0.9866373538970947, 'reward_meter_std': 0.01733180694282055, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9866373538970947, 'reward_total_composite_std': 0.01733180694282055, 'epoch': 0.12}

 90%|█████████ | 2979/3300 [4:29:18<28:42,  5.37s/it]INFO 04-12 03:06:10 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2980/3300 [4:29:23<27:37,  5.18s/it]2026-04-12 03:06:14,686 | INFO | train_grpo_train | metrics_logged mode=train step=2980

                                                     
{'loss': 0.0156, 'grad_norm': 5.659327507019043, 'learning_rate': 9.72727272727273e-07, 'num_tokens': 6768713.0, 'completions/mean_length': 68.5, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9975705742835999, 'rewards/meter/std': 0.004985023755580187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975705742835999, 'rewards/total_composite/std': 0.004985023755580187, 'reward': 0.9975705742835999, 'reward_std': 0.004985013976693153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016337888315320015, 'sampling/sampling_logp_difference/max': 1.1524267196655273, 'sampling/importance_sampling_ratio/min': 0.3158693015575409, 'sampling/importance_sampling_ratio/mean': 1.0029165744781494, 'sampling/importance_sampling_ratio/max': 1.6995705366134644, 'entropy': 0.11373988864943385, 'clip_ratio/low_mean': 0.0053571430034935474, 'clip_ratio/low_min': 0.0053571430034935474, 'clip_ratio/high_mean': 0.007247899193316698, 'clip_ratio/high_max': 0.007247899193316698, 'clip_ratio/region_mean': 0.012605042196810246, 'reward_total_mean': 0.9975705742835999, 'reward_meter_mean': 0.9975705742835999, 'reward_meter_std': 0.004985023755580187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975705742835999, 'reward_total_composite_std': 0.004985023755580187, 'epoch': 0.12}

 90%|█████████ | 2980/3300 [4:29:23<27:37,  5.18s/it]INFO 04-12 03:06:14 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2981/3300 [4:29:27<26:06,  4.91s/it]2026-04-12 03:06:18,987 | INFO | train_grpo_train | metrics_logged mode=train step=2981

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.696969696969698e-07, 'num_tokens': 6770249.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0001402015914209187, 'sampling/sampling_logp_difference/max': 0.0023617574479430914, 'sampling/importance_sampling_ratio/min': 0.9999930262565613, 'sampling/importance_sampling_ratio/mean': 1.0001403093338013, 'sampling/importance_sampling_ratio/max': 1.0023645162582397, 'entropy': 0.0012292767642065883, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 90%|█████████ | 2981/3300 [4:29:27<26:06,  4.91s/it]INFO 04-12 03:06:19 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2982/3300 [4:29:33<26:37,  5.02s/it]2026-04-12 03:06:24,257 | INFO | train_grpo_train | metrics_logged mode=train step=2982

                                                     
{'loss': -0.0015, 'grad_norm': 2.0111052989959717, 'learning_rate': 9.666666666666668e-07, 'num_tokens': 6772571.0, 'completions/mean_length': 118.25, 'completions/min_length': 117.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.25, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9989663362503052, 'rewards/meter/std': 0.0008744557853788137, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989663362503052, 'rewards/total_composite/std': 0.0008744557853788137, 'reward': 0.9989663362503052, 'reward_std': 0.0008744720835238695, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03213733434677124, 'sampling/sampling_logp_difference/max': 1.6665115356445312, 'sampling/importance_sampling_ratio/min': 0.18890489637851715, 'sampling/importance_sampling_ratio/mean': 1.0011169910430908, 'sampling/importance_sampling_ratio/max': 1.6112860441207886, 'entropy': 0.2766532879322767, 'clip_ratio/low_mean': 0.0032051282469183207, 'clip_ratio/low_min': 0.0032051282469183207, 'clip_ratio/high_mean': 0.02634434588253498, 'clip_ratio/high_max': 0.02634434588253498, 'clip_ratio/region_mean': 0.0295494741294533, 'reward_total_mean': 0.9989663362503052, 'reward_meter_mean': 0.9989663362503052, 'reward_meter_std': 0.0008744557853788137, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989663362503052, 'reward_total_composite_std': 0.0008744557853788137, 'epoch': 0.12}

 90%|█████████ | 2982/3300 [4:29:33<26:37,  5.02s/it]INFO 04-12 03:06:24 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2983/3300 [4:29:38<26:39,  5.05s/it]2026-04-12 03:06:29,367 | INFO | train_grpo_train | metrics_logged mode=train step=2983

                                                     
{'loss': 0.0056, 'grad_norm': 4.000948905944824, 'learning_rate': 9.636363636363636e-07, 'num_tokens': 6774679.0, 'completions/mean_length': 100.5, 'completions/min_length': 99.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.5, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9993005990982056, 'rewards/meter/std': 0.00027800656971521676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9249999523162842, 'rewards/repeat_penalty/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.9243372678756714, 'rewards/total_composite/std': 0.1032847911119461, 'reward': 0.9243372678756714, 'reward_std': 0.10328476876020432, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023382917046546936, 'sampling/sampling_logp_difference/max': 1.0612468719482422, 'sampling/importance_sampling_ratio/min': 0.3460240960121155, 'sampling/importance_sampling_ratio/mean': 1.0000145435333252, 'sampling/importance_sampling_ratio/max': 1.835827350616455, 'entropy': 0.17713584937155247, 'clip_ratio/low_mean': 0.0061881187139078975, 'clip_ratio/low_min': 0.0061881187139078975, 'clip_ratio/high_mean': 0.016288378508761525, 'clip_ratio/high_max': 0.016288378508761525, 'clip_ratio/region_mean': 0.022476497222669423, 'reward_total_mean': 0.9243372678756714, 'reward_meter_mean': 0.9993005990982056, 'reward_meter_std': 0.00027800656971521676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9249999523162842, 'reward_repeat_penalty_std': 0.1035098284482956, 'reward_total_composite_mean': 0.9243372678756714, 'reward_total_composite_std': 0.1032847911119461, 'epoch': 0.12}

 90%|█████████ | 2983/3300 [4:29:38<26:39,  5.05s/it]INFO 04-12 03:06:29 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2984/3300 [4:29:42<25:29,  4.84s/it]2026-04-12 03:06:33,722 | INFO | train_grpo_train | metrics_logged mode=train step=2984

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.606060606060607e-07, 'num_tokens': 6776207.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002929195179603994, 'sampling/sampling_logp_difference/max': 0.00473133847117424, 'sampling/importance_sampling_ratio/min': 0.995279848575592, 'sampling/importance_sampling_ratio/mean': 1.0001977682113647, 'sampling/importance_sampling_ratio/max': 1.004080891609192, 'entropy': 0.003427958843531087, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 90%|█████████ | 2984/3300 [4:29:42<25:29,  4.84s/it]INFO 04-12 03:06:33 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2985/3300 [4:29:48<27:19,  5.21s/it]2026-04-12 03:06:39,795 | INFO | train_grpo_train | metrics_logged mode=train step=2985

                                                     
{'loss': 0.0071, 'grad_norm': 2.747091770172119, 'learning_rate': 9.575757575757577e-07, 'num_tokens': 6779042.0, 'completions/mean_length': 167.375, 'completions/min_length': 164.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 167.375, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.9969633221626282, 'rewards/meter/std': 0.003777019679546356, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9305555820465088, 'rewards/repeat_penalty/std': 0.05750546231865883, 'rewards/total_composite/mean': 0.9275848865509033, 'rewards/total_composite/std': 0.05462638661265373, 'reward': 0.9275848865509033, 'reward_std': 0.05462638661265373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03595529869198799, 'sampling/sampling_logp_difference/max': 1.5971007347106934, 'sampling/importance_sampling_ratio/min': 0.20248271524906158, 'sampling/importance_sampling_ratio/mean': 1.004103422164917, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2825808133929968, 'clip_ratio/low_mean': 0.018017811933532357, 'clip_ratio/low_min': 0.018017811933532357, 'clip_ratio/high_mean': 0.011204644571989775, 'clip_ratio/high_max': 0.011204644571989775, 'clip_ratio/region_mean': 0.029222456505522132, 'reward_total_mean': 0.9275848865509033, 'reward_meter_mean': 0.9969633221626282, 'reward_meter_std': 0.003777019679546356, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9305555820465088, 'reward_repeat_penalty_std': 0.05750546231865883, 'reward_total_composite_mean': 0.9275848865509033, 'reward_total_composite_std': 0.05462638661265373, 'epoch': 0.12}

 90%|█████████ | 2985/3300 [4:29:48<27:19,  5.21s/it]INFO 04-12 03:06:40 [block_pool.py:378] Successfully reset prefix cache

 90%|█████████ | 2986/3300 [4:29:54<27:29,  5.25s/it]2026-04-12 03:06:45,176 | INFO | train_grpo_train | metrics_logged mode=train step=2986

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.545454545454548e-07, 'num_tokens': 6780498.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 5.2063212933717296e-05, 'sampling/sampling_logp_difference/max': 0.0009029797511175275, 'sampling/importance_sampling_ratio/min': 0.9995164275169373, 'sampling/importance_sampling_ratio/mean': 1.0000447034835815, 'sampling/importance_sampling_ratio/max': 1.0009034872055054, 'entropy': 0.00046791824570391327, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 90%|█████████ | 2986/3300 [4:29:54<27:29,  5.25s/it]INFO 04-12 03:06:45 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2987/3300 [4:29:59<27:15,  5.23s/it]2026-04-12 03:06:50,322 | INFO | train_grpo_train | metrics_logged mode=train step=2987

                                                     
{'loss': -0.004, 'grad_norm': 2.3871347904205322, 'learning_rate': 9.515151515151516e-07, 'num_tokens': 6782659.0, 'completions/mean_length': 71.125, 'completions/min_length': 70.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9993984699249268, 'rewards/meter/std': 9.4733273726888e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993984699249268, 'rewards/total_composite/std': 9.4733273726888e-05, 'reward': 0.9993984699249268, 'reward_std': 9.472780220676214e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010385324247181416, 'sampling/sampling_logp_difference/max': 0.6226983070373535, 'sampling/importance_sampling_ratio/min': 0.5364948511123657, 'sampling/importance_sampling_ratio/mean': 1.0039633512496948, 'sampling/importance_sampling_ratio/max': 1.32568359375, 'entropy': 0.08504062425345182, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0034966744715347886, 'clip_ratio/high_max': 0.0034966744715347886, 'clip_ratio/region_mean': 0.0034966744715347886, 'reward_total_mean': 0.9993984699249268, 'reward_meter_mean': 0.9993984699249268, 'reward_meter_std': 9.4733273726888e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993984699249268, 'reward_total_composite_std': 9.4733273726888e-05, 'epoch': 0.12}

 91%|█████████ | 2987/3300 [4:29:59<27:15,  5.23s/it]INFO 04-12 03:06:50 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2988/3300 [4:30:04<26:43,  5.14s/it]2026-04-12 03:06:55,255 | INFO | train_grpo_train | metrics_logged mode=train step=2988

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.484848484848485e-07, 'num_tokens': 6784427.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00022603126126341522, 'sampling/sampling_logp_difference/max': 0.003381013870239258, 'sampling/importance_sampling_ratio/min': 0.9995951652526855, 'sampling/importance_sampling_ratio/mean': 1.0002238750457764, 'sampling/importance_sampling_ratio/max': 1.0033867359161377, 'entropy': 0.0022001640463713557, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████ | 2988/3300 [4:30:04<26:43,  5.14s/it]INFO 04-12 03:06:55 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2989/3300 [4:30:10<28:10,  5.44s/it]2026-04-12 03:07:01,412 | INFO | train_grpo_train | metrics_logged mode=train step=2989

                                                     
{'loss': 0.0025, 'grad_norm': 3.191429615020752, 'learning_rate': 9.454545454545455e-07, 'num_tokens': 6786878.0, 'completions/mean_length': 133.375, 'completions/min_length': 128.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.375, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9979535937309265, 'rewards/meter/std': 0.0034614717587828636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9622775316238403, 'rewards/total_composite/std': 0.06554586440324783, 'reward': 0.9622775316238403, 'reward_std': 0.06554586440324783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03642260655760765, 'sampling/sampling_logp_difference/max': 2.4946718215942383, 'sampling/importance_sampling_ratio/min': 0.08252352476119995, 'sampling/importance_sampling_ratio/mean': 0.9995792508125305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24928219243884087, 'clip_ratio/low_mean': 0.0028013800038024783, 'clip_ratio/low_min': 0.0028013800038024783, 'clip_ratio/high_mean': 0.02337344060651958, 'clip_ratio/high_max': 0.02337344060651958, 'clip_ratio/region_mean': 0.026174820610322058, 'reward_total_mean': 0.9622775316238403, 'reward_meter_mean': 0.9979535937309265, 'reward_meter_std': 0.0034614717587828636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9622775316238403, 'reward_total_composite_std': 0.06554586440324783, 'epoch': 0.12}

 91%|█████████ | 2989/3300 [4:30:10<28:10,  5.44s/it]INFO 04-12 03:07:01 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2990/3300 [4:30:15<27:59,  5.42s/it]2026-04-12 03:07:06,755 | INFO | train_grpo_train | metrics_logged mode=train step=2990

                                                     
{'loss': -0.0015, 'grad_norm': 3.7938873767852783, 'learning_rate': 9.424242424242425e-07, 'num_tokens': 6789021.0, 'completions/mean_length': 92.875, 'completions/min_length': 91.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.875, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.997610330581665, 'rewards/meter/std': 0.00025703865685500205, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997610330581665, 'rewards/total_composite/std': 0.00025703865685500205, 'reward': 0.997610330581665, 'reward_std': 0.0002570512588135898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01364238653331995, 'sampling/sampling_logp_difference/max': 1.0927095413208008, 'sampling/importance_sampling_ratio/min': 0.33530673384666443, 'sampling/importance_sampling_ratio/mean': 0.9976533055305481, 'sampling/importance_sampling_ratio/max': 1.416971206665039, 'entropy': 0.09223943296819925, 'clip_ratio/low_mean': 0.004032257944345474, 'clip_ratio/low_min': 0.004032257944345474, 'clip_ratio/high_mean': 0.008035918464884162, 'clip_ratio/high_max': 0.008035918464884162, 'clip_ratio/region_mean': 0.012068176409229636, 'reward_total_mean': 0.997610330581665, 'reward_meter_mean': 0.997610330581665, 'reward_meter_std': 0.00025703865685500205, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997610330581665, 'reward_total_composite_std': 0.00025703865685500205, 'epoch': 0.12}

 91%|█████████ | 2990/3300 [4:30:15<27:59,  5.42s/it]INFO 04-12 03:07:07 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2991/3300 [4:30:20<26:58,  5.24s/it]2026-04-12 03:07:11,572 | INFO | train_grpo_train | metrics_logged mode=train step=2991

                                                     
{'loss': -0.0063, 'grad_norm': 2.051816463470459, 'learning_rate': 9.393939393939395e-07, 'num_tokens': 6790862.0, 'completions/mean_length': 66.125, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9928753972053528, 'rewards/meter/std': 0.0007522930391132832, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9928753972053528, 'rewards/total_composite/std': 0.0007522930391132832, 'reward': 0.9928753972053528, 'reward_std': 0.000752286403439939, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024633267894387245, 'sampling/sampling_logp_difference/max': 1.791666030883789, 'sampling/importance_sampling_ratio/min': 0.16668224334716797, 'sampling/importance_sampling_ratio/mean': 1.0033050775527954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14550988655537367, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0056535504991188645, 'clip_ratio/high_max': 0.0056535504991188645, 'clip_ratio/region_mean': 0.007547489949502051, 'reward_total_mean': 0.9928753972053528, 'reward_meter_mean': 0.9928753972053528, 'reward_meter_std': 0.0007522930391132832, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9928753972053528, 'reward_total_composite_std': 0.0007522930391132832, 'epoch': 0.12}

 91%|█████████ | 2991/3300 [4:30:20<26:58,  5.24s/it]INFO 04-12 03:07:11 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2992/3300 [4:30:25<26:19,  5.13s/it]2026-04-12 03:07:16,447 | INFO | train_grpo_train | metrics_logged mode=train step=2992

                                                     
{'loss': 0.0024, 'grad_norm': 3.857464075088501, 'learning_rate': 9.363636363636365e-07, 'num_tokens': 6792639.0, 'completions/mean_length': 61.125, 'completions/min_length': 61.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9971661567687988, 'rewards/meter/std': 0.00047978354268707335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9971661567687988, 'rewards/total_composite/std': 0.00047978354268707335, 'reward': 0.9971661567687988, 'reward_std': 0.0004797954170498997, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00791865587234497, 'sampling/sampling_logp_difference/max': 1.693861484527588, 'sampling/importance_sampling_ratio/min': 0.18380838632583618, 'sampling/importance_sampling_ratio/mean': 0.9966216087341309, 'sampling/importance_sampling_ratio/max': 1.1078962087631226, 'entropy': 0.025716731324791908, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006147540640085936, 'clip_ratio/high_max': 0.006147540640085936, 'clip_ratio/region_mean': 0.006147540640085936, 'reward_total_mean': 0.9971661567687988, 'reward_meter_mean': 0.9971661567687988, 'reward_meter_std': 0.00047978354268707335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9971661567687988, 'reward_total_composite_std': 0.00047978354268707335, 'epoch': 0.12}

 91%|█████████ | 2992/3300 [4:30:25<26:19,  5.13s/it]INFO 04-12 03:07:16 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2993/3300 [4:30:33<31:15,  6.11s/it]2026-04-12 03:07:24,856 | INFO | train_grpo_train | metrics_logged mode=train step=2993

                                                     
{'loss': 0.0031, 'grad_norm': 1.920817255973816, 'learning_rate': 9.333333333333334e-07, 'num_tokens': 6797268.0, 'completions/mean_length': 328.625, 'completions/min_length': 325.0, 'completions/max_length': 332.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 328.625, 'completions/min_terminated_length': 325.0, 'completions/max_terminated_length': 332.0, 'rewards/meter/mean': 0.9988096952438354, 'rewards/meter/std': 0.0004608924500644207, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9144736528396606, 'rewards/repeat_penalty/std': 0.05582422763109207, 'rewards/total_composite/mean': 0.9133861064910889, 'rewards/total_composite/std': 0.05579417571425438, 'reward': 0.9133861064910889, 'reward_std': 0.055794164538383484, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03226817771792412, 'sampling/sampling_logp_difference/max': 1.3698234558105469, 'sampling/importance_sampling_ratio/min': 0.2541518211364746, 'sampling/importance_sampling_ratio/mean': 1.007267951965332, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3455524481832981, 'clip_ratio/low_mean': 0.011799397761933506, 'clip_ratio/low_min': 0.011799397761933506, 'clip_ratio/high_mean': 0.014776158845052123, 'clip_ratio/high_max': 0.014776158845052123, 'clip_ratio/region_mean': 0.02657555660698563, 'reward_total_mean': 0.9133861064910889, 'reward_meter_mean': 0.9988096952438354, 'reward_meter_std': 0.0004608924500644207, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9144736528396606, 'reward_repeat_penalty_std': 0.05582422763109207, 'reward_total_composite_mean': 0.9133861064910889, 'reward_total_composite_std': 0.05579417571425438, 'epoch': 0.12}

 91%|█████████ | 2993/3300 [4:30:33<31:15,  6.11s/it]INFO 04-12 03:07:25 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2994/3300 [4:30:38<29:21,  5.76s/it]2026-04-12 03:07:29,814 | INFO | train_grpo_train | metrics_logged mode=train step=2994

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.303030303030304e-07, 'num_tokens': 6798980.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00023179415438789874, 'sampling/sampling_logp_difference/max': 0.0038368557579815388, 'sampling/importance_sampling_ratio/min': 0.99946129322052, 'sampling/importance_sampling_ratio/mean': 1.0002259016036987, 'sampling/importance_sampling_ratio/max': 1.0038442611694336, 'entropy': 0.0016921081114560366, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████ | 2994/3300 [4:30:38<29:21,  5.76s/it]INFO 04-12 03:07:30 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2995/3300 [4:30:43<27:33,  5.42s/it]2026-04-12 03:07:34,435 | INFO | train_grpo_train | metrics_logged mode=train step=2995

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.272727272727273e-07, 'num_tokens': 6800420.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00022380216978490353, 'sampling/sampling_logp_difference/max': 0.008112169802188873, 'sampling/importance_sampling_ratio/min': 0.9919206500053406, 'sampling/importance_sampling_ratio/mean': 1.0001194477081299, 'sampling/importance_sampling_ratio/max': 1.0056308507919312, 'entropy': 0.00262832005682867, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████ | 2995/3300 [4:30:43<27:33,  5.42s/it]INFO 04-12 03:07:34 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2996/3300 [4:30:49<28:04,  5.54s/it]2026-04-12 03:07:40,251 | INFO | train_grpo_train | metrics_logged mode=train step=2996

                                                     
{'loss': -0.0, 'grad_norm': 0.4544410705566406, 'learning_rate': 9.242424242424244e-07, 'num_tokens': 6802843.0, 'completions/mean_length': 131.875, 'completions/min_length': 130.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.875, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9994057416915894, 'rewards/meter/std': 4.619282117346302e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994057416915894, 'rewards/total_composite/std': 4.619282117346302e-05, 'reward': 0.9994057416915894, 'reward_std': 4.6202192606870085e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007839047349989414, 'sampling/sampling_logp_difference/max': 1.0239849090576172, 'sampling/importance_sampling_ratio/min': 0.3591609001159668, 'sampling/importance_sampling_ratio/mean': 1.0020174980163574, 'sampling/importance_sampling_ratio/max': 1.366064429283142, 'entropy': 0.0709062721580267, 'clip_ratio/low_mean': 0.000961538462433964, 'clip_ratio/low_min': 0.000961538462433964, 'clip_ratio/high_mean': 0.0009469697251915932, 'clip_ratio/high_max': 0.0009469697251915932, 'clip_ratio/region_mean': 0.0019085081876255572, 'reward_total_mean': 0.9994057416915894, 'reward_meter_mean': 0.9994057416915894, 'reward_meter_std': 4.619282117346302e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994057416915894, 'reward_total_composite_std': 4.619282117346302e-05, 'epoch': 0.12}

 91%|█████████ | 2996/3300 [4:30:49<28:04,  5.54s/it]INFO 04-12 03:07:40 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2997/3300 [4:30:59<35:23,  7.01s/it]2026-04-12 03:07:50,670 | INFO | train_grpo_train | metrics_logged mode=train step=2997

                                                     
{'loss': -0.0832, 'grad_norm': 2.213428020477295, 'learning_rate': 9.212121212121213e-07, 'num_tokens': 6806751.0, 'completions/mean_length': 507.5, 'completions/min_length': 497.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 503.0, 'completions/min_terminated_length': 497.0, 'completions/max_terminated_length': 512.0, 'rewards/meter/mean': 0.9938823580741882, 'rewards/meter/std': 0.00720559898763895, 'rewards/count_adherence/mean': 0.8472222089767456, 'rewards/count_adherence/std': 0.0257172379642725, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9873470664024353, 'rewards/repeat_penalty/std': 0.01747622899711132, 'rewards/total_composite/mean': 0.8312985897064209, 'rewards/total_composite/std': 0.026868145912885666, 'reward': 0.8312985897064209, 'reward_std': 0.026868147775530815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06921987235546112, 'sampling/sampling_logp_difference/max': 2.2429420948028564, 'sampling/importance_sampling_ratio/min': 0.10614575445652008, 'sampling/importance_sampling_ratio/mean': 1.0148636102676392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33224867284297943, 'clip_ratio/low_mean': 0.010527235455811024, 'clip_ratio/low_min': 0.010527235455811024, 'clip_ratio/high_mean': 0.012055294588208199, 'clip_ratio/high_max': 0.012055294588208199, 'clip_ratio/region_mean': 0.022582530044019222, 'reward_total_mean': 0.8312985897064209, 'reward_meter_mean': 0.9938823580741882, 'reward_meter_std': 0.00720559898763895, 'reward_count_adherence_mean': 0.8472222089767456, 'reward_count_adherence_std': 0.0257172379642725, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9873470664024353, 'reward_repeat_penalty_std': 0.01747622899711132, 'reward_total_composite_mean': 0.8312985897064209, 'reward_total_composite_std': 0.026868145912885666, 'epoch': 0.12}

 91%|█████████ | 2997/3300 [4:30:59<35:23,  7.01s/it]INFO 04-12 03:07:50 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2998/3300 [4:31:04<32:26,  6.44s/it]2026-04-12 03:07:55,796 | INFO | train_grpo_train | metrics_logged mode=train step=2998

                                                     
{'loss': -0.0063, 'grad_norm': 5.6068010330200195, 'learning_rate': 9.181818181818182e-07, 'num_tokens': 6808791.0, 'completions/mean_length': 86.0, 'completions/min_length': 84.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9505056738853455, 'rewards/meter/std': 0.01044482085853815, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9505056738853455, 'rewards/total_composite/std': 0.01044482085853815, 'reward': 0.9505056738853455, 'reward_std': 0.010444814339280128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.046680573374032974, 'sampling/sampling_logp_difference/max': 1.9859023094177246, 'sampling/importance_sampling_ratio/min': 0.1372566968202591, 'sampling/importance_sampling_ratio/mean': 1.004783272743225, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23842502385377884, 'clip_ratio/low_mean': 0.010312846396118402, 'clip_ratio/low_min': 0.010312846396118402, 'clip_ratio/high_mean': 0.03046388761140406, 'clip_ratio/high_max': 0.03046388761140406, 'clip_ratio/region_mean': 0.040776734007522464, 'reward_total_mean': 0.9505056738853455, 'reward_meter_mean': 0.9505056738853455, 'reward_meter_std': 0.01044482085853815, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9505056738853455, 'reward_total_composite_std': 0.01044482085853815, 'epoch': 0.12}

 91%|█████████ | 2998/3300 [4:31:04<32:26,  6.44s/it]INFO 04-12 03:07:56 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 2999/3300 [4:31:09<29:59,  5.98s/it]2026-04-12 03:08:00,676 | INFO | train_grpo_train | metrics_logged mode=train step=2999

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.151515151515153e-07, 'num_tokens': 6810823.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0003676189517136663, 'sampling/sampling_logp_difference/max': 0.009273192845284939, 'sampling/importance_sampling_ratio/min': 0.9973099231719971, 'sampling/importance_sampling_ratio/mean': 1.0003594160079956, 'sampling/importance_sampling_ratio/max': 1.0093164443969727, 'entropy': 0.0032808549876790494, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████ | 2999/3300 [4:31:09<29:59,  5.98s/it]INFO 04-12 03:08:00 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3000/3300 [4:31:14<27:50,  5.57s/it]2026-04-12 03:08:05,313 | INFO | train_grpo_train | metrics_logged mode=train step=3000

                                                     
{'loss': -0.003, 'grad_norm': 2.610877513885498, 'learning_rate': 9.121212121212122e-07, 'num_tokens': 6812551.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7007421255111694, 'rewards/meter/std': 0.24578078091144562, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7007421255111694, 'rewards/total_composite/std': 0.24578078091144562, 'reward': 0.7007421255111694, 'reward_std': 0.2457807958126068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0048951394855976105, 'sampling/sampling_logp_difference/max': 1.485097885131836, 'sampling/importance_sampling_ratio/min': 0.22648018598556519, 'sampling/importance_sampling_ratio/mean': 0.9997680187225342, 'sampling/importance_sampling_ratio/max': 1.2230128049850464, 'entropy': 0.01205012173159048, 'clip_ratio/low_mean': 0.002314814832061529, 'clip_ratio/low_min': 0.002314814832061529, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.002314814832061529, 'reward_total_mean': 0.7007421255111694, 'reward_meter_mean': 0.7007421255111694, 'reward_meter_std': 0.24578078091144562, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7007421255111694, 'reward_total_composite_std': 0.24578078091144562, 'epoch': 0.12}

 91%|█████████ | 3000/3300 [4:31:14<27:50,  5.57s/it]INFO 04-12 03:08:05 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:37,  3.37s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.42s/it]

 31%|███       | 4/13 [00:16<00:39,  4.41s/it]

 38%|███▊      | 5/13 [00:21<00:38,  4.83s/it]

 46%|████▌     | 6/13 [00:29<00:39,  5.64s/it]

 54%|█████▍    | 7/13 [00:34<00:33,  5.55s/it]

 62%|██████▏   | 8/13 [00:41<00:30,  6.06s/it]

 69%|██████▉   | 9/13 [00:49<00:25,  6.46s/it]

 77%|███████▋  | 10/13 [00:54<00:18,  6.04s/it]

 85%|████████▍ | 11/13 [01:00<00:12,  6.13s/it]

 92%|█████████▏| 12/13 [01:06<00:06,  6.14s/it]

100%|██████████| 13/13 [01:12<00:00,  5.98s/it]2026-04-12 03:09:24,857 | INFO | train_grpo_train | metrics_logged mode=eval step=3000

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 79.4116, 'eval_samples_per_second': 1.31, 'eval_steps_per_second': 0.164, 'eval_num_tokens': 6812551.0, 'eval_completions/mean_length': 213.93269230769232, 'eval_completions/min_length': 61.23076923076923, 'eval_completions/max_length': 421.53846153846155, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 210.74725341796875, 'eval_completions/min_terminated_length': 61.23076923076923, 'eval_completions/max_terminated_length': 412.7692307692308, 'eval_rewards/meter/mean': 0.7926333088141221, 'eval_rewards/meter/std': 0.3242100785629681, 'eval_rewards/count_adherence/mean': 0.9582245624982394, 'eval_rewards/count_adherence/std': 0.06270856238328494, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.9439093745671786, 'eval_rewards/repeat_penalty/std': 0.07702494555940995, 'eval_rewards/total_composite/mean': 0.7239083968676053, 'eval_rewards/total_composite/std': 0.3284210069821431, 'eval_reward': 0.7239083968676053, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03129683048106157, 'eval_sampling/sampling_logp_difference/max': 1.215673538354727, 'eval_sampling/importance_sampling_ratio/min': 0.3013949474463096, 'eval_sampling/importance_sampling_ratio/mean': 1.0090940640522883, 'eval_sampling/importance_sampling_ratio/max': 1.577669803912823, 'eval_entropy': 0.3522038482702695, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7239083968676053, 'eval_reward_meter_mean': 0.7926333088141221, 'eval_reward_meter_std': 0.3242100785629681, 'eval_reward_count_adherence_mean': 0.9582245624982394, 'eval_reward_count_adherence_std': 0.06270856238328494, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.9439093745671786, 'eval_reward_repeat_penalty_std': 0.07702494555940995, 'eval_reward_total_composite_mean': 0.7239083968676053, 'eval_reward_total_composite_std': 0.3284210069821431, 'epoch': 0.12}

 91%|█████████ | 3000/3300 [4:32:33<27:50,  5.57s/it]

100%|██████████| 13/13 [01:12<00:00,  5.98s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 91%|█████████ | 3001/3300 [4:32:41<2:29:50, 30.07s/it]2026-04-12 03:09:32,527 | INFO | train_grpo_train | metrics_logged mode=train step=3001

                                                       
{'loss': -0.0098, 'grad_norm': 12.373333930969238, 'learning_rate': 9.090909090909091e-07, 'num_tokens': 6814326.0, 'completions/mean_length': 62.875, 'completions/min_length': 60.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8256336450576782, 'rewards/meter/std': 0.3046148419380188, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8256336450576782, 'rewards/total_composite/std': 0.3046148419380188, 'reward': 0.8256336450576782, 'reward_std': 0.3046148419380188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06896661221981049, 'sampling/sampling_logp_difference/max': 1.8012008666992188, 'sampling/importance_sampling_ratio/min': 0.16510051488876343, 'sampling/importance_sampling_ratio/mean': 1.0055382251739502, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28370110131800175, 'clip_ratio/low_mean': 0.014480874873697758, 'clip_ratio/low_min': 0.014480874873697758, 'clip_ratio/high_mean': 0.03723430214449763, 'clip_ratio/high_max': 0.03723430214449763, 'clip_ratio/region_mean': 0.05171517701819539, 'reward_total_mean': 0.8256336450576782, 'reward_meter_mean': 0.8256336450576782, 'reward_meter_std': 0.3046148419380188, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8256336450576782, 'reward_total_composite_std': 0.3046148419380188, 'epoch': 0.12}

 91%|█████████ | 3001/3300 [4:32:41<2:29:50, 30.07s/it]INFO 04-12 03:09:32 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3002/3300 [4:32:46<1:52:10, 22.59s/it]2026-04-12 03:09:37,660 | INFO | train_grpo_train | metrics_logged mode=train step=3002

                                                       
{'loss': 0.0072, 'grad_norm': 3.2883293628692627, 'learning_rate': 9.060606060606062e-07, 'num_tokens': 6816267.0, 'completions/mean_length': 78.625, 'completions/min_length': 77.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9988808631896973, 'rewards/meter/std': 0.00034191427403129637, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988808631896973, 'rewards/total_composite/std': 0.00034191427403129637, 'reward': 0.9988808631896973, 'reward_std': 0.00034191476879641414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02113337628543377, 'sampling/sampling_logp_difference/max': 1.1642122268676758, 'sampling/importance_sampling_ratio/min': 0.3121684789657593, 'sampling/importance_sampling_ratio/mean': 1.0053126811981201, 'sampling/importance_sampling_ratio/max': 1.6066337823867798, 'entropy': 0.18326533026993275, 'clip_ratio/low_mean': 0.0031650641467422247, 'clip_ratio/low_min': 0.0031650641467422247, 'clip_ratio/high_mean': 0.01416189968585968, 'clip_ratio/high_max': 0.01416189968585968, 'clip_ratio/region_mean': 0.017326963832601905, 'reward_total_mean': 0.9988808631896973, 'reward_meter_mean': 0.9988808631896973, 'reward_meter_std': 0.00034191427403129637, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988808631896973, 'reward_total_composite_std': 0.00034191427403129637, 'epoch': 0.12}

 91%|█████████ | 3002/3300 [4:32:46<1:52:10, 22.59s/it]INFO 04-12 03:09:37 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3003/3300 [4:32:52<1:26:35, 17.49s/it]2026-04-12 03:09:43,264 | INFO | train_grpo_train | metrics_logged mode=train step=3003

                                                       
{'loss': 0.0131, 'grad_norm': 3.6394121646881104, 'learning_rate': 9.030303030303031e-07, 'num_tokens': 6818772.0, 'completions/mean_length': 137.125, 'completions/min_length': 132.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.125, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.9025046229362488, 'rewards/meter/std': 0.2555387020111084, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.8669579029083252, 'rewards/total_composite/std': 0.24918241798877716, 'reward': 0.8669579029083252, 'reward_std': 0.24918241798877716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03274257108569145, 'sampling/sampling_logp_difference/max': 1.0835933685302734, 'sampling/importance_sampling_ratio/min': 0.3383774161338806, 'sampling/importance_sampling_ratio/mean': 1.0107687711715698, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3485318683087826, 'clip_ratio/low_mean': 0.008140756515786052, 'clip_ratio/low_min': 0.008140756515786052, 'clip_ratio/high_mean': 0.012675640406087041, 'clip_ratio/high_max': 0.012675640406087041, 'clip_ratio/region_mean': 0.020816396921873093, 'reward_total_mean': 0.8669579029083252, 'reward_meter_mean': 0.9025046229362488, 'reward_meter_std': 0.2555387020111084, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.8669579029083252, 'reward_total_composite_std': 0.24918241798877716, 'epoch': 0.12}

 91%|█████████ | 3003/3300 [4:32:52<1:26:35, 17.49s/it]INFO 04-12 03:09:43 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3004/3300 [4:32:57<1:08:45, 13.94s/it]2026-04-12 03:09:48,930 | INFO | train_grpo_train | metrics_logged mode=train step=3004

                                                       
{'loss': 0.0001, 'grad_norm': 4.322078704833984, 'learning_rate': 9.000000000000001e-07, 'num_tokens': 6821058.0, 'completions/mean_length': 128.75, 'completions/min_length': 128.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.75, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.9970972537994385, 'rewards/meter/std': 0.001955085899680853, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9970972537994385, 'rewards/total_composite/std': 0.001955085899680853, 'reward': 0.9970972537994385, 'reward_std': 0.0019550775177776814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021072760224342346, 'sampling/sampling_logp_difference/max': 1.0552425384521484, 'sampling/importance_sampling_ratio/min': 0.3481079638004303, 'sampling/importance_sampling_ratio/mean': 1.0051413774490356, 'sampling/importance_sampling_ratio/max': 1.4015967845916748, 'entropy': 0.1789141520857811, 'clip_ratio/low_mean': 0.002914663462433964, 'clip_ratio/low_min': 0.002914663462433964, 'clip_ratio/high_mean': 0.00777464872226119, 'clip_ratio/high_max': 0.00777464872226119, 'clip_ratio/region_mean': 0.010689312184695154, 'reward_total_mean': 0.9970972537994385, 'reward_meter_mean': 0.9970972537994385, 'reward_meter_std': 0.001955085899680853, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9970972537994385, 'reward_total_composite_std': 0.001955085899680853, 'epoch': 0.12}

 91%|█████████ | 3004/3300 [4:32:57<1:08:45, 13.94s/it]INFO 04-12 03:09:49 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3005/3300 [4:33:04<57:32, 11.70s/it]  2026-04-12 03:09:55,415 | INFO | train_grpo_train | metrics_logged mode=train step=3005

                                                     
{'loss': 0.0018, 'grad_norm': 1.2387527227401733, 'learning_rate': 8.96969696969697e-07, 'num_tokens': 6824278.0, 'completions/mean_length': 194.5, 'completions/min_length': 190.0, 'completions/max_length': 197.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 194.5, 'completions/min_terminated_length': 190.0, 'completions/max_terminated_length': 197.0, 'rewards/meter/mean': 0.9990871548652649, 'rewards/meter/std': 8.784286910668015e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990871548652649, 'rewards/total_composite/std': 8.784286910668015e-05, 'reward': 0.9990871548652649, 'reward_std': 8.781959331827238e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03491505607962608, 'sampling/sampling_logp_difference/max': 0.9021329879760742, 'sampling/importance_sampling_ratio/min': 0.4057033956050873, 'sampling/importance_sampling_ratio/mean': 1.0082441568374634, 'sampling/importance_sampling_ratio/max': 1.4649758338928223, 'entropy': 0.3335018455982208, 'clip_ratio/low_mean': 0.009682861273176968, 'clip_ratio/low_min': 0.009682861273176968, 'clip_ratio/high_mean': 0.007038468029350042, 'clip_ratio/high_max': 0.007038468029350042, 'clip_ratio/region_mean': 0.01672132930252701, 'reward_total_mean': 0.9990871548652649, 'reward_meter_mean': 0.9990871548652649, 'reward_meter_std': 8.784286910668015e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990871548652649, 'reward_total_composite_std': 8.784286910668015e-05, 'epoch': 0.12}

 91%|█████████ | 3005/3300 [4:33:04<57:32, 11.70s/it]INFO 04-12 03:09:55 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3006/3300 [4:33:08<46:20,  9.46s/it]2026-04-12 03:09:59,631 | INFO | train_grpo_train | metrics_logged mode=train step=3006

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.93939393939394e-07, 'num_tokens': 6825838.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0001296020782319829, 'sampling/sampling_logp_difference/max': 0.003267202526330948, 'sampling/importance_sampling_ratio/min': 0.9980201125144958, 'sampling/importance_sampling_ratio/mean': 1.0001120567321777, 'sampling/importance_sampling_ratio/max': 1.003272533416748, 'entropy': 0.0011012316681444645, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████ | 3006/3300 [4:33:08<46:20,  9.46s/it]INFO 04-12 03:09:59 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3007/3300 [4:33:14<41:15,  8.45s/it]2026-04-12 03:10:05,709 | INFO | train_grpo_train | metrics_logged mode=train step=3007

                                                     
{'loss': 0.002, 'grad_norm': 3.9469776153564453, 'learning_rate': 8.90909090909091e-07, 'num_tokens': 6828661.0, 'completions/mean_length': 167.875, 'completions/min_length': 164.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 167.875, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.9986657500267029, 'rewards/meter/std': 0.0007131828460842371, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9709138870239258, 'rewards/total_composite/std': 0.05114205181598663, 'reward': 0.9709138870239258, 'reward_std': 0.05114204064011574, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04064127802848816, 'sampling/sampling_logp_difference/max': 2.306015968322754, 'sampling/importance_sampling_ratio/min': 0.09965749830007553, 'sampling/importance_sampling_ratio/mean': 1.0033116340637207, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31089969351887703, 'clip_ratio/low_mean': 0.006006060168147087, 'clip_ratio/low_min': 0.006006060168147087, 'clip_ratio/high_mean': 0.02804516162723303, 'clip_ratio/high_max': 0.02804516162723303, 'clip_ratio/region_mean': 0.034051221795380116, 'reward_total_mean': 0.9709138870239258, 'reward_meter_mean': 0.9986657500267029, 'reward_meter_std': 0.0007131828460842371, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9709138870239258, 'reward_total_composite_std': 0.05114205181598663, 'epoch': 0.12}

 91%|█████████ | 3007/3300 [4:33:14<41:15,  8.45s/it]INFO 04-12 03:10:05 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3008/3300 [4:33:23<41:49,  8.59s/it]2026-04-12 03:10:14,651 | INFO | train_grpo_train | metrics_logged mode=train step=3008

                                                     
{'loss': 0.0014, 'grad_norm': 1.8470708131790161, 'learning_rate': 8.87878787878788e-07, 'num_tokens': 6833680.0, 'completions/mean_length': 380.375, 'completions/min_length': 354.0, 'completions/max_length': 391.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 380.375, 'completions/min_terminated_length': 354.0, 'completions/max_terminated_length': 391.0, 'rewards/meter/mean': 0.9990596771240234, 'rewards/meter/std': 0.00017261884931940585, 'rewards/count_adherence/mean': 0.8863636255264282, 'rewards/count_adherence/std': 0.04208274558186531, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9404239654541016, 'rewards/repeat_penalty/std': 0.06556747853755951, 'rewards/total_composite/mean': 0.8320455551147461, 'rewards/total_composite/std': 0.06093808636069298, 'reward': 0.8320455551147461, 'reward_std': 0.060938093811273575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044413696974515915, 'sampling/sampling_logp_difference/max': 1.220409870147705, 'sampling/importance_sampling_ratio/min': 0.29510918259620667, 'sampling/importance_sampling_ratio/mean': 1.011826992034912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40278755873441696, 'clip_ratio/low_mean': 0.008936886792071164, 'clip_ratio/low_min': 0.008936886792071164, 'clip_ratio/high_mean': 0.013912130380049348, 'clip_ratio/high_max': 0.013912130380049348, 'clip_ratio/region_mean': 0.02284901717212051, 'reward_total_mean': 0.8320455551147461, 'reward_meter_mean': 0.9990596771240234, 'reward_meter_std': 0.00017261884931940585, 'reward_count_adherence_mean': 0.8863636255264282, 'reward_count_adherence_std': 0.04208274558186531, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9404239654541016, 'reward_repeat_penalty_std': 0.06556747853755951, 'reward_total_composite_mean': 0.8320455551147461, 'reward_total_composite_std': 0.06093808636069298, 'epoch': 0.12}

 91%|█████████ | 3008/3300 [4:33:23<41:49,  8.59s/it]INFO 04-12 03:10:14 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3009/3300 [4:33:28<35:44,  7.37s/it]2026-04-12 03:10:19,165 | INFO | train_grpo_train | metrics_logged mode=train step=3009

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.84848484848485e-07, 'num_tokens': 6835408.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002200069575337693, 'sampling/sampling_logp_difference/max': 0.0032224380411207676, 'sampling/importance_sampling_ratio/min': 0.9967827200889587, 'sampling/importance_sampling_ratio/mean': 1.000199556350708, 'sampling/importance_sampling_ratio/max': 1.0028958320617676, 'entropy': 0.0017017422651406378, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████ | 3009/3300 [4:33:28<35:44,  7.37s/it]INFO 04-12 03:10:19 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3010/3300 [4:33:32<32:03,  6.63s/it]2026-04-12 03:10:24,068 | INFO | train_grpo_train | metrics_logged mode=train step=3010

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.818181818181819e-07, 'num_tokens': 6836936.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 4.44791694462765e-05, 'sampling/sampling_logp_difference/max': 0.0007356764399446547, 'sampling/importance_sampling_ratio/min': 0.9992645978927612, 'sampling/importance_sampling_ratio/mean': 1.0000369548797607, 'sampling/importance_sampling_ratio/max': 1.0006588697433472, 'entropy': 0.0004426717623573495, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████ | 3010/3300 [4:33:33<32:03,  6.63s/it]INFO 04-12 03:10:24 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████ | 3011/3300 [4:33:38<29:54,  6.21s/it]2026-04-12 03:10:29,299 | INFO | train_grpo_train | metrics_logged mode=train step=3011

                                                     
{'loss': -0.0004, 'grad_norm': 0.6899365782737732, 'learning_rate': 8.787878787878788e-07, 'num_tokens': 6839162.0, 'completions/mean_length': 106.25, 'completions/min_length': 105.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.25, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.999291181564331, 'rewards/meter/std': 7.206718146335334e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999291181564331, 'rewards/total_composite/std': 7.206718146335334e-05, 'reward': 0.999291181564331, 'reward_std': 7.208617898868397e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013669301755726337, 'sampling/sampling_logp_difference/max': 0.952059268951416, 'sampling/importance_sampling_ratio/min': 0.3859454393386841, 'sampling/importance_sampling_ratio/mean': 1.0045570135116577, 'sampling/importance_sampling_ratio/max': 1.7733577489852905, 'entropy': 0.13793968316167593, 'clip_ratio/low_mean': 0.002358490601181984, 'clip_ratio/low_min': 0.002358490601181984, 'clip_ratio/high_mean': 0.003494060132652521, 'clip_ratio/high_max': 0.003494060132652521, 'clip_ratio/region_mean': 0.005852550733834505, 'reward_total_mean': 0.999291181564331, 'reward_meter_mean': 0.999291181564331, 'reward_meter_std': 7.206718146335334e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999291181564331, 'reward_total_composite_std': 7.206718146335334e-05, 'epoch': 0.12}

 91%|█████████ | 3011/3300 [4:33:38<29:54,  6.21s/it]INFO 04-12 03:10:29 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3012/3300 [4:33:43<28:13,  5.88s/it]2026-04-12 03:10:34,412 | INFO | train_grpo_train | metrics_logged mode=train step=3012

                                                     
{'loss': 0.0008, 'grad_norm': 3.873194694519043, 'learning_rate': 8.757575757575758e-07, 'num_tokens': 6841385.0, 'completions/mean_length': 99.875, 'completions/min_length': 99.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.875, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9991791248321533, 'rewards/meter/std': 0.00015561260806862265, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9741964340209961, 'rewards/total_composite/std': 0.07060983777046204, 'reward': 0.9741964340209961, 'reward_std': 0.07060984522104263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023857776075601578, 'sampling/sampling_logp_difference/max': 2.5868468284606934, 'sampling/importance_sampling_ratio/min': 0.07525696605443954, 'sampling/importance_sampling_ratio/mean': 1.0067286491394043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.159406297840178, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.008762876153923571, 'clip_ratio/high_max': 0.008762876153923571, 'clip_ratio/region_mean': 0.012550755054689944, 'reward_total_mean': 0.9741964340209961, 'reward_meter_mean': 0.9991791248321533, 'reward_meter_std': 0.00015561260806862265, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9741964340209961, 'reward_total_composite_std': 0.07060983777046204, 'epoch': 0.12}

 91%|█████████▏| 3012/3300 [4:33:43<28:13,  5.88s/it]INFO 04-12 03:10:34 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3013/3300 [4:33:48<26:49,  5.61s/it]2026-04-12 03:10:39,379 | INFO | train_grpo_train | metrics_logged mode=train step=3013

                                                     
{'loss': -0.0001, 'grad_norm': 1.6813957691192627, 'learning_rate': 8.727272727272728e-07, 'num_tokens': 6843559.0, 'completions/mean_length': 92.75, 'completions/min_length': 92.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.75, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9975481629371643, 'rewards/meter/std': 0.0004655694356188178, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975481629371643, 'rewards/total_composite/std': 0.0004655694356188178, 'reward': 0.9975481629371643, 'reward_std': 0.00046558206668123603, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01227449532598257, 'sampling/sampling_logp_difference/max': 1.0146255493164062, 'sampling/importance_sampling_ratio/min': 0.3625381588935852, 'sampling/importance_sampling_ratio/mean': 1.0000393390655518, 'sampling/importance_sampling_ratio/max': 1.3771406412124634, 'entropy': 0.06945714261382818, 'clip_ratio/low_mean': 0.0013440860202535987, 'clip_ratio/low_min': 0.0013440860202535987, 'clip_ratio/high_mean': 0.00945243111345917, 'clip_ratio/high_max': 0.00945243111345917, 'clip_ratio/region_mean': 0.010796517133712769, 'reward_total_mean': 0.9975481629371643, 'reward_meter_mean': 0.9975481629371643, 'reward_meter_std': 0.0004655694356188178, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975481629371643, 'reward_total_composite_std': 0.0004655694356188178, 'epoch': 0.12}

 91%|█████████▏| 3013/3300 [4:33:48<26:49,  5.61s/it]INFO 04-12 03:10:39 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3014/3300 [4:33:53<26:39,  5.59s/it]2026-04-12 03:10:44,940 | INFO | train_grpo_train | metrics_logged mode=train step=3014

                                                     
{'loss': -0.0004, 'grad_norm': 1.1659772396087646, 'learning_rate': 8.696969696969699e-07, 'num_tokens': 6846028.0, 'completions/mean_length': 141.625, 'completions/min_length': 140.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.625, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.999117910861969, 'rewards/meter/std': 0.00024422150454483926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999117910861969, 'rewards/total_composite/std': 0.00024422150454483926, 'reward': 0.999117910861969, 'reward_std': 0.0002442096301820129, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01364456582814455, 'sampling/sampling_logp_difference/max': 0.7532033920288086, 'sampling/importance_sampling_ratio/min': 0.47085580229759216, 'sampling/importance_sampling_ratio/mean': 1.0043392181396484, 'sampling/importance_sampling_ratio/max': 1.4682461023330688, 'entropy': 0.17108981497585773, 'clip_ratio/low_mean': 0.0017605633474886417, 'clip_ratio/low_min': 0.0017605633474886417, 'clip_ratio/high_mean': 0.007061248528771102, 'clip_ratio/high_max': 0.007061248528771102, 'clip_ratio/region_mean': 0.008821811876259744, 'reward_total_mean': 0.999117910861969, 'reward_meter_mean': 0.999117910861969, 'reward_meter_std': 0.00024422150454483926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999117910861969, 'reward_total_composite_std': 0.00024422150454483926, 'epoch': 0.12}

 91%|█████████▏| 3014/3300 [4:33:53<26:39,  5.59s/it]INFO 04-12 03:10:45 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3015/3300 [4:33:58<25:12,  5.31s/it]2026-04-12 03:10:49,569 | INFO | train_grpo_train | metrics_logged mode=train step=3015

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.666666666666668e-07, 'num_tokens': 6847788.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 9.00184822967276e-05, 'sampling/sampling_logp_difference/max': 0.0020202400628477335, 'sampling/importance_sampling_ratio/min': 0.9989049434661865, 'sampling/importance_sampling_ratio/mean': 1.0000646114349365, 'sampling/importance_sampling_ratio/max': 1.0020222663879395, 'entropy': 0.0011554057127796113, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 91%|█████████▏| 3015/3300 [4:33:58<25:12,  5.31s/it]INFO 04-12 03:10:49 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3016/3300 [4:34:04<26:16,  5.55s/it]2026-04-12 03:10:55,715 | INFO | train_grpo_train | metrics_logged mode=train step=3016

                                                     
{'loss': 0.0068, 'grad_norm': 3.0696825981140137, 'learning_rate': 8.636363636363637e-07, 'num_tokens': 6850815.0, 'completions/mean_length': 166.375, 'completions/min_length': 160.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.375, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9988635778427124, 'rewards/meter/std': 0.0006956413271836936, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988635778427124, 'rewards/total_composite/std': 0.0006956413271836936, 'reward': 0.9988635778427124, 'reward_std': 0.0006956506404094398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04618171229958534, 'sampling/sampling_logp_difference/max': 2.1471054553985596, 'sampling/importance_sampling_ratio/min': 0.11682181060314178, 'sampling/importance_sampling_ratio/mean': 1.0038132667541504, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32815663516521454, 'clip_ratio/low_mean': 0.008875739760696888, 'clip_ratio/low_min': 0.008875739760696888, 'clip_ratio/high_mean': 0.02051854378078133, 'clip_ratio/high_max': 0.02051854378078133, 'clip_ratio/region_mean': 0.029394283541478217, 'reward_total_mean': 0.9988635778427124, 'reward_meter_mean': 0.9988635778427124, 'reward_meter_std': 0.0006956413271836936, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988635778427124, 'reward_total_composite_std': 0.0006956413271836936, 'epoch': 0.12}

 91%|█████████▏| 3016/3300 [4:34:04<26:16,  5.55s/it]INFO 04-12 03:10:55 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3017/3300 [4:34:09<25:37,  5.43s/it]2026-04-12 03:11:00,849 | INFO | train_grpo_train | metrics_logged mode=train step=3017

                                                     
{'loss': 0.0083, 'grad_norm': 2.4396045207977295, 'learning_rate': 8.606060606060607e-07, 'num_tokens': 6852962.0, 'completions/mean_length': 97.375, 'completions/min_length': 96.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.375, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.992943286895752, 'rewards/meter/std': 0.0015025768661871552, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.992943286895752, 'rewards/total_composite/std': 0.0015025768661871552, 'reward': 0.992943286895752, 'reward_std': 0.0015025660395622253, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026139099150896072, 'sampling/sampling_logp_difference/max': 0.8849821090698242, 'sampling/importance_sampling_ratio/min': 0.41272157430648804, 'sampling/importance_sampling_ratio/mean': 1.0036437511444092, 'sampling/importance_sampling_ratio/max': 1.6669363975524902, 'entropy': 0.22009673342108727, 'clip_ratio/low_mean': 0.0025641699321568012, 'clip_ratio/low_min': 0.0025641699321568012, 'clip_ratio/high_mean': 0.014189229113981128, 'clip_ratio/high_max': 0.014189229113981128, 'clip_ratio/region_mean': 0.01675339904613793, 'reward_total_mean': 0.992943286895752, 'reward_meter_mean': 0.992943286895752, 'reward_meter_std': 0.0015025768661871552, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.992943286895752, 'reward_total_composite_std': 0.0015025768661871552, 'epoch': 0.12}

 91%|█████████▏| 3017/3300 [4:34:09<25:37,  5.43s/it]INFO 04-12 03:11:01 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3018/3300 [4:34:19<31:28,  6.70s/it]2026-04-12 03:11:10,518 | INFO | train_grpo_train | metrics_logged mode=train step=3018

                                                     
{'loss': -0.0211, 'grad_norm': 1.3273999691009521, 'learning_rate': 8.575757575757576e-07, 'num_tokens': 6857392.0, 'completions/mean_length': 375.75, 'completions/min_length': 348.0, 'completions/max_length': 392.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 375.75, 'completions/min_terminated_length': 348.0, 'completions/max_terminated_length': 392.0, 'rewards/meter/mean': 0.9991037845611572, 'rewards/meter/std': 9.388383477926254e-05, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.04704993963241577, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9593868255615234, 'rewards/repeat_penalty/std': 0.03774290531873703, 'rewards/total_composite/mean': 0.8386253118515015, 'rewards/total_composite/std': 0.054765164852142334, 'reward': 0.8386253118515015, 'reward_std': 0.05476517230272293, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04461507499217987, 'sampling/sampling_logp_difference/max': 1.3247604370117188, 'sampling/importance_sampling_ratio/min': 0.2658666670322418, 'sampling/importance_sampling_ratio/mean': 1.012117624282837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4122394025325775, 'clip_ratio/low_mean': 0.009317906049545854, 'clip_ratio/low_min': 0.009317906049545854, 'clip_ratio/high_mean': 0.016257623909041286, 'clip_ratio/high_max': 0.016257623909041286, 'clip_ratio/region_mean': 0.02557552995858714, 'reward_total_mean': 0.8386253118515015, 'reward_meter_mean': 0.9991037845611572, 'reward_meter_std': 9.388383477926254e-05, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.04704993963241577, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9593868255615234, 'reward_repeat_penalty_std': 0.03774290531873703, 'reward_total_composite_mean': 0.8386253118515015, 'reward_total_composite_std': 0.054765164852142334, 'epoch': 0.12}

 91%|█████████▏| 3018/3300 [4:34:19<31:28,  6.70s/it]INFO 04-12 03:11:10 [block_pool.py:378] Successfully reset prefix cache

 91%|█████████▏| 3019/3300 [4:34:24<28:49,  6.16s/it]2026-04-12 03:11:15,410 | INFO | train_grpo_train | metrics_logged mode=train step=3019

                                                     
{'loss': -0.0003, 'grad_norm': 0.24967576563358307, 'learning_rate': 8.545454545454546e-07, 'num_tokens': 6859104.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973341226577759, 'rewards/meter/std': 9.24646246858174e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973341226577759, 'rewards/total_composite/std': 9.24646246858174e-06, 'reward': 0.9973341226577759, 'reward_std': 9.240244253305718e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002222540322691202, 'sampling/sampling_logp_difference/max': 0.16164398193359375, 'sampling/importance_sampling_ratio/min': 0.8715769052505493, 'sampling/importance_sampling_ratio/mean': 1.0005959272384644, 'sampling/importance_sampling_ratio/max': 1.1754417419433594, 'entropy': 0.021813111379742622, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.9973341226577759, 'reward_meter_mean': 0.9973341226577759, 'reward_meter_std': 9.24646246858174e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973341226577759, 'reward_total_composite_std': 9.24646246858174e-06, 'epoch': 0.12}

 91%|█████████▏| 3019/3300 [4:34:24<28:49,  6.16s/it]INFO 04-12 03:11:15 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3020/3300 [4:34:29<27:25,  5.88s/it]2026-04-12 03:11:20,635 | INFO | train_grpo_train | metrics_logged mode=train step=3020

                                                     
{'loss': 0.0005, 'grad_norm': 0.15372411906719208, 'learning_rate': 8.515151515151515e-07, 'num_tokens': 6861223.0, 'completions/mean_length': 97.875, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994118809700012, 'rewards/meter/std': 1.710540527710691e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994118809700012, 'rewards/total_composite/std': 1.710540527710691e-05, 'reward': 0.9994118809700012, 'reward_std': 1.710999640636146e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006246503908187151, 'sampling/sampling_logp_difference/max': 0.6895544528961182, 'sampling/importance_sampling_ratio/min': 0.5627803802490234, 'sampling/importance_sampling_ratio/mean': 1.00173020362854, 'sampling/importance_sampling_ratio/max': 1.9928275346755981, 'entropy': 0.037544872146099806, 'clip_ratio/low_mean': 0.0012755101779475808, 'clip_ratio/low_min': 0.0012755101779475808, 'clip_ratio/high_mean': 0.0025641699321568012, 'clip_ratio/high_max': 0.0025641699321568012, 'clip_ratio/region_mean': 0.003839680110104382, 'reward_total_mean': 0.9994118809700012, 'reward_meter_mean': 0.9994118809700012, 'reward_meter_std': 1.710540527710691e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994118809700012, 'reward_total_composite_std': 1.710540527710691e-05, 'epoch': 0.12}

 92%|█████████▏| 3020/3300 [4:34:29<27:25,  5.88s/it]INFO 04-12 03:11:20 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3021/3300 [4:34:35<28:03,  6.03s/it]2026-04-12 03:11:27,036 | INFO | train_grpo_train | metrics_logged mode=train step=3021

                                                     
{'loss': 0.0012, 'grad_norm': 1.7028465270996094, 'learning_rate': 8.484848484848486e-07, 'num_tokens': 6864084.0, 'completions/mean_length': 177.625, 'completions/min_length': 177.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 177.625, 'completions/min_terminated_length': 177.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.9990875720977783, 'rewards/meter/std': 0.00012965931091457605, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9713343381881714, 'rewards/total_composite/std': 0.05137185752391815, 'reward': 0.9713343381881714, 'reward_std': 0.05137185752391815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018357736989855766, 'sampling/sampling_logp_difference/max': 0.9017171859741211, 'sampling/importance_sampling_ratio/min': 0.405872106552124, 'sampling/importance_sampling_ratio/mean': 1.004658818244934, 'sampling/importance_sampling_ratio/max': 1.6029589176177979, 'entropy': 0.20351556316018105, 'clip_ratio/low_mean': 0.0021146765793673694, 'clip_ratio/low_min': 0.0021146765793673694, 'clip_ratio/high_mean': 0.011962095857597888, 'clip_ratio/high_max': 0.011962095857597888, 'clip_ratio/region_mean': 0.014076772436965257, 'reward_total_mean': 0.9713343381881714, 'reward_meter_mean': 0.9990875720977783, 'reward_meter_std': 0.00012965931091457605, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9713343381881714, 'reward_total_composite_std': 0.05137185752391815, 'epoch': 0.12}

 92%|█████████▏| 3021/3300 [4:34:36<28:03,  6.03s/it]INFO 04-12 03:11:27 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3022/3300 [4:34:40<25:31,  5.51s/it]2026-04-12 03:11:31,300 | INFO | train_grpo_train | metrics_logged mode=train step=3022

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.454545454545456e-07, 'num_tokens': 6865556.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 5.831335874972865e-05, 'sampling/sampling_logp_difference/max': 0.0005344097153283656, 'sampling/importance_sampling_ratio/min': 0.9995866417884827, 'sampling/importance_sampling_ratio/mean': 1.0000536441802979, 'sampling/importance_sampling_ratio/max': 1.0005345344543457, 'entropy': 0.00045082847282174043, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3022/3300 [4:34:40<25:31,  5.51s/it]INFO 04-12 03:11:31 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3023/3300 [4:34:45<24:42,  5.35s/it]2026-04-12 03:11:36,290 | INFO | train_grpo_train | metrics_logged mode=train step=3023

                                                     
{'loss': 0.0024, 'grad_norm': 3.5771987438201904, 'learning_rate': 8.424242424242425e-07, 'num_tokens': 6867729.0, 'completions/mean_length': 102.625, 'completions/min_length': 101.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.625, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9883763194084167, 'rewards/meter/std': 0.009592375718057156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9883763194084167, 'rewards/total_composite/std': 0.009592375718057156, 'reward': 0.9883763194084167, 'reward_std': 0.009592383168637753, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039184827357530594, 'sampling/sampling_logp_difference/max': 0.941868782043457, 'sampling/importance_sampling_ratio/min': 0.38989853858947754, 'sampling/importance_sampling_ratio/mean': 1.0047192573547363, 'sampling/importance_sampling_ratio/max': 1.92518150806427, 'entropy': 0.359358724206686, 'clip_ratio/low_mean': 0.012183514423668385, 'clip_ratio/low_min': 0.012183514423668385, 'clip_ratio/high_mean': 0.017015473917126656, 'clip_ratio/high_max': 0.017015473917126656, 'clip_ratio/region_mean': 0.02919898834079504, 'reward_total_mean': 0.9883763194084167, 'reward_meter_mean': 0.9883763194084167, 'reward_meter_std': 0.009592375718057156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9883763194084167, 'reward_total_composite_std': 0.009592375718057156, 'epoch': 0.12}

 92%|█████████▏| 3023/3300 [4:34:45<24:42,  5.35s/it]INFO 04-12 03:11:36 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3024/3300 [4:34:49<23:51,  5.19s/it]2026-04-12 03:11:41,089 | INFO | train_grpo_train | metrics_logged mode=train step=3024

                                                     
{'loss': -0.0002, 'grad_norm': 0.03466307371854782, 'learning_rate': 8.393939393939395e-07, 'num_tokens': 6869889.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994069337844849, 'rewards/meter/std': 1.776606950443238e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994069337844849, 'rewards/total_composite/std': 1.776606950443238e-06, 'reward': 0.9994069337844849, 'reward_std': 1.775535338310874e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005284504033625126, 'sampling/sampling_logp_difference/max': 0.844066858291626, 'sampling/importance_sampling_ratio/min': 0.4993957281112671, 'sampling/importance_sampling_ratio/mean': 1.0014625787734985, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.031393167562782764, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0012755101779475808, 'clip_ratio/high_max': 0.0012755101779475808, 'clip_ratio/region_mean': 0.0012755101779475808, 'reward_total_mean': 0.9994069337844849, 'reward_meter_mean': 0.9994069337844849, 'reward_meter_std': 1.776606950443238e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994069337844849, 'reward_total_composite_std': 1.776606950443238e-06, 'epoch': 0.12}

 92%|█████████▏| 3024/3300 [4:34:50<23:51,  5.19s/it]INFO 04-12 03:11:41 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3025/3300 [4:34:55<23:46,  5.19s/it]2026-04-12 03:11:46,275 | INFO | train_grpo_train | metrics_logged mode=train step=3025

                                                     
{'loss': -0.0004, 'grad_norm': 0.8292710185050964, 'learning_rate': 8.363636363636364e-07, 'num_tokens': 6872288.0, 'completions/mean_length': 131.875, 'completions/min_length': 131.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.875, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.99941086769104, 'rewards/meter/std': 6.445148028433323e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.99941086769104, 'rewards/total_composite/std': 6.445148028433323e-05, 'reward': 0.99941086769104, 'reward_std': 6.445525650633499e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008050598204135895, 'sampling/sampling_logp_difference/max': 0.6427021026611328, 'sampling/importance_sampling_ratio/min': 0.5258695483207703, 'sampling/importance_sampling_ratio/mean': 1.0006563663482666, 'sampling/importance_sampling_ratio/max': 1.33772611618042, 'entropy': 0.0625457102432847, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006628788076341152, 'clip_ratio/high_max': 0.006628788076341152, 'clip_ratio/region_mean': 0.006628788076341152, 'reward_total_mean': 0.99941086769104, 'reward_meter_mean': 0.99941086769104, 'reward_meter_std': 6.445148028433323e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.99941086769104, 'reward_total_composite_std': 6.445148028433323e-05, 'epoch': 0.12}

 92%|█████████▏| 3025/3300 [4:34:55<23:46,  5.19s/it]INFO 04-12 03:11:46 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3026/3300 [4:35:00<23:46,  5.21s/it]2026-04-12 03:11:51,530 | INFO | train_grpo_train | metrics_logged mode=train step=3026

                                                     
{'loss': 0.0631, 'grad_norm': 6.055728435516357, 'learning_rate': 8.333333333333333e-07, 'num_tokens': 6874658.0, 'completions/mean_length': 116.25, 'completions/min_length': 107.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.25, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.9515209197998047, 'rewards/meter/std': 0.023815272375941277, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9633838534355164, 'rewards/repeat_penalty/std': 0.05091821402311325, 'rewards/total_composite/mean': 0.8265341520309448, 'rewards/total_composite/std': 0.12215512990951538, 'reward': 0.8265341520309448, 'reward_std': 0.12215511500835419, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06652670353651047, 'sampling/sampling_logp_difference/max': 2.105895757675171, 'sampling/importance_sampling_ratio/min': 0.1217365711927414, 'sampling/importance_sampling_ratio/mean': 0.9996247291564941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30979629419744015, 'clip_ratio/low_mean': 0.023373008705675602, 'clip_ratio/low_min': 0.023373008705675602, 'clip_ratio/high_mean': 0.01498432899825275, 'clip_ratio/high_max': 0.01498432899825275, 'clip_ratio/region_mean': 0.03835733770392835, 'reward_total_mean': 0.8265341520309448, 'reward_meter_mean': 0.9515209197998047, 'reward_meter_std': 0.023815272375941277, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9633838534355164, 'reward_repeat_penalty_std': 0.05091821402311325, 'reward_total_composite_mean': 0.8265341520309448, 'reward_total_composite_std': 0.12215512990951538, 'epoch': 0.12}

 92%|█████████▏| 3026/3300 [4:35:00<23:46,  5.21s/it]INFO 04-12 03:11:51 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3027/3300 [4:35:05<23:08,  5.08s/it]2026-04-12 03:11:56,324 | INFO | train_grpo_train | metrics_logged mode=train step=3027

                                                     
{'loss': -0.0003, 'grad_norm': 1.3954490423202515, 'learning_rate': 8.303030303030303e-07, 'num_tokens': 6876567.0, 'completions/mean_length': 78.625, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9990073442459106, 'rewards/meter/std': 0.000114113834570162, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990073442459106, 'rewards/total_composite/std': 0.000114113834570162, 'reward': 0.9990073442459106, 'reward_std': 0.00011411593004595488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02237360179424286, 'sampling/sampling_logp_difference/max': 0.7234420776367188, 'sampling/importance_sampling_ratio/min': 0.48507970571517944, 'sampling/importance_sampling_ratio/mean': 1.00530207157135, 'sampling/importance_sampling_ratio/max': 1.6200668811798096, 'entropy': 0.18712522648274899, 'clip_ratio/low_mean': 0.007972249411977828, 'clip_ratio/low_min': 0.007972249411977828, 'clip_ratio/high_mean': 0.014263797434978187, 'clip_ratio/high_max': 0.014263797434978187, 'clip_ratio/region_mean': 0.022236046846956015, 'reward_total_mean': 0.9990073442459106, 'reward_meter_mean': 0.9990073442459106, 'reward_meter_std': 0.000114113834570162, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990073442459106, 'reward_total_composite_std': 0.000114113834570162, 'epoch': 0.12}

 92%|█████████▏| 3027/3300 [4:35:05<23:08,  5.08s/it]INFO 04-12 03:11:56 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3028/3300 [4:35:09<22:07,  4.88s/it]2026-04-12 03:12:00,733 | INFO | train_grpo_train | metrics_logged mode=train step=3028

                                                     
{'loss': -0.0, 'grad_norm': 0.22835104167461395, 'learning_rate': 8.272727272727274e-07, 'num_tokens': 6878303.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973360300064087, 'rewards/meter/std': 8.113272997434251e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973360300064087, 'rewards/total_composite/std': 8.113272997434251e-06, 'reward': 0.9973360300064087, 'reward_std': 8.10424353403505e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0027969391085207462, 'sampling/sampling_logp_difference/max': 0.7328996658325195, 'sampling/importance_sampling_ratio/min': 0.48051363229751587, 'sampling/importance_sampling_ratio/mean': 0.9999331831932068, 'sampling/importance_sampling_ratio/max': 1.1025102138519287, 'entropy': 0.015828296658582985, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9973360300064087, 'reward_meter_mean': 0.9973360300064087, 'reward_meter_std': 8.113272997434251e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973360300064087, 'reward_total_composite_std': 8.113272997434251e-06, 'epoch': 0.12}

 92%|█████████▏| 3028/3300 [4:35:09<22:07,  4.88s/it]INFO 04-12 03:12:00 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3029/3300 [4:35:14<22:18,  4.94s/it]2026-04-12 03:12:05,800 | INFO | train_grpo_train | metrics_logged mode=train step=3029

                                                     
{'loss': 0.0018, 'grad_norm': 0.5074459314346313, 'learning_rate': 8.242424242424244e-07, 'num_tokens': 6880686.0, 'completions/mean_length': 107.875, 'completions/min_length': 107.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.875, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9992820620536804, 'rewards/meter/std': 4.010711199953221e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992820620536804, 'rewards/total_composite/std': 4.010711199953221e-05, 'reward': 0.9992820620536804, 'reward_std': 4.012084173155017e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018780170008540154, 'sampling/sampling_logp_difference/max': 1.2216401100158691, 'sampling/importance_sampling_ratio/min': 0.29474636912345886, 'sampling/importance_sampling_ratio/mean': 1.0014318227767944, 'sampling/importance_sampling_ratio/max': 1.7053853273391724, 'entropy': 0.1494951182976365, 'clip_ratio/low_mean': 0.008082556771114469, 'clip_ratio/low_min': 0.008082556771114469, 'clip_ratio/high_mean': 0.008112668758258224, 'clip_ratio/high_max': 0.008112668758258224, 'clip_ratio/region_mean': 0.016195225529372692, 'reward_total_mean': 0.9992820620536804, 'reward_meter_mean': 0.9992820620536804, 'reward_meter_std': 4.010711199953221e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992820620536804, 'reward_total_composite_std': 4.010711199953221e-05, 'epoch': 0.12}

 92%|█████████▏| 3029/3300 [4:35:14<22:18,  4.94s/it]INFO 04-12 03:12:06 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3030/3300 [4:35:22<26:19,  5.85s/it]2026-04-12 03:12:13,782 | INFO | train_grpo_train | metrics_logged mode=train step=3030

                                                     
{'loss': -0.0054, 'grad_norm': 1.8386452198028564, 'learning_rate': 8.212121212121213e-07, 'num_tokens': 6885074.0, 'completions/mean_length': 334.5, 'completions/min_length': 323.0, 'completions/max_length': 343.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 334.5, 'completions/min_terminated_length': 323.0, 'completions/max_terminated_length': 343.0, 'rewards/meter/mean': 0.9939321279525757, 'rewards/meter/std': 0.0059409006498754025, 'rewards/count_adherence/mean': 0.9090909361839294, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9621710777282715, 'rewards/repeat_penalty/std': 0.08768600970506668, 'rewards/total_composite/mean': 0.8695409297943115, 'rewards/total_composite/std': 0.08088699728250504, 'reward': 0.8695409297943115, 'reward_std': 0.08088699728250504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.054148994386196136, 'sampling/sampling_logp_difference/max': 2.918501138687134, 'sampling/importance_sampling_ratio/min': 0.05401458963751793, 'sampling/importance_sampling_ratio/mean': 1.0112972259521484, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4611641392111778, 'clip_ratio/low_mean': 0.0044835947919636965, 'clip_ratio/low_min': 0.0044835947919636965, 'clip_ratio/high_mean': 0.033679026179015636, 'clip_ratio/high_max': 0.033679026179015636, 'clip_ratio/region_mean': 0.03816262097097933, 'reward_total_mean': 0.8695409297943115, 'reward_meter_mean': 0.9939321279525757, 'reward_meter_std': 0.0059409006498754025, 'reward_count_adherence_mean': 0.9090909361839294, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9621710777282715, 'reward_repeat_penalty_std': 0.08768600970506668, 'reward_total_composite_mean': 0.8695409297943115, 'reward_total_composite_std': 0.08088699728250504, 'epoch': 0.12}

 92%|█████████▏| 3030/3300 [4:35:22<26:19,  5.85s/it]INFO 04-12 03:12:14 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3031/3300 [4:35:28<26:12,  5.84s/it]2026-04-12 03:12:19,626 | INFO | train_grpo_train | metrics_logged mode=train step=3031

                                                     
{'loss': -0.0067, 'grad_norm': 2.3149688243865967, 'learning_rate': 8.181818181818182e-07, 'num_tokens': 6887904.0, 'completions/mean_length': 154.75, 'completions/min_length': 150.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.75, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.9982709884643555, 'rewards/meter/std': 0.0024356048088520765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9982709884643555, 'rewards/total_composite/std': 0.0024356048088520765, 'reward': 0.9982709884643555, 'reward_std': 0.0024356169160455465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03701268509030342, 'sampling/sampling_logp_difference/max': 1.422220230102539, 'sampling/importance_sampling_ratio/min': 0.24117796123027802, 'sampling/importance_sampling_ratio/mean': 1.009082555770874, 'sampling/importance_sampling_ratio/max': 1.942898154258728, 'entropy': 0.33194397762417793, 'clip_ratio/low_mean': 0.0033333334140479565, 'clip_ratio/low_min': 0.0033333334140479565, 'clip_ratio/high_mean': 0.02090477745514363, 'clip_ratio/high_max': 0.02090477745514363, 'clip_ratio/region_mean': 0.024238110869191587, 'reward_total_mean': 0.9982709884643555, 'reward_meter_mean': 0.9982709884643555, 'reward_meter_std': 0.0024356048088520765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9982709884643555, 'reward_total_composite_std': 0.0024356048088520765, 'epoch': 0.12}

 92%|█████████▏| 3031/3300 [4:35:28<26:12,  5.84s/it]INFO 04-12 03:12:19 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3032/3300 [4:35:33<24:39,  5.52s/it]2026-04-12 03:12:24,387 | INFO | train_grpo_train | metrics_logged mode=train step=3032

                                                     
{'loss': -0.0001, 'grad_norm': 0.31778454780578613, 'learning_rate': 8.151515151515152e-07, 'num_tokens': 6889852.0, 'completions/mean_length': 66.5, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981399774551392, 'rewards/meter/std': 1.4936525076336693e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981399774551392, 'rewards/total_composite/std': 1.4936525076336693e-05, 'reward': 0.9981399774551392, 'reward_std': 1.4939930224500131e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007410375867038965, 'sampling/sampling_logp_difference/max': 0.7496395111083984, 'sampling/importance_sampling_ratio/min': 0.4725368618965149, 'sampling/importance_sampling_ratio/mean': 0.9997544884681702, 'sampling/importance_sampling_ratio/max': 1.3944506645202637, 'entropy': 0.05587592348456383, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.007490954245440662, 'clip_ratio/high_max': 0.007490954245440662, 'clip_ratio/region_mean': 0.009384893695823848, 'reward_total_mean': 0.9981399774551392, 'reward_meter_mean': 0.9981399774551392, 'reward_meter_std': 1.4936525076336693e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981399774551392, 'reward_total_composite_std': 1.4936525076336693e-05, 'epoch': 0.12}

 92%|█████████▏| 3032/3300 [4:35:33<24:39,  5.52s/it]INFO 04-12 03:12:24 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3033/3300 [4:35:39<25:27,  5.72s/it]2026-04-12 03:12:30,549 | INFO | train_grpo_train | metrics_logged mode=train step=3033

                                                     
{'loss': 0.0028, 'grad_norm': 2.056260347366333, 'learning_rate': 8.121212121212121e-07, 'num_tokens': 6892787.0, 'completions/mean_length': 186.875, 'completions/min_length': 183.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.875, 'completions/min_terminated_length': 183.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9618879556655884, 'rewards/meter/std': 0.06856533139944077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9090909361839294, 'rewards/repeat_penalty/std': 0.0971859022974968, 'rewards/total_composite/mean': 0.8737800121307373, 'rewards/total_composite/std': 0.10824880003929138, 'reward': 0.8737800121307373, 'reward_std': 0.10824880748987198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034711167216300964, 'sampling/sampling_logp_difference/max': 1.0839767456054688, 'sampling/importance_sampling_ratio/min': 0.33824771642684937, 'sampling/importance_sampling_ratio/mean': 1.0074325799942017, 'sampling/importance_sampling_ratio/max': 1.8432459831237793, 'entropy': 0.36450883373618126, 'clip_ratio/low_mean': 0.006038419320248067, 'clip_ratio/low_min': 0.006038419320248067, 'clip_ratio/high_mean': 0.016598281217738986, 'clip_ratio/high_max': 0.016598281217738986, 'clip_ratio/region_mean': 0.022636700537987053, 'reward_total_mean': 0.8737800121307373, 'reward_meter_mean': 0.9618879556655884, 'reward_meter_std': 0.06856533139944077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9090909361839294, 'reward_repeat_penalty_std': 0.0971859022974968, 'reward_total_composite_mean': 0.8737800121307373, 'reward_total_composite_std': 0.10824880003929138, 'epoch': 0.12}

 92%|█████████▏| 3033/3300 [4:35:39<25:27,  5.72s/it]INFO 04-12 03:12:30 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3034/3300 [4:35:43<23:11,  5.23s/it]2026-04-12 03:12:34,640 | INFO | train_grpo_train | metrics_logged mode=train step=3034

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.09090909090909e-07, 'num_tokens': 6894235.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0001638415560591966, 'sampling/sampling_logp_difference/max': 0.004102423787117004, 'sampling/importance_sampling_ratio/min': 0.9988238215446472, 'sampling/importance_sampling_ratio/mean': 1.000151515007019, 'sampling/importance_sampling_ratio/max': 1.0041109323501587, 'entropy': 0.0012948041403433308, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3034/3300 [4:35:43<23:11,  5.23s/it]INFO 04-12 03:12:34 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3035/3300 [4:35:47<21:52,  4.95s/it]2026-04-12 03:12:38,943 | INFO | train_grpo_train | metrics_logged mode=train step=3035

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.060606060606062e-07, 'num_tokens': 6895891.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002374651812715456, 'sampling/sampling_logp_difference/max': 0.0032967175357043743, 'sampling/importance_sampling_ratio/min': 0.9998509287834167, 'sampling/importance_sampling_ratio/mean': 1.0002354383468628, 'sampling/importance_sampling_ratio/max': 1.0033022165298462, 'entropy': 0.0020439386717043817, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3035/3300 [4:35:47<21:52,  4.95s/it]INFO 04-12 03:12:39 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3036/3300 [4:35:51<20:31,  4.67s/it]2026-04-12 03:12:42,938 | INFO | train_grpo_train | metrics_logged mode=train step=3036

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.030303030303031e-07, 'num_tokens': 6897435.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00020185003813821822, 'sampling/sampling_logp_difference/max': 0.008230682462453842, 'sampling/importance_sampling_ratio/min': 0.9977694749832153, 'sampling/importance_sampling_ratio/mean': 1.0001778602600098, 'sampling/importance_sampling_ratio/max': 1.0082646608352661, 'entropy': 0.0024824678112054244, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3036/3300 [4:35:51<20:31,  4.67s/it]INFO 04-12 03:12:43 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3037/3300 [4:35:56<20:55,  4.77s/it]2026-04-12 03:12:47,974 | INFO | train_grpo_train | metrics_logged mode=train step=3037

                                                     
{'loss': 0.0, 'grad_norm': 4.432851791381836, 'learning_rate': 8.000000000000001e-07, 'num_tokens': 6899483.0, 'completions/mean_length': 104.0, 'completions/min_length': 102.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.0, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9864717721939087, 'rewards/meter/std': 0.024548275396227837, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9615679979324341, 'rewards/total_composite/std': 0.07080353796482086, 'reward': 0.9615679979324341, 'reward_std': 0.07080353796482086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03678695112466812, 'sampling/sampling_logp_difference/max': 1.2483813762664795, 'sampling/importance_sampling_ratio/min': 0.2869689166545868, 'sampling/importance_sampling_ratio/mean': 1.0080146789550781, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2963132858276367, 'clip_ratio/low_mean': 0.0012254902394488454, 'clip_ratio/low_min': 0.0012254902394488454, 'clip_ratio/high_mean': 0.023894709534943104, 'clip_ratio/high_max': 0.023894709534943104, 'clip_ratio/region_mean': 0.02512019977439195, 'reward_total_mean': 0.9615679979324341, 'reward_meter_mean': 0.9864717721939087, 'reward_meter_std': 0.024548275396227837, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9615679979324341, 'reward_total_composite_std': 0.07080353796482086, 'epoch': 0.12}

 92%|█████████▏| 3037/3300 [4:35:56<20:55,  4.77s/it]INFO 04-12 03:12:48 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3038/3300 [4:36:06<27:48,  6.37s/it]2026-04-12 03:12:58,052 | INFO | train_grpo_train | metrics_logged mode=train step=3038

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.96969696969697e-07, 'num_tokens': 6900995.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.933377742767334, 'rewards/meter/std': 0.09035933762788773, 'rewards/count_adherence/mean': 0.6749999523162842, 'rewards/count_adherence/std': 0.02672613225877285, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9716880321502686, 'rewards/repeat_penalty/std': 0.01748696342110634, 'rewards/total_composite/mean': 0.6106069087982178, 'rewards/total_composite/std': 0.045595090836286545, 'reward': 0.6106069087982178, 'reward_std': 0.045595090836286545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.6106069087982178, 'reward_meter_mean': 0.933377742767334, 'reward_meter_std': 0.09035933762788773, 'reward_count_adherence_mean': 0.6749999523162842, 'reward_count_adherence_std': 0.02672613225877285, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9716880321502686, 'reward_repeat_penalty_std': 0.01748696342110634, 'reward_total_composite_mean': 0.6106069087982178, 'reward_total_composite_std': 0.045595090836286545, 'epoch': 0.12}

 92%|█████████▏| 3038/3300 [4:36:07<27:48,  6.37s/it]INFO 04-12 03:12:58 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3039/3300 [4:36:13<27:58,  6.43s/it]2026-04-12 03:13:04,626 | INFO | train_grpo_train | metrics_logged mode=train step=3039

                                                     
{'loss': 0.0015, 'grad_norm': 1.2539805173873901, 'learning_rate': 7.939393939393939e-07, 'num_tokens': 6904381.0, 'completions/mean_length': 233.25, 'completions/min_length': 230.0, 'completions/max_length': 238.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 233.25, 'completions/min_terminated_length': 230.0, 'completions/max_terminated_length': 238.0, 'rewards/meter/mean': 0.9988183975219727, 'rewards/meter/std': 0.000870219839271158, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9886363744735718, 'rewards/repeat_penalty/std': 0.03214120864868164, 'rewards/total_composite/mean': 0.987464189529419, 'rewards/total_composite/std': 0.03198371082544327, 'reward': 0.987464189529419, 'reward_std': 0.031983714550733566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03573669493198395, 'sampling/sampling_logp_difference/max': 1.3547391891479492, 'sampling/importance_sampling_ratio/min': 0.25801458954811096, 'sampling/importance_sampling_ratio/mean': 1.012012243270874, 'sampling/importance_sampling_ratio/max': 1.7767890691757202, 'entropy': 0.3743705749511719, 'clip_ratio/low_mean': 0.0021551724057644606, 'clip_ratio/low_min': 0.0021551724057644606, 'clip_ratio/high_mean': 0.022452297853305936, 'clip_ratio/high_max': 0.022452297853305936, 'clip_ratio/region_mean': 0.024607470259070396, 'reward_total_mean': 0.987464189529419, 'reward_meter_mean': 0.9988183975219727, 'reward_meter_std': 0.000870219839271158, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9886363744735718, 'reward_repeat_penalty_std': 0.03214120864868164, 'reward_total_composite_mean': 0.987464189529419, 'reward_total_composite_std': 0.03198371082544327, 'epoch': 0.12}

 92%|█████████▏| 3039/3300 [4:36:13<27:58,  6.43s/it]INFO 04-12 03:13:04 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3040/3300 [4:36:18<25:28,  5.88s/it]2026-04-12 03:13:09,225 | INFO | train_grpo_train | metrics_logged mode=train step=3040

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.909090909090909e-07, 'num_tokens': 6906301.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002939449332188815, 'sampling/sampling_logp_difference/max': 0.007810796611011028, 'sampling/importance_sampling_ratio/min': 0.9997765421867371, 'sampling/importance_sampling_ratio/mean': 1.000292420387268, 'sampling/importance_sampling_ratio/max': 1.0078413486480713, 'entropy': 0.00237001696950756, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3040/3300 [4:36:18<25:28,  5.88s/it]INFO 04-12 03:13:09 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3041/3300 [4:36:22<22:55,  5.31s/it]2026-04-12 03:13:13,204 | INFO | train_grpo_train | metrics_logged mode=train step=3041

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.878787878787879e-07, 'num_tokens': 6907797.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00019611754396464676, 'sampling/sampling_logp_difference/max': 0.007875919342041016, 'sampling/importance_sampling_ratio/min': 0.9921550154685974, 'sampling/importance_sampling_ratio/mean': 1.0001239776611328, 'sampling/importance_sampling_ratio/max': 1.0073140859603882, 'entropy': 0.0016474624135298654, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3041/3300 [4:36:22<22:55,  5.31s/it]INFO 04-12 03:13:13 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3042/3300 [4:36:26<22:03,  5.13s/it]2026-04-12 03:13:17,914 | INFO | train_grpo_train | metrics_logged mode=train step=3042

                                                     
{'loss': -0.0003, 'grad_norm': 0.7192491292953491, 'learning_rate': 7.84848484848485e-07, 'num_tokens': 6909726.0, 'completions/mean_length': 71.125, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994059801101685, 'rewards/meter/std': 3.530035974108614e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994059801101685, 'rewards/total_composite/std': 3.530035974108614e-05, 'reward': 0.9994059801101685, 'reward_std': 3.531122638378292e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008167088031768799, 'sampling/sampling_logp_difference/max': 0.9387979507446289, 'sampling/importance_sampling_ratio/min': 0.39109766483306885, 'sampling/importance_sampling_ratio/mean': 1.0006753206253052, 'sampling/importance_sampling_ratio/max': 1.1902652978897095, 'entropy': 0.0695438552647829, 'clip_ratio/low_mean': 0.0034966744715347886, 'clip_ratio/low_min': 0.0034966744715347886, 'clip_ratio/high_mean': 0.0017605633474886417, 'clip_ratio/high_max': 0.0017605633474886417, 'clip_ratio/region_mean': 0.00525723781902343, 'reward_total_mean': 0.9994059801101685, 'reward_meter_mean': 0.9994059801101685, 'reward_meter_std': 3.530035974108614e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994059801101685, 'reward_total_composite_std': 3.530035974108614e-05, 'epoch': 0.12}

 92%|█████████▏| 3042/3300 [4:36:26<22:03,  5.13s/it]INFO 04-12 03:13:18 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3043/3300 [4:36:31<20:54,  4.88s/it]2026-04-12 03:13:22,215 | INFO | train_grpo_train | metrics_logged mode=train step=3043

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.818181818181819e-07, 'num_tokens': 6911286.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002015612117247656, 'sampling/sampling_logp_difference/max': 0.002510334365069866, 'sampling/importance_sampling_ratio/min': 0.9998286962509155, 'sampling/importance_sampling_ratio/mean': 1.0002003908157349, 'sampling/importance_sampling_ratio/max': 1.0025135278701782, 'entropy': 0.0016416620055679232, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3043/3300 [4:36:31<20:54,  4.88s/it]INFO 04-12 03:13:22 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3044/3300 [4:36:35<20:19,  4.76s/it]2026-04-12 03:13:26,703 | INFO | train_grpo_train | metrics_logged mode=train step=3044

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.787878787878788e-07, 'num_tokens': 6913358.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00012283638352528214, 'sampling/sampling_logp_difference/max': 0.004916388541460037, 'sampling/importance_sampling_ratio/min': 0.9950957298278809, 'sampling/importance_sampling_ratio/mean': 1.0000611543655396, 'sampling/importance_sampling_ratio/max': 1.0022180080413818, 'entropy': 0.0007720192006672733, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 92%|█████████▏| 3044/3300 [4:36:35<20:19,  4.76s/it]INFO 04-12 03:13:26 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3045/3300 [4:36:40<20:30,  4.83s/it]2026-04-12 03:13:31,679 | INFO | train_grpo_train | metrics_logged mode=train step=3045

                                                     
{'loss': -0.0027, 'grad_norm': 1.1839877367019653, 'learning_rate': 7.757575757575758e-07, 'num_tokens': 6915686.0, 'completions/mean_length': 107.0, 'completions/min_length': 105.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.0, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9992265701293945, 'rewards/meter/std': 0.0002075855591101572, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992265701293945, 'rewards/total_composite/std': 0.0002075855591101572, 'reward': 0.9992265701293945, 'reward_std': 0.00020759036124218255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01808370277285576, 'sampling/sampling_logp_difference/max': 1.08237886428833, 'sampling/importance_sampling_ratio/min': 0.33878862857818604, 'sampling/importance_sampling_ratio/mean': 1.005001187324524, 'sampling/importance_sampling_ratio/max': 1.8381428718566895, 'entropy': 0.13631522469222546, 'clip_ratio/low_mean': 0.005886243423447013, 'clip_ratio/low_min': 0.005886243423447013, 'clip_ratio/high_mean': 0.01052544778212905, 'clip_ratio/high_max': 0.01052544778212905, 'clip_ratio/region_mean': 0.016411691205576062, 'reward_total_mean': 0.9992265701293945, 'reward_meter_mean': 0.9992265701293945, 'reward_meter_std': 0.0002075855591101572, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992265701293945, 'reward_total_composite_std': 0.0002075855591101572, 'epoch': 0.12}

 92%|█████████▏| 3045/3300 [4:36:40<20:30,  4.83s/it]INFO 04-12 03:13:31 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3046/3300 [4:36:45<20:02,  4.73s/it]2026-04-12 03:13:36,187 | INFO | train_grpo_train | metrics_logged mode=train step=3046

                                                     
{'loss': -0.0018, 'grad_norm': 1.057833194732666, 'learning_rate': 7.727272727272727e-07, 'num_tokens': 6917493.0, 'completions/mean_length': 67.875, 'completions/min_length': 67.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9994407892227173, 'rewards/meter/std': 6.79898148518987e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994407892227173, 'rewards/total_composite/std': 6.79898148518987e-05, 'reward': 0.9994407892227173, 'reward_std': 6.798980029998347e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0057174162939190865, 'sampling/sampling_logp_difference/max': 0.2612924575805664, 'sampling/importance_sampling_ratio/min': 0.7700556516647339, 'sampling/importance_sampling_ratio/mean': 1.0019086599349976, 'sampling/importance_sampling_ratio/max': 1.2558224201202393, 'entropy': 0.04532818449661136, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.0036764706019312143, 'reward_total_mean': 0.9994407892227173, 'reward_meter_mean': 0.9994407892227173, 'reward_meter_std': 6.79898148518987e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994407892227173, 'reward_total_composite_std': 6.79898148518987e-05, 'epoch': 0.12}

 92%|█████████▏| 3046/3300 [4:36:45<20:02,  4.73s/it]INFO 04-12 03:13:36 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3047/3300 [4:36:51<21:41,  5.14s/it]2026-04-12 03:13:42,295 | INFO | train_grpo_train | metrics_logged mode=train step=3047

                                                     
{'loss': 0.005, 'grad_norm': 1.7859693765640259, 'learning_rate': 7.696969696969698e-07, 'num_tokens': 6920501.0, 'completions/mean_length': 200.0, 'completions/min_length': 198.0, 'completions/max_length': 202.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 200.0, 'completions/min_terminated_length': 198.0, 'completions/max_terminated_length': 202.0, 'rewards/meter/mean': 0.9991546869277954, 'rewards/meter/std': 0.0004873749567195773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9090908765792847, 'rewards/repeat_penalty/std': 0.06872081756591797, 'rewards/total_composite/mean': 0.9083393812179565, 'rewards/total_composite/std': 0.06891340762376785, 'reward': 0.9083393812179565, 'reward_std': 0.06891340017318726, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018795771524310112, 'sampling/sampling_logp_difference/max': 1.3369550704956055, 'sampling/importance_sampling_ratio/min': 0.2626442015171051, 'sampling/importance_sampling_ratio/mean': 1.0054200887680054, 'sampling/importance_sampling_ratio/max': 1.687934398651123, 'entropy': 0.17142035998404026, 'clip_ratio/low_mean': 0.005625279794912785, 'clip_ratio/low_min': 0.005625279794912785, 'clip_ratio/high_mean': 0.005637656955514103, 'clip_ratio/high_max': 0.005637656955514103, 'clip_ratio/region_mean': 0.011262936750426888, 'reward_total_mean': 0.9083393812179565, 'reward_meter_mean': 0.9991546869277954, 'reward_meter_std': 0.0004873749567195773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9090908765792847, 'reward_repeat_penalty_std': 0.06872081756591797, 'reward_total_composite_mean': 0.9083393812179565, 'reward_total_composite_std': 0.06891340762376785, 'epoch': 0.12}

 92%|█████████▏| 3047/3300 [4:36:51<21:41,  5.14s/it]INFO 04-12 03:13:42 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3048/3300 [4:36:55<20:23,  4.85s/it]2026-04-12 03:13:46,470 | INFO | train_grpo_train | metrics_logged mode=train step=3048

                                                     
{'loss': -0.0015, 'grad_norm': 5.802453994750977, 'learning_rate': 7.666666666666667e-07, 'num_tokens': 6921887.0, 'completions/mean_length': 34.25, 'completions/min_length': 34.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9926164150238037, 'rewards/meter/std': 0.0012579105095937848, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9926164150238037, 'rewards/total_composite/std': 0.0012579105095937848, 'reward': 0.9926164150238037, 'reward_std': 0.001257919822819531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014383560046553612, 'sampling/sampling_logp_difference/max': 1.1232624053955078, 'sampling/importance_sampling_ratio/min': 0.325217068195343, 'sampling/importance_sampling_ratio/mean': 0.9971778988838196, 'sampling/importance_sampling_ratio/max': 1.4339221715927124, 'entropy': 0.09313629940152168, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.0035714285913854837, 'clip_ratio/high_max': 0.0035714285913854837, 'clip_ratio/region_mean': 0.0071428571827709675, 'reward_total_mean': 0.9926164150238037, 'reward_meter_mean': 0.9926164150238037, 'reward_meter_std': 0.0012579105095937848, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9926164150238037, 'reward_total_composite_std': 0.0012579105095937848, 'epoch': 0.12}

 92%|█████████▏| 3048/3300 [4:36:55<20:23,  4.85s/it]INFO 04-12 03:13:46 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3049/3300 [4:37:00<20:57,  5.01s/it]2026-04-12 03:13:51,841 | INFO | train_grpo_train | metrics_logged mode=train step=3049

                                                     
{'loss': -0.0064, 'grad_norm': 3.5305817127227783, 'learning_rate': 7.636363636363637e-07, 'num_tokens': 6924412.0, 'completions/mean_length': 137.625, 'completions/min_length': 134.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.625, 'completions/min_terminated_length': 134.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.9496752023696899, 'rewards/meter/std': 0.0848766416311264, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.8786302804946899, 'rewards/total_composite/std': 0.0753149688243866, 'reward': 0.8786302804946899, 'reward_std': 0.0753149539232254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03166425973176956, 'sampling/sampling_logp_difference/max': 1.181009292602539, 'sampling/importance_sampling_ratio/min': 0.3069687485694885, 'sampling/importance_sampling_ratio/mean': 1.0010079145431519, 'sampling/importance_sampling_ratio/max': 1.5370413064956665, 'entropy': 0.2945951074361801, 'clip_ratio/low_mean': 0.011909950990229845, 'clip_ratio/low_min': 0.011909950990229845, 'clip_ratio/high_mean': 0.01253445539623499, 'clip_ratio/high_max': 0.01253445539623499, 'clip_ratio/region_mean': 0.024444406386464834, 'reward_total_mean': 0.8786302804946899, 'reward_meter_mean': 0.9496752023696899, 'reward_meter_std': 0.0848766416311264, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.8786302804946899, 'reward_total_composite_std': 0.0753149688243866, 'epoch': 0.12}

 92%|█████████▏| 3049/3300 [4:37:00<20:57,  5.01s/it]INFO 04-12 03:13:52 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3050/3300 [4:37:07<22:54,  5.50s/it]2026-04-12 03:13:58,488 | INFO | train_grpo_train | metrics_logged mode=train step=3050

                                                     
{'loss': -0.0011, 'grad_norm': 1.4543901681900024, 'learning_rate': 7.606060606060607e-07, 'num_tokens': 6928012.0, 'completions/mean_length': 234.0, 'completions/min_length': 228.0, 'completions/max_length': 238.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 234.0, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 238.0, 'rewards/meter/mean': 0.9989113807678223, 'rewards/meter/std': 0.0005675117135979235, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989113807678223, 'rewards/total_composite/std': 0.0005675117135979235, 'reward': 0.9989113807678223, 'reward_std': 0.0005675173015333712, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03521440923213959, 'sampling/sampling_logp_difference/max': 1.4896221160888672, 'sampling/importance_sampling_ratio/min': 0.22545784711837769, 'sampling/importance_sampling_ratio/mean': 1.007856011390686, 'sampling/importance_sampling_ratio/max': 1.7903379201889038, 'entropy': 0.3369157761335373, 'clip_ratio/low_mean': 0.004255396313965321, 'clip_ratio/low_min': 0.004255396313965321, 'clip_ratio/high_mean': 0.016509592533111572, 'clip_ratio/high_max': 0.016509592533111572, 'clip_ratio/region_mean': 0.020764988847076893, 'reward_total_mean': 0.9989113807678223, 'reward_meter_mean': 0.9989113807678223, 'reward_meter_std': 0.0005675117135979235, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989113807678223, 'reward_total_composite_std': 0.0005675117135979235, 'epoch': 0.12}

 92%|█████████▏| 3050/3300 [4:37:07<22:54,  5.50s/it]INFO 04-12 03:13:58 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:38,  3.54s/it]

 23%|██▎       | 3/13 [00:10<00:35,  3.53s/it]

 31%|███       | 4/13 [00:16<00:40,  4.50s/it]

 38%|███▊      | 5/13 [00:22<00:38,  4.77s/it]

 46%|████▌     | 6/13 [00:28<00:37,  5.40s/it]

 54%|█████▍    | 7/13 [00:34<00:32,  5.38s/it]

 62%|██████▏   | 8/13 [00:40<00:29,  5.84s/it]

 69%|██████▉   | 9/13 [00:48<00:25,  6.30s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.92s/it]

 85%|████████▍ | 11/13 [00:58<00:11,  5.86s/it]

 92%|█████████▏| 12/13 [01:05<00:05,  5.94s/it]

100%|██████████| 13/13 [01:10<00:00,  5.82s/it]2026-04-12 03:15:16,123 | INFO | train_grpo_train | metrics_logged mode=eval step=3050

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 77.5007, 'eval_samples_per_second': 1.342, 'eval_steps_per_second': 0.168, 'eval_num_tokens': 6928012.0, 'eval_completions/mean_length': 212.6346153846154, 'eval_completions/min_length': 60.76923076923077, 'eval_completions/max_length': 416.3076923076923, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 206.1826934814453, 'eval_completions/min_terminated_length': 60.76923076923077, 'eval_completions/max_terminated_length': 395.53846153846155, 'eval_rewards/meter/mean': 0.7826263767022353, 'eval_rewards/meter/std': 0.3530325018442594, 'eval_rewards/count_adherence/mean': 0.9593305908716642, 'eval_rewards/count_adherence/std': 0.06269322364376141, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.9508223029283377, 'eval_rewards/repeat_penalty/std': 0.06783326672246823, 'eval_rewards/total_composite/mean': 0.722142334167774, 'eval_rewards/total_composite/std': 0.3449985155692467, 'eval_reward': 0.722142334167774, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.030635818409231994, 'eval_sampling/sampling_logp_difference/max': 1.1543501340425932, 'eval_sampling/importance_sampling_ratio/min': 0.3223212957382202, 'eval_sampling/importance_sampling_ratio/mean': 1.0082989380909846, 'eval_sampling/importance_sampling_ratio/max': 1.5470959773430457, 'eval_entropy': 0.34184010556110966, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.722142334167774, 'eval_reward_meter_mean': 0.7826263767022353, 'eval_reward_meter_std': 0.3530325018442594, 'eval_reward_count_adherence_mean': 0.9593305908716642, 'eval_reward_count_adherence_std': 0.06269322364376141, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.9508223029283377, 'eval_reward_repeat_penalty_std': 0.06783326672246823, 'eval_reward_total_composite_mean': 0.722142334167774, 'eval_reward_total_composite_std': 0.3449985155692467, 'epoch': 0.12}

 92%|█████████▏| 3050/3300 [4:38:25<22:54,  5.50s/it]

100%|██████████| 13/13 [01:10<00:00,  5.82s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 92%|█████████▏| 3051/3300 [4:38:34<2:04:34, 30.02s/it]2026-04-12 03:15:25,738 | INFO | train_grpo_train | metrics_logged mode=train step=3051

                                                       
{'loss': 0.0034, 'grad_norm': 1.5024135112762451, 'learning_rate': 7.575757575757576e-07, 'num_tokens': 6931580.0, 'completions/mean_length': 233.0, 'completions/min_length': 231.0, 'completions/max_length': 235.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 233.0, 'completions/min_terminated_length': 231.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.999167799949646, 'rewards/meter/std': 7.227841706480831e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9886363744735718, 'rewards/repeat_penalty/std': 0.03214120864868164, 'rewards/total_composite/mean': 0.9878134727478027, 'rewards/total_composite/std': 0.03211057558655739, 'reward': 0.9878134727478027, 'reward_std': 0.03211057186126709, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.037654075771570206, 'sampling/sampling_logp_difference/max': 1.2263059616088867, 'sampling/importance_sampling_ratio/min': 0.29337432980537415, 'sampling/importance_sampling_ratio/mean': 1.0094155073165894, 'sampling/importance_sampling_ratio/max': 1.6730788946151733, 'entropy': 0.3387974835932255, 'clip_ratio/low_mean': 0.0026709402445703745, 'clip_ratio/low_min': 0.0026709402445703745, 'clip_ratio/high_mean': 0.02257009909953922, 'clip_ratio/high_max': 0.02257009909953922, 'clip_ratio/region_mean': 0.025241039344109595, 'reward_total_mean': 0.9878134727478027, 'reward_meter_mean': 0.999167799949646, 'reward_meter_std': 7.227841706480831e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9886363744735718, 'reward_repeat_penalty_std': 0.03214120864868164, 'reward_total_composite_mean': 0.9878134727478027, 'reward_total_composite_std': 0.03211057558655739, 'epoch': 0.12}

 92%|█████████▏| 3051/3300 [4:38:34<2:04:34, 30.02s/it]INFO 04-12 03:15:26 [block_pool.py:378] Successfully reset prefix cache

 92%|█████████▏| 3052/3300 [4:38:42<1:37:01, 23.47s/it]2026-04-12 03:15:33,935 | INFO | train_grpo_train | metrics_logged mode=train step=3052

                                                       
{'loss': 0.0461, 'grad_norm': 2.0073533058166504, 'learning_rate': 7.545454545454546e-07, 'num_tokens': 6935794.0, 'completions/mean_length': 322.75, 'completions/min_length': 308.0, 'completions/max_length': 356.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 322.75, 'completions/min_terminated_length': 308.0, 'completions/max_terminated_length': 356.0, 'rewards/meter/mean': 0.9991399645805359, 'rewards/meter/std': 0.00012341544788796455, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9916666746139526, 'rewards/repeat_penalty/std': 0.0235702246427536, 'rewards/total_composite/mean': 0.9595891237258911, 'rewards/total_composite/std': 0.05744243040680885, 'reward': 0.9595891237258911, 'reward_std': 0.057442426681518555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04169188067317009, 'sampling/sampling_logp_difference/max': 1.2492733001708984, 'sampling/importance_sampling_ratio/min': 0.2867130637168884, 'sampling/importance_sampling_ratio/mean': 1.0115916728973389, 'sampling/importance_sampling_ratio/max': 1.883813500404358, 'entropy': 0.41726846620440483, 'clip_ratio/low_mean': 0.009418385569006205, 'clip_ratio/low_min': 0.009418385569006205, 'clip_ratio/high_mean': 0.016292808344587684, 'clip_ratio/high_max': 0.016292808344587684, 'clip_ratio/region_mean': 0.02571119391359389, 'reward_total_mean': 0.9595891237258911, 'reward_meter_mean': 0.9991399645805359, 'reward_meter_std': 0.00012341544788796455, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9916666746139526, 'reward_repeat_penalty_std': 0.0235702246427536, 'reward_total_composite_mean': 0.9595891237258911, 'reward_total_composite_std': 0.05744243040680885, 'epoch': 0.12}

 92%|█████████▏| 3052/3300 [4:38:42<1:37:01, 23.47s/it]INFO 04-12 03:15:34 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3053/3300 [4:38:47<1:13:27, 17.85s/it]2026-04-12 03:15:38,663 | INFO | train_grpo_train | metrics_logged mode=train step=3053

                                                       
{'loss': 0.0049, 'grad_norm': 2.259249448776245, 'learning_rate': 7.515151515151516e-07, 'num_tokens': 6937691.0, 'completions/mean_length': 66.125, 'completions/min_length': 65.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9928057193756104, 'rewards/meter/std': 0.0009291421738453209, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9928057193756104, 'rewards/total_composite/std': 0.0009291421738453209, 'reward': 0.9928057193756104, 'reward_std': 0.0009291406604461372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015979096293449402, 'sampling/sampling_logp_difference/max': 1.00278902053833, 'sampling/importance_sampling_ratio/min': 0.36685484647750854, 'sampling/importance_sampling_ratio/mean': 1.0060573816299438, 'sampling/importance_sampling_ratio/max': 1.8138948678970337, 'entropy': 0.10556314932182431, 'clip_ratio/low_mean': 0.013010540511459112, 'clip_ratio/low_min': 0.013010540511459112, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.013010540511459112, 'reward_total_mean': 0.9928057193756104, 'reward_meter_mean': 0.9928057193756104, 'reward_meter_std': 0.0009291421738453209, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9928057193756104, 'reward_total_composite_std': 0.0009291421738453209, 'epoch': 0.12}

 93%|█████████▎| 3053/3300 [4:38:47<1:13:27, 17.85s/it]INFO 04-12 03:15:38 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3054/3300 [4:38:53<58:05, 14.17s/it]  2026-04-12 03:15:44,242 | INFO | train_grpo_train | metrics_logged mode=train step=3054

                                                     
{'loss': 0.0032, 'grad_norm': 2.697394847869873, 'learning_rate': 7.484848484848485e-07, 'num_tokens': 6940138.0, 'completions/mean_length': 137.875, 'completions/min_length': 135.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.875, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9923188090324402, 'rewards/meter/std': 0.004795641638338566, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9745643138885498, 'rewards/total_composite/std': 0.04966309294104576, 'reward': 0.9745643138885498, 'reward_std': 0.04966310039162636, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02996024861931801, 'sampling/sampling_logp_difference/max': 1.230236291885376, 'sampling/importance_sampling_ratio/min': 0.2922235131263733, 'sampling/importance_sampling_ratio/mean': 1.0037504434585571, 'sampling/importance_sampling_ratio/max': 1.498881459236145, 'entropy': 0.28184244222939014, 'clip_ratio/low_mean': 0.0018115942366421223, 'clip_ratio/low_min': 0.0018115942366421223, 'clip_ratio/high_mean': 0.026310671470128, 'clip_ratio/high_max': 0.026310671470128, 'clip_ratio/region_mean': 0.028122265706770122, 'reward_total_mean': 0.9745643138885498, 'reward_meter_mean': 0.9923188090324402, 'reward_meter_std': 0.004795641638338566, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9745643138885498, 'reward_total_composite_std': 0.04966309294104576, 'epoch': 0.12}

 93%|█████████▎| 3054/3300 [4:38:53<58:05, 14.17s/it]INFO 04-12 03:15:44 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3055/3300 [4:38:59<48:27, 11.87s/it]2026-04-12 03:15:50,716 | INFO | train_grpo_train | metrics_logged mode=train step=3055

                                                     
{'loss': 0.0153, 'grad_norm': 2.5886197090148926, 'learning_rate': 7.454545454545455e-07, 'num_tokens': 6943559.0, 'completions/mean_length': 233.625, 'completions/min_length': 228.0, 'completions/max_length': 238.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 233.625, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 238.0, 'rewards/meter/mean': 0.9977272748947144, 'rewards/meter/std': 0.0019398098811507225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.932692289352417, 'rewards/repeat_penalty/std': 0.08661473542451859, 'rewards/total_composite/mean': 0.9304856061935425, 'rewards/total_composite/std': 0.08541877567768097, 'reward': 0.9304856061935425, 'reward_std': 0.08541877567768097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.045273296535015106, 'sampling/sampling_logp_difference/max': 1.65692138671875, 'sampling/importance_sampling_ratio/min': 0.19072523713111877, 'sampling/importance_sampling_ratio/mean': 1.0053743124008179, 'sampling/importance_sampling_ratio/max': 1.9301133155822754, 'entropy': 0.3986949659883976, 'clip_ratio/low_mean': 0.010666976682841778, 'clip_ratio/low_min': 0.010666976682841778, 'clip_ratio/high_mean': 0.025692227762192488, 'clip_ratio/high_max': 0.025692227762192488, 'clip_ratio/region_mean': 0.036359204445034266, 'reward_total_mean': 0.9304856061935425, 'reward_meter_mean': 0.9977272748947144, 'reward_meter_std': 0.0019398098811507225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.932692289352417, 'reward_repeat_penalty_std': 0.08661473542451859, 'reward_total_composite_mean': 0.9304856061935425, 'reward_total_composite_std': 0.08541877567768097, 'epoch': 0.12}

 93%|█████████▎| 3055/3300 [4:38:59<48:27, 11.87s/it]INFO 04-12 03:15:50 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3056/3300 [4:39:04<40:20,  9.92s/it]2026-04-12 03:15:56,105 | INFO | train_grpo_train | metrics_logged mode=train step=3056

                                                     
{'loss': -0.0003, 'grad_norm': 1.2885347604751587, 'learning_rate': 7.424242424242425e-07, 'num_tokens': 6946121.0, 'completions/mean_length': 142.25, 'completions/min_length': 138.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.25, 'completions/min_terminated_length': 138.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9991947412490845, 'rewards/meter/std': 7.931316940812394e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9813523292541504, 'rewards/total_composite/std': 0.050475094467401505, 'reward': 0.9813523292541504, 'reward_std': 0.05047507956624031, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015199076384305954, 'sampling/sampling_logp_difference/max': 0.8146209716796875, 'sampling/importance_sampling_ratio/min': 0.4428071081638336, 'sampling/importance_sampling_ratio/mean': 1.004903793334961, 'sampling/importance_sampling_ratio/max': 1.6997216939926147, 'entropy': 0.18231415934860706, 'clip_ratio/low_mean': 0.0017605633474886417, 'clip_ratio/low_min': 0.0017605633474886417, 'clip_ratio/high_mean': 0.01225746376439929, 'clip_ratio/high_max': 0.01225746376439929, 'clip_ratio/region_mean': 0.014018027111887932, 'reward_total_mean': 0.9813523292541504, 'reward_meter_mean': 0.9991947412490845, 'reward_meter_std': 7.931316940812394e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9813523292541504, 'reward_total_composite_std': 0.050475094467401505, 'epoch': 0.12}

 93%|█████████▎| 3056/3300 [4:39:05<40:20,  9.92s/it]INFO 04-12 03:15:56 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3057/3300 [4:39:09<33:32,  8.28s/it]2026-04-12 03:16:00,577 | INFO | train_grpo_train | metrics_logged mode=train step=3057

                                                     
{'loss': -0.0002, 'grad_norm': 0.10690882056951523, 'learning_rate': 7.393939393939395e-07, 'num_tokens': 6948039.0, 'completions/mean_length': 66.75, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981523752212524, 'rewards/meter/std': 3.7664287901861826e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981523752212524, 'rewards/total_composite/std': 3.7664287901861826e-06, 'reward': 0.9981523752212524, 'reward_std': 3.7686852465412812e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006302142050117254, 'sampling/sampling_logp_difference/max': 0.39133310317993164, 'sampling/importance_sampling_ratio/min': 0.8453654646873474, 'sampling/importance_sampling_ratio/mean': 1.0037543773651123, 'sampling/importance_sampling_ratio/max': 1.4789509773254395, 'entropy': 0.05063098296523094, 'clip_ratio/low_mean': 0.0056535504991188645, 'clip_ratio/low_min': 0.0056535504991188645, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0056535504991188645, 'reward_total_mean': 0.9981523752212524, 'reward_meter_mean': 0.9981523752212524, 'reward_meter_std': 3.7664287901861826e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981523752212524, 'reward_total_composite_std': 3.7664287901861826e-06, 'epoch': 0.12}

 93%|█████████▎| 3057/3300 [4:39:09<33:32,  8.28s/it]INFO 04-12 03:16:00 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3058/3300 [4:39:13<28:45,  7.13s/it]2026-04-12 03:16:05,017 | INFO | train_grpo_train | metrics_logged mode=train step=3058

                                                     
{'loss': 0.0021, 'grad_norm': 2.2398717403411865, 'learning_rate': 7.363636363636364e-07, 'num_tokens': 6949869.0, 'completions/mean_length': 65.75, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9914587736129761, 'rewards/meter/std': 0.00509608956053853, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9914587736129761, 'rewards/total_composite/std': 0.00509608956053853, 'reward': 0.9914587736129761, 'reward_std': 0.005096097476780415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011036393232643604, 'sampling/sampling_logp_difference/max': 0.8546088933944702, 'sampling/importance_sampling_ratio/min': 0.42544955015182495, 'sampling/importance_sampling_ratio/mean': 1.003832221031189, 'sampling/importance_sampling_ratio/max': 1.2646030187606812, 'entropy': 0.07432558294385672, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.0038170163752511144, 'clip_ratio/high_max': 0.0038170163752511144, 'clip_ratio/region_mean': 0.005710955825634301, 'reward_total_mean': 0.9914587736129761, 'reward_meter_mean': 0.9914587736129761, 'reward_meter_std': 0.00509608956053853, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9914587736129761, 'reward_total_composite_std': 0.00509608956053853, 'epoch': 0.12}

 93%|█████████▎| 3058/3300 [4:39:14<28:45,  7.13s/it]INFO 04-12 03:16:05 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3059/3300 [4:39:19<26:14,  6.53s/it]2026-04-12 03:16:10,160 | INFO | train_grpo_train | metrics_logged mode=train step=3059

                                                     
{'loss': -0.0012, 'grad_norm': 2.9161179065704346, 'learning_rate': 7.333333333333334e-07, 'num_tokens': 6952388.0, 'completions/mean_length': 125.875, 'completions/min_length': 123.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.875, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.938283383846283, 'rewards/meter/std': 0.15279191732406616, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9285714626312256, 'rewards/repeat_penalty/std': 0.07636035233736038, 'rewards/total_composite/mean': 0.8751063346862793, 'rewards/total_composite/std': 0.17424653470516205, 'reward': 0.8751063346862793, 'reward_std': 0.17424653470516205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025627773255109787, 'sampling/sampling_logp_difference/max': 0.9624452590942383, 'sampling/importance_sampling_ratio/min': 0.3819577693939209, 'sampling/importance_sampling_ratio/mean': 1.0073537826538086, 'sampling/importance_sampling_ratio/max': 1.7322660684585571, 'entropy': 0.24840066954493523, 'clip_ratio/low_mean': 0.008024971117265522, 'clip_ratio/low_min': 0.008024971117265522, 'clip_ratio/high_mean': 0.006889763753861189, 'clip_ratio/high_max': 0.006889763753861189, 'clip_ratio/region_mean': 0.014914734871126711, 'reward_total_mean': 0.8751063346862793, 'reward_meter_mean': 0.938283383846283, 'reward_meter_std': 0.15279191732406616, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9285714626312256, 'reward_repeat_penalty_std': 0.07636035233736038, 'reward_total_composite_mean': 0.8751063346862793, 'reward_total_composite_std': 0.17424653470516205, 'epoch': 0.12}

 93%|█████████▎| 3059/3300 [4:39:19<26:14,  6.53s/it]INFO 04-12 03:16:10 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3060/3300 [4:39:24<24:51,  6.22s/it]2026-04-12 03:16:15,622 | INFO | train_grpo_train | metrics_logged mode=train step=3060

                                                     
{'loss': 0.0128, 'grad_norm': 3.135732650756836, 'learning_rate': 7.303030303030304e-07, 'num_tokens': 6954749.0, 'completions/mean_length': 138.125, 'completions/min_length': 133.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.125, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.9924278259277344, 'rewards/meter/std': 0.0035289754159748554, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924278259277344, 'rewards/total_composite/std': 0.0035289754159748554, 'reward': 0.9924278259277344, 'reward_std': 0.0035289691295474768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03332981467247009, 'sampling/sampling_logp_difference/max': 0.9910516738891602, 'sampling/importance_sampling_ratio/min': 0.37118610739707947, 'sampling/importance_sampling_ratio/mean': 1.0079230070114136, 'sampling/importance_sampling_ratio/max': 1.7354730367660522, 'entropy': 0.32595139369368553, 'clip_ratio/low_mean': 0.010748314904049039, 'clip_ratio/low_min': 0.010748314904049039, 'clip_ratio/high_mean': 0.01723209215560928, 'clip_ratio/high_max': 0.01723209215560928, 'clip_ratio/region_mean': 0.02798040705965832, 'reward_total_mean': 0.9924278259277344, 'reward_meter_mean': 0.9924278259277344, 'reward_meter_std': 0.0035289754159748554, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924278259277344, 'reward_total_composite_std': 0.0035289754159748554, 'epoch': 0.12}

 93%|█████████▎| 3060/3300 [4:39:24<24:51,  6.22s/it]INFO 04-12 03:16:15 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3061/3300 [4:39:28<22:12,  5.57s/it]2026-04-12 03:16:19,701 | INFO | train_grpo_train | metrics_logged mode=train step=3061

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.272727272727273e-07, 'num_tokens': 6956349.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 5.019194941269234e-05, 'sampling/sampling_logp_difference/max': 0.0008984014857560396, 'sampling/importance_sampling_ratio/min': 0.9998447299003601, 'sampling/importance_sampling_ratio/mean': 1.0000483989715576, 'sampling/importance_sampling_ratio/max': 1.000898838043213, 'entropy': 0.000393166919820942, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 93%|█████████▎| 3061/3300 [4:39:28<22:12,  5.57s/it]INFO 04-12 03:16:19 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3062/3300 [4:39:33<21:38,  5.46s/it]2026-04-12 03:16:24,892 | INFO | train_grpo_train | metrics_logged mode=train step=3062

                                                     
{'loss': 0.0008, 'grad_norm': 1.910351037979126, 'learning_rate': 7.242424242424243e-07, 'num_tokens': 6958746.0, 'completions/mean_length': 124.625, 'completions/min_length': 123.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.625, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.997795820236206, 'rewards/meter/std': 0.0004729351494461298, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997795820236206, 'rewards/total_composite/std': 0.0004729351494461298, 'reward': 0.997795820236206, 'reward_std': 0.0004729465872514993, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018115131184458733, 'sampling/sampling_logp_difference/max': 1.9803028106689453, 'sampling/importance_sampling_ratio/min': 0.13802744448184967, 'sampling/importance_sampling_ratio/mean': 1.0041706562042236, 'sampling/importance_sampling_ratio/max': 1.4670944213867188, 'entropy': 0.1536051593720913, 'clip_ratio/low_mean': 0.0010000000474974513, 'clip_ratio/low_min': 0.0010000000474974513, 'clip_ratio/high_mean': 0.01407297421246767, 'clip_ratio/high_max': 0.01407297421246767, 'clip_ratio/region_mean': 0.015072974259965122, 'reward_total_mean': 0.997795820236206, 'reward_meter_mean': 0.997795820236206, 'reward_meter_std': 0.0004729351494461298, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997795820236206, 'reward_total_composite_std': 0.0004729351494461298, 'epoch': 0.12}

 93%|█████████▎| 3062/3300 [4:39:33<21:38,  5.46s/it]INFO 04-12 03:16:25 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3063/3300 [4:39:39<21:23,  5.41s/it]2026-04-12 03:16:30,203 | INFO | train_grpo_train | metrics_logged mode=train step=3063

                                                     
{'loss': 0.0173, 'grad_norm': 3.4471962451934814, 'learning_rate': 7.212121212121213e-07, 'num_tokens': 6961175.0, 'completions/mean_length': 137.625, 'completions/min_length': 134.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.625, 'completions/min_terminated_length': 134.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9925941228866577, 'rewards/meter/std': 0.0024486577603965998, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9925941228866577, 'rewards/total_composite/std': 0.0024486577603965998, 'reward': 0.9925941228866577, 'reward_std': 0.0024486577603965998, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04126367345452309, 'sampling/sampling_logp_difference/max': 1.3208866119384766, 'sampling/importance_sampling_ratio/min': 0.2668985426425934, 'sampling/importance_sampling_ratio/mean': 1.0079936981201172, 'sampling/importance_sampling_ratio/max': 1.8253576755523682, 'entropy': 0.3563910871744156, 'clip_ratio/low_mean': 0.014415563317015767, 'clip_ratio/low_min': 0.014415563317015767, 'clip_ratio/high_mean': 0.018338565016165376, 'clip_ratio/high_max': 0.018338565016165376, 'clip_ratio/region_mean': 0.03275412833318114, 'reward_total_mean': 0.9925941228866577, 'reward_meter_mean': 0.9925941228866577, 'reward_meter_std': 0.0024486577603965998, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9925941228866577, 'reward_total_composite_std': 0.0024486577603965998, 'epoch': 0.12}

 93%|█████████▎| 3063/3300 [4:39:39<21:23,  5.41s/it]INFO 04-12 03:16:30 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3064/3300 [4:39:43<20:18,  5.16s/it]2026-04-12 03:16:34,778 | INFO | train_grpo_train | metrics_logged mode=train step=3064

                                                     
{'loss': 0.0097, 'grad_norm': 3.6814091205596924, 'learning_rate': 7.181818181818182e-07, 'num_tokens': 6963041.0, 'completions/mean_length': 65.25, 'completions/min_length': 63.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.25, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9920260310173035, 'rewards/meter/std': 0.0032561845146119595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9920260310173035, 'rewards/total_composite/std': 0.0032561845146119595, 'reward': 0.9920260310173035, 'reward_std': 0.0032561977859586477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031295765191316605, 'sampling/sampling_logp_difference/max': 1.3808412551879883, 'sampling/importance_sampling_ratio/min': 0.2513670027256012, 'sampling/importance_sampling_ratio/mean': 1.0053472518920898, 'sampling/importance_sampling_ratio/max': 1.6320992708206177, 'entropy': 0.20045749936252832, 'clip_ratio/low_mean': 0.009565269108861685, 'clip_ratio/low_min': 0.009565269108861685, 'clip_ratio/high_mean': 0.009527972200885415, 'clip_ratio/high_max': 0.009527972200885415, 'clip_ratio/region_mean': 0.0190932413097471, 'reward_total_mean': 0.9920260310173035, 'reward_meter_mean': 0.9920260310173035, 'reward_meter_std': 0.0032561845146119595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9920260310173035, 'reward_total_composite_std': 0.0032561845146119595, 'epoch': 0.12}

 93%|█████████▎| 3064/3300 [4:39:43<20:18,  5.16s/it]INFO 04-12 03:16:34 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3065/3300 [4:39:51<23:33,  6.01s/it]2026-04-12 03:16:42,782 | INFO | train_grpo_train | metrics_logged mode=train step=3065

                                                     
{'loss': 0.0003, 'grad_norm': 1.8719370365142822, 'learning_rate': 7.151515151515153e-07, 'num_tokens': 6967523.0, 'completions/mean_length': 350.25, 'completions/min_length': 343.0, 'completions/max_length': 362.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 350.25, 'completions/min_terminated_length': 343.0, 'completions/max_terminated_length': 362.0, 'rewards/meter/mean': 0.9957402944564819, 'rewards/meter/std': 0.009437782689929008, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.8961663246154785, 'rewards/total_composite/std': 0.008493990637362003, 'reward': 0.8961663246154785, 'reward_std': 0.008493990637362003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05361352488398552, 'sampling/sampling_logp_difference/max': 2.0558176040649414, 'sampling/importance_sampling_ratio/min': 0.12798814475536346, 'sampling/importance_sampling_ratio/mean': 1.0094969272613525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4838559664785862, 'clip_ratio/low_mean': 0.004285714123398066, 'clip_ratio/low_min': 0.004285714123398066, 'clip_ratio/high_mean': 0.032739987364038825, 'clip_ratio/high_max': 0.032739987364038825, 'clip_ratio/region_mean': 0.03702570148743689, 'reward_total_mean': 0.8961663246154785, 'reward_meter_mean': 0.9957402944564819, 'reward_meter_std': 0.009437782689929008, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.8961663246154785, 'reward_total_composite_std': 0.008493990637362003, 'epoch': 0.12}

 93%|█████████▎| 3065/3300 [4:39:51<23:33,  6.01s/it]INFO 04-12 03:16:42 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3066/3300 [4:39:55<21:10,  5.43s/it]2026-04-12 03:16:46,846 | INFO | train_grpo_train | metrics_logged mode=train step=3066

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.121212121212122e-07, 'num_tokens': 6969011.0, 'completions/mean_length': 36.0, 'completions/min_length': 36.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9996045231819153, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9996045231819153, 'rewards/total_composite/std': 0.0, 'reward': 0.9996045231819153, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0014014473417773843, 'sampling/sampling_logp_difference/max': 0.034461501985788345, 'sampling/importance_sampling_ratio/min': 0.9661255478858948, 'sampling/importance_sampling_ratio/mean': 1.0010969638824463, 'sampling/importance_sampling_ratio/max': 1.0258433818817139, 'entropy': 0.013131682993844151, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9996045231819153, 'reward_meter_mean': 0.9996045231819153, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9996045231819153, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 93%|█████████▎| 3066/3300 [4:39:55<21:10,  5.43s/it]INFO 04-12 03:16:47 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3067/3300 [4:40:00<20:03,  5.17s/it]2026-04-12 03:16:51,396 | INFO | train_grpo_train | metrics_logged mode=train step=3067

                                                     
{'loss': -0.0038, 'grad_norm': 3.1952340602874756, 'learning_rate': 7.090909090909092e-07, 'num_tokens': 6970813.0, 'completions/mean_length': 69.25, 'completions/min_length': 68.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9955199956893921, 'rewards/meter/std': 0.0008645570487715304, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9955199956893921, 'rewards/total_composite/std': 0.0008645570487715304, 'reward': 0.9955199956893921, 'reward_std': 0.0008645570487715304, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02916860766708851, 'sampling/sampling_logp_difference/max': 1.2169954776763916, 'sampling/importance_sampling_ratio/min': 0.29611852765083313, 'sampling/importance_sampling_ratio/mean': 1.0018832683563232, 'sampling/importance_sampling_ratio/max': 1.5354810953140259, 'entropy': 0.21143894083797932, 'clip_ratio/low_mean': 0.0036498295376077294, 'clip_ratio/low_min': 0.0036498295376077294, 'clip_ratio/high_mean': 0.010746001382358372, 'clip_ratio/high_max': 0.010746001382358372, 'clip_ratio/region_mean': 0.014395830919966102, 'reward_total_mean': 0.9955199956893921, 'reward_meter_mean': 0.9955199956893921, 'reward_meter_std': 0.0008645570487715304, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9955199956893921, 'reward_total_composite_std': 0.0008645570487715304, 'epoch': 0.12}

 93%|█████████▎| 3067/3300 [4:40:00<20:03,  5.17s/it]INFO 04-12 03:16:51 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3068/3300 [4:40:06<20:44,  5.36s/it]2026-04-12 03:16:57,217 | INFO | train_grpo_train | metrics_logged mode=train step=3068

                                                     
{'loss': 0.0041, 'grad_norm': 1.5282530784606934, 'learning_rate': 7.060606060606061e-07, 'num_tokens': 6973664.0, 'completions/mean_length': 177.375, 'completions/min_length': 176.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 177.375, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.9991249442100525, 'rewards/meter/std': 0.00017947136075235903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9861111044883728, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.9852460026741028, 'rewards/total_composite/std': 0.039191942662000656, 'reward': 0.9852460026741028, 'reward_std': 0.03919193148612976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02057509496808052, 'sampling/sampling_logp_difference/max': 1.0894832611083984, 'sampling/importance_sampling_ratio/min': 0.33639028668403625, 'sampling/importance_sampling_ratio/mean': 1.003198266029358, 'sampling/importance_sampling_ratio/max': 1.8010437488555908, 'entropy': 0.22719617374241352, 'clip_ratio/low_mean': 0.0007022471982054412, 'clip_ratio/low_min': 0.0007022471982054412, 'clip_ratio/high_mean': 0.014811165689025074, 'clip_ratio/high_max': 0.014811165689025074, 'clip_ratio/region_mean': 0.015513412887230515, 'reward_total_mean': 0.9852460026741028, 'reward_meter_mean': 0.9991249442100525, 'reward_meter_std': 0.00017947136075235903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9861111044883728, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.9852460026741028, 'reward_total_composite_std': 0.039191942662000656, 'epoch': 0.12}

 93%|█████████▎| 3068/3300 [4:40:06<20:44,  5.36s/it]INFO 04-12 03:16:57 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3069/3300 [4:40:13<22:40,  5.89s/it]2026-04-12 03:17:04,347 | INFO | train_grpo_train | metrics_logged mode=train step=3069

                                                     
{'loss': 0.0046, 'grad_norm': 1.41837477684021, 'learning_rate': 7.03030303030303e-07, 'num_tokens': 6977475.0, 'completions/mean_length': 285.375, 'completions/min_length': 282.0, 'completions/max_length': 288.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 285.375, 'completions/min_terminated_length': 282.0, 'completions/max_terminated_length': 288.0, 'rewards/meter/mean': 0.9989954233169556, 'rewards/meter/std': 0.00014736292359884828, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9500000476837158, 'rewards/repeat_penalty/std': 0.0471404492855072, 'rewards/total_composite/mean': 0.9490443468093872, 'rewards/total_composite/std': 0.04706759378314018, 'reward': 0.9490443468093872, 'reward_std': 0.04706757515668869, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.026732422411441803, 'sampling/sampling_logp_difference/max': 1.257065773010254, 'sampling/importance_sampling_ratio/min': 0.2844875454902649, 'sampling/importance_sampling_ratio/mean': 1.005640983581543, 'sampling/importance_sampling_ratio/max': 1.655118703842163, 'entropy': 0.3014299161732197, 'clip_ratio/low_mean': 0.005277339863823727, 'clip_ratio/low_min': 0.005277339863823727, 'clip_ratio/high_mean': 0.007437994237989187, 'clip_ratio/high_max': 0.007437994237989187, 'clip_ratio/region_mean': 0.012715334101812914, 'reward_total_mean': 0.9490443468093872, 'reward_meter_mean': 0.9989954233169556, 'reward_meter_std': 0.00014736292359884828, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9500000476837158, 'reward_repeat_penalty_std': 0.0471404492855072, 'reward_total_composite_mean': 0.9490443468093872, 'reward_total_composite_std': 0.04706759378314018, 'epoch': 0.12}

 93%|█████████▎| 3069/3300 [4:40:13<22:40,  5.89s/it]INFO 04-12 03:17:04 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3070/3300 [4:40:19<23:28,  6.12s/it]2026-04-12 03:17:11,012 | INFO | train_grpo_train | metrics_logged mode=train step=3070

                                                     
{'loss': 0.0073, 'grad_norm': 3.581768751144409, 'learning_rate': 7.000000000000001e-07, 'num_tokens': 6981013.0, 'completions/mean_length': 234.25, 'completions/min_length': 224.0, 'completions/max_length': 250.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 234.25, 'completions/min_terminated_length': 224.0, 'completions/max_terminated_length': 250.0, 'rewards/meter/mean': 0.9958624839782715, 'rewards/meter/std': 0.004256864078342915, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9807692170143127, 'rewards/repeat_penalty/std': 0.03560846298933029, 'rewards/total_composite/mean': 0.9766829013824463, 'rewards/total_composite/std': 0.034849490970373154, 'reward': 0.9766829013824463, 'reward_std': 0.03484949469566345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05148389935493469, 'sampling/sampling_logp_difference/max': 3.614051103591919, 'sampling/importance_sampling_ratio/min': 0.026942478492856026, 'sampling/importance_sampling_ratio/mean': 1.0083545446395874, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43540840223431587, 'clip_ratio/low_mean': 0.005974947940558195, 'clip_ratio/low_min': 0.005974947940558195, 'clip_ratio/high_mean': 0.031633416656404734, 'clip_ratio/high_max': 0.031633416656404734, 'clip_ratio/region_mean': 0.03760836459696293, 'reward_total_mean': 0.9766829013824463, 'reward_meter_mean': 0.9958624839782715, 'reward_meter_std': 0.004256864078342915, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9807692170143127, 'reward_repeat_penalty_std': 0.03560846298933029, 'reward_total_composite_mean': 0.9766829013824463, 'reward_total_composite_std': 0.034849490970373154, 'epoch': 0.12}

 93%|█████████▎| 3070/3300 [4:40:20<23:28,  6.12s/it]INFO 04-12 03:17:11 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3071/3300 [4:40:24<21:48,  5.71s/it]2026-04-12 03:17:15,757 | INFO | train_grpo_train | metrics_logged mode=train step=3071

                                                     
{'loss': 0.0005, 'grad_norm': 0.05822064355015755, 'learning_rate': 6.969696969696971e-07, 'num_tokens': 6982996.0, 'completions/mean_length': 97.875, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994089603424072, 'rewards/meter/std': 7.70060796639882e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994089603424072, 'rewards/total_composite/std': 7.70060796639882e-06, 'reward': 0.9994089603424072, 'reward_std': 7.685384844080545e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004502264782786369, 'sampling/sampling_logp_difference/max': 0.6515674591064453, 'sampling/importance_sampling_ratio/min': 0.5212281346321106, 'sampling/importance_sampling_ratio/mean': 1.0012834072113037, 'sampling/importance_sampling_ratio/max': 1.2397805452346802, 'entropy': 0.02978663402609527, 'clip_ratio/low_mean': 0.0025510203558951616, 'clip_ratio/low_min': 0.0025510203558951616, 'clip_ratio/high_mean': 0.0025641699321568012, 'clip_ratio/high_max': 0.0025641699321568012, 'clip_ratio/region_mean': 0.005115190288051963, 'reward_total_mean': 0.9994089603424072, 'reward_meter_mean': 0.9994089603424072, 'reward_meter_std': 7.70060796639882e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994089603424072, 'reward_total_composite_std': 7.70060796639882e-06, 'epoch': 0.12}

 93%|█████████▎| 3071/3300 [4:40:24<21:48,  5.71s/it]INFO 04-12 03:17:15 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3072/3300 [4:40:28<20:09,  5.30s/it]2026-04-12 03:17:20,110 | INFO | train_grpo_train | metrics_logged mode=train step=3072

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.939393939393941e-07, 'num_tokens': 6984748.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00010368959920015186, 'sampling/sampling_logp_difference/max': 0.002008092822507024, 'sampling/importance_sampling_ratio/min': 0.9984797239303589, 'sampling/importance_sampling_ratio/mean': 1.000089168548584, 'sampling/importance_sampling_ratio/max': 1.0020101070404053, 'entropy': 0.0010117705096490681, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 93%|█████████▎| 3072/3300 [4:40:29<20:09,  5.30s/it]INFO 04-12 03:17:20 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3073/3300 [4:40:33<19:15,  5.09s/it]2026-04-12 03:17:24,715 | INFO | train_grpo_train | metrics_logged mode=train step=3073

                                                     
{'loss': 0.0005, 'grad_norm': 0.6883857846260071, 'learning_rate': 6.90909090909091e-07, 'num_tokens': 6986610.0, 'completions/mean_length': 66.75, 'completions/min_length': 66.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9981359243392944, 'rewards/meter/std': 3.569047839846462e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981359243392944, 'rewards/total_composite/std': 3.569047839846462e-05, 'reward': 0.9981359243392944, 'reward_std': 3.56822092726361e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013685759156942368, 'sampling/sampling_logp_difference/max': 1.1063222885131836, 'sampling/importance_sampling_ratio/min': 0.3307732045650482, 'sampling/importance_sampling_ratio/mean': 0.9993820190429688, 'sampling/importance_sampling_ratio/max': 1.4870829582214355, 'entropy': 0.056920765433460474, 'clip_ratio/low_mean': 0.0018939394503831863, 'clip_ratio/low_min': 0.0018939394503831863, 'clip_ratio/high_mean': 0.007463517948053777, 'clip_ratio/high_max': 0.007463517948053777, 'clip_ratio/region_mean': 0.009357457398436964, 'reward_total_mean': 0.9981359243392944, 'reward_meter_mean': 0.9981359243392944, 'reward_meter_std': 3.569047839846462e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981359243392944, 'reward_total_composite_std': 3.569047839846462e-05, 'epoch': 0.12}

 93%|█████████▎| 3073/3300 [4:40:33<19:15,  5.09s/it]INFO 04-12 03:17:24 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3074/3300 [4:40:38<19:26,  5.16s/it]2026-04-12 03:17:30,024 | INFO | train_grpo_train | metrics_logged mode=train step=3074

                                                     
{'loss': 0.0032, 'grad_norm': 1.0535986423492432, 'learning_rate': 6.878787878787879e-07, 'num_tokens': 6989111.0, 'completions/mean_length': 142.625, 'completions/min_length': 140.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.625, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9991422891616821, 'rewards/meter/std': 0.0001949365541804582, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9813013076782227, 'rewards/total_composite/std': 0.05048074945807457, 'reward': 0.9813013076782227, 'reward_std': 0.050480738282203674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018873082473874092, 'sampling/sampling_logp_difference/max': 1.6918869018554688, 'sampling/importance_sampling_ratio/min': 0.18417169153690338, 'sampling/importance_sampling_ratio/mean': 1.0043222904205322, 'sampling/importance_sampling_ratio/max': 1.7058764696121216, 'entropy': 0.18659264594316483, 'clip_ratio/low_mean': 0.003496503457427025, 'clip_ratio/low_min': 0.003496503457427025, 'clip_ratio/high_mean': 0.009702706767711788, 'clip_ratio/high_max': 0.009702706767711788, 'clip_ratio/region_mean': 0.013199210225138813, 'reward_total_mean': 0.9813013076782227, 'reward_meter_mean': 0.9991422891616821, 'reward_meter_std': 0.0001949365541804582, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9813013076782227, 'reward_total_composite_std': 0.05048074945807457, 'epoch': 0.12}

 93%|█████████▎| 3074/3300 [4:40:39<19:26,  5.16s/it]INFO 04-12 03:17:30 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3075/3300 [4:40:43<18:42,  4.99s/it]2026-04-12 03:17:34,617 | INFO | train_grpo_train | metrics_logged mode=train step=3075

                                                     
{'loss': 0.0244, 'grad_norm': 3.619894504547119, 'learning_rate': 6.848484848484849e-07, 'num_tokens': 6990969.0, 'completions/mean_length': 69.25, 'completions/min_length': 66.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9823685884475708, 'rewards/meter/std': 0.03472607955336571, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9823685884475708, 'rewards/total_composite/std': 0.03472607955336571, 'reward': 0.9823685884475708, 'reward_std': 0.034726083278656006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030778199434280396, 'sampling/sampling_logp_difference/max': 1.412278175354004, 'sampling/importance_sampling_ratio/min': 0.24358773231506348, 'sampling/importance_sampling_ratio/mean': 1.006119728088379, 'sampling/importance_sampling_ratio/max': 1.8515444993972778, 'entropy': 0.24566587805747986, 'clip_ratio/low_mean': 0.0034246575087308884, 'clip_ratio/low_min': 0.0034246575087308884, 'clip_ratio/high_mean': 0.021908456226810813, 'clip_ratio/high_max': 0.021908456226810813, 'clip_ratio/region_mean': 0.0253331137355417, 'reward_total_mean': 0.9823685884475708, 'reward_meter_mean': 0.9823685884475708, 'reward_meter_std': 0.03472607955336571, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9823685884475708, 'reward_total_composite_std': 0.03472607955336571, 'epoch': 0.12}

 93%|█████████▎| 3075/3300 [4:40:43<18:42,  4.99s/it]INFO 04-12 03:17:34 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3076/3300 [4:40:47<17:56,  4.81s/it]2026-04-12 03:17:38,999 | INFO | train_grpo_train | metrics_logged mode=train step=3076

                                                     
{'loss': 0.0003, 'grad_norm': 0.014152090065181255, 'learning_rate': 6.818181818181818e-07, 'num_tokens': 6992713.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.997339129447937, 'rewards/meter/std': 3.875939285080676e-07, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997339129447937, 'rewards/total_composite/std': 3.875939285080676e-07, 'reward': 0.997339129447937, 'reward_std': 3.8628223819614504e-07, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004271084442734718, 'sampling/sampling_logp_difference/max': 0.5116375088691711, 'sampling/importance_sampling_ratio/min': 0.5995131134986877, 'sampling/importance_sampling_ratio/mean': 0.9999513030052185, 'sampling/importance_sampling_ratio/max': 1.2609317302703857, 'entropy': 0.02187307458370924, 'clip_ratio/low_mean': 0.004098360426723957, 'clip_ratio/low_min': 0.004098360426723957, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.997339129447937, 'reward_meter_mean': 0.997339129447937, 'reward_meter_std': 3.875939285080676e-07, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997339129447937, 'reward_total_composite_std': 3.875939285080676e-07, 'epoch': 0.12}

 93%|█████████▎| 3076/3300 [4:40:47<17:56,  4.81s/it]INFO 04-12 03:17:39 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3077/3300 [4:40:52<17:37,  4.74s/it]2026-04-12 03:17:43,587 | INFO | train_grpo_train | metrics_logged mode=train step=3077

                                                     
{'loss': 0.0002, 'grad_norm': 0.27777397632598877, 'learning_rate': 6.78787878787879e-07, 'num_tokens': 6994530.0, 'completions/mean_length': 68.125, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9994935989379883, 'rewards/meter/std': 8.932632226787973e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994935989379883, 'rewards/total_composite/std': 8.932632226787973e-06, 'reward': 0.9994935989379883, 'reward_std': 8.932575838116463e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00869977567344904, 'sampling/sampling_logp_difference/max': 1.2481575012207031, 'sampling/importance_sampling_ratio/min': 0.28703317046165466, 'sampling/importance_sampling_ratio/mean': 1.0012462139129639, 'sampling/importance_sampling_ratio/max': 1.1400055885314941, 'entropy': 0.050312931183725595, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.0018115942366421223, 'clip_ratio/high_max': 0.0018115942366421223, 'clip_ratio/region_mean': 0.0036498295376077294, 'reward_total_mean': 0.9994935989379883, 'reward_meter_mean': 0.9994935989379883, 'reward_meter_std': 8.932632226787973e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994935989379883, 'reward_total_composite_std': 8.932632226787973e-06, 'epoch': 0.12}

 93%|█████████▎| 3077/3300 [4:40:52<17:37,  4.74s/it]INFO 04-12 03:17:43 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3078/3300 [4:40:57<17:34,  4.75s/it]2026-04-12 03:17:48,356 | INFO | train_grpo_train | metrics_logged mode=train step=3078

                                                     
{'loss': -0.0002, 'grad_norm': 0.11041238158941269, 'learning_rate': 6.757575757575759e-07, 'num_tokens': 6996666.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994069337844849, 'rewards/meter/std': 4.966230790159898e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994069337844849, 'rewards/total_composite/std': 4.966230790159898e-06, 'reward': 0.9994069337844849, 'reward_std': 4.970342160959262e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003691497491672635, 'sampling/sampling_logp_difference/max': 0.6290616989135742, 'sampling/importance_sampling_ratio/min': 0.5330917835235596, 'sampling/importance_sampling_ratio/mean': 0.9998006820678711, 'sampling/importance_sampling_ratio/max': 1.217292070388794, 'entropy': 0.032520961947739124, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9994069337844849, 'reward_meter_mean': 0.9994069337844849, 'reward_meter_std': 4.966230790159898e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994069337844849, 'reward_total_composite_std': 4.966230790159898e-06, 'epoch': 0.12}

 93%|█████████▎| 3078/3300 [4:40:57<17:34,  4.75s/it]INFO 04-12 03:17:48 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3079/3300 [4:41:03<18:40,  5.07s/it]2026-04-12 03:17:54,173 | INFO | train_grpo_train | metrics_logged mode=train step=3079

                                                     
{'loss': 0.0004, 'grad_norm': 1.676236867904663, 'learning_rate': 6.727272727272728e-07, 'num_tokens': 6999685.0, 'completions/mean_length': 178.375, 'completions/min_length': 177.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.375, 'completions/min_terminated_length': 177.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.9991593360900879, 'rewards/meter/std': 8.344750676769763e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9583333134651184, 'rewards/repeat_penalty/std': 0.08266931027173996, 'rewards/total_composite/mean': 0.9575244188308716, 'rewards/total_composite/std': 0.08255957812070847, 'reward': 0.9575244188308716, 'reward_std': 0.08255956321954727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020297784358263016, 'sampling/sampling_logp_difference/max': 1.107102394104004, 'sampling/importance_sampling_ratio/min': 0.3305152654647827, 'sampling/importance_sampling_ratio/mean': 1.0064208507537842, 'sampling/importance_sampling_ratio/max': 1.5720078945159912, 'entropy': 0.23156088776886463, 'clip_ratio/low_mean': 0.004197834758087993, 'clip_ratio/low_min': 0.004197834758087993, 'clip_ratio/high_mean': 0.009800075204111636, 'clip_ratio/high_max': 0.009800075204111636, 'clip_ratio/region_mean': 0.013997909962199628, 'reward_total_mean': 0.9575244188308716, 'reward_meter_mean': 0.9991593360900879, 'reward_meter_std': 8.344750676769763e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9583333134651184, 'reward_repeat_penalty_std': 0.08266931027173996, 'reward_total_composite_mean': 0.9575244188308716, 'reward_total_composite_std': 0.08255957812070847, 'epoch': 0.12}

 93%|█████████▎| 3079/3300 [4:41:03<18:40,  5.07s/it]INFO 04-12 03:17:54 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3080/3300 [4:41:07<18:02,  4.92s/it]2026-04-12 03:17:58,744 | INFO | train_grpo_train | metrics_logged mode=train step=3080

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.696969696969698e-07, 'num_tokens': 7001620.0, 'completions/mean_length': 79.875, 'completions/min_length': 79.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0004583710106089711, 'sampling/sampling_logp_difference/max': 0.014358794316649437, 'sampling/importance_sampling_ratio/min': 0.9983435869216919, 'sampling/importance_sampling_ratio/mean': 1.0004509687423706, 'sampling/importance_sampling_ratio/max': 1.0144623517990112, 'entropy': 0.0046348033065442, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 93%|█████████▎| 3080/3300 [4:41:07<18:02,  4.92s/it]INFO 04-12 03:17:58 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3081/3300 [4:41:12<18:18,  5.02s/it]2026-04-12 03:18:03,991 | INFO | train_grpo_train | metrics_logged mode=train step=3081

                                                     
{'loss': 0.0079, 'grad_norm': 2.4082579612731934, 'learning_rate': 6.666666666666667e-07, 'num_tokens': 7004021.0, 'completions/mean_length': 126.125, 'completions/min_length': 124.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.125, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9883460402488708, 'rewards/meter/std': 0.006720090284943581, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9464285969734192, 'rewards/repeat_penalty/std': 0.07393559068441391, 'rewards/total_composite/mean': 0.9353639483451843, 'rewards/total_composite/std': 0.07287098467350006, 'reward': 0.9353639483451843, 'reward_std': 0.07287097722291946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03183456510305405, 'sampling/sampling_logp_difference/max': 1.410654067993164, 'sampling/importance_sampling_ratio/min': 0.243983656167984, 'sampling/importance_sampling_ratio/mean': 1.0118496417999268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36210910603404045, 'clip_ratio/low_mean': 0.005876097013242543, 'clip_ratio/low_min': 0.005876097013242543, 'clip_ratio/high_mean': 0.02169638266786933, 'clip_ratio/high_max': 0.02169638266786933, 'clip_ratio/region_mean': 0.027572479681111872, 'reward_total_mean': 0.9353639483451843, 'reward_meter_mean': 0.9883460402488708, 'reward_meter_std': 0.006720090284943581, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9464285969734192, 'reward_repeat_penalty_std': 0.07393559068441391, 'reward_total_composite_mean': 0.9353639483451843, 'reward_total_composite_std': 0.07287098467350006, 'epoch': 0.12}

 93%|█████████▎| 3081/3300 [4:41:12<18:18,  5.02s/it]INFO 04-12 03:18:04 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3082/3300 [4:41:20<21:12,  5.84s/it]2026-04-12 03:18:11,754 | INFO | train_grpo_train | metrics_logged mode=train step=3082

                                                     
{'loss': 0.0066, 'grad_norm': 3.0218822956085205, 'learning_rate': 6.636363636363636e-07, 'num_tokens': 7008316.0, 'completions/mean_length': 331.875, 'completions/min_length': 316.0, 'completions/max_length': 340.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 331.875, 'completions/min_terminated_length': 316.0, 'completions/max_terminated_length': 340.0, 'rewards/meter/mean': 0.9956144094467163, 'rewards/meter/std': 0.0046219234354794025, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9868420958518982, 'rewards/repeat_penalty/std': 0.024363677948713303, 'rewards/total_composite/mean': 0.8187316656112671, 'rewards/total_composite/std': 0.019191961735486984, 'reward': 0.8187316656112671, 'reward_std': 0.019191952422261238, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06327174603939056, 'sampling/sampling_logp_difference/max': 4.011778354644775, 'sampling/importance_sampling_ratio/min': 0.018101176247000694, 'sampling/importance_sampling_ratio/mean': 1.016357183456421, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.594636969268322, 'clip_ratio/low_mean': 0.011420327704399824, 'clip_ratio/low_min': 0.011420327704399824, 'clip_ratio/high_mean': 0.03181355516426265, 'clip_ratio/high_max': 0.03181355516426265, 'clip_ratio/region_mean': 0.04323388286866248, 'reward_total_mean': 0.8187316656112671, 'reward_meter_mean': 0.9956144094467163, 'reward_meter_std': 0.0046219234354794025, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9868420958518982, 'reward_repeat_penalty_std': 0.024363677948713303, 'reward_total_composite_mean': 0.8187316656112671, 'reward_total_composite_std': 0.019191961735486984, 'epoch': 0.12}

 93%|█████████▎| 3082/3300 [4:41:20<21:12,  5.84s/it]INFO 04-12 03:18:11 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3083/3300 [4:41:25<19:55,  5.51s/it]2026-04-12 03:18:16,519 | INFO | train_grpo_train | metrics_logged mode=train step=3083

                                                     
{'loss': -0.001, 'grad_norm': 3.886202335357666, 'learning_rate': 6.606060606060606e-07, 'num_tokens': 7010295.0, 'completions/mean_length': 68.375, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9993771314620972, 'rewards/meter/std': 0.00016229395987465978, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993771314620972, 'rewards/total_composite/std': 0.00016229395987465978, 'reward': 0.9993771314620972, 'reward_std': 0.00016229994071181864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0466373972594738, 'sampling/sampling_logp_difference/max': 3.3516950607299805, 'sampling/importance_sampling_ratio/min': 0.03502493351697922, 'sampling/importance_sampling_ratio/mean': 0.9933637380599976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.12113892659544945, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.023763853823766112, 'clip_ratio/high_max': 0.023763853823766112, 'clip_ratio/region_mean': 0.02560208912473172, 'reward_total_mean': 0.9993771314620972, 'reward_meter_mean': 0.9993771314620972, 'reward_meter_std': 0.00016229395987465978, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993771314620972, 'reward_total_composite_std': 0.00016229395987465978, 'epoch': 0.12}

 93%|█████████▎| 3083/3300 [4:41:25<19:55,  5.51s/it]INFO 04-12 03:18:16 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3084/3300 [4:41:30<19:02,  5.29s/it]2026-04-12 03:18:21,274 | INFO | train_grpo_train | metrics_logged mode=train step=3084

                                                     
{'loss': -0.0038, 'grad_norm': 7.438833713531494, 'learning_rate': 6.575757575757575e-07, 'num_tokens': 7012093.0, 'completions/mean_length': 70.75, 'completions/min_length': 69.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9911935925483704, 'rewards/meter/std': 0.011583052575588226, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9911935925483704, 'rewards/total_composite/std': 0.011583052575588226, 'reward': 0.9911935925483704, 'reward_std': 0.011583039537072182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034917399287223816, 'sampling/sampling_logp_difference/max': 1.0490436553955078, 'sampling/importance_sampling_ratio/min': 0.3502725660800934, 'sampling/importance_sampling_ratio/mean': 0.999201238155365, 'sampling/importance_sampling_ratio/max': 1.5489575862884521, 'entropy': 0.2923265937715769, 'clip_ratio/low_mean': 0.0036231884732842445, 'clip_ratio/low_min': 0.0036231884732842445, 'clip_ratio/high_mean': 0.03329520847182721, 'clip_ratio/high_max': 0.03329520847182721, 'clip_ratio/region_mean': 0.036918396945111454, 'reward_total_mean': 0.9911935925483704, 'reward_meter_mean': 0.9911935925483704, 'reward_meter_std': 0.011583052575588226, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9911935925483704, 'reward_total_composite_std': 0.011583052575588226, 'epoch': 0.12}

 93%|█████████▎| 3084/3300 [4:41:30<19:02,  5.29s/it]INFO 04-12 03:18:21 [block_pool.py:378] Successfully reset prefix cache

 93%|█████████▎| 3085/3300 [4:41:36<19:43,  5.50s/it]2026-04-12 03:18:27,304 | INFO | train_grpo_train | metrics_logged mode=train step=3085

                                                     
{'loss': 0.0045, 'grad_norm': 1.1208101511001587, 'learning_rate': 6.545454545454547e-07, 'num_tokens': 7014791.0, 'completions/mean_length': 156.25, 'completions/min_length': 150.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 156.25, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.9991309642791748, 'rewards/meter/std': 0.00010435016884002835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991309642791748, 'rewards/total_composite/std': 0.00010435016884002835, 'reward': 0.9991309642791748, 'reward_std': 0.00010435186413815245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03277255594730377, 'sampling/sampling_logp_difference/max': 0.7845733165740967, 'sampling/importance_sampling_ratio/min': 0.456314355134964, 'sampling/importance_sampling_ratio/mean': 1.0100115537643433, 'sampling/importance_sampling_ratio/max': 1.801207184791565, 'entropy': 0.3303154893219471, 'clip_ratio/low_mean': 0.006369811948388815, 'clip_ratio/low_min': 0.006369811948388815, 'clip_ratio/high_mean': 0.01920302864164114, 'clip_ratio/high_max': 0.01920302864164114, 'clip_ratio/region_mean': 0.025572840590029955, 'reward_total_mean': 0.9991309642791748, 'reward_meter_mean': 0.9991309642791748, 'reward_meter_std': 0.00010435016884002835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991309642791748, 'reward_total_composite_std': 0.00010435016884002835, 'epoch': 0.12}

 93%|█████████▎| 3085/3300 [4:41:36<19:43,  5.50s/it]INFO 04-12 03:18:27 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3086/3300 [4:41:41<19:40,  5.52s/it]2026-04-12 03:18:32,829 | INFO | train_grpo_train | metrics_logged mode=train step=3086

                                                     
{'loss': 0.0065, 'grad_norm': 4.402876377105713, 'learning_rate': 6.515151515151516e-07, 'num_tokens': 7017132.0, 'completions/mean_length': 133.625, 'completions/min_length': 132.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.625, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.998715341091156, 'rewards/meter/std': 0.0006201790529303253, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998715341091156, 'rewards/total_composite/std': 0.0006201790529303253, 'reward': 0.998715341091156, 'reward_std': 0.00062018126482144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03640494868159294, 'sampling/sampling_logp_difference/max': 1.8164355754852295, 'sampling/importance_sampling_ratio/min': 0.16260430216789246, 'sampling/importance_sampling_ratio/mean': 1.0022952556610107, 'sampling/importance_sampling_ratio/max': 1.980418086051941, 'entropy': 0.2722382918000221, 'clip_ratio/low_mean': 0.010255418019369245, 'clip_ratio/low_min': 0.010255418019369245, 'clip_ratio/high_mean': 0.028133108280599117, 'clip_ratio/high_max': 0.028133108280599117, 'clip_ratio/region_mean': 0.03838852629996836, 'reward_total_mean': 0.998715341091156, 'reward_meter_mean': 0.998715341091156, 'reward_meter_std': 0.0006201790529303253, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998715341091156, 'reward_total_composite_std': 0.0006201790529303253, 'epoch': 0.12}

 94%|█████████▎| 3086/3300 [4:41:41<19:40,  5.52s/it]INFO 04-12 03:18:33 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3087/3300 [4:41:46<18:38,  5.25s/it]2026-04-12 03:18:37,451 | INFO | train_grpo_train | metrics_logged mode=train step=3087

                                                     
{'loss': -0.0, 'grad_norm': 0.165548175573349, 'learning_rate': 6.484848484848485e-07, 'num_tokens': 7018852.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973361492156982, 'rewards/meter/std': 8.160675861290656e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973361492156982, 'rewards/total_composite/std': 8.160675861290656e-06, 'reward': 0.9973361492156982, 'reward_std': 8.157768206729088e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003540643723681569, 'sampling/sampling_logp_difference/max': 0.9303553104400635, 'sampling/importance_sampling_ratio/min': 0.394413560628891, 'sampling/importance_sampling_ratio/mean': 0.9994197487831116, 'sampling/importance_sampling_ratio/max': 1.119980812072754, 'entropy': 0.016273654997348785, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.0020491802133619785, 'reward_total_mean': 0.9973361492156982, 'reward_meter_mean': 0.9973361492156982, 'reward_meter_std': 8.160675861290656e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973361492156982, 'reward_total_composite_std': 8.160675861290656e-06, 'epoch': 0.12}

 94%|█████████▎| 3087/3300 [4:41:46<18:38,  5.25s/it]INFO 04-12 03:18:37 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3088/3300 [4:41:51<18:26,  5.22s/it]2026-04-12 03:18:42,590 | INFO | train_grpo_train | metrics_logged mode=train step=3088

                                                     
{'loss': 0.0063, 'grad_norm': 5.205334186553955, 'learning_rate': 6.454545454545455e-07, 'num_tokens': 7021320.0, 'completions/mean_length': 123.5, 'completions/min_length': 121.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.5, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9974862337112427, 'rewards/meter/std': 0.00045686395606026053, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9796685576438904, 'rewards/total_composite/std': 0.05027516558766365, 'reward': 0.9796685576438904, 'reward_std': 0.050275154411792755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0200046356767416, 'sampling/sampling_logp_difference/max': 1.3565974235534668, 'sampling/importance_sampling_ratio/min': 0.25753557682037354, 'sampling/importance_sampling_ratio/mean': 1.004548192024231, 'sampling/importance_sampling_ratio/max': 1.8565744161605835, 'entropy': 0.1684292033314705, 'clip_ratio/low_mean': 0.0020000000949949026, 'clip_ratio/low_min': 0.0020000000949949026, 'clip_ratio/high_mean': 0.014212732203304768, 'clip_ratio/high_max': 0.014212732203304768, 'clip_ratio/region_mean': 0.01621273229829967, 'reward_total_mean': 0.9796685576438904, 'reward_meter_mean': 0.9974862337112427, 'reward_meter_std': 0.00045686395606026053, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9796685576438904, 'reward_total_composite_std': 0.05027516558766365, 'epoch': 0.12}

 94%|█████████▎| 3088/3300 [4:41:51<18:26,  5.22s/it]INFO 04-12 03:18:42 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3089/3300 [4:41:59<21:40,  6.16s/it]2026-04-12 03:18:50,962 | INFO | train_grpo_train | metrics_logged mode=train step=3089

                                                     
{'loss': -0.0405, 'grad_norm': 1.6393530368804932, 'learning_rate': 6.424242424242424e-07, 'num_tokens': 7025761.0, 'completions/mean_length': 360.125, 'completions/min_length': 344.0, 'completions/max_length': 397.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 360.125, 'completions/min_terminated_length': 344.0, 'completions/max_terminated_length': 397.0, 'rewards/meter/mean': 0.9991971254348755, 'rewards/meter/std': 0.00017116445815190673, 'rewards/count_adherence/mean': 0.8409091234207153, 'rewards/count_adherence/std': 0.04208274558186531, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9860681295394897, 'rewards/repeat_penalty/std': 0.02584986388683319, 'rewards/total_composite/mean': 0.8282498717308044, 'rewards/total_composite/std': 0.040605854243040085, 'reward': 0.8282498717308044, 'reward_std': 0.04060585796833038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04586139693856239, 'sampling/sampling_logp_difference/max': 1.3152236938476562, 'sampling/importance_sampling_ratio/min': 0.2684142589569092, 'sampling/importance_sampling_ratio/mean': 1.0105684995651245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4755103662610054, 'clip_ratio/low_mean': 0.01749891194049269, 'clip_ratio/low_min': 0.01749891194049269, 'clip_ratio/high_mean': 0.007637429982423782, 'clip_ratio/high_max': 0.007637429982423782, 'clip_ratio/region_mean': 0.025136341922916472, 'reward_total_mean': 0.8282498717308044, 'reward_meter_mean': 0.9991971254348755, 'reward_meter_std': 0.00017116445815190673, 'reward_count_adherence_mean': 0.8409091234207153, 'reward_count_adherence_std': 0.04208274558186531, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9860681295394897, 'reward_repeat_penalty_std': 0.02584986388683319, 'reward_total_composite_mean': 0.8282498717308044, 'reward_total_composite_std': 0.040605854243040085, 'epoch': 0.12}

 94%|█████████▎| 3089/3300 [4:41:59<21:40,  6.16s/it]INFO 04-12 03:18:51 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3090/3300 [4:42:04<19:56,  5.70s/it]2026-04-12 03:18:55,566 | INFO | train_grpo_train | metrics_logged mode=train step=3090

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.393939393939394e-07, 'num_tokens': 7027640.0, 'completions/mean_length': 79.875, 'completions/min_length': 79.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.001453942502848804, 'sampling/sampling_logp_difference/max': 0.6242718696594238, 'sampling/importance_sampling_ratio/min': 0.5356513261795044, 'sampling/importance_sampling_ratio/mean': 0.9997469782829285, 'sampling/importance_sampling_ratio/max': 1.015737533569336, 'entropy': 0.004133307666052133, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 94%|█████████▎| 3090/3300 [4:42:04<19:56,  5.70s/it]INFO 04-12 03:18:55 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3091/3300 [4:42:08<18:39,  5.36s/it]2026-04-12 03:19:00,137 | INFO | train_grpo_train | metrics_logged mode=train step=3091

                                                     
{'loss': 0.0015, 'grad_norm': 0.5760459899902344, 'learning_rate': 6.363636363636364e-07, 'num_tokens': 7029612.0, 'completions/mean_length': 71.5, 'completions/min_length': 70.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.5, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994015097618103, 'rewards/meter/std': 4.198703754809685e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994015097618103, 'rewards/total_composite/std': 4.198703754809685e-05, 'reward': 0.9994015097618103, 'reward_std': 4.200140756438486e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010645559057593346, 'sampling/sampling_logp_difference/max': 0.8983035087585449, 'sampling/importance_sampling_ratio/min': 0.4072599709033966, 'sampling/importance_sampling_ratio/mean': 1.0001534223556519, 'sampling/importance_sampling_ratio/max': 1.4178498983383179, 'entropy': 0.09367726277559996, 'clip_ratio/low_mean': 0.0034722222480922937, 'clip_ratio/low_min': 0.0034722222480922937, 'clip_ratio/high_mean': 0.007018499891273677, 'clip_ratio/high_max': 0.007018499891273677, 'clip_ratio/region_mean': 0.010490722139365971, 'reward_total_mean': 0.9994015097618103, 'reward_meter_mean': 0.9994015097618103, 'reward_meter_std': 4.198703754809685e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994015097618103, 'reward_total_composite_std': 4.198703754809685e-05, 'epoch': 0.12}

 94%|█████████▎| 3091/3300 [4:42:09<18:39,  5.36s/it]INFO 04-12 03:19:00 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3092/3300 [4:42:14<18:26,  5.32s/it]2026-04-12 03:19:05,357 | INFO | train_grpo_train | metrics_logged mode=train step=3092

                                                     
{'loss': -0.0027, 'grad_norm': 2.74900484085083, 'learning_rate': 6.333333333333334e-07, 'num_tokens': 7031807.0, 'completions/mean_length': 117.375, 'completions/min_length': 116.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.375, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.9986400604248047, 'rewards/meter/std': 0.0013286188477650285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9737505912780762, 'rewards/total_composite/std': 0.07163625955581665, 'reward': 0.9737505912780762, 'reward_std': 0.07163627445697784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03260769695043564, 'sampling/sampling_logp_difference/max': 1.1829195022583008, 'sampling/importance_sampling_ratio/min': 0.306382954120636, 'sampling/importance_sampling_ratio/mean': 1.0021259784698486, 'sampling/importance_sampling_ratio/max': 1.3800021409988403, 'entropy': 0.2854463141411543, 'clip_ratio/low_mean': 0.0032327587250620127, 'clip_ratio/low_min': 0.0032327587250620127, 'clip_ratio/high_mean': 0.01596410130150616, 'clip_ratio/high_max': 0.01596410130150616, 'clip_ratio/region_mean': 0.019196860026568174, 'reward_total_mean': 0.9737505912780762, 'reward_meter_mean': 0.9986400604248047, 'reward_meter_std': 0.0013286188477650285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9737505912780762, 'reward_total_composite_std': 0.07163625955581665, 'epoch': 0.12}

 94%|█████████▎| 3092/3300 [4:42:14<18:26,  5.32s/it]INFO 04-12 03:19:05 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▎| 3093/3300 [4:42:18<17:29,  5.07s/it]2026-04-12 03:19:09,860 | INFO | train_grpo_train | metrics_logged mode=train step=3093

                                                     
{'loss': 0.0001, 'grad_norm': 0.078668013215065, 'learning_rate': 6.303030303030304e-07, 'num_tokens': 7033567.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973331689834595, 'rewards/meter/std': 1.0622773515933659e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973331689834595, 'rewards/total_composite/std': 1.0622773515933659e-05, 'reward': 0.9973331689834595, 'reward_std': 1.0613573067530524e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0038269164506345987, 'sampling/sampling_logp_difference/max': 0.5654642581939697, 'sampling/importance_sampling_ratio/min': 0.7023022770881653, 'sampling/importance_sampling_ratio/mean': 1.0025733709335327, 'sampling/importance_sampling_ratio/max': 1.7602647542953491, 'entropy': 0.020028789876960218, 'clip_ratio/low_mean': 0.004098360426723957, 'clip_ratio/low_min': 0.004098360426723957, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.9973331689834595, 'reward_meter_mean': 0.9973331689834595, 'reward_meter_std': 1.0622773515933659e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973331689834595, 'reward_total_composite_std': 1.0622773515933659e-05, 'epoch': 0.12}

 94%|█████████▎| 3093/3300 [4:42:18<17:29,  5.07s/it]INFO 04-12 03:19:10 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3094/3300 [4:42:23<16:48,  4.90s/it]2026-04-12 03:19:14,341 | INFO | train_grpo_train | metrics_logged mode=train step=3094

                                                     
{'loss': 0.0003, 'grad_norm': 0.07660355418920517, 'learning_rate': 6.272727272727273e-07, 'num_tokens': 7035327.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973366260528564, 'rewards/meter/std': 8.457681360596325e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973366260528564, 'rewards/total_composite/std': 8.457681360596325e-06, 'reward': 0.9973366260528564, 'reward_std': 8.46226976136677e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008617924526333809, 'sampling/sampling_logp_difference/max': 1.087322473526001, 'sampling/importance_sampling_ratio/min': 0.3371179401874542, 'sampling/importance_sampling_ratio/mean': 0.9982660412788391, 'sampling/importance_sampling_ratio/max': 1.83155357837677, 'entropy': 0.024718819418922067, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.010245901066809893, 'clip_ratio/high_max': 0.010245901066809893, 'clip_ratio/region_mean': 0.012295081280171871, 'reward_total_mean': 0.9973366260528564, 'reward_meter_mean': 0.9973366260528564, 'reward_meter_std': 8.457681360596325e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973366260528564, 'reward_total_composite_std': 8.457681360596325e-06, 'epoch': 0.12}

 94%|█████████▍| 3094/3300 [4:42:23<16:48,  4.90s/it]INFO 04-12 03:19:14 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3095/3300 [4:42:28<17:11,  5.03s/it]2026-04-12 03:19:19,683 | INFO | train_grpo_train | metrics_logged mode=train step=3095

                                                     
{'loss': 0.0003, 'grad_norm': 0.8752101063728333, 'learning_rate': 6.242424242424243e-07, 'num_tokens': 7037833.0, 'completions/mean_length': 131.25, 'completions/min_length': 130.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.25, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.9993793964385986, 'rewards/meter/std': 9.467204654356465e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993793964385986, 'rewards/total_composite/std': 9.467204654356465e-05, 'reward': 0.9993793964385986, 'reward_std': 9.466917254030704e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008658221922814846, 'sampling/sampling_logp_difference/max': 0.8032641410827637, 'sampling/importance_sampling_ratio/min': 0.44786468148231506, 'sampling/importance_sampling_ratio/mean': 1.0029832124710083, 'sampling/importance_sampling_ratio/max': 1.359288215637207, 'entropy': 0.08324715122580528, 'clip_ratio/low_mean': 0.0028625954291783273, 'clip_ratio/low_min': 0.0028625954291783273, 'clip_ratio/high_mean': 0.0019085081876255572, 'clip_ratio/high_max': 0.0019085081876255572, 'clip_ratio/region_mean': 0.0047711036168038845, 'reward_total_mean': 0.9993793964385986, 'reward_meter_mean': 0.9993793964385986, 'reward_meter_std': 9.467204654356465e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993793964385986, 'reward_total_composite_std': 9.467204654356465e-05, 'epoch': 0.12}

 94%|█████████▍| 3095/3300 [4:42:28<17:11,  5.03s/it]INFO 04-12 03:19:19 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3096/3300 [4:42:33<17:00,  5.00s/it]2026-04-12 03:19:24,623 | INFO | train_grpo_train | metrics_logged mode=train step=3096

                                                     
{'loss': -0.0004, 'grad_norm': 0.717937171459198, 'learning_rate': 6.212121212121212e-07, 'num_tokens': 7039952.0, 'completions/mean_length': 97.875, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9993786811828613, 'rewards/meter/std': 8.659059676574543e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993786811828613, 'rewards/total_composite/std': 8.659059676574543e-05, 'reward': 0.9993786811828613, 'reward_std': 8.658833394292742e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006720404606312513, 'sampling/sampling_logp_difference/max': 0.9676318168640137, 'sampling/importance_sampling_ratio/min': 0.3799818456172943, 'sampling/importance_sampling_ratio/mean': 0.9991164803504944, 'sampling/importance_sampling_ratio/max': 1.574755072593689, 'entropy': 0.03362013725563884, 'clip_ratio/low_mean': 0.0025510203558951616, 'clip_ratio/low_min': 0.0025510203558951616, 'clip_ratio/high_mean': 0.006390700465999544, 'clip_ratio/high_max': 0.006390700465999544, 'clip_ratio/region_mean': 0.008941720821894705, 'reward_total_mean': 0.9993786811828613, 'reward_meter_mean': 0.9993786811828613, 'reward_meter_std': 8.659059676574543e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993786811828613, 'reward_total_composite_std': 8.659059676574543e-05, 'epoch': 0.12}

 94%|█████████▍| 3096/3300 [4:42:33<17:00,  5.00s/it]INFO 04-12 03:19:24 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3097/3300 [4:42:38<16:32,  4.89s/it]2026-04-12 03:19:29,237 | INFO | train_grpo_train | metrics_logged mode=train step=3097

                                                     
{'loss': -0.0005, 'grad_norm': 0.17411062121391296, 'learning_rate': 6.181818181818182e-07, 'num_tokens': 7041771.0, 'completions/mean_length': 66.375, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981486797332764, 'rewards/meter/std': 1.2580998372868635e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981486797332764, 'rewards/total_composite/std': 1.2580998372868635e-05, 'reward': 0.9981486797332764, 'reward_std': 1.2579364010889549e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00830040592700243, 'sampling/sampling_logp_difference/max': 0.5602436065673828, 'sampling/importance_sampling_ratio/min': 0.5710699558258057, 'sampling/importance_sampling_ratio/mean': 1.0030303001403809, 'sampling/importance_sampling_ratio/max': 1.580132246017456, 'entropy': 0.05762590793892741, 'clip_ratio/low_mean': 0.005681818351149559, 'clip_ratio/low_min': 0.005681818351149559, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.007547489949502051, 'reward_total_mean': 0.9981486797332764, 'reward_meter_mean': 0.9981486797332764, 'reward_meter_std': 1.2580998372868635e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981486797332764, 'reward_total_composite_std': 1.2580998372868635e-05, 'epoch': 0.12}

 94%|█████████▍| 3097/3300 [4:42:38<16:32,  4.89s/it]INFO 04-12 03:19:29 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3098/3300 [4:42:46<19:53,  5.91s/it]2026-04-12 03:19:37,538 | INFO | train_grpo_train | metrics_logged mode=train step=3098

                                                     
{'loss': -0.0254, 'grad_norm': 2.0594077110290527, 'learning_rate': 6.151515151515152e-07, 'num_tokens': 7045692.0, 'completions/mean_length': 283.125, 'completions/min_length': 262.0, 'completions/max_length': 310.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 283.125, 'completions/min_terminated_length': 262.0, 'completions/max_terminated_length': 310.0, 'rewards/meter/mean': 0.9590431451797485, 'rewards/meter/std': 0.07114315778017044, 'rewards/count_adherence/mean': 0.9722222089767456, 'rewards/count_adherence/std': 0.05143444985151291, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9272365570068359, 'rewards/repeat_penalty/std': 0.04068033769726753, 'rewards/total_composite/mean': 0.7604771852493286, 'rewards/total_composite/std': 0.31678301095962524, 'reward': 0.7604771852493286, 'reward_std': 0.31678304076194763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05060786381363869, 'sampling/sampling_logp_difference/max': 1.4762558937072754, 'sampling/importance_sampling_ratio/min': 0.22849158942699432, 'sampling/importance_sampling_ratio/mean': 1.0093165636062622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.522564172744751, 'clip_ratio/low_mean': 0.006949802860617638, 'clip_ratio/low_min': 0.006949802860617638, 'clip_ratio/high_mean': 0.025927380891516805, 'clip_ratio/high_max': 0.025927380891516805, 'clip_ratio/region_mean': 0.03287718375213444, 'reward_total_mean': 0.7604771852493286, 'reward_meter_mean': 0.9590431451797485, 'reward_meter_std': 0.07114315778017044, 'reward_count_adherence_mean': 0.9722222089767456, 'reward_count_adherence_std': 0.05143444985151291, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9272365570068359, 'reward_repeat_penalty_std': 0.04068033769726753, 'reward_total_composite_mean': 0.7604771852493286, 'reward_total_composite_std': 0.31678301095962524, 'epoch': 0.12}

 94%|█████████▍| 3098/3300 [4:42:46<19:53,  5.91s/it]INFO 04-12 03:19:37 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3099/3300 [4:42:51<19:28,  5.82s/it]2026-04-12 03:19:43,127 | INFO | train_grpo_train | metrics_logged mode=train step=3099

                                                     
{'loss': 0.0008, 'grad_norm': 0.2977856993675232, 'learning_rate': 6.121212121212121e-07, 'num_tokens': 7047543.0, 'completions/mean_length': 71.375, 'completions/min_length': 71.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.375, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9994341135025024, 'rewards/meter/std': 2.3516733563155867e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994341135025024, 'rewards/total_composite/std': 2.3516733563155867e-05, 'reward': 0.9994341135025024, 'reward_std': 2.3526572476839647e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010309785604476929, 'sampling/sampling_logp_difference/max': 1.0300507545471191, 'sampling/importance_sampling_ratio/min': 0.3569888472557068, 'sampling/importance_sampling_ratio/mean': 0.999859631061554, 'sampling/importance_sampling_ratio/max': 1.2391035556793213, 'entropy': 0.07044319622218609, 'clip_ratio/low_mean': 0.005138941807672381, 'clip_ratio/low_min': 0.005138941807672381, 'clip_ratio/high_mean': 0.0035211266949772835, 'clip_ratio/high_max': 0.0035211266949772835, 'clip_ratio/region_mean': 0.008660068502649665, 'reward_total_mean': 0.9994341135025024, 'reward_meter_mean': 0.9994341135025024, 'reward_meter_std': 2.3516733563155867e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994341135025024, 'reward_total_composite_std': 2.3516733563155867e-05, 'epoch': 0.12}

 94%|█████████▍| 3099/3300 [4:42:52<19:28,  5.82s/it]INFO 04-12 03:19:43 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3100/3300 [4:42:57<19:00,  5.70s/it]2026-04-12 03:19:48,567 | INFO | train_grpo_train | metrics_logged mode=train step=3100

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.090909090909092e-07, 'num_tokens': 7049375.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 8.230307139456272e-05, 'sampling/sampling_logp_difference/max': 0.0053017400205135345, 'sampling/importance_sampling_ratio/min': 0.9947122931480408, 'sampling/importance_sampling_ratio/mean': 1.0000481605529785, 'sampling/importance_sampling_ratio/max': 1.0023247003555298, 'entropy': 0.0007583037222502753, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 94%|█████████▍| 3100/3300 [4:42:57<19:00,  5.70s/it]INFO 04-12 03:19:48 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:38,  3.53s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.50s/it]

 31%|███       | 4/13 [00:16<00:39,  4.34s/it]

 38%|███▊      | 5/13 [00:21<00:37,  4.72s/it]

 46%|████▌     | 6/13 [00:28<00:37,  5.30s/it]

 54%|█████▍    | 7/13 [00:33<00:31,  5.26s/it]

 62%|██████▏   | 8/13 [00:40<00:29,  5.86s/it]

 69%|██████▉   | 9/13 [00:47<00:25,  6.30s/it]

 77%|███████▋  | 10/13 [00:52<00:17,  5.83s/it]

 85%|████████▍ | 11/13 [00:58<00:11,  5.90s/it]

 92%|█████████▏| 12/13 [01:04<00:05,  5.90s/it]

100%|██████████| 13/13 [01:10<00:00,  5.90s/it]2026-04-12 03:21:06,357 | INFO | train_grpo_train | metrics_logged mode=eval step=3100

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 77.6639, 'eval_samples_per_second': 1.339, 'eval_steps_per_second': 0.167, 'eval_num_tokens': 7049375.0, 'eval_completions/mean_length': 210.7403846153846, 'eval_completions/min_length': 61.07692307692308, 'eval_completions/max_length': 415.0, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 204.17994689941406, 'eval_completions/min_terminated_length': 61.07692307692308, 'eval_completions/max_terminated_length': 392.53846153846155, 'eval_rewards/meter/mean': 0.8045936226844788, 'eval_rewards/meter/std': 0.3283666269137309, 'eval_rewards/count_adherence/mean': 0.9585646253365737, 'eval_rewards/count_adherence/std': 0.065789727637401, 'eval_rewards/arabic_clean/mean': 0.9903846153846154, 'eval_rewards/arabic_clean/std': 0.027196414195574246, 'eval_rewards/repeat_penalty/mean': 0.9568017308528607, 'eval_rewards/repeat_penalty/std': 0.065969582217244, 'eval_rewards/total_composite/mean': 0.7422462472548852, 'eval_rewards/total_composite/std': 0.3302810134795996, 'eval_reward': 0.7422462472548852, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03276929044379638, 'eval_sampling/sampling_logp_difference/max': 1.120217965199397, 'eval_sampling/importance_sampling_ratio/min': 0.3349529756949498, 'eval_sampling/importance_sampling_ratio/mean': 1.009482246178847, 'eval_sampling/importance_sampling_ratio/max': 1.5006029880963838, 'eval_entropy': 0.37355728218188655, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7422462472548852, 'eval_reward_meter_mean': 0.8045936226844788, 'eval_reward_meter_std': 0.3283666269137309, 'eval_reward_count_adherence_mean': 0.9585646253365737, 'eval_reward_count_adherence_std': 0.065789727637401, 'eval_reward_arabic_clean_mean': 0.9903846153846154, 'eval_reward_arabic_clean_std': 0.027196414195574246, 'eval_reward_repeat_penalty_mean': 0.9568017308528607, 'eval_reward_repeat_penalty_std': 0.065969582217244, 'eval_reward_total_composite_mean': 0.7422462472548852, 'eval_reward_total_composite_std': 0.3302810134795996, 'epoch': 0.12}

 94%|█████████▍| 3100/3300 [4:44:15<19:00,  5.70s/it]

100%|██████████| 13/13 [01:10<00:00,  5.90s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 94%|█████████▍| 3101/3300 [4:44:25<1:40:45, 30.38s/it]2026-04-12 03:21:16,534 | INFO | train_grpo_train | metrics_logged mode=train step=3101

                                                       
{'loss': 0.0255, 'grad_norm': 2.886378765106201, 'learning_rate': 6.060606060606061e-07, 'num_tokens': 7052434.0, 'completions/mean_length': 209.375, 'completions/min_length': 198.0, 'completions/max_length': 223.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 209.375, 'completions/min_terminated_length': 198.0, 'completions/max_terminated_length': 223.0, 'rewards/meter/mean': 0.9796897768974304, 'rewards/meter/std': 0.013765236362814903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9545454978942871, 'rewards/repeat_penalty/std': 0.0971859022974968, 'rewards/total_composite/mean': 0.9347763061523438, 'rewards/total_composite/std': 0.0924786925315857, 'reward': 0.9347763061523438, 'reward_std': 0.0924786925315857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03803591802716255, 'sampling/sampling_logp_difference/max': 0.9048151969909668, 'sampling/importance_sampling_ratio/min': 0.40461668372154236, 'sampling/importance_sampling_ratio/mean': 1.0072675943374634, 'sampling/importance_sampling_ratio/max': 1.8678045272827148, 'entropy': 0.42885980382561684, 'clip_ratio/low_mean': 0.005714374827221036, 'clip_ratio/low_min': 0.005714374827221036, 'clip_ratio/high_mean': 0.03715289803221822, 'clip_ratio/high_max': 0.03715289803221822, 'clip_ratio/region_mean': 0.042867272859439254, 'reward_total_mean': 0.9347763061523438, 'reward_meter_mean': 0.9796897768974304, 'reward_meter_std': 0.013765236362814903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9545454978942871, 'reward_repeat_penalty_std': 0.0971859022974968, 'reward_total_composite_mean': 0.9347763061523438, 'reward_total_composite_std': 0.0924786925315857, 'epoch': 0.12}

 94%|█████████▍| 3101/3300 [4:44:25<1:40:45, 30.38s/it]INFO 04-12 03:21:16 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3102/3300 [4:44:31<1:16:08, 23.07s/it]2026-04-12 03:21:22,556 | INFO | train_grpo_train | metrics_logged mode=train step=3102

                                                       
{'loss': 0.0064, 'grad_norm': 3.9274234771728516, 'learning_rate': 6.03030303030303e-07, 'num_tokens': 7054118.0, 'completions/mean_length': 65.5, 'completions/min_length': 64.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9924110770225525, 'rewards/meter/std': 0.00258190231397748, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9924110770225525, 'rewards/total_composite/std': 0.00258190231397748, 'reward': 0.9924110770225525, 'reward_std': 0.0025818964932113886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.023297905921936035, 'sampling/sampling_logp_difference/max': 1.1882240772247314, 'sampling/importance_sampling_ratio/min': 0.30476200580596924, 'sampling/importance_sampling_ratio/mean': 1.002340316772461, 'sampling/importance_sampling_ratio/max': 1.3347103595733643, 'entropy': 0.14634791854768991, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005800189450383186, 'clip_ratio/high_max': 0.005800189450383186, 'clip_ratio/region_mean': 0.005800189450383186, 'reward_total_mean': 0.9924110770225525, 'reward_meter_mean': 0.9924110770225525, 'reward_meter_std': 0.00258190231397748, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9924110770225525, 'reward_total_composite_std': 0.00258190231397748, 'epoch': 0.12}

 94%|█████████▍| 3102/3300 [4:44:31<1:16:08, 23.07s/it]INFO 04-12 03:21:22 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3103/3300 [4:44:37<58:57, 17.96s/it]  2026-04-12 03:21:28,585 | INFO | train_grpo_train | metrics_logged mode=train step=3103

                                                     
{'loss': 0.0107, 'grad_norm': 5.046589374542236, 'learning_rate': 6.000000000000001e-07, 'num_tokens': 7056293.0, 'completions/mean_length': 85.875, 'completions/min_length': 83.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.875, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9477604627609253, 'rewards/meter/std': 0.021915482357144356, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.930683970451355, 'rewards/total_composite/std': 0.04980318620800972, 'reward': 0.930683970451355, 'reward_std': 0.04980317875742912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06328704953193665, 'sampling/sampling_logp_difference/max': 1.8605287075042725, 'sampling/importance_sampling_ratio/min': 0.15559034049510956, 'sampling/importance_sampling_ratio/mean': 0.993628203868866, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26804009079933167, 'clip_ratio/low_mean': 0.017914682626724243, 'clip_ratio/low_min': 0.017914682626724243, 'clip_ratio/high_mean': 0.03492570295929909, 'clip_ratio/high_max': 0.03492570295929909, 'clip_ratio/region_mean': 0.05284038558602333, 'reward_total_mean': 0.930683970451355, 'reward_meter_mean': 0.9477604627609253, 'reward_meter_std': 0.021915482357144356, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.930683970451355, 'reward_total_composite_std': 0.04980318620800972, 'epoch': 0.12}

 94%|█████████▍| 3103/3300 [4:44:37<58:57, 17.96s/it]INFO 04-12 03:21:28 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3104/3300 [4:44:43<47:12, 14.45s/it]2026-04-12 03:21:34,867 | INFO | train_grpo_train | metrics_logged mode=train step=3104

                                                     
{'loss': 0.002, 'grad_norm': 3.8720288276672363, 'learning_rate': 5.96969696969697e-07, 'num_tokens': 7058098.0, 'completions/mean_length': 77.625, 'completions/min_length': 77.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.625, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9990447163581848, 'rewards/meter/std': 0.0004974787007085979, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990447163581848, 'rewards/total_composite/std': 0.0004974787007085979, 'reward': 0.9990447163581848, 'reward_std': 0.0004974827752448618, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.019970104098320007, 'sampling/sampling_logp_difference/max': 0.8766874074935913, 'sampling/importance_sampling_ratio/min': 0.4161592125892639, 'sampling/importance_sampling_ratio/mean': 1.0045377016067505, 'sampling/importance_sampling_ratio/max': 1.4500758647918701, 'entropy': 0.18993182107806206, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.00967782223597169, 'clip_ratio/high_max': 0.00967782223597169, 'clip_ratio/region_mean': 0.00967782223597169, 'reward_total_mean': 0.9990447163581848, 'reward_meter_mean': 0.9990447163581848, 'reward_meter_std': 0.0004974787007085979, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990447163581848, 'reward_total_composite_std': 0.0004974787007085979, 'epoch': 0.12}

 94%|█████████▍| 3104/3300 [4:44:43<47:12, 14.45s/it]INFO 04-12 03:21:35 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3105/3300 [4:44:49<38:57, 11.98s/it]2026-04-12 03:21:41,086 | INFO | train_grpo_train | metrics_logged mode=train step=3105

                                                     
{'loss': -0.0052, 'grad_norm': 6.449222087860107, 'learning_rate': 5.93939393939394e-07, 'num_tokens': 7059998.0, 'completions/mean_length': 66.5, 'completions/min_length': 64.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9450548887252808, 'rewards/meter/std': 0.015829402953386307, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9450548887252808, 'rewards/total_composite/std': 0.015829402953386307, 'reward': 0.9450548887252808, 'reward_std': 0.015829408541321754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03087073564529419, 'sampling/sampling_logp_difference/max': 1.6123805046081543, 'sampling/importance_sampling_ratio/min': 0.19941234588623047, 'sampling/importance_sampling_ratio/mean': 1.0009596347808838, 'sampling/importance_sampling_ratio/max': 1.7957602739334106, 'entropy': 0.13373739924281836, 'clip_ratio/low_mean': 0.007753314450383186, 'clip_ratio/low_min': 0.007753314450383186, 'clip_ratio/high_mean': 0.02611940260976553, 'clip_ratio/high_max': 0.02611940260976553, 'clip_ratio/region_mean': 0.033872717060148716, 'reward_total_mean': 0.9450548887252808, 'reward_meter_mean': 0.9450548887252808, 'reward_meter_std': 0.015829402953386307, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9450548887252808, 'reward_total_composite_std': 0.015829402953386307, 'epoch': 0.12}

 94%|█████████▍| 3105/3300 [4:44:50<38:57, 11.98s/it]INFO 04-12 03:21:41 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3106/3300 [4:44:55<32:57, 10.19s/it]2026-04-12 03:21:47,086 | INFO | train_grpo_train | metrics_logged mode=train step=3106

                                                     
{'loss': 0.0012, 'grad_norm': 1.8398211002349854, 'learning_rate': 5.90909090909091e-07, 'num_tokens': 7062542.0, 'completions/mean_length': 155.0, 'completions/min_length': 153.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 155.0, 'completions/min_terminated_length': 153.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.9989175200462341, 'rewards/meter/std': 0.00033335990156047046, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989175200462341, 'rewards/total_composite/std': 0.00033335990156047046, 'reward': 0.9989175200462341, 'reward_std': 0.0003333572531118989, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03886400908231735, 'sampling/sampling_logp_difference/max': 2.4166221618652344, 'sampling/importance_sampling_ratio/min': 0.0892224907875061, 'sampling/importance_sampling_ratio/mean': 1.0071746110916138, 'sampling/importance_sampling_ratio/max': 1.7075154781341553, 'entropy': 0.3304919973015785, 'clip_ratio/low_mean': 0.004859996202867478, 'clip_ratio/low_min': 0.004859996202867478, 'clip_ratio/high_mean': 0.0121077821822837, 'clip_ratio/high_max': 0.0121077821822837, 'clip_ratio/region_mean': 0.016967778385151178, 'reward_total_mean': 0.9989175200462341, 'reward_meter_mean': 0.9989175200462341, 'reward_meter_std': 0.00033335990156047046, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989175200462341, 'reward_total_composite_std': 0.00033335990156047046, 'epoch': 0.12}

 94%|█████████▍| 3106/3300 [4:44:56<32:57, 10.19s/it]INFO 04-12 03:21:47 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3107/3300 [4:45:05<32:01,  9.95s/it]2026-04-12 03:21:56,486 | INFO | train_grpo_train | metrics_logged mode=train step=3107

                                                     
{'loss': -0.0223, 'grad_norm': 1.4729881286621094, 'learning_rate': 5.878787878787879e-07, 'num_tokens': 7067381.0, 'completions/mean_length': 403.875, 'completions/min_length': 383.0, 'completions/max_length': 426.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 403.875, 'completions/min_terminated_length': 383.0, 'completions/max_terminated_length': 426.0, 'rewards/meter/mean': 0.9989637136459351, 'rewards/meter/std': 0.00037519802572205663, 'rewards/count_adherence/mean': 0.7980769276618958, 'rewards/count_adherence/std': 0.039811473339796066, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9746710062026978, 'rewards/repeat_penalty/std': 0.038055676966905594, 'rewards/total_composite/mean': 0.7777890563011169, 'rewards/total_composite/std': 0.061021268367767334, 'reward': 0.7777890563011169, 'reward_std': 0.061021264642477036, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04732804745435715, 'sampling/sampling_logp_difference/max': 1.5902900695800781, 'sampling/importance_sampling_ratio/min': 0.20386646687984467, 'sampling/importance_sampling_ratio/mean': 1.0116568803787231, 'sampling/importance_sampling_ratio/max': 1.9504014253616333, 'entropy': 0.4669685959815979, 'clip_ratio/low_mean': 0.01397907012142241, 'clip_ratio/low_min': 0.01397907012142241, 'clip_ratio/high_mean': 0.014827882871031761, 'clip_ratio/high_max': 0.014827882871031761, 'clip_ratio/region_mean': 0.02880695299245417, 'reward_total_mean': 0.7777890563011169, 'reward_meter_mean': 0.9989637136459351, 'reward_meter_std': 0.00037519802572205663, 'reward_count_adherence_mean': 0.7980769276618958, 'reward_count_adherence_std': 0.039811473339796066, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9746710062026978, 'reward_repeat_penalty_std': 0.038055676966905594, 'reward_total_composite_mean': 0.7777890563011169, 'reward_total_composite_std': 0.061021268367767334, 'epoch': 0.12}

 94%|█████████▍| 3107/3300 [4:45:05<32:01,  9.95s/it]INFO 04-12 03:21:56 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3108/3300 [4:45:10<27:24,  8.56s/it]2026-04-12 03:22:01,832 | INFO | train_grpo_train | metrics_logged mode=train step=3108

                                                     
{'loss': 0.0018, 'grad_norm': 0.7012980580329895, 'learning_rate': 5.848484848484849e-07, 'num_tokens': 7069454.0, 'completions/mean_length': 93.125, 'completions/min_length': 93.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.125, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9972705841064453, 'rewards/meter/std': 0.0013248658506199718, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972705841064453, 'rewards/total_composite/std': 0.0013248658506199718, 'reward': 0.9972705841064453, 'reward_std': 0.0013248566538095474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009307874366641045, 'sampling/sampling_logp_difference/max': 0.49368953704833984, 'sampling/importance_sampling_ratio/min': 0.6166864037513733, 'sampling/importance_sampling_ratio/mean': 1.004794716835022, 'sampling/importance_sampling_ratio/max': 1.6383498907089233, 'entropy': 0.0752844586968422, 'clip_ratio/low_mean': 0.003989361692219973, 'clip_ratio/low_min': 0.003989361692219973, 'clip_ratio/high_mean': 0.0067204301012679935, 'clip_ratio/high_max': 0.0067204301012679935, 'clip_ratio/region_mean': 0.010709791793487966, 'reward_total_mean': 0.9972705841064453, 'reward_meter_mean': 0.9972705841064453, 'reward_meter_std': 0.0013248658506199718, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972705841064453, 'reward_total_composite_std': 0.0013248658506199718, 'epoch': 0.12}

 94%|█████████▍| 3108/3300 [4:45:10<27:24,  8.56s/it]INFO 04-12 03:22:02 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3109/3300 [4:45:17<25:15,  7.93s/it]2026-04-12 03:22:08,295 | INFO | train_grpo_train | metrics_logged mode=train step=3109

                                                     
{'loss': 0.0016, 'grad_norm': 3.0335049629211426, 'learning_rate': 5.818181818181819e-07, 'num_tokens': 7072207.0, 'completions/mean_length': 172.125, 'completions/min_length': 165.0, 'completions/max_length': 178.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 172.125, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 178.0, 'rewards/meter/mean': 0.961860179901123, 'rewards/meter/std': 0.08823229372501373, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9342338442802429, 'rewards/total_composite/std': 0.0913098156452179, 'reward': 0.9342338442802429, 'reward_std': 0.09130982309579849, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03732313588261604, 'sampling/sampling_logp_difference/max': 2.1083364486694336, 'sampling/importance_sampling_ratio/min': 0.12143982201814651, 'sampling/importance_sampling_ratio/mean': 1.0083178281784058, 'sampling/importance_sampling_ratio/max': 1.8755393028259277, 'entropy': 0.37529650889337063, 'clip_ratio/low_mean': 0.011014241958037019, 'clip_ratio/low_min': 0.011014241958037019, 'clip_ratio/high_mean': 0.029623867012560368, 'clip_ratio/high_max': 0.029623867012560368, 'clip_ratio/region_mean': 0.040638108970597386, 'reward_total_mean': 0.9342338442802429, 'reward_meter_mean': 0.961860179901123, 'reward_meter_std': 0.08823229372501373, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9342338442802429, 'reward_total_composite_std': 0.0913098156452179, 'epoch': 0.12}

 94%|█████████▍| 3109/3300 [4:45:17<25:15,  7.93s/it]INFO 04-12 03:22:08 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3110/3300 [4:45:23<23:10,  7.32s/it]2026-04-12 03:22:14,163 | INFO | train_grpo_train | metrics_logged mode=train step=3110

                                                     
{'loss': 0.0026, 'grad_norm': 2.6455962657928467, 'learning_rate': 5.787878787878789e-07, 'num_tokens': 7074959.0, 'completions/mean_length': 166.0, 'completions/min_length': 162.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.0, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9984124898910522, 'rewards/meter/std': 0.0009795344667509198, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984124898910522, 'rewards/total_composite/std': 0.0009795344667509198, 'reward': 0.9984124898910522, 'reward_std': 0.0009795371443033218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04640757292509079, 'sampling/sampling_logp_difference/max': 3.5209038257598877, 'sampling/importance_sampling_ratio/min': 0.02957269363105297, 'sampling/importance_sampling_ratio/mean': 1.0051828622817993, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3597927987575531, 'clip_ratio/low_mean': 0.004545454401522875, 'clip_ratio/low_min': 0.004545454401522875, 'clip_ratio/high_mean': 0.03839668887667358, 'clip_ratio/high_max': 0.03839668887667358, 'clip_ratio/region_mean': 0.042942143278196454, 'reward_total_mean': 0.9984124898910522, 'reward_meter_mean': 0.9984124898910522, 'reward_meter_std': 0.0009795344667509198, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984124898910522, 'reward_total_composite_std': 0.0009795344667509198, 'epoch': 0.12}

 94%|█████████▍| 3110/3300 [4:45:23<23:10,  7.32s/it]INFO 04-12 03:22:14 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3111/3300 [4:45:27<20:21,  6.46s/it]2026-04-12 03:22:18,627 | INFO | train_grpo_train | metrics_logged mode=train step=3111

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.757575757575758e-07, 'num_tokens': 7076575.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00021090851805638522, 'sampling/sampling_logp_difference/max': 0.0025518755428493023, 'sampling/importance_sampling_ratio/min': 0.9997379779815674, 'sampling/importance_sampling_ratio/mean': 1.0002080202102661, 'sampling/importance_sampling_ratio/max': 1.0025551319122314, 'entropy': 0.0017129705229308456, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.12}

 94%|█████████▍| 3111/3300 [4:45:27<20:21,  6.46s/it]INFO 04-12 03:22:18 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3112/3300 [4:45:35<21:22,  6.82s/it]2026-04-12 03:22:26,275 | INFO | train_grpo_train | metrics_logged mode=train step=3112

                                                     
{'loss': 0.0147, 'grad_norm': 2.5759596824645996, 'learning_rate': 5.727272727272728e-07, 'num_tokens': 7080735.0, 'completions/mean_length': 303.0, 'completions/min_length': 286.0, 'completions/max_length': 313.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 303.0, 'completions/min_terminated_length': 286.0, 'completions/max_terminated_length': 313.0, 'rewards/meter/mean': 0.9937278628349304, 'rewards/meter/std': 0.004818979650735855, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9937278628349304, 'rewards/total_composite/std': 0.004818979650735855, 'reward': 0.9937278628349304, 'reward_std': 0.004818981513381004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0645064041018486, 'sampling/sampling_logp_difference/max': 6.46595573425293, 'sampling/importance_sampling_ratio/min': 0.0015555039281025529, 'sampling/importance_sampling_ratio/mean': 1.0059781074523926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5345265790820122, 'clip_ratio/low_mean': 0.022525052540004253, 'clip_ratio/low_min': 0.022525052540004253, 'clip_ratio/high_mean': 0.026788415852934122, 'clip_ratio/high_max': 0.026788415852934122, 'clip_ratio/region_mean': 0.049313468392938375, 'reward_total_mean': 0.9937278628349304, 'reward_meter_mean': 0.9937278628349304, 'reward_meter_std': 0.004818979650735855, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9937278628349304, 'reward_total_composite_std': 0.004818979650735855, 'epoch': 0.12}

 94%|█████████▍| 3112/3300 [4:45:35<21:22,  6.82s/it]INFO 04-12 03:22:26 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3113/3300 [4:45:39<18:48,  6.04s/it]2026-04-12 03:22:30,484 | INFO | train_grpo_train | metrics_logged mode=train step=3113

                                                     
{'loss': -0.0003, 'grad_norm': 0.42850980162620544, 'learning_rate': 5.696969696969698e-07, 'num_tokens': 7082156.0, 'completions/mean_length': 40.625, 'completions/min_length': 39.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.998562216758728, 'rewards/meter/std': 0.0005881982506252825, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998562216758728, 'rewards/total_composite/std': 0.0005881982506252825, 'reward': 0.998562216758728, 'reward_std': 0.0005882052937522531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009273692034184933, 'sampling/sampling_logp_difference/max': 0.7357292175292969, 'sampling/importance_sampling_ratio/min': 0.47915592789649963, 'sampling/importance_sampling_ratio/mean': 1.0016939640045166, 'sampling/importance_sampling_ratio/max': 1.4733530282974243, 'entropy': 0.09460031241178513, 'clip_ratio/low_mean': 0.0062500000931322575, 'clip_ratio/low_min': 0.0062500000931322575, 'clip_ratio/high_mean': 0.006097560748457909, 'clip_ratio/high_max': 0.006097560748457909, 'clip_ratio/region_mean': 0.012347560841590166, 'reward_total_mean': 0.998562216758728, 'reward_meter_mean': 0.998562216758728, 'reward_meter_std': 0.0005881982506252825, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998562216758728, 'reward_total_composite_std': 0.0005881982506252825, 'epoch': 0.13}

 94%|█████████▍| 3113/3300 [4:45:39<18:48,  6.04s/it]INFO 04-12 03:22:30 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3114/3300 [4:45:44<17:45,  5.73s/it]2026-04-12 03:22:35,492 | INFO | train_grpo_train | metrics_logged mode=train step=3114

                                                     
{'loss': -0.0061, 'grad_norm': 1.817691683769226, 'learning_rate': 5.666666666666667e-07, 'num_tokens': 7084326.0, 'completions/mean_length': 97.25, 'completions/min_length': 95.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.25, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9978147745132446, 'rewards/meter/std': 0.00044735919800587, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9978147745132446, 'rewards/total_composite/std': 0.00044735919800587, 'reward': 0.9978147745132446, 'reward_std': 0.0004473614681046456, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013976364396512508, 'sampling/sampling_logp_difference/max': 0.9773850440979004, 'sampling/importance_sampling_ratio/min': 0.3762938380241394, 'sampling/importance_sampling_ratio/mean': 1.0041371583938599, 'sampling/importance_sampling_ratio/max': 1.3954778909683228, 'entropy': 0.10958225093781948, 'clip_ratio/low_mean': 0.0013157895300537348, 'clip_ratio/low_min': 0.0013157895300537348, 'clip_ratio/high_mean': 0.008981169550679624, 'clip_ratio/high_max': 0.008981169550679624, 'clip_ratio/region_mean': 0.010296959080733359, 'reward_total_mean': 0.9978147745132446, 'reward_meter_mean': 0.9978147745132446, 'reward_meter_std': 0.00044735919800587, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9978147745132446, 'reward_total_composite_std': 0.00044735919800587, 'epoch': 0.13}

 94%|█████████▍| 3114/3300 [4:45:44<17:45,  5.73s/it]INFO 04-12 03:22:35 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3115/3300 [4:45:48<16:36,  5.39s/it]2026-04-12 03:22:40,080 | INFO | train_grpo_train | metrics_logged mode=train step=3115

                                                     
{'loss': 0.0004, 'grad_norm': 0.015785139054059982, 'learning_rate': 5.636363636363638e-07, 'num_tokens': 7086062.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973392486572266, 'rewards/meter/std': 4.345127990745823e-07, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973392486572266, 'rewards/total_composite/std': 4.345127990745823e-07, 'reward': 0.9973392486572266, 'reward_std': 4.280403231859964e-07, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004269226919859648, 'sampling/sampling_logp_difference/max': 0.5515934228897095, 'sampling/importance_sampling_ratio/min': 0.5760312080383301, 'sampling/importance_sampling_ratio/mean': 0.9995636940002441, 'sampling/importance_sampling_ratio/max': 1.2580742835998535, 'entropy': 0.02565961377695203, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.9973392486572266, 'reward_meter_mean': 0.9973392486572266, 'reward_meter_std': 4.345127990745823e-07, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973392486572266, 'reward_total_composite_std': 4.345127990745823e-07, 'epoch': 0.13}

 94%|█████████▍| 3115/3300 [4:45:49<16:36,  5.39s/it]INFO 04-12 03:22:40 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3116/3300 [4:45:59<21:02,  6.86s/it]2026-04-12 03:22:50,387 | INFO | train_grpo_train | metrics_logged mode=train step=3116

                                                     
{'loss': 0.0232, 'grad_norm': 1.0452237129211426, 'learning_rate': 5.606060606060607e-07, 'num_tokens': 7091420.0, 'completions/mean_length': 490.75, 'completions/min_length': 461.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 487.71429443359375, 'completions/min_terminated_length': 461.0, 'completions/max_terminated_length': 509.0, 'rewards/meter/mean': 0.9985963106155396, 'rewards/meter/std': 0.0007045524544082582, 'rewards/count_adherence/mean': 0.8515625, 'rewards/count_adherence/std': 0.032346826046705246, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9714829921722412, 'rewards/repeat_penalty/std': 0.027092305943369865, 'rewards/total_composite/mean': 0.8266471028327942, 'rewards/total_composite/std': 0.049881432205438614, 'reward': 0.8266471028327942, 'reward_std': 0.049881432205438614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034601159393787384, 'sampling/sampling_logp_difference/max': 1.3683700561523438, 'sampling/importance_sampling_ratio/min': 0.2545214593410492, 'sampling/importance_sampling_ratio/mean': 1.0090035200119019, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3404971808195114, 'clip_ratio/low_mean': 0.0044667941983789206, 'clip_ratio/low_min': 0.0044667941983789206, 'clip_ratio/high_mean': 0.013212713180109859, 'clip_ratio/high_max': 0.013212713180109859, 'clip_ratio/region_mean': 0.01767950737848878, 'reward_total_mean': 0.8266471028327942, 'reward_meter_mean': 0.9985963106155396, 'reward_meter_std': 0.0007045524544082582, 'reward_count_adherence_mean': 0.8515625, 'reward_count_adherence_std': 0.032346826046705246, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9714829921722412, 'reward_repeat_penalty_std': 0.027092305943369865, 'reward_total_composite_mean': 0.8266471028327942, 'reward_total_composite_std': 0.049881432205438614, 'epoch': 0.13}

 94%|█████████▍| 3116/3300 [4:45:59<21:02,  6.86s/it]INFO 04-12 03:22:50 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3117/3300 [4:46:03<18:36,  6.10s/it]2026-04-12 03:22:54,727 | INFO | train_grpo_train | metrics_logged mode=train step=3117

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.575757575757576e-07, 'num_tokens': 7092924.0, 'completions/mean_length': 34.0, 'completions/min_length': 34.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9923644065856934, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9923644065856934, 'rewards/total_composite/std': 0.0, 'reward': 0.9923644065856934, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.004286302253603935, 'sampling/sampling_logp_difference/max': 0.12234780192375183, 'sampling/importance_sampling_ratio/min': 0.9824780821800232, 'sampling/importance_sampling_ratio/mean': 1.0040427446365356, 'sampling/importance_sampling_ratio/max': 1.1301469802856445, 'entropy': 0.03289771918207407, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9923644065856934, 'reward_meter_mean': 0.9923644065856934, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9923644065856934, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 94%|█████████▍| 3117/3300 [4:46:03<18:36,  6.10s/it]INFO 04-12 03:22:54 [block_pool.py:378] Successfully reset prefix cache

 94%|█████████▍| 3118/3300 [4:46:11<20:29,  6.75s/it]2026-04-12 03:23:02,984 | INFO | train_grpo_train | metrics_logged mode=train step=3118

                                                     
{'loss': 0.0017, 'grad_norm': 1.1702804565429688, 'learning_rate': 5.545454545454547e-07, 'num_tokens': 7097281.0, 'completions/mean_length': 354.625, 'completions/min_length': 350.0, 'completions/max_length': 359.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 354.625, 'completions/min_terminated_length': 350.0, 'completions/max_terminated_length': 359.0, 'rewards/meter/mean': 0.9987514019012451, 'rewards/meter/std': 0.00020078917441423982, 'rewards/count_adherence/mean': 0.9090909361839294, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9671052694320679, 'rewards/repeat_penalty/std': 0.027239417657256126, 'rewards/total_composite/mean': 0.8780869841575623, 'rewards/total_composite/std': 0.024653173983097076, 'reward': 0.8780869841575623, 'reward_std': 0.02465316839516163, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029794396832585335, 'sampling/sampling_logp_difference/max': 1.3524599075317383, 'sampling/importance_sampling_ratio/min': 0.2586033344268799, 'sampling/importance_sampling_ratio/mean': 1.0095971822738647, 'sampling/importance_sampling_ratio/max': 1.7627272605895996, 'entropy': 0.3398951441049576, 'clip_ratio/low_mean': 0.008492077584378421, 'clip_ratio/low_min': 0.008492077584378421, 'clip_ratio/high_mean': 0.008788066916167736, 'clip_ratio/high_max': 0.008788066916167736, 'clip_ratio/region_mean': 0.017280144500546157, 'reward_total_mean': 0.8780869841575623, 'reward_meter_mean': 0.9987514019012451, 'reward_meter_std': 0.00020078917441423982, 'reward_count_adherence_mean': 0.9090909361839294, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9671052694320679, 'reward_repeat_penalty_std': 0.027239417657256126, 'reward_total_composite_mean': 0.8780869841575623, 'reward_total_composite_std': 0.024653173983097076, 'epoch': 0.13}

 94%|█████████▍| 3118/3300 [4:46:11<20:29,  6.75s/it]INFO 04-12 03:23:03 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3119/3300 [4:46:16<18:31,  6.14s/it]2026-04-12 03:23:07,718 | INFO | train_grpo_train | metrics_logged mode=train step=3119

                                                     
{'loss': -0.0005, 'grad_norm': 5.017796516418457, 'learning_rate': 5.515151515151516e-07, 'num_tokens': 7098968.0, 'completions/mean_length': 60.875, 'completions/min_length': 60.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9972956776618958, 'rewards/meter/std': 0.0001135577549575828, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9972956776618958, 'rewards/total_composite/std': 0.0001135577549575828, 'reward': 0.9972956776618958, 'reward_std': 0.0001135657585109584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005037278868257999, 'sampling/sampling_logp_difference/max': 1.2046854496002197, 'sampling/importance_sampling_ratio/min': 0.2997862994670868, 'sampling/importance_sampling_ratio/mean': 0.9997225999832153, 'sampling/importance_sampling_ratio/max': 1.1826646327972412, 'entropy': 0.02145107788965106, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.004098360426723957, 'clip_ratio/high_max': 0.004098360426723957, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.9972956776618958, 'reward_meter_mean': 0.9972956776618958, 'reward_meter_std': 0.0001135577549575828, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9972956776618958, 'reward_total_composite_std': 0.0001135577549575828, 'epoch': 0.13}

 95%|█████████▍| 3119/3300 [4:46:16<18:31,  6.14s/it]INFO 04-12 03:23:07 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3120/3300 [4:46:21<17:04,  5.69s/it]2026-04-12 03:23:12,338 | INFO | train_grpo_train | metrics_logged mode=train step=3120

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.484848484848485e-07, 'num_tokens': 7100936.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00018837934476323426, 'sampling/sampling_logp_difference/max': 0.0016895392909646034, 'sampling/importance_sampling_ratio/min': 0.998882532119751, 'sampling/importance_sampling_ratio/mean': 1.0001810789108276, 'sampling/importance_sampling_ratio/max': 1.0016909837722778, 'entropy': 0.0014699531602673233, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 95%|█████████▍| 3120/3300 [4:46:21<17:04,  5.69s/it]INFO 04-12 03:23:12 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3121/3300 [4:46:30<20:05,  6.73s/it]2026-04-12 03:23:21,514 | INFO | train_grpo_train | metrics_logged mode=train step=3121

                                                     
{'loss': 0.0239, 'grad_norm': 1.84979248046875, 'learning_rate': 5.454545454545455e-07, 'num_tokens': 7105325.0, 'completions/mean_length': 354.625, 'completions/min_length': 342.0, 'completions/max_length': 394.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 354.625, 'completions/min_terminated_length': 342.0, 'completions/max_terminated_length': 394.0, 'rewards/meter/mean': 0.9991722106933594, 'rewards/meter/std': 0.0001384944043820724, 'rewards/count_adherence/mean': 0.9124999642372131, 'rewards/count_adherence/std': 0.0353553481400013, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.6744744777679443, 'rewards/total_composite/std': 0.41629472374916077, 'reward': 0.6744744777679443, 'reward_std': 0.4162946939468384, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04996606335043907, 'sampling/sampling_logp_difference/max': 2.3657424449920654, 'sampling/importance_sampling_ratio/min': 0.0938795730471611, 'sampling/importance_sampling_ratio/mean': 1.0077815055847168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47248122096061707, 'clip_ratio/low_mean': 0.007510315626859665, 'clip_ratio/low_min': 0.007510315626859665, 'clip_ratio/high_mean': 0.02431198745034635, 'clip_ratio/high_max': 0.02431198745034635, 'clip_ratio/region_mean': 0.031822303077206016, 'reward_total_mean': 0.6744744777679443, 'reward_meter_mean': 0.9991722106933594, 'reward_meter_std': 0.0001384944043820724, 'reward_count_adherence_mean': 0.9124999642372131, 'reward_count_adherence_std': 0.0353553481400013, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.6744744777679443, 'reward_total_composite_std': 0.41629472374916077, 'epoch': 0.13}

 95%|█████████▍| 3121/3300 [4:46:30<20:05,  6.73s/it]INFO 04-12 03:23:21 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3122/3300 [4:46:35<18:49,  6.35s/it]2026-04-12 03:23:26,950 | INFO | train_grpo_train | metrics_logged mode=train step=3122

                                                     
{'loss': 0.0001, 'grad_norm': 0.5849205851554871, 'learning_rate': 5.424242424242425e-07, 'num_tokens': 7107085.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973349571228027, 'rewards/meter/std': 1.3741479051532224e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973349571228027, 'rewards/total_composite/std': 1.3741479051532224e-05, 'reward': 0.9973349571228027, 'reward_std': 1.3751973710895982e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005648355465382338, 'sampling/sampling_logp_difference/max': 0.814976692199707, 'sampling/importance_sampling_ratio/min': 0.44264963269233704, 'sampling/importance_sampling_ratio/mean': 0.9968710541725159, 'sampling/importance_sampling_ratio/max': 1.0885894298553467, 'entropy': 0.019400187535211444, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006147540640085936, 'clip_ratio/high_max': 0.006147540640085936, 'clip_ratio/region_mean': 0.006147540640085936, 'reward_total_mean': 0.9973349571228027, 'reward_meter_mean': 0.9973349571228027, 'reward_meter_std': 1.3741479051532224e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973349571228027, 'reward_total_composite_std': 1.3741479051532224e-05, 'epoch': 0.13}

 95%|█████████▍| 3122/3300 [4:46:35<18:49,  6.35s/it]INFO 04-12 03:23:27 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3123/3300 [4:46:41<17:43,  6.01s/it]2026-04-12 03:23:32,251 | INFO | train_grpo_train | metrics_logged mode=train step=3123

                                                     
{'loss': 0.0044, 'grad_norm': 2.6911187171936035, 'learning_rate': 5.393939393939395e-07, 'num_tokens': 7109065.0, 'completions/mean_length': 100.5, 'completions/min_length': 99.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.5, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9990350008010864, 'rewards/meter/std': 0.0004222920979373157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990350008010864, 'rewards/total_composite/std': 0.0004222920979373157, 'reward': 0.9990350008010864, 'reward_std': 0.0004223074938636273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025933468714356422, 'sampling/sampling_logp_difference/max': 0.8458013534545898, 'sampling/importance_sampling_ratio/min': 0.5429390072822571, 'sampling/importance_sampling_ratio/mean': 1.0030157566070557, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1784425787627697, 'clip_ratio/low_mean': 0.007426470518112183, 'clip_ratio/low_min': 0.007426470518112183, 'clip_ratio/high_mean': 0.02493911876808852, 'clip_ratio/high_max': 0.02493911876808852, 'clip_ratio/region_mean': 0.0323655892862007, 'reward_total_mean': 0.9990350008010864, 'reward_meter_mean': 0.9990350008010864, 'reward_meter_std': 0.0004222920979373157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990350008010864, 'reward_total_composite_std': 0.0004222920979373157, 'epoch': 0.13}

 95%|█████████▍| 3123/3300 [4:46:41<17:43,  6.01s/it]INFO 04-12 03:23:32 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3124/3300 [4:46:51<21:27,  7.32s/it]2026-04-12 03:23:42,525 | INFO | train_grpo_train | metrics_logged mode=train step=3124

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.363636363636364e-07, 'num_tokens': 7110721.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9872919917106628, 'rewards/meter/std': 0.03354829549789429, 'rewards/count_adherence/mean': 0.7361111044883728, 'rewards/count_adherence/std': 0.0257172379642725, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9615384340286255, 'rewards/repeat_penalty/std': 0.029074188321828842, 'rewards/total_composite/mean': 0.6985058784484863, 'rewards/total_composite/std': 0.03456052392721176, 'reward': 0.6985058784484863, 'reward_std': 0.03456050530076027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.6985058784484863, 'reward_meter_mean': 0.9872919917106628, 'reward_meter_std': 0.03354829549789429, 'reward_count_adherence_mean': 0.7361111044883728, 'reward_count_adherence_std': 0.0257172379642725, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9615384340286255, 'reward_repeat_penalty_std': 0.029074188321828842, 'reward_total_composite_mean': 0.6985058784484863, 'reward_total_composite_std': 0.03456052392721176, 'epoch': 0.13}

 95%|█████████▍| 3124/3300 [4:46:51<21:27,  7.32s/it]INFO 04-12 03:23:42 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3125/3300 [4:46:56<18:59,  6.51s/it]2026-04-12 03:23:47,170 | INFO | train_grpo_train | metrics_logged mode=train step=3125

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.333333333333335e-07, 'num_tokens': 7112433.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00018432487559039146, 'sampling/sampling_logp_difference/max': 0.0014849056024104357, 'sampling/importance_sampling_ratio/min': 0.9985162019729614, 'sampling/importance_sampling_ratio/mean': 1.0001718997955322, 'sampling/importance_sampling_ratio/max': 1.0014132261276245, 'entropy': 0.0014961521519580856, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 95%|█████████▍| 3125/3300 [4:46:56<18:59,  6.51s/it]INFO 04-12 03:23:47 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3126/3300 [4:47:05<21:04,  7.27s/it]2026-04-12 03:23:56,223 | INFO | train_grpo_train | metrics_logged mode=train step=3126

                                                     
{'loss': -0.0003, 'grad_norm': 1.4976996183395386, 'learning_rate': 5.303030303030304e-07, 'num_tokens': 7117396.0, 'completions/mean_length': 418.375, 'completions/min_length': 411.0, 'completions/max_length': 423.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 418.375, 'completions/min_terminated_length': 411.0, 'completions/max_terminated_length': 423.0, 'rewards/meter/mean': 0.9984115362167358, 'rewards/meter/std': 0.0010889971163123846, 'rewards/count_adherence/mean': 0.8365384340286255, 'rewards/count_adherence/std': 0.027196412906050682, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.988095223903656, 'rewards/repeat_penalty/std': 0.033671751618385315, 'rewards/total_composite/mean': 0.8251492977142334, 'rewards/total_composite/std': 0.03634733706712723, 'reward': 0.8251492977142334, 'reward_std': 0.036347318440675735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05572379007935524, 'sampling/sampling_logp_difference/max': 1.5548510551452637, 'sampling/importance_sampling_ratio/min': 0.21122083067893982, 'sampling/importance_sampling_ratio/mean': 1.0117367506027222, 'sampling/importance_sampling_ratio/max': 1.753551959991455, 'entropy': 0.51447868719697, 'clip_ratio/low_mean': 0.007205356378108263, 'clip_ratio/low_min': 0.007205356378108263, 'clip_ratio/high_mean': 0.026507085654884577, 'clip_ratio/high_max': 0.026507085654884577, 'clip_ratio/region_mean': 0.03371244203299284, 'reward_total_mean': 0.8251492977142334, 'reward_meter_mean': 0.9984115362167358, 'reward_meter_std': 0.0010889971163123846, 'reward_count_adherence_mean': 0.8365384340286255, 'reward_count_adherence_std': 0.027196412906050682, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.988095223903656, 'reward_repeat_penalty_std': 0.033671751618385315, 'reward_total_composite_mean': 0.8251492977142334, 'reward_total_composite_std': 0.03634733706712723, 'epoch': 0.13}

 95%|█████████▍| 3126/3300 [4:47:05<21:04,  7.27s/it]INFO 04-12 03:23:56 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3127/3300 [4:47:15<23:24,  8.12s/it]2026-04-12 03:24:06,293 | INFO | train_grpo_train | metrics_logged mode=train step=3127

                                                     
{'loss': -0.0212, 'grad_norm': 1.312803030014038, 'learning_rate': 5.272727272727273e-07, 'num_tokens': 7122682.0, 'completions/mean_length': 429.75, 'completions/min_length': 418.0, 'completions/max_length': 468.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 429.75, 'completions/min_terminated_length': 418.0, 'completions/max_terminated_length': 468.0, 'rewards/meter/mean': 0.9992569088935852, 'rewards/meter/std': 7.736113184364513e-05, 'rewards/count_adherence/mean': 0.7946428060531616, 'rewards/count_adherence/std': 0.025253823027014732, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9824134111404419, 'rewards/repeat_penalty/std': 0.024280980229377747, 'rewards/total_composite/mean': 0.7802447080612183, 'rewards/total_composite/std': 0.035804178565740585, 'reward': 0.7802447080612183, 'reward_std': 0.03580417484045029, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05896211043000221, 'sampling/sampling_logp_difference/max': 10.541126251220703, 'sampling/importance_sampling_ratio/min': 2.6426951080793515e-05, 'sampling/importance_sampling_ratio/mean': 1.0106730461120605, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49032482132315636, 'clip_ratio/low_mean': 0.00941323209553957, 'clip_ratio/low_min': 0.00941323209553957, 'clip_ratio/high_mean': 0.022230799309909344, 'clip_ratio/high_max': 0.022230799309909344, 'clip_ratio/region_mean': 0.031644031405448914, 'reward_total_mean': 0.7802447080612183, 'reward_meter_mean': 0.9992569088935852, 'reward_meter_std': 7.736113184364513e-05, 'reward_count_adherence_mean': 0.7946428060531616, 'reward_count_adherence_std': 0.025253823027014732, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9824134111404419, 'reward_repeat_penalty_std': 0.024280980229377747, 'reward_total_composite_mean': 0.7802447080612183, 'reward_total_composite_std': 0.035804178565740585, 'epoch': 0.13}

 95%|█████████▍| 3127/3300 [4:47:15<23:24,  8.12s/it]INFO 04-12 03:24:06 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3128/3300 [4:47:21<21:45,  7.59s/it]2026-04-12 03:24:12,668 | INFO | train_grpo_train | metrics_logged mode=train step=3128

                                                     
{'loss': 0.0112, 'grad_norm': 1.9481908082962036, 'learning_rate': 5.242424242424243e-07, 'num_tokens': 7125754.0, 'completions/mean_length': 186.0, 'completions/min_length': 182.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.0, 'completions/min_terminated_length': 182.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9975753426551819, 'rewards/meter/std': 0.000392713351175189, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9318182468414307, 'rewards/repeat_penalty/std': 0.04208271950483322, 'rewards/total_composite/mean': 0.9295545816421509, 'rewards/total_composite/std': 0.041869208216667175, 'reward': 0.9295545816421509, 'reward_std': 0.04186920449137688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031066881492733955, 'sampling/sampling_logp_difference/max': 1.1641087532043457, 'sampling/importance_sampling_ratio/min': 0.31220078468322754, 'sampling/importance_sampling_ratio/mean': 1.00235915184021, 'sampling/importance_sampling_ratio/max': 1.669154405593872, 'entropy': 0.2618657611310482, 'clip_ratio/low_mean': 0.010725616884883493, 'clip_ratio/low_min': 0.010725616884883493, 'clip_ratio/high_mean': 0.010851842351257801, 'clip_ratio/high_max': 0.010851842351257801, 'clip_ratio/region_mean': 0.021577459236141294, 'reward_total_mean': 0.9295545816421509, 'reward_meter_mean': 0.9975753426551819, 'reward_meter_std': 0.000392713351175189, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9318182468414307, 'reward_repeat_penalty_std': 0.04208271950483322, 'reward_total_composite_mean': 0.9295545816421509, 'reward_total_composite_std': 0.041869208216667175, 'epoch': 0.13}

 95%|█████████▍| 3128/3300 [4:47:21<21:45,  7.59s/it]INFO 04-12 03:24:12 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3129/3300 [4:47:27<19:53,  6.98s/it]2026-04-12 03:24:18,217 | INFO | train_grpo_train | metrics_logged mode=train step=3129

                                                     
{'loss': 0.0067, 'grad_norm': 2.561624526977539, 'learning_rate': 5.212121212121213e-07, 'num_tokens': 7128173.0, 'completions/mean_length': 127.375, 'completions/min_length': 123.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.375, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.9965527057647705, 'rewards/meter/std': 0.0038294827099889517, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9787326455116272, 'rewards/total_composite/std': 0.04999339208006859, 'reward': 0.9787326455116272, 'reward_std': 0.04999341070652008, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017186593264341354, 'sampling/sampling_logp_difference/max': 1.2021827697753906, 'sampling/importance_sampling_ratio/min': 0.30053746700286865, 'sampling/importance_sampling_ratio/mean': 1.0033414363861084, 'sampling/importance_sampling_ratio/max': 1.5255719423294067, 'entropy': 0.14594370871782303, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.011735557112842798, 'clip_ratio/high_max': 0.011735557112842798, 'clip_ratio/region_mean': 0.011735557112842798, 'reward_total_mean': 0.9787326455116272, 'reward_meter_mean': 0.9965527057647705, 'reward_meter_std': 0.0038294827099889517, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9787326455116272, 'reward_total_composite_std': 0.04999339208006859, 'epoch': 0.13}

 95%|█████████▍| 3129/3300 [4:47:27<19:53,  6.98s/it]INFO 04-12 03:24:18 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3130/3300 [4:47:31<17:58,  6.35s/it]2026-04-12 03:24:23,067 | INFO | train_grpo_train | metrics_logged mode=train step=3130

                                                     
{'loss': 0.0179, 'grad_norm': 3.1293835639953613, 'learning_rate': 5.181818181818182e-07, 'num_tokens': 7129971.0, 'completions/mean_length': 65.75, 'completions/min_length': 64.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9895097613334656, 'rewards/meter/std': 0.006478929426521063, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9895097613334656, 'rewards/total_composite/std': 0.006478929426521063, 'reward': 0.9895097613334656, 'reward_std': 0.006478935480117798, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02998344786465168, 'sampling/sampling_logp_difference/max': 0.9678926467895508, 'sampling/importance_sampling_ratio/min': 0.37988272309303284, 'sampling/importance_sampling_ratio/mean': 1.008150577545166, 'sampling/importance_sampling_ratio/max': 1.572881817817688, 'entropy': 0.23753815703094006, 'clip_ratio/low_mean': 0.020387701224535704, 'clip_ratio/low_min': 0.020387701224535704, 'clip_ratio/high_mean': 0.007722355774603784, 'clip_ratio/high_max': 0.007722355774603784, 'clip_ratio/region_mean': 0.028110056999139488, 'reward_total_mean': 0.9895097613334656, 'reward_meter_mean': 0.9895097613334656, 'reward_meter_std': 0.006478929426521063, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9895097613334656, 'reward_total_composite_std': 0.006478929426521063, 'epoch': 0.13}

 95%|█████████▍| 3130/3300 [4:47:32<17:58,  6.35s/it]INFO 04-12 03:24:23 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3131/3300 [4:47:36<16:30,  5.86s/it]2026-04-12 03:24:27,805 | INFO | train_grpo_train | metrics_logged mode=train step=3131

                                                     
{'loss': 0.0024, 'grad_norm': 3.762568473815918, 'learning_rate': 5.151515151515152e-07, 'num_tokens': 7131734.0, 'completions/mean_length': 65.375, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9759130477905273, 'rewards/meter/std': 0.04954683780670166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9759130477905273, 'rewards/total_composite/std': 0.04954683780670166, 'reward': 0.9759130477905273, 'reward_std': 0.04954684525728226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020955001935362816, 'sampling/sampling_logp_difference/max': 1.4279863834381104, 'sampling/importance_sampling_ratio/min': 0.32048511505126953, 'sampling/importance_sampling_ratio/mean': 1.0054841041564941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10212477203458548, 'clip_ratio/low_mean': 0.005769230891019106, 'clip_ratio/low_min': 0.005769230891019106, 'clip_ratio/high_mean': 0.009557109675370157, 'clip_ratio/high_max': 0.009557109675370157, 'clip_ratio/region_mean': 0.015326340566389263, 'reward_total_mean': 0.9759130477905273, 'reward_meter_mean': 0.9759130477905273, 'reward_meter_std': 0.04954683780670166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9759130477905273, 'reward_total_composite_std': 0.04954683780670166, 'epoch': 0.13}

 95%|█████████▍| 3131/3300 [4:47:36<16:30,  5.86s/it]INFO 04-12 03:24:28 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3132/3300 [4:47:41<15:33,  5.56s/it]2026-04-12 03:24:32,669 | INFO | train_grpo_train | metrics_logged mode=train step=3132

                                                     
{'loss': 0.0002, 'grad_norm': 0.13032668828964233, 'learning_rate': 5.121212121212121e-07, 'num_tokens': 7133676.0, 'completions/mean_length': 66.75, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981508255004883, 'rewards/meter/std': 6.344345820252784e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981508255004883, 'rewards/total_composite/std': 6.344345820252784e-06, 'reward': 0.9981508255004883, 'reward_std': 6.347765065584099e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007171308156102896, 'sampling/sampling_logp_difference/max': 0.5747629404067993, 'sampling/importance_sampling_ratio/min': 0.5628383159637451, 'sampling/importance_sampling_ratio/mean': 1.0008666515350342, 'sampling/importance_sampling_ratio/max': 1.516728162765503, 'entropy': 0.05024628387764096, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9981508255004883, 'reward_meter_mean': 0.9981508255004883, 'reward_meter_std': 6.344345820252784e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981508255004883, 'reward_total_composite_std': 6.344345820252784e-06, 'epoch': 0.13}

 95%|█████████▍| 3132/3300 [4:47:41<15:33,  5.56s/it]INFO 04-12 03:24:32 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3133/3300 [4:47:46<15:11,  5.46s/it]2026-04-12 03:24:37,877 | INFO | train_grpo_train | metrics_logged mode=train step=3133

                                                     
{'loss': 0.0045, 'grad_norm': 2.7664103507995605, 'learning_rate': 5.090909090909092e-07, 'num_tokens': 7135896.0, 'completions/mean_length': 100.5, 'completions/min_length': 99.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.5, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.999111533164978, 'rewards/meter/std': 0.00020586077880579978, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999111533164978, 'rewards/total_composite/std': 0.00020586077880579978, 'reward': 0.999111533164978, 'reward_std': 0.00020586182654369622, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025510529056191444, 'sampling/sampling_logp_difference/max': 1.5977736711502075, 'sampling/importance_sampling_ratio/min': 0.20234650373458862, 'sampling/importance_sampling_ratio/mean': 1.0036976337432861, 'sampling/importance_sampling_ratio/max': 1.8379536867141724, 'entropy': 0.17687924578785896, 'clip_ratio/low_mean': 0.00993861141614616, 'clip_ratio/low_min': 0.00993861141614616, 'clip_ratio/high_mean': 0.006225490127690136, 'clip_ratio/high_max': 0.006225490127690136, 'clip_ratio/region_mean': 0.016164101543836296, 'reward_total_mean': 0.999111533164978, 'reward_meter_mean': 0.999111533164978, 'reward_meter_std': 0.00020586077880579978, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999111533164978, 'reward_total_composite_std': 0.00020586077880579978, 'epoch': 0.13}

 95%|█████████▍| 3133/3300 [4:47:46<15:11,  5.46s/it]INFO 04-12 03:24:38 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▍| 3134/3300 [4:47:51<14:49,  5.36s/it]2026-04-12 03:24:43,019 | INFO | train_grpo_train | metrics_logged mode=train step=3134

                                                     
{'loss': 0.0178, 'grad_norm': 5.516884803771973, 'learning_rate': 5.060606060606061e-07, 'num_tokens': 7138289.0, 'completions/mean_length': 110.125, 'completions/min_length': 107.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.125, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9559026956558228, 'rewards/meter/std': 0.015433711931109428, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.9024706482887268, 'rewards/total_composite/std': 0.052703484892845154, 'reward': 0.9024706482887268, 'reward_std': 0.05270349606871605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05570008605718613, 'sampling/sampling_logp_difference/max': 2.0624022483825684, 'sampling/importance_sampling_ratio/min': 0.12714816629886627, 'sampling/importance_sampling_ratio/mean': 1.0074241161346436, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2725033536553383, 'clip_ratio/low_mean': 0.01586226187646389, 'clip_ratio/low_min': 0.01586226187646389, 'clip_ratio/high_mean': 0.020418287720531225, 'clip_ratio/high_max': 0.020418287720531225, 'clip_ratio/region_mean': 0.036280549596995115, 'reward_total_mean': 0.9024706482887268, 'reward_meter_mean': 0.9559026956558228, 'reward_meter_std': 0.015433711931109428, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.9024706482887268, 'reward_total_composite_std': 0.052703484892845154, 'epoch': 0.13}

 95%|█████████▍| 3134/3300 [4:47:52<14:49,  5.36s/it]INFO 04-12 03:24:43 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3135/3300 [4:47:56<14:30,  5.28s/it]2026-04-12 03:24:48,093 | INFO | train_grpo_train | metrics_logged mode=train step=3135

                                                     
{'loss': 0.0022, 'grad_norm': 1.680987000465393, 'learning_rate': 5.03030303030303e-07, 'num_tokens': 7140487.0, 'completions/mean_length': 100.75, 'completions/min_length': 100.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.75, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9992322325706482, 'rewards/meter/std': 0.00010666978050721809, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992322325706482, 'rewards/total_composite/std': 0.00010666978050721809, 'reward': 0.9992322325706482, 'reward_std': 0.00010666977323126048, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029712356626987457, 'sampling/sampling_logp_difference/max': 1.3842487335205078, 'sampling/importance_sampling_ratio/min': 0.25051194429397583, 'sampling/importance_sampling_ratio/mean': 1.006977915763855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16831262316554785, 'clip_ratio/low_mean': 0.011116024106740952, 'clip_ratio/low_min': 0.011116024106740952, 'clip_ratio/high_mean': 0.006212871172465384, 'clip_ratio/high_max': 0.006212871172465384, 'clip_ratio/region_mean': 0.017328895279206336, 'reward_total_mean': 0.9992322325706482, 'reward_meter_mean': 0.9992322325706482, 'reward_meter_std': 0.00010666978050721809, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992322325706482, 'reward_total_composite_std': 0.00010666978050721809, 'epoch': 0.13}

 95%|█████████▌| 3135/3300 [4:47:57<14:30,  5.28s/it]INFO 04-12 03:24:48 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3136/3300 [4:48:01<13:54,  5.09s/it]2026-04-12 03:24:52,756 | INFO | train_grpo_train | metrics_logged mode=train step=3136

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 5.000000000000001e-07, 'num_tokens': 7142015.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 7.456184539478272e-05, 'sampling/sampling_logp_difference/max': 0.0015591848641633987, 'sampling/importance_sampling_ratio/min': 0.999688982963562, 'sampling/importance_sampling_ratio/mean': 1.0000699758529663, 'sampling/importance_sampling_ratio/max': 1.0015604496002197, 'entropy': 0.0006697150638501626, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 95%|█████████▌| 3136/3300 [4:48:01<13:54,  5.09s/it]INFO 04-12 03:24:53 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3137/3300 [4:48:06<13:41,  5.04s/it]2026-04-12 03:24:57,681 | INFO | train_grpo_train | metrics_logged mode=train step=3137

                                                     
{'loss': 0.0016, 'grad_norm': 1.265021800994873, 'learning_rate': 4.96969696969697e-07, 'num_tokens': 7143751.0, 'completions/mean_length': 68.0, 'completions/min_length': 68.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9994433522224426, 'rewards/meter/std': 8.680017344886437e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994433522224426, 'rewards/total_composite/std': 8.680017344886437e-05, 'reward': 0.9994433522224426, 'reward_std': 8.68034694576636e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01834726519882679, 'sampling/sampling_logp_difference/max': 1.1407603025436401, 'sampling/importance_sampling_ratio/min': 0.4000520408153534, 'sampling/importance_sampling_ratio/mean': 1.0005712509155273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.058580921962857246, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.011029411805793643, 'clip_ratio/high_max': 0.011029411805793643, 'clip_ratio/region_mean': 0.014705882407724857, 'reward_total_mean': 0.9994433522224426, 'reward_meter_mean': 0.9994433522224426, 'reward_meter_std': 8.680017344886437e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994433522224426, 'reward_total_composite_std': 8.680017344886437e-05, 'epoch': 0.13}

 95%|█████████▌| 3137/3300 [4:48:06<13:41,  5.04s/it]INFO 04-12 03:24:57 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3138/3300 [4:48:10<13:01,  4.82s/it]2026-04-12 03:25:01,999 | INFO | train_grpo_train | metrics_logged mode=train step=3138

                                                     
{'loss': -0.0077, 'grad_norm': 2.8887453079223633, 'learning_rate': 4.93939393939394e-07, 'num_tokens': 7145214.0, 'completions/mean_length': 40.875, 'completions/min_length': 40.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9984819889068604, 'rewards/meter/std': 0.0004429934488143772, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9984819889068604, 'rewards/total_composite/std': 0.0004429934488143772, 'reward': 0.9984819889068604, 'reward_std': 0.00044299106230027974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012076342478394508, 'sampling/sampling_logp_difference/max': 0.41540956497192383, 'sampling/importance_sampling_ratio/min': 0.6600699424743652, 'sampling/importance_sampling_ratio/mean': 1.0034947395324707, 'sampling/importance_sampling_ratio/max': 1.4185824394226074, 'entropy': 0.12436437420547009, 'clip_ratio/low_mean': 0.0030487803742289543, 'clip_ratio/low_min': 0.0030487803742289543, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0030487803742289543, 'reward_total_mean': 0.9984819889068604, 'reward_meter_mean': 0.9984819889068604, 'reward_meter_std': 0.0004429934488143772, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9984819889068604, 'reward_total_composite_std': 0.0004429934488143772, 'epoch': 0.13}

 95%|█████████▌| 3138/3300 [4:48:10<13:01,  4.82s/it]INFO 04-12 03:25:02 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3139/3300 [4:48:15<12:39,  4.72s/it]2026-04-12 03:25:06,447 | INFO | train_grpo_train | metrics_logged mode=train step=3139

                                                     
{'loss': -0.0001, 'grad_norm': 0.1556718647480011, 'learning_rate': 4.909090909090909e-07, 'num_tokens': 7146949.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981504678726196, 'rewards/meter/std': 9.697491805127356e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981504678726196, 'rewards/total_composite/std': 9.697491805127356e-06, 'reward': 0.9981504678726196, 'reward_std': 9.708711331768427e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007338312920182943, 'sampling/sampling_logp_difference/max': 0.5803370475769043, 'sampling/importance_sampling_ratio/min': 0.5597096681594849, 'sampling/importance_sampling_ratio/mean': 1.0014523267745972, 'sampling/importance_sampling_ratio/max': 1.3643027544021606, 'entropy': 0.04900990845635533, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9981504678726196, 'reward_meter_mean': 0.9981504678726196, 'reward_meter_std': 9.697491805127356e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981504678726196, 'reward_total_composite_std': 9.697491805127356e-06, 'epoch': 0.13}

 95%|█████████▌| 3139/3300 [4:48:15<12:39,  4.72s/it]INFO 04-12 03:25:06 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3140/3300 [4:48:23<14:59,  5.62s/it]2026-04-12 03:25:14,182 | INFO | train_grpo_train | metrics_logged mode=train step=3140

                                                     
{'loss': -0.0074, 'grad_norm': 2.1409451961517334, 'learning_rate': 4.878787878787879e-07, 'num_tokens': 7151106.0, 'completions/mean_length': 311.625, 'completions/min_length': 305.0, 'completions/max_length': 322.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 311.625, 'completions/min_terminated_length': 305.0, 'completions/max_terminated_length': 322.0, 'rewards/meter/mean': 0.9965137243270874, 'rewards/meter/std': 0.00763244554400444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9965137243270874, 'rewards/total_composite/std': 0.00763244554400444, 'reward': 0.9965137243270874, 'reward_std': 0.0076324427500367165, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0488644503057003, 'sampling/sampling_logp_difference/max': 1.3517093658447266, 'sampling/importance_sampling_ratio/min': 0.2587974965572357, 'sampling/importance_sampling_ratio/mean': 1.0111055374145508, 'sampling/importance_sampling_ratio/max': 1.9414808750152588, 'entropy': 0.447547797113657, 'clip_ratio/low_mean': 0.005737704690545797, 'clip_ratio/low_min': 0.005737704690545797, 'clip_ratio/high_mean': 0.02794860501307994, 'clip_ratio/high_max': 0.02794860501307994, 'clip_ratio/region_mean': 0.03368630970362574, 'reward_total_mean': 0.9965137243270874, 'reward_meter_mean': 0.9965137243270874, 'reward_meter_std': 0.00763244554400444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9965137243270874, 'reward_total_composite_std': 0.00763244554400444, 'epoch': 0.13}

 95%|█████████▌| 3140/3300 [4:48:23<14:59,  5.62s/it]INFO 04-12 03:25:14 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3141/3300 [4:48:27<13:44,  5.18s/it]2026-04-12 03:25:18,329 | INFO | train_grpo_train | metrics_logged mode=train step=3141

                                                     
{'loss': -0.0163, 'grad_norm': 5.089842319488525, 'learning_rate': 4.848484848484849e-07, 'num_tokens': 7152584.0, 'completions/mean_length': 33.75, 'completions/min_length': 32.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9182018041610718, 'rewards/meter/std': 0.2099631130695343, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9182018041610718, 'rewards/total_composite/std': 0.2099631130695343, 'reward': 0.9182018041610718, 'reward_std': 0.2099630981683731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012079699896275997, 'sampling/sampling_logp_difference/max': 0.8862104415893555, 'sampling/importance_sampling_ratio/min': 0.41221490502357483, 'sampling/importance_sampling_ratio/mean': 1.0063190460205078, 'sampling/importance_sampling_ratio/max': 1.7678229808807373, 'entropy': 0.10679818876087666, 'clip_ratio/low_mean': 0.0078125, 'clip_ratio/low_min': 0.0078125, 'clip_ratio/high_mean': 0.0036764706019312143, 'clip_ratio/high_max': 0.0036764706019312143, 'clip_ratio/region_mean': 0.011488970601931214, 'reward_total_mean': 0.9182018041610718, 'reward_meter_mean': 0.9182018041610718, 'reward_meter_std': 0.2099631130695343, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9182018041610718, 'reward_total_composite_std': 0.2099631130695343, 'epoch': 0.13}

 95%|█████████▌| 3141/3300 [4:48:27<13:44,  5.18s/it]INFO 04-12 03:25:18 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3142/3300 [4:48:31<13:10,  5.00s/it]2026-04-12 03:25:22,959 | INFO | train_grpo_train | metrics_logged mode=train step=3142

                                                     
{'loss': -0.0134, 'grad_norm': 3.1886298656463623, 'learning_rate': 4.818181818181818e-07, 'num_tokens': 7154301.0, 'completions/mean_length': 70.625, 'completions/min_length': 68.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.625, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9989311099052429, 'rewards/meter/std': 0.0014027439756318927, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989311099052429, 'rewards/total_composite/std': 0.0014027439756318927, 'reward': 0.9989311099052429, 'reward_std': 0.0014027438592165709, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013442294672131538, 'sampling/sampling_logp_difference/max': 0.9266881942749023, 'sampling/importance_sampling_ratio/min': 0.3958625793457031, 'sampling/importance_sampling_ratio/mean': 1.0028244256973267, 'sampling/importance_sampling_ratio/max': 1.341347575187683, 'entropy': 0.09765590541064739, 'clip_ratio/low_mean': 0.0036764706019312143, 'clip_ratio/low_min': 0.0036764706019312143, 'clip_ratio/high_mean': 0.0035211266949772835, 'clip_ratio/high_max': 0.0035211266949772835, 'clip_ratio/region_mean': 0.007197597296908498, 'reward_total_mean': 0.9989311099052429, 'reward_meter_mean': 0.9989311099052429, 'reward_meter_std': 0.0014027439756318927, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989311099052429, 'reward_total_composite_std': 0.0014027439756318927, 'epoch': 0.13}

 95%|█████████▌| 3142/3300 [4:48:31<13:10,  5.00s/it]INFO 04-12 03:25:23 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3143/3300 [4:48:37<13:21,  5.11s/it]2026-04-12 03:25:28,302 | INFO | train_grpo_train | metrics_logged mode=train step=3143

                                                     
{'loss': 0.0034, 'grad_norm': 1.7062424421310425, 'learning_rate': 4.787878787878789e-07, 'num_tokens': 7156840.0, 'completions/mean_length': 133.375, 'completions/min_length': 128.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.375, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.9990321397781372, 'rewards/meter/std': 0.00018018556875176728, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990321397781372, 'rewards/total_composite/std': 0.00018018556875176728, 'reward': 0.9990321397781372, 'reward_std': 0.00018020268180407584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035543572157621384, 'sampling/sampling_logp_difference/max': 1.157923698425293, 'sampling/importance_sampling_ratio/min': 0.3141377568244934, 'sampling/importance_sampling_ratio/mean': 1.0070942640304565, 'sampling/importance_sampling_ratio/max': 1.9823683500289917, 'entropy': 0.29481677152216434, 'clip_ratio/low_mean': 0.013188510201871395, 'clip_ratio/low_min': 0.013188510201871395, 'clip_ratio/high_mean': 0.01591338007710874, 'clip_ratio/high_max': 0.01591338007710874, 'clip_ratio/region_mean': 0.029101890278980136, 'reward_total_mean': 0.9990321397781372, 'reward_meter_mean': 0.9990321397781372, 'reward_meter_std': 0.00018018556875176728, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990321397781372, 'reward_total_composite_std': 0.00018018556875176728, 'epoch': 0.13}

 95%|█████████▌| 3143/3300 [4:48:37<13:21,  5.11s/it]INFO 04-12 03:25:28 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3144/3300 [4:48:41<12:53,  4.96s/it]2026-04-12 03:25:32,878 | INFO | train_grpo_train | metrics_logged mode=train step=3144

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.757575757575758e-07, 'num_tokens': 7158856.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00035932144965045154, 'sampling/sampling_logp_difference/max': 0.012860830873250961, 'sampling/importance_sampling_ratio/min': 0.9996905326843262, 'sampling/importance_sampling_ratio/mean': 1.0003576278686523, 'sampling/importance_sampling_ratio/max': 1.0129438638687134, 'entropy': 0.0032840893836691976, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 95%|█████████▌| 3144/3300 [4:48:41<12:53,  4.96s/it]INFO 04-12 03:25:33 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3145/3300 [4:48:46<12:43,  4.93s/it]2026-04-12 03:25:37,738 | INFO | train_grpo_train | metrics_logged mode=train step=3145

                                                     
{'loss': 0.0038, 'grad_norm': 3.3756773471832275, 'learning_rate': 4.7272727272727273e-07, 'num_tokens': 7161044.0, 'completions/mean_length': 96.5, 'completions/min_length': 95.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.5, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9917601346969604, 'rewards/meter/std': 0.004016265273094177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9917601346969604, 'rewards/total_composite/std': 0.004016265273094177, 'reward': 0.9917601346969604, 'reward_std': 0.004016268998384476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031790200620889664, 'sampling/sampling_logp_difference/max': 1.1975455284118652, 'sampling/importance_sampling_ratio/min': 0.3019343912601471, 'sampling/importance_sampling_ratio/mean': 1.0063228607177734, 'sampling/importance_sampling_ratio/max': 1.6533253192901611, 'entropy': 0.23462153412401676, 'clip_ratio/low_mean': 0.014030612306669354, 'clip_ratio/low_min': 0.014030612306669354, 'clip_ratio/high_mean': 0.014405153575353324, 'clip_ratio/high_max': 0.014405153575353324, 'clip_ratio/region_mean': 0.02843576588202268, 'reward_total_mean': 0.9917601346969604, 'reward_meter_mean': 0.9917601346969604, 'reward_meter_std': 0.004016265273094177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9917601346969604, 'reward_total_composite_std': 0.004016265273094177, 'epoch': 0.13}

 95%|█████████▌| 3145/3300 [4:48:46<12:43,  4.93s/it]INFO 04-12 03:25:37 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3146/3300 [4:48:50<11:56,  4.65s/it]2026-04-12 03:25:41,760 | INFO | train_grpo_train | metrics_logged mode=train step=3146

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.696969696969697e-07, 'num_tokens': 7162452.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00029772642301395535, 'sampling/sampling_logp_difference/max': 0.008134791627526283, 'sampling/importance_sampling_ratio/min': 0.9950645565986633, 'sampling/importance_sampling_ratio/mean': 1.0002386569976807, 'sampling/importance_sampling_ratio/max': 1.0081679821014404, 'entropy': 0.00371554572484456, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 95%|█████████▌| 3146/3300 [4:48:50<11:56,  4.65s/it]INFO 04-12 03:25:41 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3147/3300 [4:48:58<14:15,  5.59s/it]2026-04-12 03:25:49,542 | INFO | train_grpo_train | metrics_logged mode=train step=3147

                                                     
{'loss': 0.0288, 'grad_norm': 1.647410273551941, 'learning_rate': 4.666666666666667e-07, 'num_tokens': 7166497.0, 'completions/mean_length': 315.625, 'completions/min_length': 309.0, 'completions/max_length': 339.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 315.625, 'completions/min_terminated_length': 309.0, 'completions/max_terminated_length': 339.0, 'rewards/meter/mean': 0.9990444779396057, 'rewards/meter/std': 0.00020862040400970727, 'rewards/count_adherence/mean': 0.984375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9843137264251709, 'rewards/repeat_penalty/std': 0.029120875522494316, 'rewards/total_composite/mean': 0.9686882495880127, 'rewards/total_composite/std': 0.06351150572299957, 'reward': 0.9686882495880127, 'reward_std': 0.06351151317358017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04411856085062027, 'sampling/sampling_logp_difference/max': 1.3515138626098633, 'sampling/importance_sampling_ratio/min': 0.25884810090065, 'sampling/importance_sampling_ratio/mean': 1.0113532543182373, 'sampling/importance_sampling_ratio/max': 1.7542108297348022, 'entropy': 0.41412847116589546, 'clip_ratio/low_mean': 0.004168422543443739, 'clip_ratio/low_min': 0.004168422543443739, 'clip_ratio/high_mean': 0.017632243572734296, 'clip_ratio/high_max': 0.017632243572734296, 'clip_ratio/region_mean': 0.021800666116178036, 'reward_total_mean': 0.9686882495880127, 'reward_meter_mean': 0.9990444779396057, 'reward_meter_std': 0.00020862040400970727, 'reward_count_adherence_mean': 0.984375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9843137264251709, 'reward_repeat_penalty_std': 0.029120875522494316, 'reward_total_composite_mean': 0.9686882495880127, 'reward_total_composite_std': 0.06351150572299957, 'epoch': 0.13}

 95%|█████████▌| 3147/3300 [4:48:58<14:15,  5.59s/it]INFO 04-12 03:25:49 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3148/3300 [4:49:03<13:40,  5.40s/it]2026-04-12 03:25:54,491 | INFO | train_grpo_train | metrics_logged mode=train step=3148

                                                     
{'loss': 0.0009, 'grad_norm': 0.7614604234695435, 'learning_rate': 4.6363636363636365e-07, 'num_tokens': 7168719.0, 'completions/mean_length': 106.75, 'completions/min_length': 106.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.75, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9992802143096924, 'rewards/meter/std': 0.0001533372706035152, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992802143096924, 'rewards/total_composite/std': 0.0001533372706035152, 'reward': 0.9992802143096924, 'reward_std': 0.00015333489864133298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013726499862968922, 'sampling/sampling_logp_difference/max': 1.1793193817138672, 'sampling/importance_sampling_ratio/min': 0.30748796463012695, 'sampling/importance_sampling_ratio/mean': 1.0032622814178467, 'sampling/importance_sampling_ratio/max': 1.5043672323226929, 'entropy': 0.14908354356884956, 'clip_ratio/low_mean': 0.008188590873032808, 'clip_ratio/low_min': 0.008188590873032808, 'clip_ratio/high_mean': 0.00821063295006752, 'clip_ratio/high_max': 0.00821063295006752, 'clip_ratio/region_mean': 0.01639922382310033, 'reward_total_mean': 0.9992802143096924, 'reward_meter_mean': 0.9992802143096924, 'reward_meter_std': 0.0001533372706035152, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992802143096924, 'reward_total_composite_std': 0.0001533372706035152, 'epoch': 0.13}

 95%|█████████▌| 3148/3300 [4:49:03<13:40,  5.40s/it]INFO 04-12 03:25:54 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3149/3300 [4:49:13<17:11,  6.83s/it]2026-04-12 03:26:04,658 | INFO | train_grpo_train | metrics_logged mode=train step=3149

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.6060606060606064e-07, 'num_tokens': 7170575.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9272494316101074, 'rewards/meter/std': 0.10411977022886276, 'rewards/count_adherence/mean': 0.6687499284744263, 'rewards/count_adherence/std': 0.025877466425299644, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9813033938407898, 'rewards/repeat_penalty/std': 0.03969252109527588, 'rewards/total_composite/mean': 0.6074195504188538, 'rewards/total_composite/std': 0.06437568366527557, 'reward': 0.6074195504188538, 'reward_std': 0.06437568366527557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.6074195504188538, 'reward_meter_mean': 0.9272494316101074, 'reward_meter_std': 0.10411977022886276, 'reward_count_adherence_mean': 0.6687499284744263, 'reward_count_adherence_std': 0.025877466425299644, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9813033938407898, 'reward_repeat_penalty_std': 0.03969252109527588, 'reward_total_composite_mean': 0.6074195504188538, 'reward_total_composite_std': 0.06437568366527557, 'epoch': 0.13}

 95%|█████████▌| 3149/3300 [4:49:13<17:11,  6.83s/it]INFO 04-12 03:26:04 [block_pool.py:378] Successfully reset prefix cache

 95%|█████████▌| 3150/3300 [4:49:17<15:13,  6.09s/it]2026-04-12 03:26:09,022 | INFO | train_grpo_train | metrics_logged mode=train step=3150

                                                     
{'loss': 0.0003, 'grad_norm': 0.012665612623095512, 'learning_rate': 4.5757575757575764e-07, 'num_tokens': 7172439.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.997339129447937, 'rewards/meter/std': 3.875939285080676e-07, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.997339129447937, 'rewards/total_composite/std': 3.875939285080676e-07, 'reward': 0.997339129447937, 'reward_std': 3.8628223819614504e-07, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.004391301888972521, 'sampling/sampling_logp_difference/max': 0.7720130681991577, 'sampling/importance_sampling_ratio/min': 0.4620819389820099, 'sampling/importance_sampling_ratio/mean': 0.9999790787696838, 'sampling/importance_sampling_ratio/max': 1.506388783454895, 'entropy': 0.021077871089801192, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.0020491802133619785, 'reward_total_mean': 0.997339129447937, 'reward_meter_mean': 0.997339129447937, 'reward_meter_std': 3.875939285080676e-07, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.997339129447937, 'reward_total_composite_std': 3.875939285080676e-07, 'epoch': 0.13}

 95%|█████████▌| 3150/3300 [4:49:18<15:13,  6.09s/it]INFO 04-12 03:26:09 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:38,  3.53s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.45s/it]

 31%|███       | 4/13 [00:16<00:39,  4.38s/it]

 38%|███▊      | 5/13 [00:23<00:42,  5.33s/it]

 46%|████▌     | 6/13 [00:30<00:41,  5.88s/it]

 54%|█████▍    | 7/13 [00:36<00:34,  5.80s/it]

 62%|██████▏   | 8/13 [00:43<00:31,  6.28s/it]

 69%|██████▉   | 9/13 [00:50<00:26,  6.59s/it]

 77%|███████▋  | 10/13 [00:55<00:18,  6.08s/it]

 85%|████████▍ | 11/13 [01:01<00:12,  6.02s/it]

 92%|█████████▏| 12/13 [01:07<00:06,  6.12s/it]

100%|██████████| 13/13 [01:13<00:00,  5.93s/it]2026-04-12 03:27:29,553 | INFO | train_grpo_train | metrics_logged mode=eval step=3150

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 80.4009, 'eval_samples_per_second': 1.294, 'eval_steps_per_second': 0.162, 'eval_num_tokens': 7172439.0, 'eval_completions/mean_length': 217.70192307692307, 'eval_completions/min_length': 61.69230769230769, 'eval_completions/max_length': 436.2307692307692, 'eval_completions/clipped_ratio': 0.038461538461538464, 'eval_completions/mean_terminated_length': 206.14835533728967, 'eval_completions/min_terminated_length': 61.69230769230769, 'eval_completions/max_terminated_length': 401.84615384615387, 'eval_rewards/meter/mean': 0.7973346756054804, 'eval_rewards/meter/std': 0.3138508295210508, 'eval_rewards/count_adherence/mean': 0.963410904774299, 'eval_rewards/count_adherence/std': 0.054318822060640044, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.9485706686973572, 'eval_rewards/repeat_penalty/std': 0.08086307346820831, 'eval_rewards/total_composite/mean': 0.7355112204184899, 'eval_rewards/total_composite/std': 0.31246042595459866, 'eval_reward': 0.7355112204184899, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0324688284442975, 'eval_sampling/sampling_logp_difference/max': 1.1065230736365685, 'eval_sampling/importance_sampling_ratio/min': 0.33571984332341415, 'eval_sampling/importance_sampling_ratio/mean': 1.0091725221047034, 'eval_sampling/importance_sampling_ratio/max': 1.4886829761358409, 'eval_entropy': 0.37681426910253674, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7355112204184899, 'eval_reward_meter_mean': 0.7973346756054804, 'eval_reward_meter_std': 0.3138508295210508, 'eval_reward_count_adherence_mean': 0.963410904774299, 'eval_reward_count_adherence_std': 0.054318822060640044, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.9485706686973572, 'eval_reward_repeat_penalty_std': 0.08086307346820831, 'eval_reward_total_composite_mean': 0.7355112204184899, 'eval_reward_total_composite_std': 0.31246042595459866, 'epoch': 0.13}

 95%|█████████▌| 3150/3300 [4:50:38<15:13,  6.09s/it]

100%|██████████| 13/13 [01:13<00:00,  5.93s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 95%|█████████▌| 3151/3300 [4:50:46<1:16:15, 30.71s/it]2026-04-12 03:27:37,178 | INFO | train_grpo_train | metrics_logged mode=train step=3151

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.5454545454545457e-07, 'num_tokens': 7174111.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00020680490706581622, 'sampling/sampling_logp_difference/max': 0.002479560673236847, 'sampling/importance_sampling_ratio/min': 0.9986385703086853, 'sampling/importance_sampling_ratio/mean': 1.000199317932129, 'sampling/importance_sampling_ratio/max': 1.0024826526641846, 'entropy': 0.0016731252981116995, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 95%|█████████▌| 3151/3300 [4:50:46<1:16:15, 30.71s/it]INFO 04-12 03:27:37 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3152/3300 [4:50:50<56:36, 22.95s/it]  2026-04-12 03:27:42,019 | INFO | train_grpo_train | metrics_logged mode=train step=3152

                                                     
{'loss': 0.0014, 'grad_norm': 1.8619328737258911, 'learning_rate': 4.5151515151515156e-07, 'num_tokens': 7175864.0, 'completions/mean_length': 65.125, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9934186935424805, 'rewards/meter/std': 0.00011813976016128436, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9934186935424805, 'rewards/total_composite/std': 0.00011813976016128436, 'reward': 0.9934186935424805, 'reward_std': 0.00011814001481980085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010398104786872864, 'sampling/sampling_logp_difference/max': 0.37988075613975525, 'sampling/importance_sampling_ratio/min': 0.683942973613739, 'sampling/importance_sampling_ratio/mean': 1.0027790069580078, 'sampling/importance_sampling_ratio/max': 1.3604105710983276, 'entropy': 0.06057531526312232, 'clip_ratio/low_mean': 0.007634032750502229, 'clip_ratio/low_min': 0.007634032750502229, 'clip_ratio/high_mean': 0.003846153849735856, 'clip_ratio/high_max': 0.003846153849735856, 'clip_ratio/region_mean': 0.011480186600238085, 'reward_total_mean': 0.9934186935424805, 'reward_meter_mean': 0.9934186935424805, 'reward_meter_std': 0.00011813976016128436, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9934186935424805, 'reward_total_composite_std': 0.00011813976016128436, 'epoch': 0.13}

 96%|█████████▌| 3152/3300 [4:50:51<56:36, 22.95s/it]INFO 04-12 03:27:42 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3153/3300 [4:50:57<43:58, 17.95s/it]2026-04-12 03:27:48,298 | INFO | train_grpo_train | metrics_logged mode=train step=3153

                                                     
{'loss': 0.0065, 'grad_norm': 2.150918483734131, 'learning_rate': 4.484848484848485e-07, 'num_tokens': 7179187.0, 'completions/mean_length': 199.375, 'completions/min_length': 197.0, 'completions/max_length': 204.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 199.375, 'completions/min_terminated_length': 197.0, 'completions/max_terminated_length': 204.0, 'rewards/meter/mean': 0.9993048906326294, 'rewards/meter/std': 0.0001919995847856626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9318181276321411, 'rewards/repeat_penalty/std': 0.08058229833841324, 'rewards/total_composite/mean': 0.9311637878417969, 'rewards/total_composite/std': 0.08044037967920303, 'reward': 0.9311637878417969, 'reward_std': 0.08044035732746124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020360471680760384, 'sampling/sampling_logp_difference/max': 1.146202564239502, 'sampling/importance_sampling_ratio/min': 0.31784144043922424, 'sampling/importance_sampling_ratio/mean': 1.0063588619232178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18566425889730453, 'clip_ratio/low_mean': 0.0037722690613009036, 'clip_ratio/low_min': 0.0037722690613009036, 'clip_ratio/high_mean': 0.006870172452181578, 'clip_ratio/high_max': 0.006870172452181578, 'clip_ratio/region_mean': 0.010642441513482481, 'reward_total_mean': 0.9311637878417969, 'reward_meter_mean': 0.9993048906326294, 'reward_meter_std': 0.0001919995847856626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9318181276321411, 'reward_repeat_penalty_std': 0.08058229833841324, 'reward_total_composite_mean': 0.9311637878417969, 'reward_total_composite_std': 0.08044037967920303, 'epoch': 0.13}

 96%|█████████▌| 3153/3300 [4:50:57<43:58, 17.95s/it]INFO 04-12 03:27:48 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3154/3300 [4:51:02<34:25, 14.14s/it]2026-04-12 03:27:53,571 | INFO | train_grpo_train | metrics_logged mode=train step=3154

                                                     
{'loss': 0.0047, 'grad_norm': 3.9876134395599365, 'learning_rate': 4.454545454545455e-07, 'num_tokens': 7181304.0, 'completions/mean_length': 103.625, 'completions/min_length': 102.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.625, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9898449182510376, 'rewards/meter/std': 0.012184920720756054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9898449182510376, 'rewards/total_composite/std': 0.012184920720756054, 'reward': 0.9898449182510376, 'reward_std': 0.01218490768224001, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036951709538698196, 'sampling/sampling_logp_difference/max': 1.5696239471435547, 'sampling/importance_sampling_ratio/min': 0.20812343060970306, 'sampling/importance_sampling_ratio/mean': 1.005436658859253, 'sampling/importance_sampling_ratio/max': 1.9000133275985718, 'entropy': 0.30003978684544563, 'clip_ratio/low_mean': 0.007282239850610495, 'clip_ratio/low_min': 0.007282239850610495, 'clip_ratio/high_mean': 0.021670067915692925, 'clip_ratio/high_max': 0.021670067915692925, 'clip_ratio/region_mean': 0.02895230776630342, 'reward_total_mean': 0.9898449182510376, 'reward_meter_mean': 0.9898449182510376, 'reward_meter_std': 0.012184920720756054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9898449182510376, 'reward_total_composite_std': 0.012184920720756054, 'epoch': 0.13}

 96%|█████████▌| 3154/3300 [4:51:02<34:25, 14.14s/it]INFO 04-12 03:27:53 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3155/3300 [4:51:06<27:12, 11.26s/it]2026-04-12 03:27:58,078 | INFO | train_grpo_train | metrics_logged mode=train step=3155

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.424242424242425e-07, 'num_tokens': 7183088.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973388910293579, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973388910293579, 'rewards/total_composite/std': 0.0, 'reward': 0.9973388910293579, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0017873203614726663, 'sampling/sampling_logp_difference/max': 0.15021102130413055, 'sampling/importance_sampling_ratio/min': 0.8605263233184814, 'sampling/importance_sampling_ratio/mean': 1.0003093481063843, 'sampling/importance_sampling_ratio/max': 1.055625557899475, 'entropy': 0.013011510367505252, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9973388910293579, 'reward_meter_mean': 0.9973388910293579, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973388910293579, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3155/3300 [4:51:07<27:12, 11.26s/it]INFO 04-12 03:27:58 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3156/3300 [4:51:11<22:01,  9.18s/it]2026-04-12 03:28:02,401 | INFO | train_grpo_train | metrics_logged mode=train step=3156

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.393939393939394e-07, 'num_tokens': 7184864.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002461693366058171, 'sampling/sampling_logp_difference/max': 0.0035922862589359283, 'sampling/importance_sampling_ratio/min': 0.9997955560684204, 'sampling/importance_sampling_ratio/mean': 1.000244379043579, 'sampling/importance_sampling_ratio/max': 1.0035988092422485, 'entropy': 0.0018204424268333241, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3156/3300 [4:51:11<22:01,  9.18s/it]INFO 04-12 03:28:02 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3157/3300 [4:51:15<18:21,  7.70s/it]2026-04-12 03:28:06,665 | INFO | train_grpo_train | metrics_logged mode=train step=3157

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.363636363636364e-07, 'num_tokens': 7186584.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00020154265803284943, 'sampling/sampling_logp_difference/max': 0.002117213560268283, 'sampling/importance_sampling_ratio/min': 0.9994971752166748, 'sampling/importance_sampling_ratio/mean': 1.0001963376998901, 'sampling/importance_sampling_ratio/max': 1.0021194219589233, 'entropy': 0.0016821113676996902, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3157/3300 [4:51:15<18:21,  7.70s/it]INFO 04-12 03:28:06 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3158/3300 [4:51:20<15:58,  6.75s/it]2026-04-12 03:28:11,203 | INFO | train_grpo_train | metrics_logged mode=train step=3158

                                                     
{'loss': 0.0133, 'grad_norm': 4.2719950675964355, 'learning_rate': 4.333333333333334e-07, 'num_tokens': 7188342.0, 'completions/mean_length': 68.75, 'completions/min_length': 66.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9841428399085999, 'rewards/meter/std': 0.03312673419713974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9841428399085999, 'rewards/total_composite/std': 0.03312673419713974, 'reward': 0.9841428399085999, 'reward_std': 0.03312673047184944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03139575570821762, 'sampling/sampling_logp_difference/max': 1.0852737426757812, 'sampling/importance_sampling_ratio/min': 0.337809294462204, 'sampling/importance_sampling_ratio/mean': 1.007124900817871, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23758242093026638, 'clip_ratio/low_mean': 0.0035211266949772835, 'clip_ratio/low_min': 0.0035211266949772835, 'clip_ratio/high_mean': 0.009170901728793979, 'clip_ratio/high_max': 0.009170901728793979, 'clip_ratio/region_mean': 0.012692028423771262, 'reward_total_mean': 0.9841428399085999, 'reward_meter_mean': 0.9841428399085999, 'reward_meter_std': 0.03312673419713974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9841428399085999, 'reward_total_composite_std': 0.03312673419713974, 'epoch': 0.13}

 96%|█████████▌| 3158/3300 [4:51:20<15:58,  6.75s/it]INFO 04-12 03:28:11 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3159/3300 [4:51:24<14:28,  6.16s/it]2026-04-12 03:28:15,983 | INFO | train_grpo_train | metrics_logged mode=train step=3159

                                                     
{'loss': 0.0047, 'grad_norm': 6.7398481369018555, 'learning_rate': 4.3030303030303034e-07, 'num_tokens': 7190330.0, 'completions/mean_length': 86.5, 'completions/min_length': 83.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.5, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.8662310242652893, 'rewards/meter/std': 0.18711434304714203, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.8583756685256958, 'rewards/total_composite/std': 0.20775045454502106, 'reward': 0.8583756685256958, 'reward_std': 0.20775045454502106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05684678256511688, 'sampling/sampling_logp_difference/max': 1.4105424880981445, 'sampling/importance_sampling_ratio/min': 0.3277885615825653, 'sampling/importance_sampling_ratio/mean': 1.0018366575241089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24941212497651577, 'clip_ratio/low_mean': 0.01434953324496746, 'clip_ratio/low_min': 0.01434953324496746, 'clip_ratio/high_mean': 0.04191699391230941, 'clip_ratio/high_max': 0.04191699391230941, 'clip_ratio/region_mean': 0.05626652715727687, 'reward_total_mean': 0.8583756685256958, 'reward_meter_mean': 0.8662310242652893, 'reward_meter_std': 0.18711434304714203, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.8583756685256958, 'reward_total_composite_std': 0.20775045454502106, 'epoch': 0.13}

 96%|█████████▌| 3159/3300 [4:51:24<14:28,  6.16s/it]INFO 04-12 03:28:16 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3160/3300 [4:51:34<17:09,  7.35s/it]2026-04-12 03:28:26,108 | INFO | train_grpo_train | metrics_logged mode=train step=3160

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.272727272727273e-07, 'num_tokens': 7192010.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.9984534978866577, 'rewards/meter/std': 0.0011360765201970935, 'rewards/count_adherence/mean': 0.6624999642372131, 'rewards/count_adherence/std': 0.023145509883761406, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9809472560882568, 'rewards/repeat_penalty/std': 0.020373545587062836, 'rewards/total_composite/mean': 0.5677040219306946, 'rewards/total_composite/std': 0.2308495044708252, 'reward': 0.5677040219306946, 'reward_std': 0.230849489569664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.5677040219306946, 'reward_meter_mean': 0.9984534978866577, 'reward_meter_std': 0.0011360765201970935, 'reward_count_adherence_mean': 0.6624999642372131, 'reward_count_adherence_std': 0.023145509883761406, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9809472560882568, 'reward_repeat_penalty_std': 0.020373545587062836, 'reward_total_composite_mean': 0.5677040219306946, 'reward_total_composite_std': 0.2308495044708252, 'epoch': 0.13}

 96%|█████████▌| 3160/3300 [4:51:35<17:09,  7.35s/it]INFO 04-12 03:28:26 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3161/3300 [4:51:43<17:32,  7.57s/it]2026-04-12 03:28:34,194 | INFO | train_grpo_train | metrics_logged mode=train step=3161

                                                     
{'loss': -0.0163, 'grad_norm': 1.4183757305145264, 'learning_rate': 4.242424242424243e-07, 'num_tokens': 7196839.0, 'completions/mean_length': 349.625, 'completions/min_length': 338.0, 'completions/max_length': 359.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 349.625, 'completions/min_terminated_length': 338.0, 'completions/max_terminated_length': 359.0, 'rewards/meter/mean': 0.9987608194351196, 'rewards/meter/std': 0.00043402795563451946, 'rewards/count_adherence/mean': 0.8020833134651184, 'rewards/count_adherence/std': 0.0431290864944458, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9601608514785767, 'rewards/repeat_penalty/std': 0.02460995875298977, 'rewards/total_composite/mean': 0.7685060501098633, 'rewards/total_composite/std': 0.03033612295985222, 'reward': 0.7685060501098633, 'reward_std': 0.030336128547787666, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.034961603581905365, 'sampling/sampling_logp_difference/max': 1.671947956085205, 'sampling/importance_sampling_ratio/min': 0.1878807246685028, 'sampling/importance_sampling_ratio/mean': 1.0060912370681763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35866962373256683, 'clip_ratio/low_mean': 0.008485189639031887, 'clip_ratio/low_min': 0.008485189639031887, 'clip_ratio/high_mean': 0.013319394318386912, 'clip_ratio/high_max': 0.013319394318386912, 'clip_ratio/region_mean': 0.0218045839574188, 'reward_total_mean': 0.7685060501098633, 'reward_meter_mean': 0.9987608194351196, 'reward_meter_std': 0.00043402795563451946, 'reward_count_adherence_mean': 0.8020833134651184, 'reward_count_adherence_std': 0.0431290864944458, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9601608514785767, 'reward_repeat_penalty_std': 0.02460995875298977, 'reward_total_composite_mean': 0.7685060501098633, 'reward_total_composite_std': 0.03033612295985222, 'epoch': 0.13}

 96%|█████████▌| 3161/3300 [4:51:43<17:32,  7.57s/it]INFO 04-12 03:28:34 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3162/3300 [4:51:47<15:22,  6.69s/it]2026-04-12 03:28:38,825 | INFO | train_grpo_train | metrics_logged mode=train step=3162

                                                     
{'loss': -0.0002, 'grad_norm': 0.1628890037536621, 'learning_rate': 4.2121212121212126e-07, 'num_tokens': 7198784.0, 'completions/mean_length': 71.125, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994406700134277, 'rewards/meter/std': 1.3193358427088242e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994406700134277, 'rewards/total_composite/std': 1.3193358427088242e-05, 'reward': 0.9994406700134277, 'reward_std': 1.3199281056586187e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0068392762914299965, 'sampling/sampling_logp_difference/max': 0.3337416648864746, 'sampling/importance_sampling_ratio/min': 0.716238796710968, 'sampling/importance_sampling_ratio/mean': 1.003661036491394, 'sampling/importance_sampling_ratio/max': 1.3171619176864624, 'entropy': 0.07914929743856192, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0035211266949772835, 'clip_ratio/high_max': 0.0035211266949772835, 'clip_ratio/region_mean': 0.0035211266949772835, 'reward_total_mean': 0.9994406700134277, 'reward_meter_mean': 0.9994406700134277, 'reward_meter_std': 1.3193358427088242e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994406700134277, 'reward_total_composite_std': 1.3193358427088242e-05, 'epoch': 0.13}

 96%|█████████▌| 3162/3300 [4:51:47<15:22,  6.69s/it]INFO 04-12 03:28:39 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3163/3300 [4:51:52<13:50,  6.06s/it]2026-04-12 03:28:43,424 | INFO | train_grpo_train | metrics_logged mode=train step=3163

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.181818181818182e-07, 'num_tokens': 7200688.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0004915767349302769, 'sampling/sampling_logp_difference/max': 0.022466091439127922, 'sampling/importance_sampling_ratio/min': 0.9993630647659302, 'sampling/importance_sampling_ratio/mean': 1.0004879236221313, 'sampling/importance_sampling_ratio/max': 1.0227203369140625, 'entropy': 0.003859572723740712, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3163/3300 [4:51:52<13:50,  6.06s/it]INFO 04-12 03:28:43 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3164/3300 [4:51:56<12:34,  5.55s/it]2026-04-12 03:28:47,771 | INFO | train_grpo_train | metrics_logged mode=train step=3164

                                                     
{'loss': -0.0002, 'grad_norm': 7.426182270050049, 'learning_rate': 4.1515151515151513e-07, 'num_tokens': 7202591.0, 'completions/mean_length': 45.875, 'completions/min_length': 45.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9413228631019592, 'rewards/meter/std': 0.004475572612136602, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9413228631019592, 'rewards/total_composite/std': 0.004475572612136602, 'reward': 0.9413228631019592, 'reward_std': 0.004475583788007498, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033416956663131714, 'sampling/sampling_logp_difference/max': 2.256293296813965, 'sampling/importance_sampling_ratio/min': 0.10473800450563431, 'sampling/importance_sampling_ratio/mean': 0.9981755614280701, 'sampling/importance_sampling_ratio/max': 1.8759413957595825, 'entropy': 0.10774505604058504, 'clip_ratio/low_mean': 0.0027173913549631834, 'clip_ratio/low_min': 0.0027173913549631834, 'clip_ratio/high_mean': 0.019021739484742284, 'clip_ratio/high_max': 0.019021739484742284, 'clip_ratio/region_mean': 0.021739130839705467, 'reward_total_mean': 0.9413228631019592, 'reward_meter_mean': 0.9413228631019592, 'reward_meter_std': 0.004475572612136602, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9413228631019592, 'reward_total_composite_std': 0.004475572612136602, 'epoch': 0.13}

 96%|█████████▌| 3164/3300 [4:51:56<12:34,  5.55s/it]INFO 04-12 03:28:48 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3165/3300 [4:52:01<11:46,  5.23s/it]2026-04-12 03:28:52,279 | INFO | train_grpo_train | metrics_logged mode=train step=3165

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.121212121212122e-07, 'num_tokens': 7204447.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00010137181379832327, 'sampling/sampling_logp_difference/max': 0.0022658593952655792, 'sampling/importance_sampling_ratio/min': 0.9977367520332336, 'sampling/importance_sampling_ratio/mean': 1.000064730644226, 'sampling/importance_sampling_ratio/max': 1.0019134283065796, 'entropy': 0.0008098809266812168, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3165/3300 [4:52:01<11:46,  5.23s/it]INFO 04-12 03:28:52 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3166/3300 [4:52:05<11:12,  5.02s/it]2026-04-12 03:28:56,800 | INFO | train_grpo_train | metrics_logged mode=train step=3166

                                                     
{'loss': -0.0001, 'grad_norm': 0.4401102066040039, 'learning_rate': 4.090909090909091e-07, 'num_tokens': 7206119.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981397390365601, 'rewards/meter/std': 1.352704748569522e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981397390365601, 'rewards/total_composite/std': 1.352704748569522e-05, 'reward': 0.9981397390365601, 'reward_std': 1.3527009286917746e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006174927577376366, 'sampling/sampling_logp_difference/max': 0.37103140354156494, 'sampling/importance_sampling_ratio/min': 0.6950086951255798, 'sampling/importance_sampling_ratio/mean': 1.0013184547424316, 'sampling/importance_sampling_ratio/max': 1.4492285251617432, 'entropy': 0.052269697189331055, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0018656715983524919, 'clip_ratio/high_max': 0.0018656715983524919, 'clip_ratio/region_mean': 0.0037313431967049837, 'reward_total_mean': 0.9981397390365601, 'reward_meter_mean': 0.9981397390365601, 'reward_meter_std': 1.352704748569522e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981397390365601, 'reward_total_composite_std': 1.352704748569522e-05, 'epoch': 0.13}

 96%|█████████▌| 3166/3300 [4:52:05<11:12,  5.02s/it]INFO 04-12 03:28:57 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3167/3300 [4:52:10<11:13,  5.07s/it]2026-04-12 03:29:01,966 | INFO | train_grpo_train | metrics_logged mode=train step=3167

                                                     
{'loss': 0.0009, 'grad_norm': 1.5397974252700806, 'learning_rate': 4.0606060606060605e-07, 'num_tokens': 7208456.0, 'completions/mean_length': 125.125, 'completions/min_length': 124.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.125, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.9850419759750366, 'rewards/meter/std': 0.024076560512185097, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9673014283180237, 'rewards/total_composite/std': 0.052494850009679794, 'reward': 0.9673014283180237, 'reward_std': 0.0524948425590992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02755284309387207, 'sampling/sampling_logp_difference/max': 1.6096177101135254, 'sampling/importance_sampling_ratio/min': 0.19996404647827148, 'sampling/importance_sampling_ratio/mean': 1.0092271566390991, 'sampling/importance_sampling_ratio/max': 1.744919776916504, 'entropy': 0.20761017128825188, 'clip_ratio/low_mean': 0.006024193484336138, 'clip_ratio/low_min': 0.006024193484336138, 'clip_ratio/high_mean': 0.012976191123016179, 'clip_ratio/high_max': 0.012976191123016179, 'clip_ratio/region_mean': 0.019000384607352316, 'reward_total_mean': 0.9673014283180237, 'reward_meter_mean': 0.9850419759750366, 'reward_meter_std': 0.024076560512185097, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9673014283180237, 'reward_total_composite_std': 0.052494850009679794, 'epoch': 0.13}

 96%|█████████▌| 3167/3300 [4:52:10<11:13,  5.07s/it]INFO 04-12 03:29:02 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3168/3300 [4:52:15<10:44,  4.89s/it]2026-04-12 03:29:06,438 | INFO | train_grpo_train | metrics_logged mode=train step=3168

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.030303030303031e-07, 'num_tokens': 7210280.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 8.32213117973879e-05, 'sampling/sampling_logp_difference/max': 0.004544587340205908, 'sampling/importance_sampling_ratio/min': 0.9954656958580017, 'sampling/importance_sampling_ratio/mean': 1.000028371810913, 'sampling/importance_sampling_ratio/max': 1.0013597011566162, 'entropy': 0.0007030887718428858, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3168/3300 [4:52:15<10:44,  4.89s/it]INFO 04-12 03:29:06 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3169/3300 [4:52:20<10:58,  5.03s/it]2026-04-12 03:29:11,798 | INFO | train_grpo_train | metrics_logged mode=train step=3169

                                                     
{'loss': 0.0016, 'grad_norm': 2.7400386333465576, 'learning_rate': 4.0000000000000003e-07, 'num_tokens': 7212758.0, 'completions/mean_length': 137.75, 'completions/min_length': 132.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.75, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.9911104440689087, 'rewards/meter/std': 0.00600889977067709, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.973473310470581, 'rewards/total_composite/std': 0.05160484462976456, 'reward': 0.973473310470581, 'reward_std': 0.051604848355054855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04144033044576645, 'sampling/sampling_logp_difference/max': 1.1764769554138184, 'sampling/importance_sampling_ratio/min': 0.3083631992340088, 'sampling/importance_sampling_ratio/mean': 1.0084030628204346, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3814433552324772, 'clip_ratio/low_mean': 0.0036231884732842445, 'clip_ratio/low_min': 0.0036231884732842445, 'clip_ratio/high_mean': 0.028263273648917675, 'clip_ratio/high_max': 0.028263273648917675, 'clip_ratio/region_mean': 0.03188646212220192, 'reward_total_mean': 0.973473310470581, 'reward_meter_mean': 0.9911104440689087, 'reward_meter_std': 0.00600889977067709, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.973473310470581, 'reward_total_composite_std': 0.05160484462976456, 'epoch': 0.13}

 96%|█████████▌| 3169/3300 [4:52:20<10:58,  5.03s/it]INFO 04-12 03:29:12 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3170/3300 [4:52:27<12:01,  5.55s/it]2026-04-12 03:29:18,580 | INFO | train_grpo_train | metrics_logged mode=train step=3170

                                                     
{'loss': 0.0069, 'grad_norm': 1.578904628753662, 'learning_rate': 3.9696969696969697e-07, 'num_tokens': 7216446.0, 'completions/mean_length': 248.0, 'completions/min_length': 244.0, 'completions/max_length': 251.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 248.0, 'completions/min_terminated_length': 244.0, 'completions/max_terminated_length': 251.0, 'rewards/meter/mean': 0.9988981485366821, 'rewards/meter/std': 0.0001572021865285933, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9615384340286255, 'rewards/repeat_penalty/std': 0.041117113083601, 'rewards/total_composite/mean': 0.9604784250259399, 'rewards/total_composite/std': 0.04105609655380249, 'reward': 0.9604784250259399, 'reward_std': 0.041056081652641296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.027040995657444, 'sampling/sampling_logp_difference/max': 1.3942060470581055, 'sampling/importance_sampling_ratio/min': 0.248029887676239, 'sampling/importance_sampling_ratio/mean': 1.0076178312301636, 'sampling/importance_sampling_ratio/max': 1.6617929935455322, 'entropy': 0.3082658052444458, 'clip_ratio/low_mean': 0.004030561889521778, 'clip_ratio/low_min': 0.004030561889521778, 'clip_ratio/high_mean': 0.0075775557197630405, 'clip_ratio/high_max': 0.0075775557197630405, 'clip_ratio/region_mean': 0.011608117609284818, 'reward_total_mean': 0.9604784250259399, 'reward_meter_mean': 0.9988981485366821, 'reward_meter_std': 0.0001572021865285933, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9615384340286255, 'reward_repeat_penalty_std': 0.041117113083601, 'reward_total_composite_mean': 0.9604784250259399, 'reward_total_composite_std': 0.04105609655380249, 'epoch': 0.13}

 96%|█████████▌| 3170/3300 [4:52:27<12:01,  5.55s/it]INFO 04-12 03:29:18 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3171/3300 [4:52:32<11:19,  5.27s/it]2026-04-12 03:29:23,166 | INFO | train_grpo_train | metrics_logged mode=train step=3171

                                                     
{'loss': -0.0007, 'grad_norm': 2.6443123817443848, 'learning_rate': 3.9393939393939396e-07, 'num_tokens': 7218213.0, 'completions/mean_length': 69.875, 'completions/min_length': 68.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9944202899932861, 'rewards/meter/std': 0.002229472389444709, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9944202899932861, 'rewards/total_composite/std': 0.002229472389444709, 'reward': 0.9944202899932861, 'reward_std': 0.002229455392807722, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02252027578651905, 'sampling/sampling_logp_difference/max': 1.6969127655029297, 'sampling/importance_sampling_ratio/min': 0.183248370885849, 'sampling/importance_sampling_ratio/mean': 1.0053551197052002, 'sampling/importance_sampling_ratio/max': 1.854751706123352, 'entropy': 0.1526617044582963, 'clip_ratio/low_mean': 0.005335517227649689, 'clip_ratio/low_min': 0.005335517227649689, 'clip_ratio/high_mean': 0.009086863487027586, 'clip_ratio/high_max': 0.009086863487027586, 'clip_ratio/region_mean': 0.014422380714677274, 'reward_total_mean': 0.9944202899932861, 'reward_meter_mean': 0.9944202899932861, 'reward_meter_std': 0.002229472389444709, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9944202899932861, 'reward_total_composite_std': 0.002229472389444709, 'epoch': 0.13}

 96%|█████████▌| 3171/3300 [4:52:32<11:19,  5.27s/it]INFO 04-12 03:29:23 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3172/3300 [4:52:36<10:39,  5.00s/it]2026-04-12 03:29:27,549 | INFO | train_grpo_train | metrics_logged mode=train step=3172

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.9090909090909095e-07, 'num_tokens': 7219917.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 9.710989979794249e-05, 'sampling/sampling_logp_difference/max': 0.0017959647811949253, 'sampling/importance_sampling_ratio/min': 0.9985963106155396, 'sampling/importance_sampling_ratio/mean': 1.0000759363174438, 'sampling/importance_sampling_ratio/max': 1.0017976760864258, 'entropy': 0.0011070799009758048, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3172/3300 [4:52:36<10:39,  5.00s/it]INFO 04-12 03:29:27 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3173/3300 [4:52:43<11:36,  5.48s/it]2026-04-12 03:29:34,152 | INFO | train_grpo_train | metrics_logged mode=train step=3173

                                                     
{'loss': 0.0057, 'grad_norm': 1.6205352544784546, 'learning_rate': 3.878787878787879e-07, 'num_tokens': 7223459.0, 'completions/mean_length': 230.75, 'completions/min_length': 226.0, 'completions/max_length': 234.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 230.75, 'completions/min_terminated_length': 226.0, 'completions/max_terminated_length': 234.0, 'rewards/meter/mean': 0.9989705085754395, 'rewards/meter/std': 0.00055393495131284, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9615384340286255, 'rewards/repeat_penalty/std': 0.058148376643657684, 'rewards/total_composite/mean': 0.9605352282524109, 'rewards/total_composite/std': 0.05785057693719864, 'reward': 0.9605352282524109, 'reward_std': 0.05785060301423073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.025763755664229393, 'sampling/sampling_logp_difference/max': 1.1109724044799805, 'sampling/importance_sampling_ratio/min': 0.3292386531829834, 'sampling/importance_sampling_ratio/mean': 1.0043230056762695, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2557064164429903, 'clip_ratio/low_mean': 0.006500809220597148, 'clip_ratio/low_min': 0.006500809220597148, 'clip_ratio/high_mean': 0.016817589348647743, 'clip_ratio/high_max': 0.016817589348647743, 'clip_ratio/region_mean': 0.02331839856924489, 'reward_total_mean': 0.9605352282524109, 'reward_meter_mean': 0.9989705085754395, 'reward_meter_std': 0.00055393495131284, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9615384340286255, 'reward_repeat_penalty_std': 0.058148376643657684, 'reward_total_composite_mean': 0.9605352282524109, 'reward_total_composite_std': 0.05785057693719864, 'epoch': 0.13}

 96%|█████████▌| 3173/3300 [4:52:43<11:36,  5.48s/it]INFO 04-12 03:29:34 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3174/3300 [4:52:47<10:42,  5.10s/it]2026-04-12 03:29:38,352 | INFO | train_grpo_train | metrics_logged mode=train step=3174

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.848484848484849e-07, 'num_tokens': 7224971.0, 'completions/mean_length': 36.0, 'completions/min_length': 36.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9996045231819153, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9996045231819153, 'rewards/total_composite/std': 0.0, 'reward': 0.9996045231819153, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0017466507852077484, 'sampling/sampling_logp_difference/max': 0.04659882187843323, 'sampling/importance_sampling_ratio/min': 0.9647499918937683, 'sampling/importance_sampling_ratio/mean': 1.0015071630477905, 'sampling/importance_sampling_ratio/max': 1.0477015972137451, 'entropy': 0.016203245264478028, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9996045231819153, 'reward_meter_mean': 0.9996045231819153, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9996045231819153, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▌| 3174/3300 [4:52:47<10:42,  5.10s/it]INFO 04-12 03:29:38 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3175/3300 [4:52:52<10:44,  5.16s/it]2026-04-12 03:29:43,661 | INFO | train_grpo_train | metrics_logged mode=train step=3175

                                                     
{'loss': -0.0001, 'grad_norm': 0.08287062495946884, 'learning_rate': 3.8181818181818187e-07, 'num_tokens': 7227420.0, 'completions/mean_length': 132.125, 'completions/min_length': 132.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.125, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9994284510612488, 'rewards/meter/std': 6.205716090335045e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994284510612488, 'rewards/total_composite/std': 6.205716090335045e-06, 'reward': 0.9994284510612488, 'reward_std': 6.210763785929885e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008646445348858833, 'sampling/sampling_logp_difference/max': 0.9603474140167236, 'sampling/importance_sampling_ratio/min': 0.3827598989009857, 'sampling/importance_sampling_ratio/mean': 1.0014747381210327, 'sampling/importance_sampling_ratio/max': 1.300461769104004, 'entropy': 0.08139224071055651, 'clip_ratio/low_mean': 0.0028337890980765224, 'clip_ratio/low_min': 0.0028337890980765224, 'clip_ratio/high_mean': 0.0018939394503831863, 'clip_ratio/high_max': 0.0018939394503831863, 'clip_ratio/region_mean': 0.004727728548459709, 'reward_total_mean': 0.9994284510612488, 'reward_meter_mean': 0.9994284510612488, 'reward_meter_std': 6.205716090335045e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994284510612488, 'reward_total_composite_std': 6.205716090335045e-06, 'epoch': 0.13}

 96%|█████████▌| 3175/3300 [4:52:52<10:44,  5.16s/it]INFO 04-12 03:29:43 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▌| 3176/3300 [4:53:00<12:32,  6.07s/it]2026-04-12 03:29:51,838 | INFO | train_grpo_train | metrics_logged mode=train step=3176

                                                     
{'loss': -0.0238, 'grad_norm': 1.365169644355774, 'learning_rate': 3.787878787878788e-07, 'num_tokens': 7232148.0, 'completions/mean_length': 351.0, 'completions/min_length': 317.0, 'completions/max_length': 371.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 351.0, 'completions/min_terminated_length': 317.0, 'completions/max_terminated_length': 371.0, 'rewards/meter/mean': 0.9988387823104858, 'rewards/meter/std': 0.0003177140897605568, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.04629101976752281, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9736841917037964, 'rewards/repeat_penalty/std': 0.03978573530912399, 'rewards/total_composite/mean': 0.9475728273391724, 'rewards/total_composite/std': 0.04686007648706436, 'reward': 0.9475728273391724, 'reward_std': 0.04686008021235466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03374847397208214, 'sampling/sampling_logp_difference/max': 4.901726722717285, 'sampling/importance_sampling_ratio/min': 0.007433735765516758, 'sampling/importance_sampling_ratio/mean': 1.008555293083191, 'sampling/importance_sampling_ratio/max': 1.7764185667037964, 'entropy': 0.3763493411242962, 'clip_ratio/low_mean': 0.008392618678044528, 'clip_ratio/low_min': 0.008392618678044528, 'clip_ratio/high_mean': 0.012750733643770218, 'clip_ratio/high_max': 0.012750733643770218, 'clip_ratio/region_mean': 0.021143352321814746, 'reward_total_mean': 0.9475728273391724, 'reward_meter_mean': 0.9988387823104858, 'reward_meter_std': 0.0003177140897605568, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.04629101976752281, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9736841917037964, 'reward_repeat_penalty_std': 0.03978573530912399, 'reward_total_composite_mean': 0.9475728273391724, 'reward_total_composite_std': 0.04686007648706436, 'epoch': 0.13}

 96%|█████████▌| 3176/3300 [4:53:00<12:32,  6.07s/it]INFO 04-12 03:29:52 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3177/3300 [4:53:05<11:22,  5.55s/it]2026-04-12 03:29:56,181 | INFO | train_grpo_train | metrics_logged mode=train step=3177

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.757575757575758e-07, 'num_tokens': 7233836.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00024669314734637737, 'sampling/sampling_logp_difference/max': 0.0033810948953032494, 'sampling/importance_sampling_ratio/min': 0.9966246485710144, 'sampling/importance_sampling_ratio/mean': 1.0002270936965942, 'sampling/importance_sampling_ratio/max': 1.002587914466858, 'entropy': 0.001803019520593807, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▋| 3177/3300 [4:53:05<11:22,  5.55s/it]INFO 04-12 03:29:56 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3178/3300 [4:53:09<10:42,  5.27s/it]2026-04-12 03:30:00,793 | INFO | train_grpo_train | metrics_logged mode=train step=3178

                                                     
{'loss': 0.0007, 'grad_norm': 1.7416008710861206, 'learning_rate': 3.7272727272727274e-07, 'num_tokens': 7235742.0, 'completions/mean_length': 71.25, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994281530380249, 'rewards/meter/std': 4.47747042926494e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994281530380249, 'rewards/total_composite/std': 4.47747042926494e-05, 'reward': 0.9994281530380249, 'reward_std': 4.4774700654670596e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007580549921840429, 'sampling/sampling_logp_difference/max': 0.3693375587463379, 'sampling/importance_sampling_ratio/min': 0.691192090511322, 'sampling/importance_sampling_ratio/mean': 1.004219651222229, 'sampling/importance_sampling_ratio/max': 1.2836189270019531, 'entropy': 0.07951132394373417, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0034966744715347886, 'clip_ratio/high_max': 0.0034966744715347886, 'clip_ratio/region_mean': 0.0034966744715347886, 'reward_total_mean': 0.9994281530380249, 'reward_meter_mean': 0.9994281530380249, 'reward_meter_std': 4.47747042926494e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994281530380249, 'reward_total_composite_std': 4.47747042926494e-05, 'epoch': 0.13}

 96%|█████████▋| 3178/3300 [4:53:09<10:42,  5.27s/it]INFO 04-12 03:30:01 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3179/3300 [4:53:17<11:54,  5.91s/it]2026-04-12 03:30:08,188 | INFO | train_grpo_train | metrics_logged mode=train step=3179

                                                     
{'loss': 0.0176, 'grad_norm': 2.992551565170288, 'learning_rate': 3.6969696969696973e-07, 'num_tokens': 7239068.0, 'completions/mean_length': 234.75, 'completions/min_length': 224.0, 'completions/max_length': 241.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 234.75, 'completions/min_terminated_length': 224.0, 'completions/max_terminated_length': 241.0, 'rewards/meter/mean': 0.998660683631897, 'rewards/meter/std': 0.0007363483309745789, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9807692170143127, 'rewards/repeat_penalty/std': 0.03560846298933029, 'rewards/total_composite/mean': 0.9794449806213379, 'rewards/total_composite/std': 0.03524334356188774, 'reward': 0.9794449806213379, 'reward_std': 0.03524333983659744, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.050825342535972595, 'sampling/sampling_logp_difference/max': 3.3192310333251953, 'sampling/importance_sampling_ratio/min': 0.03618064150214195, 'sampling/importance_sampling_ratio/mean': 1.0093677043914795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37664885073900223, 'clip_ratio/low_mean': 0.00682031805627048, 'clip_ratio/low_min': 0.00682031805627048, 'clip_ratio/high_mean': 0.03241610305849463, 'clip_ratio/high_max': 0.03241610305849463, 'clip_ratio/region_mean': 0.03923642111476511, 'reward_total_mean': 0.9794449806213379, 'reward_meter_mean': 0.998660683631897, 'reward_meter_std': 0.0007363483309745789, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9807692170143127, 'reward_repeat_penalty_std': 0.03560846298933029, 'reward_total_composite_mean': 0.9794449806213379, 'reward_total_composite_std': 0.03524334356188774, 'epoch': 0.13}

 96%|█████████▋| 3179/3300 [4:53:17<11:54,  5.91s/it]INFO 04-12 03:30:08 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3180/3300 [4:53:22<11:29,  5.75s/it]2026-04-12 03:30:13,574 | INFO | train_grpo_train | metrics_logged mode=train step=3180

                                                     
{'loss': -0.0004, 'grad_norm': 0.22532886266708374, 'learning_rate': 3.666666666666667e-07, 'num_tokens': 7241707.0, 'completions/mean_length': 131.875, 'completions/min_length': 131.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.875, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9994198083877563, 'rewards/meter/std': 2.31738686125027e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994198083877563, 'rewards/total_composite/std': 2.31738686125027e-05, 'reward': 0.9994198083877563, 'reward_std': 2.3179616619017906e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010175392031669617, 'sampling/sampling_logp_difference/max': 1.3818755149841309, 'sampling/importance_sampling_ratio/min': 0.2511071562767029, 'sampling/importance_sampling_ratio/mean': 1.0032905340194702, 'sampling/importance_sampling_ratio/max': 1.3655613660812378, 'entropy': 0.08243910130113363, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.002848137926775962, 'clip_ratio/high_max': 0.002848137926775962, 'clip_ratio/region_mean': 0.002848137926775962, 'reward_total_mean': 0.9994198083877563, 'reward_meter_mean': 0.9994198083877563, 'reward_meter_std': 2.31738686125027e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994198083877563, 'reward_total_composite_std': 2.31738686125027e-05, 'epoch': 0.13}

 96%|█████████▋| 3180/3300 [4:53:22<11:29,  5.75s/it]INFO 04-12 03:30:13 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3181/3300 [4:53:28<11:27,  5.78s/it]2026-04-12 03:30:19,400 | INFO | train_grpo_train | metrics_logged mode=train step=3181

                                                     
{'loss': -0.0034, 'grad_norm': 0.8209965825080872, 'learning_rate': 3.6363636363636366e-07, 'num_tokens': 7244435.0, 'completions/mean_length': 165.0, 'completions/min_length': 162.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.0, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9993767738342285, 'rewards/meter/std': 0.00010731098882388324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993767738342285, 'rewards/total_composite/std': 0.00010731098882388324, 'reward': 0.9993767738342285, 'reward_std': 0.00010731603106250986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014372244477272034, 'sampling/sampling_logp_difference/max': 1.1659832000732422, 'sampling/importance_sampling_ratio/min': 0.31161612272262573, 'sampling/importance_sampling_ratio/mean': 1.0039643049240112, 'sampling/importance_sampling_ratio/max': 1.6672699451446533, 'entropy': 0.12881212681531906, 'clip_ratio/low_mean': 0.0030676000751554966, 'clip_ratio/low_min': 0.0030676000751554966, 'clip_ratio/high_mean': 0.004531931248493493, 'clip_ratio/high_max': 0.004531931248493493, 'clip_ratio/region_mean': 0.007599531323648989, 'reward_total_mean': 0.9993767738342285, 'reward_meter_mean': 0.9993767738342285, 'reward_meter_std': 0.00010731098882388324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993767738342285, 'reward_total_composite_std': 0.00010731098882388324, 'epoch': 0.13}

 96%|█████████▋| 3181/3300 [4:53:28<11:27,  5.78s/it]INFO 04-12 03:30:19 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3182/3300 [4:53:32<10:40,  5.43s/it]2026-04-12 03:30:24,043 | INFO | train_grpo_train | metrics_logged mode=train step=3182

                                                     
{'loss': -0.0002, 'grad_norm': 1.5534006357192993, 'learning_rate': 3.6060606060606065e-07, 'num_tokens': 7246300.0, 'completions/mean_length': 68.125, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9994250535964966, 'rewards/meter/std': 0.00022647925652563572, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994250535964966, 'rewards/total_composite/std': 0.00022647925652563572, 'reward': 0.9994250535964966, 'reward_std': 0.0002264756039949134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007597931195050478, 'sampling/sampling_logp_difference/max': 0.6480650901794434, 'sampling/importance_sampling_ratio/min': 0.5230568647384644, 'sampling/importance_sampling_ratio/mean': 1.0001282691955566, 'sampling/importance_sampling_ratio/max': 1.234535813331604, 'entropy': 0.05426653893664479, 'clip_ratio/low_mean': 0.005434782709926367, 'clip_ratio/low_min': 0.005434782709926367, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.005434782709926367, 'reward_total_mean': 0.9994250535964966, 'reward_meter_mean': 0.9994250535964966, 'reward_meter_std': 0.00022647925652563572, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994250535964966, 'reward_total_composite_std': 0.00022647925652563572, 'epoch': 0.13}

 96%|█████████▋| 3182/3300 [4:53:33<10:40,  5.43s/it]INFO 04-12 03:30:24 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3183/3300 [4:53:36<09:47,  5.02s/it]2026-04-12 03:30:28,090 | INFO | train_grpo_train | metrics_logged mode=train step=3183

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.5757575757575764e-07, 'num_tokens': 7247884.0, 'completions/mean_length': 35.0, 'completions/min_length': 35.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9980231523513794, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9980231523513794, 'rewards/total_composite/std': 0.0, 'reward': 0.9980231523513794, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00043650183943100274, 'sampling/sampling_logp_difference/max': 0.008478658273816109, 'sampling/importance_sampling_ratio/min': 1.000000238418579, 'sampling/importance_sampling_ratio/mean': 1.0004373788833618, 'sampling/importance_sampling_ratio/max': 1.0085147619247437, 'entropy': 0.0031260672258213162, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9980231523513794, 'reward_meter_mean': 0.9980231523513794, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9980231523513794, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 96%|█████████▋| 3183/3300 [4:53:37<09:47,  5.02s/it]INFO 04-12 03:30:28 [block_pool.py:378] Successfully reset prefix cache

 96%|█████████▋| 3184/3300 [4:53:41<09:20,  4.83s/it]2026-04-12 03:30:32,493 | INFO | train_grpo_train | metrics_logged mode=train step=3184

                                                     
{'loss': -0.0004, 'grad_norm': 0.10775468498468399, 'learning_rate': 3.545454545454546e-07, 'num_tokens': 7249747.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981533288955688, 'rewards/meter/std': 3.6652184007834876e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981533288955688, 'rewards/total_composite/std': 3.6652184007834876e-06, 'reward': 0.9981533288955688, 'reward_std': 3.6725155041494872e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007368480786681175, 'sampling/sampling_logp_difference/max': 0.9039099216461182, 'sampling/importance_sampling_ratio/min': 0.4049831032752991, 'sampling/importance_sampling_ratio/mean': 1.000977635383606, 'sampling/importance_sampling_ratio/max': 1.437882661819458, 'entropy': 0.04700829554349184, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018656715983524919, 'reward_total_mean': 0.9981533288955688, 'reward_meter_mean': 0.9981533288955688, 'reward_meter_std': 3.6652184007834876e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981533288955688, 'reward_total_composite_std': 3.6652184007834876e-06, 'epoch': 0.13}

 96%|█████████▋| 3184/3300 [4:53:41<09:20,  4.83s/it]INFO 04-12 03:30:32 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3185/3300 [4:53:45<09:08,  4.77s/it]2026-04-12 03:30:37,108 | INFO | train_grpo_train | metrics_logged mode=train step=3185

                                                     
{'loss': 0.0221, 'grad_norm': 5.8911333084106445, 'learning_rate': 3.515151515151515e-07, 'num_tokens': 7251597.0, 'completions/mean_length': 66.25, 'completions/min_length': 65.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9895763397216797, 'rewards/meter/std': 0.009416628628969193, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9895763397216797, 'rewards/total_composite/std': 0.009416628628969193, 'reward': 0.9895763397216797, 'reward_std': 0.00941664818674326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02673407271504402, 'sampling/sampling_logp_difference/max': 1.822418212890625, 'sampling/importance_sampling_ratio/min': 0.1616344153881073, 'sampling/importance_sampling_ratio/mean': 1.0064246654510498, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15871062874794006, 'clip_ratio/low_mean': 0.0071428571827709675, 'clip_ratio/low_min': 0.0071428571827709675, 'clip_ratio/high_mean': 0.013145374483428895, 'clip_ratio/high_max': 0.013145374483428895, 'clip_ratio/region_mean': 0.020288231666199863, 'reward_total_mean': 0.9895763397216797, 'reward_meter_mean': 0.9895763397216797, 'reward_meter_std': 0.009416628628969193, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9895763397216797, 'reward_total_composite_std': 0.009416628628969193, 'epoch': 0.13}

 97%|█████████▋| 3185/3300 [4:53:46<09:08,  4.77s/it]INFO 04-12 03:30:37 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3186/3300 [4:53:50<08:49,  4.64s/it]2026-04-12 03:30:41,449 | INFO | train_grpo_train | metrics_logged mode=train step=3186

                                                     
{'loss': 0.0207, 'grad_norm': 7.721897125244141, 'learning_rate': 3.4848484848484856e-07, 'num_tokens': 7253219.0, 'completions/mean_length': 45.75, 'completions/min_length': 43.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9423834681510925, 'rewards/meter/std': 0.005602226592600346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9423834681510925, 'rewards/total_composite/std': 0.005602226592600346, 'reward': 0.9423834681510925, 'reward_std': 0.005602239165455103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047884196043014526, 'sampling/sampling_logp_difference/max': 1.721750020980835, 'sampling/importance_sampling_ratio/min': 0.1787530481815338, 'sampling/importance_sampling_ratio/mean': 1.0041768550872803, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14682762045413256, 'clip_ratio/low_mean': 0.005434782709926367, 'clip_ratio/low_min': 0.005434782709926367, 'clip_ratio/high_mean': 0.022629241226240993, 'clip_ratio/high_max': 0.022629241226240993, 'clip_ratio/region_mean': 0.02806402393616736, 'reward_total_mean': 0.9423834681510925, 'reward_meter_mean': 0.9423834681510925, 'reward_meter_std': 0.005602226592600346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9423834681510925, 'reward_total_composite_std': 0.005602226592600346, 'epoch': 0.13}

 97%|█████████▋| 3186/3300 [4:53:50<08:49,  4.64s/it]INFO 04-12 03:30:41 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3187/3300 [4:53:55<08:45,  4.65s/it]2026-04-12 03:30:46,133 | INFO | train_grpo_train | metrics_logged mode=train step=3187

                                                     
{'loss': 0.0047, 'grad_norm': 1.325953483581543, 'learning_rate': 3.454545454545455e-07, 'num_tokens': 7255248.0, 'completions/mean_length': 78.625, 'completions/min_length': 78.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9991061091423035, 'rewards/meter/std': 8.766540122451261e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991061091423035, 'rewards/total_composite/std': 8.766540122451261e-05, 'reward': 0.9991061091423035, 'reward_std': 8.767841063672677e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01794569194316864, 'sampling/sampling_logp_difference/max': 0.9910240173339844, 'sampling/importance_sampling_ratio/min': 0.3711963891983032, 'sampling/importance_sampling_ratio/mean': 1.0035501718521118, 'sampling/importance_sampling_ratio/max': 1.395779013633728, 'entropy': 0.17396301217377186, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.00957481365185231, 'clip_ratio/high_max': 0.00957481365185231, 'clip_ratio/region_mean': 0.00957481365185231, 'reward_total_mean': 0.9991061091423035, 'reward_meter_mean': 0.9991061091423035, 'reward_meter_std': 8.766540122451261e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991061091423035, 'reward_total_composite_std': 8.766540122451261e-05, 'epoch': 0.13}

 97%|█████████▋| 3187/3300 [4:53:55<08:45,  4.65s/it]INFO 04-12 03:30:46 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3188/3300 [4:54:00<08:53,  4.76s/it]2026-04-12 03:30:51,148 | INFO | train_grpo_train | metrics_logged mode=train step=3188

                                                     
{'loss': 0.0022, 'grad_norm': 2.738718032836914, 'learning_rate': 3.4242424242424243e-07, 'num_tokens': 7257350.0, 'completions/mean_length': 100.75, 'completions/min_length': 99.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.75, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.9992691278457642, 'rewards/meter/std': 0.00017112072964664549, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9742865562438965, 'rewards/total_composite/std': 0.07064837217330933, 'reward': 0.9742865562438965, 'reward_std': 0.07064837962388992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018476763740181923, 'sampling/sampling_logp_difference/max': 1.5925981998443604, 'sampling/importance_sampling_ratio/min': 0.20339646935462952, 'sampling/importance_sampling_ratio/mean': 1.0054959058761597, 'sampling/importance_sampling_ratio/max': 1.7092481851577759, 'entropy': 0.1316568348556757, 'clip_ratio/low_mean': 0.0012499999720603228, 'clip_ratio/low_min': 0.0012499999720603228, 'clip_ratio/high_mean': 0.014866774436086416, 'clip_ratio/high_max': 0.014866774436086416, 'clip_ratio/region_mean': 0.01611677440814674, 'reward_total_mean': 0.9742865562438965, 'reward_meter_mean': 0.9992691278457642, 'reward_meter_std': 0.00017112072964664549, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9742865562438965, 'reward_total_composite_std': 0.07064837217330933, 'epoch': 0.13}

 97%|█████████▋| 3188/3300 [4:54:00<08:53,  4.76s/it]INFO 04-12 03:30:51 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3189/3300 [4:54:04<08:27,  4.57s/it]2026-04-12 03:30:55,267 | INFO | train_grpo_train | metrics_logged mode=train step=3189

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.393939393939395e-07, 'num_tokens': 7258606.0, 'completions/mean_length': 34.0, 'completions/min_length': 34.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9923644065856934, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9923644065856934, 'rewards/total_composite/std': 0.0, 'reward': 0.9923644065856934, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.003486229106783867, 'sampling/sampling_logp_difference/max': 0.09297092258930206, 'sampling/importance_sampling_ratio/min': 0.961505115032196, 'sampling/importance_sampling_ratio/mean': 1.0026298761367798, 'sampling/importance_sampling_ratio/max': 1.097429871559143, 'entropy': 0.03649881505407393, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9923644065856934, 'reward_meter_mean': 0.9923644065856934, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9923644065856934, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 97%|█████████▋| 3189/3300 [4:54:04<08:27,  4.57s/it]INFO 04-12 03:30:55 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3190/3300 [4:54:08<08:21,  4.56s/it]2026-04-12 03:30:59,819 | INFO | train_grpo_train | metrics_logged mode=train step=3190

                                                     
{'loss': 0.0015, 'grad_norm': 2.772718906402588, 'learning_rate': 3.363636363636364e-07, 'num_tokens': 7260528.0, 'completions/mean_length': 68.25, 'completions/min_length': 68.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9994656443595886, 'rewards/meter/std': 8.461968536721542e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994656443595886, 'rewards/total_composite/std': 8.461968536721542e-05, 'reward': 0.9994656443595886, 'reward_std': 8.462517871521413e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.00850264448672533, 'sampling/sampling_logp_difference/max': 0.919212818145752, 'sampling/importance_sampling_ratio/min': 0.39883288741111755, 'sampling/importance_sampling_ratio/mean': 1.003930926322937, 'sampling/importance_sampling_ratio/max': 1.5606142282485962, 'entropy': 0.052496086340397596, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0018382353009656072, 'reward_total_mean': 0.9994656443595886, 'reward_meter_mean': 0.9994656443595886, 'reward_meter_std': 8.461968536721542e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994656443595886, 'reward_total_composite_std': 8.461968536721542e-05, 'epoch': 0.13}

 97%|█████████▋| 3190/3300 [4:54:08<08:21,  4.56s/it]INFO 04-12 03:31:00 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3191/3300 [4:54:13<08:14,  4.54s/it]2026-04-12 03:31:04,290 | INFO | train_grpo_train | metrics_logged mode=train step=3191

                                                     
{'loss': 0.0002, 'grad_norm': 0.23446327447891235, 'learning_rate': 3.3333333333333335e-07, 'num_tokens': 7262336.0, 'completions/mean_length': 71.0, 'completions/min_length': 71.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9994317293167114, 'rewards/meter/std': 1.128076564782532e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994317293167114, 'rewards/total_composite/std': 1.128076564782532e-05, 'reward': 0.9994317293167114, 'reward_std': 1.127942687162431e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007548379711806774, 'sampling/sampling_logp_difference/max': 0.2933727502822876, 'sampling/importance_sampling_ratio/min': 0.7457441091537476, 'sampling/importance_sampling_ratio/mean': 1.0047414302825928, 'sampling/importance_sampling_ratio/max': 1.223402500152588, 'entropy': 0.08830153662711382, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9994317293167114, 'reward_meter_mean': 0.9994317293167114, 'reward_meter_std': 1.128076564782532e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994317293167114, 'reward_total_composite_std': 1.128076564782532e-05, 'epoch': 0.13}

 97%|█████████▋| 3191/3300 [4:54:13<08:14,  4.54s/it]INFO 04-12 03:31:04 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3192/3300 [4:54:17<08:01,  4.45s/it]2026-04-12 03:31:08,560 | INFO | train_grpo_train | metrics_logged mode=train step=3192

                                                     
{'loss': 0.0136, 'grad_norm': 5.140721321105957, 'learning_rate': 3.303030303030303e-07, 'num_tokens': 7263826.0, 'completions/mean_length': 34.25, 'completions/min_length': 34.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9904930591583252, 'rewards/meter/std': 0.02051379904150963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9904930591583252, 'rewards/total_composite/std': 0.02051379904150963, 'reward': 0.9904930591583252, 'reward_std': 0.02051379904150963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006564823444932699, 'sampling/sampling_logp_difference/max': 0.8152861595153809, 'sampling/importance_sampling_ratio/min': 0.4425126910209656, 'sampling/importance_sampling_ratio/mean': 1.0013935565948486, 'sampling/importance_sampling_ratio/max': 1.33383309841156, 'entropy': 0.05123408907093108, 'clip_ratio/low_mean': 0.0069444444961845875, 'clip_ratio/low_min': 0.0069444444961845875, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0069444444961845875, 'reward_total_mean': 0.9904930591583252, 'reward_meter_mean': 0.9904930591583252, 'reward_meter_std': 0.02051379904150963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9904930591583252, 'reward_total_composite_std': 0.02051379904150963, 'epoch': 0.13}

 97%|█████████▋| 3192/3300 [4:54:17<08:01,  4.45s/it]INFO 04-12 03:31:08 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3193/3300 [4:54:21<07:55,  4.44s/it]2026-04-12 03:31:12,967 | INFO | train_grpo_train | metrics_logged mode=train step=3193

                                                     
{'loss': 0.0001, 'grad_norm': 0.0006765525904484093, 'learning_rate': 3.2727272727272733e-07, 'num_tokens': 7265538.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973390102386475, 'rewards/meter/std': 2.954577098535083e-07, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973390102386475, 'rewards/total_composite/std': 2.954577098535083e-07, 'reward': 0.9973390102386475, 'reward_std': 3.0140995477268007e-07, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0017861899686977267, 'sampling/sampling_logp_difference/max': 0.09160566329956055, 'sampling/importance_sampling_ratio/min': 0.9242770671844482, 'sampling/importance_sampling_ratio/mean': 1.0007272958755493, 'sampling/importance_sampling_ratio/max': 1.0959326028823853, 'entropy': 0.013490123208612204, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.0020491802133619785, 'reward_total_mean': 0.9973390102386475, 'reward_meter_mean': 0.9973390102386475, 'reward_meter_std': 2.954577098535083e-07, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973390102386475, 'reward_total_composite_std': 2.954577098535083e-07, 'epoch': 0.13}

 97%|█████████▋| 3193/3300 [4:54:21<07:55,  4.44s/it]INFO 04-12 03:31:13 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3194/3300 [4:54:27<08:17,  4.70s/it]2026-04-12 03:31:18,253 | INFO | train_grpo_train | metrics_logged mode=train step=3194

                                                     
{'loss': -0.0017, 'grad_norm': 0.48764580488204956, 'learning_rate': 3.2424242424242427e-07, 'num_tokens': 7267985.0, 'completions/mean_length': 131.875, 'completions/min_length': 131.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.875, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.999407172203064, 'rewards/meter/std': 5.462394256028347e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999407172203064, 'rewards/total_composite/std': 5.462394256028347e-05, 'reward': 0.999407172203064, 'reward_std': 5.462476474349387e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0077711548656225204, 'sampling/sampling_logp_difference/max': 0.6230001449584961, 'sampling/importance_sampling_ratio/min': 0.5363329648971558, 'sampling/importance_sampling_ratio/mean': 1.003739833831787, 'sampling/importance_sampling_ratio/max': 1.5405017137527466, 'entropy': 0.0831528240814805, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.006628788076341152, 'clip_ratio/high_max': 0.006628788076341152, 'clip_ratio/region_mean': 0.006628788076341152, 'reward_total_mean': 0.999407172203064, 'reward_meter_mean': 0.999407172203064, 'reward_meter_std': 5.462394256028347e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999407172203064, 'reward_total_composite_std': 5.462394256028347e-05, 'epoch': 0.13}

 97%|█████████▋| 3194/3300 [4:54:27<08:17,  4.70s/it]INFO 04-12 03:31:18 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3195/3300 [4:54:31<07:55,  4.53s/it]2026-04-12 03:31:22,414 | INFO | train_grpo_train | metrics_logged mode=train step=3195

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.212121212121212e-07, 'num_tokens': 7269473.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 6.992343696765602e-05, 'sampling/sampling_logp_difference/max': 0.0013757887063547969, 'sampling/importance_sampling_ratio/min': 0.9994438290596008, 'sampling/importance_sampling_ratio/mean': 1.0000648498535156, 'sampling/importance_sampling_ratio/max': 1.001376748085022, 'entropy': 0.000906125656911172, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 97%|█████████▋| 3195/3300 [4:54:31<07:55,  4.53s/it]INFO 04-12 03:31:22 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3196/3300 [4:54:40<10:07,  5.84s/it]2026-04-12 03:31:31,302 | INFO | train_grpo_train | metrics_logged mode=train step=3196

                                                     
{'loss': -0.0218, 'grad_norm': 1.7959932088851929, 'learning_rate': 3.181818181818182e-07, 'num_tokens': 7274421.0, 'completions/mean_length': 397.5, 'completions/min_length': 379.0, 'completions/max_length': 424.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 397.5, 'completions/min_terminated_length': 379.0, 'completions/max_terminated_length': 424.0, 'rewards/meter/mean': 0.9988911747932434, 'rewards/meter/std': 0.0005708896787837148, 'rewards/count_adherence/mean': 0.7884615659713745, 'rewards/count_adherence/std': 0.03560846298933029, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9937499761581421, 'rewards/repeat_penalty/std': 0.01767767407000065, 'rewards/total_composite/mean': 0.782776951789856, 'rewards/total_composite/std': 0.04057204723358154, 'reward': 0.782776951789856, 'reward_std': 0.04057204723358154, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05620747059583664, 'sampling/sampling_logp_difference/max': 1.8958301544189453, 'sampling/importance_sampling_ratio/min': 0.15019360184669495, 'sampling/importance_sampling_ratio/mean': 1.0130902528762817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5010012574493885, 'clip_ratio/low_mean': 0.017221659421920776, 'clip_ratio/low_min': 0.017221659421920776, 'clip_ratio/high_mean': 0.007785308640450239, 'clip_ratio/high_max': 0.007785308640450239, 'clip_ratio/region_mean': 0.025006968062371016, 'reward_total_mean': 0.782776951789856, 'reward_meter_mean': 0.9988911747932434, 'reward_meter_std': 0.0005708896787837148, 'reward_count_adherence_mean': 0.7884615659713745, 'reward_count_adherence_std': 0.03560846298933029, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9937499761581421, 'reward_repeat_penalty_std': 0.01767767407000065, 'reward_total_composite_mean': 0.782776951789856, 'reward_total_composite_std': 0.04057204723358154, 'epoch': 0.13}

 97%|█████████▋| 3196/3300 [4:54:40<10:07,  5.84s/it]INFO 04-12 03:31:31 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3197/3300 [4:54:47<10:51,  6.33s/it]2026-04-12 03:31:38,772 | INFO | train_grpo_train | metrics_logged mode=train step=3197

                                                     
{'loss': 0.0732, 'grad_norm': 3.9520342350006104, 'learning_rate': 3.151515151515152e-07, 'num_tokens': 7278424.0, 'completions/mean_length': 277.375, 'completions/min_length': 254.0, 'completions/max_length': 313.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 277.375, 'completions/min_terminated_length': 254.0, 'completions/max_terminated_length': 313.0, 'rewards/meter/mean': 0.9972818493843079, 'rewards/meter/std': 0.001952509512193501, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9852941036224365, 'rewards/repeat_penalty/std': 0.04159451276063919, 'rewards/total_composite/mean': 0.953273355960846, 'rewards/total_composite/std': 0.08584806323051453, 'reward': 0.953273355960846, 'reward_std': 0.08584806323051453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06301229447126389, 'sampling/sampling_logp_difference/max': 1.7296597957611084, 'sampling/importance_sampling_ratio/min': 0.22898982465267181, 'sampling/importance_sampling_ratio/mean': 1.01344633102417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.570828665047884, 'clip_ratio/low_mean': 0.012125813402235508, 'clip_ratio/low_min': 0.012125813402235508, 'clip_ratio/high_mean': 0.036240562330931425, 'clip_ratio/high_max': 0.036240562330931425, 'clip_ratio/region_mean': 0.04836637573316693, 'reward_total_mean': 0.953273355960846, 'reward_meter_mean': 0.9972818493843079, 'reward_meter_std': 0.001952509512193501, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9852941036224365, 'reward_repeat_penalty_std': 0.04159451276063919, 'reward_total_composite_mean': 0.953273355960846, 'reward_total_composite_std': 0.08584806323051453, 'epoch': 0.13}

 97%|█████████▋| 3197/3300 [4:54:47<10:51,  6.33s/it]INFO 04-12 03:31:38 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3198/3300 [4:54:51<09:44,  5.73s/it]2026-04-12 03:31:43,120 | INFO | train_grpo_train | metrics_logged mode=train step=3198

                                                     
{'loss': 0.0029, 'grad_norm': 11.894042015075684, 'learning_rate': 3.1212121212121213e-07, 'num_tokens': 7280073.0, 'completions/mean_length': 46.125, 'completions/min_length': 44.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9368916749954224, 'rewards/meter/std': 0.018891815096139908, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9368916749954224, 'rewards/total_composite/std': 0.018891815096139908, 'reward': 0.9368916749954224, 'reward_std': 0.018891818821430206, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03293498232960701, 'sampling/sampling_logp_difference/max': 2.211425304412842, 'sampling/importance_sampling_ratio/min': 0.10954440385103226, 'sampling/importance_sampling_ratio/mean': 1.0053000450134277, 'sampling/importance_sampling_ratio/max': 1.6582306623458862, 'entropy': 0.14615605771541595, 'clip_ratio/low_mean': 0.013334879651665688, 'clip_ratio/low_min': 0.013334879651665688, 'clip_ratio/high_mean': 0.019021739484742284, 'clip_ratio/high_max': 0.019021739484742284, 'clip_ratio/region_mean': 0.03235661913640797, 'reward_total_mean': 0.9368916749954224, 'reward_meter_mean': 0.9368916749954224, 'reward_meter_std': 0.018891815096139908, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9368916749954224, 'reward_total_composite_std': 0.018891815096139908, 'epoch': 0.13}

 97%|█████████▋| 3198/3300 [4:54:52<09:44,  5.73s/it]INFO 04-12 03:31:43 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3199/3300 [4:54:58<09:56,  5.91s/it]2026-04-12 03:31:49,442 | INFO | train_grpo_train | metrics_logged mode=train step=3199

                                                     
{'loss': 0.0104, 'grad_norm': 1.7074098587036133, 'learning_rate': 3.090909090909091e-07, 'num_tokens': 7283464.0, 'completions/mean_length': 201.875, 'completions/min_length': 189.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 201.875, 'completions/min_terminated_length': 189.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.8017416000366211, 'rewards/meter/std': 0.2793266177177429, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.96875, 'rewards/repeat_penalty/std': 0.0431290864944458, 'rewards/total_composite/mean': 0.7810623049736023, 'rewards/total_composite/std': 0.2836529314517975, 'reward': 0.7810623049736023, 'reward_std': 0.2836529314517975, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.032916199415922165, 'sampling/sampling_logp_difference/max': 4.614783763885498, 'sampling/importance_sampling_ratio/min': 0.009904325008392334, 'sampling/importance_sampling_ratio/mean': 1.007271647453308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31216721422970295, 'clip_ratio/low_mean': 0.00802644295617938, 'clip_ratio/low_min': 0.00802644295617938, 'clip_ratio/high_mean': 0.010650901531334966, 'clip_ratio/high_max': 0.010650901531334966, 'clip_ratio/region_mean': 0.018677344487514347, 'reward_total_mean': 0.7810623049736023, 'reward_meter_mean': 0.8017416000366211, 'reward_meter_std': 0.2793266177177429, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.96875, 'reward_repeat_penalty_std': 0.0431290864944458, 'reward_total_composite_mean': 0.7810623049736023, 'reward_total_composite_std': 0.2836529314517975, 'epoch': 0.13}

 97%|█████████▋| 3199/3300 [4:54:58<09:56,  5.91s/it]INFO 04-12 03:31:49 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3200/3300 [4:55:03<09:17,  5.57s/it]2026-04-12 03:31:54,214 | INFO | train_grpo_train | metrics_logged mode=train step=3200

                                                     
{'loss': 0.0, 'grad_norm': 0.4144695997238159, 'learning_rate': 3.0606060606060606e-07, 'num_tokens': 7285520.0, 'completions/mean_length': 93.0, 'completions/min_length': 93.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.0, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9977416396141052, 'rewards/meter/std': 3.231431037420407e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977416396141052, 'rewards/total_composite/std': 3.231431037420407e-05, 'reward': 0.9977416396141052, 'reward_std': 3.2314139389200136e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010148909874260426, 'sampling/sampling_logp_difference/max': 0.6211330890655518, 'sampling/importance_sampling_ratio/min': 0.5517995953559875, 'sampling/importance_sampling_ratio/mean': 1.0042681694030762, 'sampling/importance_sampling_ratio/max': 1.8610354661941528, 'entropy': 0.07502706721425056, 'clip_ratio/low_mean': 0.0026881720405071974, 'clip_ratio/low_min': 0.0026881720405071974, 'clip_ratio/high_mean': 0.004032258060760796, 'clip_ratio/high_max': 0.004032258060760796, 'clip_ratio/region_mean': 0.0067204301012679935, 'reward_total_mean': 0.9977416396141052, 'reward_meter_mean': 0.9977416396141052, 'reward_meter_std': 3.231431037420407e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977416396141052, 'reward_total_composite_std': 3.231431037420407e-05, 'epoch': 0.13}

 97%|█████████▋| 3200/3300 [4:55:03<09:17,  5.57s/it]INFO 04-12 03:31:54 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:38,  3.51s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.41s/it]

 31%|███       | 4/13 [00:16<00:38,  4.30s/it]

 38%|███▊      | 5/13 [00:21<00:38,  4.77s/it]

 46%|████▌     | 6/13 [00:28<00:37,  5.37s/it]

 54%|█████▍    | 7/13 [00:33<00:32,  5.43s/it]

 62%|██████▏   | 8/13 [00:41<00:30,  6.02s/it]

 69%|██████▉   | 9/13 [00:48<00:25,  6.41s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.97s/it]

 85%|████████▍ | 11/13 [00:59<00:11,  5.91s/it]

 92%|█████████▏| 12/13 [01:05<00:05,  5.98s/it]

100%|██████████| 13/13 [01:10<00:00,  5.85s/it]2026-04-12 03:33:12,688 | INFO | train_grpo_train | metrics_logged mode=eval step=3200

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 78.3425, 'eval_samples_per_second': 1.328, 'eval_steps_per_second': 0.166, 'eval_num_tokens': 7285520.0, 'eval_completions/mean_length': 217.47115384615384, 'eval_completions/min_length': 60.76923076923077, 'eval_completions/max_length': 424.15384615384613, 'eval_completions/clipped_ratio': 0.04807692307692308, 'eval_completions/mean_terminated_length': 202.3791222205529, 'eval_completions/min_terminated_length': 60.76923076923077, 'eval_completions/max_terminated_length': 381.38461538461536, 'eval_rewards/meter/mean': 0.7966891756424537, 'eval_rewards/meter/std': 0.3122874472576838, 'eval_rewards/count_adherence/mean': 0.962191457931812, 'eval_rewards/count_adherence/std': 0.056447364103335604, 'eval_rewards/arabic_clean/mean': 0.9711538461538461, 'eval_rewards/arabic_clean/std': 0.06280488005051246, 'eval_rewards/repeat_penalty/mean': 0.9550704268308786, 'eval_rewards/repeat_penalty/std': 0.07092011161148548, 'eval_rewards/total_composite/mean': 0.716111809015274, 'eval_rewards/total_composite/std': 0.33144921350937623, 'eval_reward': 0.716111809015274, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03270428670713535, 'eval_sampling/sampling_logp_difference/max': 1.224547532888559, 'eval_sampling/importance_sampling_ratio/min': 0.3024700696651752, 'eval_sampling/importance_sampling_ratio/mean': 1.0096190892733061, 'eval_sampling/importance_sampling_ratio/max': 1.5303410291671753, 'eval_entropy': 0.38261800087415254, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.716111809015274, 'eval_reward_meter_mean': 0.7966891756424537, 'eval_reward_meter_std': 0.3122874472576838, 'eval_reward_count_adherence_mean': 0.962191457931812, 'eval_reward_count_adherence_std': 0.056447364103335604, 'eval_reward_arabic_clean_mean': 0.9711538461538461, 'eval_reward_arabic_clean_std': 0.06280488005051246, 'eval_reward_repeat_penalty_mean': 0.9550704268308786, 'eval_reward_repeat_penalty_std': 0.07092011161148548, 'eval_reward_total_composite_mean': 0.716111809015274, 'eval_reward_total_composite_std': 0.33144921350937623, 'epoch': 0.13}

 97%|█████████▋| 3200/3300 [4:56:21<09:17,  5.57s/it]

100%|██████████| 13/13 [01:11<00:00,  5.85s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 97%|█████████▋| 3201/3300 [4:56:29<49:18, 29.88s/it]2026-04-12 03:33:20,802 | INFO | train_grpo_train | metrics_logged mode=train step=3201

                                                     
{'loss': 0.0005, 'grad_norm': 0.3707689046859741, 'learning_rate': 3.0303030303030305e-07, 'num_tokens': 7287416.0, 'completions/mean_length': 71.0, 'completions/min_length': 71.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9994514584541321, 'rewards/meter/std': 2.538687112974003e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994514584541321, 'rewards/total_composite/std': 2.538687112974003e-05, 'reward': 0.9994514584541321, 'reward_std': 2.539291381253861e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009306100197136402, 'sampling/sampling_logp_difference/max': 0.6520857810974121, 'sampling/importance_sampling_ratio/min': 0.5209580659866333, 'sampling/importance_sampling_ratio/mean': 1.0068362951278687, 'sampling/importance_sampling_ratio/max': 1.66619074344635, 'entropy': 0.07758000679314137, 'clip_ratio/low_mean': 0.005281690042465925, 'clip_ratio/low_min': 0.005281690042465925, 'clip_ratio/high_mean': 0.0017605633474886417, 'clip_ratio/high_max': 0.0017605633474886417, 'clip_ratio/region_mean': 0.007042253389954567, 'reward_total_mean': 0.9994514584541321, 'reward_meter_mean': 0.9994514584541321, 'reward_meter_std': 2.538687112974003e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994514584541321, 'reward_total_composite_std': 2.538687112974003e-05, 'epoch': 0.13}

 97%|█████████▋| 3201/3300 [4:56:29<49:18, 29.88s/it]INFO 04-12 03:33:21 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3202/3300 [4:56:35<36:47, 22.52s/it]2026-04-12 03:33:26,158 | INFO | train_grpo_train | metrics_logged mode=train step=3202

                                                     
{'loss': -0.0017, 'grad_norm': 2.368190050125122, 'learning_rate': 3.0000000000000004e-07, 'num_tokens': 7289899.0, 'completions/mean_length': 128.375, 'completions/min_length': 127.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.375, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9979463815689087, 'rewards/meter/std': 3.4329503250773996e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.962304949760437, 'rewards/total_composite/std': 0.06598580628633499, 'reward': 0.962304949760437, 'reward_std': 0.0659857988357544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.017233658581972122, 'sampling/sampling_logp_difference/max': 1.2134406566619873, 'sampling/importance_sampling_ratio/min': 0.29717305302619934, 'sampling/importance_sampling_ratio/mean': 1.0038442611694336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.16156750172376633, 'clip_ratio/low_mean': 0.003929318394511938, 'clip_ratio/low_min': 0.003929318394511938, 'clip_ratio/high_mean': 0.01261215889826417, 'clip_ratio/high_max': 0.01261215889826417, 'clip_ratio/region_mean': 0.016541477292776108, 'reward_total_mean': 0.962304949760437, 'reward_meter_mean': 0.9979463815689087, 'reward_meter_std': 3.4329503250773996e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.962304949760437, 'reward_total_composite_std': 0.06598580628633499, 'epoch': 0.13}

 97%|█████████▋| 3202/3300 [4:56:35<36:47, 22.52s/it]INFO 04-12 03:33:26 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3203/3300 [4:56:42<29:03, 17.97s/it]2026-04-12 03:33:33,516 | INFO | train_grpo_train | metrics_logged mode=train step=3203

                                                     
{'loss': 0.0342, 'grad_norm': 2.0183863639831543, 'learning_rate': 2.96969696969697e-07, 'num_tokens': 7293996.0, 'completions/mean_length': 281.125, 'completions/min_length': 260.0, 'completions/max_length': 294.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 281.125, 'completions/min_terminated_length': 260.0, 'completions/max_terminated_length': 294.0, 'rewards/meter/mean': 0.9827648401260376, 'rewards/meter/std': 0.0459895133972168, 'rewards/count_adherence/mean': 0.921875, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9612745046615601, 'rewards/repeat_penalty/std': 0.045027956366539, 'rewards/total_composite/mean': 0.8708112835884094, 'rewards/total_composite/std': 0.08316829800605774, 'reward': 0.8708112835884094, 'reward_std': 0.08316829800605774, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040366996079683304, 'sampling/sampling_logp_difference/max': 2.540407419204712, 'sampling/importance_sampling_ratio/min': 0.07883427292108536, 'sampling/importance_sampling_ratio/mean': 1.0026259422302246, 'sampling/importance_sampling_ratio/max': 1.6892542839050293, 'entropy': 0.34676245227456093, 'clip_ratio/low_mean': 0.006470522261224687, 'clip_ratio/low_min': 0.006470522261224687, 'clip_ratio/high_mean': 0.01714910357259214, 'clip_ratio/high_max': 0.01714910357259214, 'clip_ratio/region_mean': 0.023619625833816826, 'reward_total_mean': 0.8708112835884094, 'reward_meter_mean': 0.9827648401260376, 'reward_meter_std': 0.0459895133972168, 'reward_count_adherence_mean': 0.921875, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9612745046615601, 'reward_repeat_penalty_std': 0.045027956366539, 'reward_total_composite_mean': 0.8708112835884094, 'reward_total_composite_std': 0.08316829800605774, 'epoch': 0.13}

 97%|█████████▋| 3203/3300 [4:56:42<29:03, 17.97s/it]INFO 04-12 03:33:33 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3204/3300 [4:56:47<22:20, 13.96s/it]2026-04-12 03:33:38,133 | INFO | train_grpo_train | metrics_logged mode=train step=3204

                                                     
{'loss': 0.0001, 'grad_norm': 0.02218056656420231, 'learning_rate': 2.9393939393939397e-07, 'num_tokens': 7295796.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973360300064087, 'rewards/meter/std': 8.113272997434251e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973360300064087, 'rewards/total_composite/std': 8.113272997434251e-06, 'reward': 0.9973360300064087, 'reward_std': 8.10424353403505e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002050558337941766, 'sampling/sampling_logp_difference/max': 0.22490382194519043, 'sampling/importance_sampling_ratio/min': 0.8462150692939758, 'sampling/importance_sampling_ratio/mean': 1.0012545585632324, 'sampling/importance_sampling_ratio/max': 1.2522022724151611, 'entropy': 0.014257052214816213, 'clip_ratio/low_mean': 0.0020491802133619785, 'clip_ratio/low_min': 0.0020491802133619785, 'clip_ratio/high_mean': 0.0020491802133619785, 'clip_ratio/high_max': 0.0020491802133619785, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.9973360300064087, 'reward_meter_mean': 0.9973360300064087, 'reward_meter_std': 8.113272997434251e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973360300064087, 'reward_total_composite_std': 8.113272997434251e-06, 'epoch': 0.13}

 97%|█████████▋| 3204/3300 [4:56:47<22:20, 13.96s/it]INFO 04-12 03:33:38 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3205/3300 [4:56:51<17:24, 11.00s/it]2026-04-12 03:33:42,207 | INFO | train_grpo_train | metrics_logged mode=train step=3205

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.9090909090909096e-07, 'num_tokens': 7297276.0, 'completions/mean_length': 28.0, 'completions/min_length': 28.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9943599104881287, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9943599104881287, 'rewards/total_composite/std': 0.0, 'reward': 0.9943599104881287, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0001377786829834804, 'sampling/sampling_logp_difference/max': 0.0014927273150533438, 'sampling/importance_sampling_ratio/min': 0.9986100196838379, 'sampling/importance_sampling_ratio/mean': 1.0001215934753418, 'sampling/importance_sampling_ratio/max': 1.0014938116073608, 'entropy': 0.0011580012142076157, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9943599104881287, 'reward_meter_mean': 0.9943599104881287, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9943599104881287, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 97%|█████████▋| 3205/3300 [4:56:51<17:24, 11.00s/it]INFO 04-12 03:33:42 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3206/3300 [4:56:56<14:31,  9.27s/it]2026-04-12 03:33:47,442 | INFO | train_grpo_train | metrics_logged mode=train step=3206

                                                     
{'loss': 0.0027, 'grad_norm': 2.458156108856201, 'learning_rate': 2.878787878787879e-07, 'num_tokens': 7299630.0, 'completions/mean_length': 127.25, 'completions/min_length': 125.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.25, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9723162651062012, 'rewards/meter/std': 0.04477069899439812, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9546165466308594, 'rewards/total_composite/std': 0.06121518090367317, 'reward': 0.9546165466308594, 'reward_std': 0.06121518462896347, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.033622466027736664, 'sampling/sampling_logp_difference/max': 1.334040641784668, 'sampling/importance_sampling_ratio/min': 0.2634107768535614, 'sampling/importance_sampling_ratio/mean': 1.0102778673171997, 'sampling/importance_sampling_ratio/max': 1.842721700668335, 'entropy': 0.31193940714001656, 'clip_ratio/low_mean': 0.005936880130320787, 'clip_ratio/low_min': 0.005936880130320787, 'clip_ratio/high_mean': 0.012698987615294755, 'clip_ratio/high_max': 0.012698987615294755, 'clip_ratio/region_mean': 0.018635867745615542, 'reward_total_mean': 0.9546165466308594, 'reward_meter_mean': 0.9723162651062012, 'reward_meter_std': 0.04477069899439812, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9546165466308594, 'reward_total_composite_std': 0.06121518090367317, 'epoch': 0.13}

 97%|█████████▋| 3206/3300 [4:56:56<14:31,  9.27s/it]INFO 04-12 03:33:47 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3207/3300 [4:57:01<12:16,  7.92s/it]2026-04-12 03:33:52,230 | INFO | train_grpo_train | metrics_logged mode=train step=3207

                                                     
{'loss': 0.001, 'grad_norm': 3.696357250213623, 'learning_rate': 2.848484848484849e-07, 'num_tokens': 7301645.0, 'completions/mean_length': 92.875, 'completions/min_length': 92.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.875, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9976005554199219, 'rewards/meter/std': 0.0004543311079032719, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9976005554199219, 'rewards/total_composite/std': 0.0004543311079032719, 'reward': 0.9976005554199219, 'reward_std': 0.00045434184721671045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009967589750885963, 'sampling/sampling_logp_difference/max': 0.724400520324707, 'sampling/importance_sampling_ratio/min': 0.4846150279045105, 'sampling/importance_sampling_ratio/mean': 1.0031743049621582, 'sampling/importance_sampling_ratio/max': 1.5881195068359375, 'entropy': 0.07594481343403459, 'clip_ratio/low_mean': 0.0013440860202535987, 'clip_ratio/low_min': 0.0013440860202535987, 'clip_ratio/high_mean': 0.004061477375216782, 'clip_ratio/high_max': 0.004061477375216782, 'clip_ratio/region_mean': 0.005405563395470381, 'reward_total_mean': 0.9976005554199219, 'reward_meter_mean': 0.9976005554199219, 'reward_meter_std': 0.0004543311079032719, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9976005554199219, 'reward_total_composite_std': 0.0004543311079032719, 'epoch': 0.13}

 97%|█████████▋| 3207/3300 [4:57:01<12:16,  7.92s/it]INFO 04-12 03:33:52 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3208/3300 [4:57:05<10:42,  6.98s/it]2026-04-12 03:33:57,016 | INFO | train_grpo_train | metrics_logged mode=train step=3208

                                                     
{'loss': -0.0039, 'grad_norm': 2.1962215900421143, 'learning_rate': 2.818181818181819e-07, 'num_tokens': 7303658.0, 'completions/mean_length': 92.625, 'completions/min_length': 91.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.625, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9975332021713257, 'rewards/meter/std': 0.0004231163766235113, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9975332021713257, 'rewards/total_composite/std': 0.0004231163766235113, 'reward': 0.9975332021713257, 'reward_std': 0.00042310450226068497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010751576162874699, 'sampling/sampling_logp_difference/max': 0.9464244842529297, 'sampling/importance_sampling_ratio/min': 0.38812631368637085, 'sampling/importance_sampling_ratio/mean': 1.0051023960113525, 'sampling/importance_sampling_ratio/max': 1.477807641029358, 'entropy': 0.08394959289580584, 'clip_ratio/low_mean': 0.006838270695880055, 'clip_ratio/low_min': 0.006838270695880055, 'clip_ratio/high_mean': 0.0013440860202535987, 'clip_ratio/high_max': 0.0013440860202535987, 'clip_ratio/region_mean': 0.008182356716133654, 'reward_total_mean': 0.9975332021713257, 'reward_meter_mean': 0.9975332021713257, 'reward_meter_std': 0.0004231163766235113, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9975332021713257, 'reward_total_composite_std': 0.0004231163766235113, 'epoch': 0.13}

 97%|█████████▋| 3208/3300 [4:57:06<10:42,  6.98s/it]INFO 04-12 03:33:57 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3209/3300 [4:57:11<09:58,  6.58s/it]2026-04-12 03:34:02,646 | INFO | train_grpo_train | metrics_logged mode=train step=3209

                                                     
{'loss': 0.0075, 'grad_norm': 1.3126479387283325, 'learning_rate': 2.787878787878788e-07, 'num_tokens': 7306277.0, 'completions/mean_length': 159.375, 'completions/min_length': 158.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.375, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.9976499080657959, 'rewards/meter/std': 0.0005515830125659704, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9444444179534912, 'rewards/repeat_penalty/std': 0.059391383081674576, 'rewards/total_composite/mean': 0.9422216415405273, 'rewards/total_composite/std': 0.05919419974088669, 'reward': 0.9422216415405273, 'reward_std': 0.059194166213274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020425979048013687, 'sampling/sampling_logp_difference/max': 1.9696846008300781, 'sampling/importance_sampling_ratio/min': 0.13950084149837494, 'sampling/importance_sampling_ratio/mean': 1.0039074420928955, 'sampling/importance_sampling_ratio/max': 1.4307658672332764, 'entropy': 0.20060089603066444, 'clip_ratio/low_mean': 0.004707215121015906, 'clip_ratio/low_min': 0.004707215121015906, 'clip_ratio/high_mean': 0.00788172468310222, 'clip_ratio/high_max': 0.00788172468310222, 'clip_ratio/region_mean': 0.012588939804118127, 'reward_total_mean': 0.9422216415405273, 'reward_meter_mean': 0.9976499080657959, 'reward_meter_std': 0.0005515830125659704, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9444444179534912, 'reward_repeat_penalty_std': 0.059391383081674576, 'reward_total_composite_mean': 0.9422216415405273, 'reward_total_composite_std': 0.05919419974088669, 'epoch': 0.13}

 97%|█████████▋| 3209/3300 [4:57:11<09:58,  6.58s/it]INFO 04-12 03:34:02 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3210/3300 [4:57:16<08:56,  5.96s/it]2026-04-12 03:34:07,176 | INFO | train_grpo_train | metrics_logged mode=train step=3210

                                                     
{'loss': -0.0001, 'grad_norm': 0.4994388222694397, 'learning_rate': 2.757575757575758e-07, 'num_tokens': 7308194.0, 'completions/mean_length': 71.625, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994117021560669, 'rewards/meter/std': 7.800360617693514e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994117021560669, 'rewards/total_composite/std': 7.800360617693514e-05, 'reward': 0.9994117021560669, 'reward_std': 7.79872716520913e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.010017759166657925, 'sampling/sampling_logp_difference/max': 0.9018526077270508, 'sampling/importance_sampling_ratio/min': 0.4058171510696411, 'sampling/importance_sampling_ratio/mean': 1.0026097297668457, 'sampling/importance_sampling_ratio/max': 1.3537853956222534, 'entropy': 0.08111674338579178, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.0052327855955809355, 'clip_ratio/high_max': 0.0052327855955809355, 'clip_ratio/region_mean': 0.006968896719627082, 'reward_total_mean': 0.9994117021560669, 'reward_meter_mean': 0.9994117021560669, 'reward_meter_std': 7.800360617693514e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994117021560669, 'reward_total_composite_std': 7.800360617693514e-05, 'epoch': 0.13}

 97%|█████████▋| 3210/3300 [4:57:16<08:56,  5.96s/it]INFO 04-12 03:34:07 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3211/3300 [4:57:20<08:13,  5.55s/it]2026-04-12 03:34:11,751 | INFO | train_grpo_train | metrics_logged mode=train step=3211

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.7272727272727274e-07, 'num_tokens': 7310146.0, 'completions/mean_length': 80.0, 'completions/min_length': 80.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7190229296684265, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.7190229296684265, 'rewards/total_composite/std': 0.0, 'reward': 0.7190229296684265, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00030469015473499894, 'sampling/sampling_logp_difference/max': 0.007013600319623947, 'sampling/importance_sampling_ratio/min': 0.9983328580856323, 'sampling/importance_sampling_ratio/mean': 1.0002856254577637, 'sampling/importance_sampling_ratio/max': 1.0070383548736572, 'entropy': 0.0023747361556161195, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7190229296684265, 'reward_meter_mean': 0.7190229296684265, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.7190229296684265, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 97%|█████████▋| 3211/3300 [4:57:20<08:13,  5.55s/it]INFO 04-12 03:34:11 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3212/3300 [4:57:25<07:39,  5.22s/it]2026-04-12 03:34:16,200 | INFO | train_grpo_train | metrics_logged mode=train step=3212

                                                     
{'loss': -0.003, 'grad_norm': 0.7813161611557007, 'learning_rate': 2.6969696969696973e-07, 'num_tokens': 7311944.0, 'completions/mean_length': 66.75, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.998114824295044, 'rewards/meter/std': 6.83319303789176e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998114824295044, 'rewards/total_composite/std': 6.83319303789176e-05, 'reward': 0.998114824295044, 'reward_std': 6.832612416474149e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014992273412644863, 'sampling/sampling_logp_difference/max': 1.2304191589355469, 'sampling/importance_sampling_ratio/min': 0.29217007756233215, 'sampling/importance_sampling_ratio/mean': 0.9998994469642639, 'sampling/importance_sampling_ratio/max': 1.6796454191207886, 'entropy': 0.052677970845252275, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.009384893695823848, 'clip_ratio/high_max': 0.009384893695823848, 'clip_ratio/region_mean': 0.013172772596590221, 'reward_total_mean': 0.998114824295044, 'reward_meter_mean': 0.998114824295044, 'reward_meter_std': 6.83319303789176e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998114824295044, 'reward_total_composite_std': 6.83319303789176e-05, 'epoch': 0.13}

 97%|█████████▋| 3212/3300 [4:57:25<07:39,  5.22s/it]INFO 04-12 03:34:16 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3213/3300 [4:57:32<08:40,  5.98s/it]2026-04-12 03:34:23,953 | INFO | train_grpo_train | metrics_logged mode=train step=3213

                                                     
{'loss': 0.0308, 'grad_norm': 1.7816143035888672, 'learning_rate': 2.666666666666667e-07, 'num_tokens': 7315948.0, 'completions/mean_length': 315.5, 'completions/min_length': 307.0, 'completions/max_length': 338.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 315.5, 'completions/min_terminated_length': 307.0, 'completions/max_terminated_length': 338.0, 'rewards/meter/mean': 0.9991602897644043, 'rewards/meter/std': 0.00017255292914342135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9921875, 'rewards/repeat_penalty/std': 0.022097086533904076, 'rewards/total_composite/mean': 0.9913522005081177, 'rewards/total_composite/std': 0.021975873038172722, 'reward': 0.9913522005081177, 'reward_std': 0.021975882351398468, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04883609339594841, 'sampling/sampling_logp_difference/max': 1.6247730255126953, 'sampling/importance_sampling_ratio/min': 0.19695636630058289, 'sampling/importance_sampling_ratio/mean': 1.0101171731948853, 'sampling/importance_sampling_ratio/max': 1.7197778224945068, 'entropy': 0.4627891331911087, 'clip_ratio/low_mean': 0.002218934940174222, 'clip_ratio/low_min': 0.002218934940174222, 'clip_ratio/high_mean': 0.030003347201272845, 'clip_ratio/high_max': 0.030003347201272845, 'clip_ratio/region_mean': 0.03222228214144707, 'reward_total_mean': 0.9913522005081177, 'reward_meter_mean': 0.9991602897644043, 'reward_meter_std': 0.00017255292914342135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9921875, 'reward_repeat_penalty_std': 0.022097086533904076, 'reward_total_composite_mean': 0.9913522005081177, 'reward_total_composite_std': 0.021975873038172722, 'epoch': 0.13}

 97%|█████████▋| 3213/3300 [4:57:32<08:40,  5.98s/it]INFO 04-12 03:34:24 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3214/3300 [4:57:38<08:31,  5.95s/it]2026-04-12 03:34:29,820 | INFO | train_grpo_train | metrics_logged mode=train step=3214

                                                     
{'loss': 0.0001, 'grad_norm': 0.841699481010437, 'learning_rate': 2.6363636363636366e-07, 'num_tokens': 7318839.0, 'completions/mean_length': 178.375, 'completions/min_length': 176.0, 'completions/max_length': 181.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.375, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.9990983009338379, 'rewards/meter/std': 0.0001289558131247759, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990983009338379, 'rewards/total_composite/std': 0.0001289558131247759, 'reward': 0.9990983009338379, 'reward_std': 0.00012895492545794696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018457449972629547, 'sampling/sampling_logp_difference/max': 1.2622642517089844, 'sampling/importance_sampling_ratio/min': 0.2830124795436859, 'sampling/importance_sampling_ratio/mean': 1.0072314739227295, 'sampling/importance_sampling_ratio/max': 1.48737633228302, 'entropy': 0.2543965280056, 'clip_ratio/low_mean': 0.003487827139906585, 'clip_ratio/low_min': 0.003487827139906585, 'clip_ratio/high_mean': 0.009855842916294932, 'clip_ratio/high_max': 0.009855842916294932, 'clip_ratio/region_mean': 0.013343670056201518, 'reward_total_mean': 0.9990983009338379, 'reward_meter_mean': 0.9990983009338379, 'reward_meter_std': 0.0001289558131247759, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990983009338379, 'reward_total_composite_std': 0.0001289558131247759, 'epoch': 0.13}

 97%|█████████▋| 3214/3300 [4:57:38<08:31,  5.95s/it]INFO 04-12 03:34:30 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3215/3300 [4:57:43<07:50,  5.54s/it]2026-04-12 03:34:34,394 | INFO | train_grpo_train | metrics_logged mode=train step=3215

                                                     
{'loss': 0.0005, 'grad_norm': 0.2847835123538971, 'learning_rate': 2.6060606060606065e-07, 'num_tokens': 7320585.0, 'completions/mean_length': 71.25, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994255304336548, 'rewards/meter/std': 1.9644281564978883e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994255304336548, 'rewards/total_composite/std': 1.9644281564978883e-05, 'reward': 0.9994255304336548, 'reward_std': 1.9626990251708776e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.012419765815138817, 'sampling/sampling_logp_difference/max': 1.4161691665649414, 'sampling/importance_sampling_ratio/min': 0.24264176189899445, 'sampling/importance_sampling_ratio/mean': 1.001387357711792, 'sampling/importance_sampling_ratio/max': 1.3071258068084717, 'entropy': 0.08346446044743061, 'clip_ratio/low_mean': 0.0035211266949772835, 'clip_ratio/low_min': 0.0035211266949772835, 'clip_ratio/high_mean': 0.007017801166512072, 'clip_ratio/high_max': 0.007017801166512072, 'clip_ratio/region_mean': 0.010538927861489356, 'reward_total_mean': 0.9994255304336548, 'reward_meter_mean': 0.9994255304336548, 'reward_meter_std': 1.9644281564978883e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994255304336548, 'reward_total_composite_std': 1.9644281564978883e-05, 'epoch': 0.13}

 97%|█████████▋| 3215/3300 [4:57:43<07:50,  5.54s/it]INFO 04-12 03:34:34 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3216/3300 [4:57:47<07:07,  5.09s/it]2026-04-12 03:34:38,445 | INFO | train_grpo_train | metrics_logged mode=train step=3216

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.575757575757576e-07, 'num_tokens': 7321969.0, 'completions/mean_length': 29.0, 'completions/min_length': 29.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9957436919212341, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9957436919212341, 'rewards/total_composite/std': 0.0, 'reward': 0.9957436919212341, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00043711578473448753, 'sampling/sampling_logp_difference/max': 0.022034302353858948, 'sampling/importance_sampling_ratio/min': 0.9782066345214844, 'sampling/importance_sampling_ratio/mean': 1.0002456903457642, 'sampling/importance_sampling_ratio/max': 1.0145117044448853, 'entropy': 0.003761589468922466, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9957436919212341, 'reward_meter_mean': 0.9957436919212341, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9957436919212341, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 97%|█████████▋| 3216/3300 [4:57:47<07:07,  5.09s/it]INFO 04-12 03:34:38 [block_pool.py:378] Successfully reset prefix cache

 97%|█████████▋| 3217/3300 [4:57:52<06:54,  5.00s/it]2026-04-12 03:34:43,226 | INFO | train_grpo_train | metrics_logged mode=train step=3217

                                                     
{'loss': 0.0004, 'grad_norm': 0.06737548857927322, 'learning_rate': 2.545454545454546e-07, 'num_tokens': 7324185.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994082450866699, 'rewards/meter/std': 3.3991077543760184e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994082450866699, 'rewards/total_composite/std': 3.3991077543760184e-06, 'reward': 0.9994082450866699, 'reward_std': 3.404888502700487e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003399990499019623, 'sampling/sampling_logp_difference/max': 0.3715987205505371, 'sampling/importance_sampling_ratio/min': 0.689630925655365, 'sampling/importance_sampling_ratio/mean': 1.0018640756607056, 'sampling/importance_sampling_ratio/max': 1.321240782737732, 'entropy': 0.035912956576794386, 'clip_ratio/low_mean': 0.0012755101779475808, 'clip_ratio/low_min': 0.0012755101779475808, 'clip_ratio/high_mean': 0.0012755101779475808, 'clip_ratio/high_max': 0.0012755101779475808, 'clip_ratio/region_mean': 0.0025510203558951616, 'reward_total_mean': 0.9994082450866699, 'reward_meter_mean': 0.9994082450866699, 'reward_meter_std': 3.3991077543760184e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994082450866699, 'reward_total_composite_std': 3.3991077543760184e-06, 'epoch': 0.13}

 97%|█████████▋| 3217/3300 [4:57:52<06:54,  5.00s/it]INFO 04-12 03:34:43 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3218/3300 [4:57:57<06:58,  5.11s/it]2026-04-12 03:34:48,590 | INFO | train_grpo_train | metrics_logged mode=train step=3218

                                                     
{'loss': -0.0009, 'grad_norm': 1.3049845695495605, 'learning_rate': 2.515151515151515e-07, 'num_tokens': 7326588.0, 'completions/mean_length': 141.375, 'completions/min_length': 139.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.375, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9991159439086914, 'rewards/meter/std': 0.00015829414769541472, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991159439086914, 'rewards/total_composite/std': 0.00015829414769541472, 'reward': 0.9991159439086914, 'reward_std': 0.00015829637413844466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020292429253458977, 'sampling/sampling_logp_difference/max': 0.8954935073852539, 'sampling/importance_sampling_ratio/min': 0.40840601921081543, 'sampling/importance_sampling_ratio/mean': 1.0080312490463257, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21374017372727394, 'clip_ratio/low_mean': 0.009702177019789815, 'clip_ratio/low_min': 0.009702177019789815, 'clip_ratio/high_mean': 0.0061869441997259855, 'clip_ratio/high_max': 0.0061869441997259855, 'clip_ratio/region_mean': 0.0158891212195158, 'reward_total_mean': 0.9991159439086914, 'reward_meter_mean': 0.9991159439086914, 'reward_meter_std': 0.00015829414769541472, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991159439086914, 'reward_total_composite_std': 0.00015829414769541472, 'epoch': 0.13}

 98%|█████████▊| 3218/3300 [4:57:57<06:58,  5.11s/it]INFO 04-12 03:34:48 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3219/3300 [4:58:02<06:53,  5.10s/it]2026-04-12 03:34:53,682 | INFO | train_grpo_train | metrics_logged mode=train step=3219

                                                     
{'loss': 0.0279, 'grad_norm': 5.071619510650635, 'learning_rate': 2.484848484848485e-07, 'num_tokens': 7328757.0, 'completions/mean_length': 105.125, 'completions/min_length': 102.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.125, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9915302991867065, 'rewards/meter/std': 0.006453692447394133, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9915302991867065, 'rewards/total_composite/std': 0.006453692447394133, 'reward': 0.9915302991867065, 'reward_std': 0.006453694310039282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0352078415453434, 'sampling/sampling_logp_difference/max': 0.7839181423187256, 'sampling/importance_sampling_ratio/min': 0.4566134214401245, 'sampling/importance_sampling_ratio/mean': 1.0104427337646484, 'sampling/importance_sampling_ratio/max': 1.9355528354644775, 'entropy': 0.3442672826349735, 'clip_ratio/low_mean': 0.0102413734421134, 'clip_ratio/low_min': 0.0102413734421134, 'clip_ratio/high_mean': 0.017989110434427857, 'clip_ratio/high_max': 0.017989110434427857, 'clip_ratio/region_mean': 0.028230483876541257, 'reward_total_mean': 0.9915302991867065, 'reward_meter_mean': 0.9915302991867065, 'reward_meter_std': 0.006453692447394133, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9915302991867065, 'reward_total_composite_std': 0.006453692447394133, 'epoch': 0.13}

 98%|█████████▊| 3219/3300 [4:58:02<06:53,  5.10s/it]INFO 04-12 03:34:53 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3220/3300 [4:58:07<06:35,  4.94s/it]2026-04-12 03:34:58,252 | INFO | train_grpo_train | metrics_logged mode=train step=3220

                                                     
{'loss': 0.0085, 'grad_norm': 3.3606433868408203, 'learning_rate': 2.4545454545454545e-07, 'num_tokens': 7330682.0, 'completions/mean_length': 68.625, 'completions/min_length': 65.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9950666427612305, 'rewards/meter/std': 0.0027368878945708275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9950666427612305, 'rewards/total_composite/std': 0.0027368878945708275, 'reward': 0.9950666427612305, 'reward_std': 0.0027369032613933086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.024690676480531693, 'sampling/sampling_logp_difference/max': 0.7779765129089355, 'sampling/importance_sampling_ratio/min': 0.45933452248573303, 'sampling/importance_sampling_ratio/mean': 1.0102657079696655, 'sampling/importance_sampling_ratio/max': 1.609100103378296, 'entropy': 0.21459667198359966, 'clip_ratio/low_mean': 0.005359299597330391, 'clip_ratio/low_min': 0.005359299597330391, 'clip_ratio/high_mean': 0.014640397042967379, 'clip_ratio/high_max': 0.014640397042967379, 'clip_ratio/region_mean': 0.01999969664029777, 'reward_total_mean': 0.9950666427612305, 'reward_meter_mean': 0.9950666427612305, 'reward_meter_std': 0.0027368878945708275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9950666427612305, 'reward_total_composite_std': 0.0027368878945708275, 'epoch': 0.13}

 98%|█████████▊| 3220/3300 [4:58:07<06:35,  4.94s/it]INFO 04-12 03:34:58 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3221/3300 [4:58:11<06:22,  4.85s/it]2026-04-12 03:35:02,871 | INFO | train_grpo_train | metrics_logged mode=train step=3221

                                                     
{'loss': 0.0021, 'grad_norm': 3.8582441806793213, 'learning_rate': 2.4242424242424244e-07, 'num_tokens': 7332402.0, 'completions/mean_length': 71.0, 'completions/min_length': 69.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9949785470962524, 'rewards/meter/std': 0.0019115530885756016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9949785470962524, 'rewards/total_composite/std': 0.0019115530885756016, 'reward': 0.9949785470962524, 'reward_std': 0.0019115599570795894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03391466289758682, 'sampling/sampling_logp_difference/max': 1.395737648010254, 'sampling/importance_sampling_ratio/min': 0.24765028059482574, 'sampling/importance_sampling_ratio/mean': 1.0079292058944702, 'sampling/importance_sampling_ratio/max': 1.7042900323867798, 'entropy': 0.2918657101690769, 'clip_ratio/low_mean': 0.024722653324715793, 'clip_ratio/low_min': 0.024722653324715793, 'clip_ratio/high_mean': 0.012232337845489383, 'clip_ratio/high_max': 0.012232337845489383, 'clip_ratio/region_mean': 0.036954991170205176, 'reward_total_mean': 0.9949785470962524, 'reward_meter_mean': 0.9949785470962524, 'reward_meter_std': 0.0019115530885756016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9949785470962524, 'reward_total_composite_std': 0.0019115530885756016, 'epoch': 0.13}

 98%|█████████▊| 3221/3300 [4:58:11<06:22,  4.85s/it]INFO 04-12 03:35:03 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3222/3300 [4:58:17<06:48,  5.24s/it]2026-04-12 03:35:09,039 | INFO | train_grpo_train | metrics_logged mode=train step=3222

                                                     
{'loss': 0.0076, 'grad_norm': 2.1812803745269775, 'learning_rate': 2.3939393939393943e-07, 'num_tokens': 7335408.0, 'completions/mean_length': 197.75, 'completions/min_length': 194.0, 'completions/max_length': 205.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.75, 'completions/min_terminated_length': 194.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.9989155530929565, 'rewards/meter/std': 0.00026041135424748063, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989155530929565, 'rewards/total_composite/std': 0.00026041135424748063, 'reward': 0.9989155530929565, 'reward_std': 0.0002604102483019233, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04675072431564331, 'sampling/sampling_logp_difference/max': 1.7516670227050781, 'sampling/importance_sampling_ratio/min': 0.17348448932170868, 'sampling/importance_sampling_ratio/mean': 1.0105738639831543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36750227957963943, 'clip_ratio/low_mean': 0.018319292226806283, 'clip_ratio/low_min': 0.018319292226806283, 'clip_ratio/high_mean': 0.01708377245813608, 'clip_ratio/high_max': 0.01708377245813608, 'clip_ratio/region_mean': 0.035403064684942365, 'reward_total_mean': 0.9989155530929565, 'reward_meter_mean': 0.9989155530929565, 'reward_meter_std': 0.00026041135424748063, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989155530929565, 'reward_total_composite_std': 0.00026041135424748063, 'epoch': 0.13}

 98%|█████████▊| 3222/3300 [4:58:18<06:48,  5.24s/it]INFO 04-12 03:35:09 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3223/3300 [4:58:22<06:21,  4.95s/it]2026-04-12 03:35:13,312 | INFO | train_grpo_train | metrics_logged mode=train step=3223

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.3636363636363637e-07, 'num_tokens': 7336960.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00022982771042734385, 'sampling/sampling_logp_difference/max': 0.006922472268342972, 'sampling/importance_sampling_ratio/min': 0.9931014776229858, 'sampling/importance_sampling_ratio/mean': 1.0001906156539917, 'sampling/importance_sampling_ratio/max': 1.0026967525482178, 'entropy': 0.0017184649041155353, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 98%|█████████▊| 3223/3300 [4:58:22<06:21,  4.95s/it]INFO 04-12 03:35:13 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3224/3300 [4:58:26<06:09,  4.86s/it]2026-04-12 03:35:17,971 | INFO | train_grpo_train | metrics_logged mode=train step=3224

                                                     
{'loss': -0.0014, 'grad_norm': 2.2346813678741455, 'learning_rate': 2.3333333333333336e-07, 'num_tokens': 7339001.0, 'completions/mean_length': 78.125, 'completions/min_length': 77.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.125, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9987478256225586, 'rewards/meter/std': 0.000937741482630372, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987478256225586, 'rewards/total_composite/std': 0.000937741482630372, 'reward': 0.9987478256225586, 'reward_std': 0.0009377306560054421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016266027465462685, 'sampling/sampling_logp_difference/max': 0.7468376159667969, 'sampling/importance_sampling_ratio/min': 0.4738627076148987, 'sampling/importance_sampling_ratio/mean': 1.0025646686553955, 'sampling/importance_sampling_ratio/max': 1.5145747661590576, 'entropy': 0.1890299841761589, 'clip_ratio/low_mean': 0.004746835213154554, 'clip_ratio/low_min': 0.004746835213154554, 'clip_ratio/high_mean': 0.006410256493836641, 'clip_ratio/high_max': 0.006410256493836641, 'clip_ratio/region_mean': 0.011157091706991196, 'reward_total_mean': 0.9987478256225586, 'reward_meter_mean': 0.9987478256225586, 'reward_meter_std': 0.000937741482630372, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987478256225586, 'reward_total_composite_std': 0.000937741482630372, 'epoch': 0.13}

 98%|█████████▊| 3224/3300 [4:58:26<06:09,  4.86s/it]INFO 04-12 03:35:18 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3225/3300 [4:58:31<05:55,  4.74s/it]2026-04-12 03:35:22,417 | INFO | train_grpo_train | metrics_logged mode=train step=3225

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.3030303030303032e-07, 'num_tokens': 7340889.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 6.631435098825023e-05, 'sampling/sampling_logp_difference/max': 0.005175924859941006, 'sampling/importance_sampling_ratio/min': 0.9948374629020691, 'sampling/importance_sampling_ratio/mean': 1.0000407695770264, 'sampling/importance_sampling_ratio/max': 1.0014703273773193, 'entropy': 0.0005534949777938891, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 98%|█████████▊| 3225/3300 [4:58:31<05:55,  4.74s/it]INFO 04-12 03:35:22 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3226/3300 [4:58:35<05:41,  4.61s/it]2026-04-12 03:35:26,738 | INFO | train_grpo_train | metrics_logged mode=train step=3226

                                                     
{'loss': 0.0043, 'grad_norm': 5.847517967224121, 'learning_rate': 2.2727272727272729e-07, 'num_tokens': 7342490.0, 'completions/mean_length': 46.125, 'completions/min_length': 46.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9431781768798828, 'rewards/meter/std': 0.0021464419551193714, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9431781768798828, 'rewards/total_composite/std': 0.0021464419551193714, 'reward': 0.9431781768798828, 'reward_std': 0.002146440092474222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020598582923412323, 'sampling/sampling_logp_difference/max': 1.1673927307128906, 'sampling/importance_sampling_ratio/min': 0.31117722392082214, 'sampling/importance_sampling_ratio/mean': 1.005907416343689, 'sampling/importance_sampling_ratio/max': 1.9484820365905762, 'entropy': 0.09778465609997511, 'clip_ratio/low_mean': 0.019021739484742284, 'clip_ratio/low_min': 0.019021739484742284, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.019021739484742284, 'reward_total_mean': 0.9431781768798828, 'reward_meter_mean': 0.9431781768798828, 'reward_meter_std': 0.0021464419551193714, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9431781768798828, 'reward_total_composite_std': 0.0021464419551193714, 'epoch': 0.13}

 98%|█████████▊| 3226/3300 [4:58:35<05:41,  4.61s/it]INFO 04-12 03:35:26 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3227/3300 [4:58:40<05:31,  4.55s/it]2026-04-12 03:35:31,137 | INFO | train_grpo_train | metrics_logged mode=train step=3227

                                                     
{'loss': -0.0002, 'grad_norm': 0.25633835792541504, 'learning_rate': 2.2424242424242425e-07, 'num_tokens': 7344170.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981385469436646, 'rewards/meter/std': 1.2708615031442605e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981385469436646, 'rewards/total_composite/std': 1.2708615031442605e-05, 'reward': 0.9981385469436646, 'reward_std': 1.2711200724879745e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.007955299690365791, 'sampling/sampling_logp_difference/max': 0.6916778087615967, 'sampling/importance_sampling_ratio/min': 0.6211740374565125, 'sampling/importance_sampling_ratio/mean': 1.002211332321167, 'sampling/importance_sampling_ratio/max': 1.997063398361206, 'entropy': 0.047321764286607504, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.0037313431967049837, 'clip_ratio/high_max': 0.0037313431967049837, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9981385469436646, 'reward_meter_mean': 0.9981385469436646, 'reward_meter_std': 1.2708615031442605e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981385469436646, 'reward_total_composite_std': 1.2708615031442605e-05, 'epoch': 0.13}

 98%|█████████▊| 3227/3300 [4:58:40<05:31,  4.55s/it]INFO 04-12 03:35:31 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3228/3300 [4:58:44<05:21,  4.47s/it]2026-04-12 03:35:35,414 | INFO | train_grpo_train | metrics_logged mode=train step=3228

                                                     
{'loss': 0.0145, 'grad_norm': 8.390018463134766, 'learning_rate': 2.2121212121212124e-07, 'num_tokens': 7345827.0, 'completions/mean_length': 46.125, 'completions/min_length': 46.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9406979084014893, 'rewards/meter/std': 0.006258478853851557, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9406979084014893, 'rewards/total_composite/std': 0.006258478853851557, 'reward': 0.9406979084014893, 'reward_std': 0.0062584723345935345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02605927735567093, 'sampling/sampling_logp_difference/max': 2.4645028114318848, 'sampling/importance_sampling_ratio/min': 0.08505111932754517, 'sampling/importance_sampling_ratio/mean': 0.996314287185669, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.07639594282954931, 'clip_ratio/low_mean': 0.002659574383869767, 'clip_ratio/low_min': 0.002659574383869767, 'clip_ratio/high_mean': 0.00815217406488955, 'clip_ratio/high_max': 0.00815217406488955, 'clip_ratio/region_mean': 0.010811748448759317, 'reward_total_mean': 0.9406979084014893, 'reward_meter_mean': 0.9406979084014893, 'reward_meter_std': 0.006258478853851557, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9406979084014893, 'reward_total_composite_std': 0.006258478853851557, 'epoch': 0.13}

 98%|█████████▊| 3228/3300 [4:58:44<05:21,  4.47s/it]INFO 04-12 03:35:35 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3229/3300 [4:58:49<05:24,  4.57s/it]2026-04-12 03:35:40,228 | INFO | train_grpo_train | metrics_logged mode=train step=3229

                                                     
{'loss': -0.0021, 'grad_norm': 0.8496602773666382, 'learning_rate': 2.181818181818182e-07, 'num_tokens': 7348082.0, 'completions/mean_length': 97.875, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9979760646820068, 'rewards/meter/std': 7.630670006619766e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9979760646820068, 'rewards/total_composite/std': 7.630670006619766e-05, 'reward': 0.9979760646820068, 'reward_std': 7.630588515894488e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01204567588865757, 'sampling/sampling_logp_difference/max': 0.7632923126220703, 'sampling/importance_sampling_ratio/min': 0.46612924337387085, 'sampling/importance_sampling_ratio/mean': 1.0022556781768799, 'sampling/importance_sampling_ratio/max': 1.3757518529891968, 'entropy': 0.10427160188555717, 'clip_ratio/low_mean': 0.008941720938310027, 'clip_ratio/low_min': 0.008941720938310027, 'clip_ratio/high_mean': 0.005102040828205645, 'clip_ratio/high_max': 0.005102040828205645, 'clip_ratio/region_mean': 0.014043761766515672, 'reward_total_mean': 0.9979760646820068, 'reward_meter_mean': 0.9979760646820068, 'reward_meter_std': 7.630670006619766e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9979760646820068, 'reward_total_composite_std': 7.630670006619766e-05, 'epoch': 0.13}

 98%|█████████▊| 3229/3300 [4:58:49<05:24,  4.57s/it]INFO 04-12 03:35:40 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3230/3300 [4:58:53<05:17,  4.53s/it]2026-04-12 03:35:44,663 | INFO | train_grpo_train | metrics_logged mode=train step=3230

                                                     
{'loss': -0.0006, 'grad_norm': 0.018652301281690598, 'learning_rate': 2.1515151515151517e-07, 'num_tokens': 7349856.0, 'completions/mean_length': 66.75, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981513023376465, 'rewards/meter/std': 1.5048592558741802e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981513023376465, 'rewards/total_composite/std': 1.5048592558741802e-06, 'reward': 0.9981513023376465, 'reward_std': 1.5035095657367492e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.006907826755195856, 'sampling/sampling_logp_difference/max': 0.5330893993377686, 'sampling/importance_sampling_ratio/min': 0.5867893099784851, 'sampling/importance_sampling_ratio/mean': 1.0017226934432983, 'sampling/importance_sampling_ratio/max': 1.4481335878372192, 'entropy': 0.04502238845452666, 'clip_ratio/low_mean': 0.0018656715983524919, 'clip_ratio/low_min': 0.0018656715983524919, 'clip_ratio/high_mean': 0.003759611048735678, 'clip_ratio/high_max': 0.003759611048735678, 'clip_ratio/region_mean': 0.00562528264708817, 'reward_total_mean': 0.9981513023376465, 'reward_meter_mean': 0.9981513023376465, 'reward_meter_std': 1.5048592558741802e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981513023376465, 'reward_total_composite_std': 1.5048592558741802e-06, 'epoch': 0.13}

 98%|█████████▊| 3230/3300 [4:58:53<05:17,  4.53s/it]INFO 04-12 03:35:44 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3231/3300 [4:58:58<05:22,  4.68s/it]2026-04-12 03:35:49,679 | INFO | train_grpo_train | metrics_logged mode=train step=3231

                                                     
{'loss': -0.0038, 'grad_norm': 1.449289321899414, 'learning_rate': 2.1212121212121216e-07, 'num_tokens': 7352000.0, 'completions/mean_length': 107.0, 'completions/min_length': 106.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.0, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9992383718490601, 'rewards/meter/std': 0.00011975476081715897, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992383718490601, 'rewards/total_composite/std': 0.00011975476081715897, 'reward': 0.9992383718490601, 'reward_std': 0.0001197509845951572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.015148649923503399, 'sampling/sampling_logp_difference/max': 0.9686565399169922, 'sampling/importance_sampling_ratio/min': 0.3795926868915558, 'sampling/importance_sampling_ratio/mean': 1.0047478675842285, 'sampling/importance_sampling_ratio/max': 1.8299890756607056, 'entropy': 0.15556516498327255, 'clip_ratio/low_mean': 0.004716981202363968, 'clip_ratio/low_min': 0.004716981202363968, 'clip_ratio/high_mean': 0.006999526987783611, 'clip_ratio/high_max': 0.006999526987783611, 'clip_ratio/region_mean': 0.011716508190147579, 'reward_total_mean': 0.9992383718490601, 'reward_meter_mean': 0.9992383718490601, 'reward_meter_std': 0.00011975476081715897, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992383718490601, 'reward_total_composite_std': 0.00011975476081715897, 'epoch': 0.13}

 98%|█████████▊| 3231/3300 [4:58:58<05:22,  4.68s/it]INFO 04-12 03:35:49 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3232/3300 [4:59:03<05:18,  4.69s/it]2026-04-12 03:35:54,388 | INFO | train_grpo_train | metrics_logged mode=train step=3232

                                                     
{'loss': -0.0004, 'grad_norm': 2.6132781505584717, 'learning_rate': 2.090909090909091e-07, 'num_tokens': 7354036.0, 'completions/mean_length': 77.5, 'completions/min_length': 75.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9991195797920227, 'rewards/meter/std': 0.0001867539540398866, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991195797920227, 'rewards/total_composite/std': 0.0001867539540398866, 'reward': 0.9991195797920227, 'reward_std': 0.00018676480976864696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.022667229175567627, 'sampling/sampling_logp_difference/max': 0.9247050285339355, 'sampling/importance_sampling_ratio/min': 0.39664843678474426, 'sampling/importance_sampling_ratio/mean': 1.005802035331726, 'sampling/importance_sampling_ratio/max': 1.5898345708847046, 'entropy': 0.19562415033578873, 'clip_ratio/low_mean': 0.0032681134762242436, 'clip_ratio/low_min': 0.0032681134762242436, 'clip_ratio/high_mean': 0.009743589907884598, 'clip_ratio/high_max': 0.009743589907884598, 'clip_ratio/region_mean': 0.013011703384108841, 'reward_total_mean': 0.9991195797920227, 'reward_meter_mean': 0.9991195797920227, 'reward_meter_std': 0.0001867539540398866, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991195797920227, 'reward_total_composite_std': 0.0001867539540398866, 'epoch': 0.13}

 98%|█████████▊| 3232/3300 [4:59:03<05:18,  4.69s/it]INFO 04-12 03:35:54 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3233/3300 [4:59:08<05:23,  4.83s/it]2026-04-12 03:35:59,576 | INFO | train_grpo_train | metrics_logged mode=train step=3233

                                                     
{'loss': -0.0061, 'grad_norm': 3.281510591506958, 'learning_rate': 2.060606060606061e-07, 'num_tokens': 7356649.0, 'completions/mean_length': 128.625, 'completions/min_length': 127.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.625, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9830908179283142, 'rewards/meter/std': 0.04200791195034981, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9830908179283142, 'rewards/total_composite/std': 0.04200791195034981, 'reward': 0.9830908179283142, 'reward_std': 0.042007915675640106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016228584572672844, 'sampling/sampling_logp_difference/max': 0.8928537368774414, 'sampling/importance_sampling_ratio/min': 0.4094855487346649, 'sampling/importance_sampling_ratio/mean': 1.00603187084198, 'sampling/importance_sampling_ratio/max': 1.564560890197754, 'entropy': 0.15214181691408157, 'clip_ratio/low_mean': 0.0009842519648373127, 'clip_ratio/low_min': 0.0009842519648373127, 'clip_ratio/high_mean': 0.00582152372226119, 'clip_ratio/high_max': 0.00582152372226119, 'clip_ratio/region_mean': 0.006805775687098503, 'reward_total_mean': 0.9830908179283142, 'reward_meter_mean': 0.9830908179283142, 'reward_meter_std': 0.04200791195034981, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9830908179283142, 'reward_total_composite_std': 0.04200791195034981, 'epoch': 0.13}

 98%|█████████▊| 3233/3300 [4:59:08<05:23,  4.83s/it]INFO 04-12 03:35:59 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3234/3300 [4:59:12<05:08,  4.67s/it]2026-04-12 03:36:03,854 | INFO | train_grpo_train | metrics_logged mode=train step=3234

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.0303030303030303e-07, 'num_tokens': 7358433.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002218023146269843, 'sampling/sampling_logp_difference/max': 0.0019890288822352886, 'sampling/importance_sampling_ratio/min': 0.9995585680007935, 'sampling/importance_sampling_ratio/mean': 1.0002152919769287, 'sampling/importance_sampling_ratio/max': 1.0019910335540771, 'entropy': 0.001778529680450447, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 98%|█████████▊| 3234/3300 [4:59:12<05:08,  4.67s/it]INFO 04-12 03:36:04 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3235/3300 [4:59:18<05:16,  4.88s/it]2026-04-12 03:36:09,210 | INFO | train_grpo_train | metrics_logged mode=train step=3235

                                                     
{'loss': 0.0024, 'grad_norm': 2.066392660140991, 'learning_rate': 2.0000000000000002e-07, 'num_tokens': 7361110.0, 'completions/mean_length': 133.625, 'completions/min_length': 130.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.625, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9991873502731323, 'rewards/meter/std': 0.00011400806397432461, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9821428656578064, 'rewards/repeat_penalty/std': 0.05050762742757797, 'rewards/total_composite/mean': 0.9813442230224609, 'rewards/total_composite/std': 0.05045589804649353, 'reward': 0.9813442230224609, 'reward_std': 0.05045590177178383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031857188791036606, 'sampling/sampling_logp_difference/max': 2.231940269470215, 'sampling/importance_sampling_ratio/min': 0.10731999576091766, 'sampling/importance_sampling_ratio/mean': 1.0075880289077759, 'sampling/importance_sampling_ratio/max': 1.8417998552322388, 'entropy': 0.2376344669610262, 'clip_ratio/low_mean': 0.001879699295386672, 'clip_ratio/low_min': 0.001879699295386672, 'clip_ratio/high_mean': 0.025164754129946232, 'clip_ratio/high_max': 0.025164754129946232, 'clip_ratio/region_mean': 0.027044453425332904, 'reward_total_mean': 0.9813442230224609, 'reward_meter_mean': 0.9991873502731323, 'reward_meter_std': 0.00011400806397432461, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9821428656578064, 'reward_repeat_penalty_std': 0.05050762742757797, 'reward_total_composite_mean': 0.9813442230224609, 'reward_total_composite_std': 0.05045589804649353, 'epoch': 0.13}

 98%|█████████▊| 3235/3300 [4:59:18<05:16,  4.88s/it]INFO 04-12 03:36:09 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3236/3300 [4:59:25<05:58,  5.60s/it]2026-04-12 03:36:16,489 | INFO | train_grpo_train | metrics_logged mode=train step=3236

                                                     
{'loss': -0.0018, 'grad_norm': 2.294574022293091, 'learning_rate': 1.9696969696969698e-07, 'num_tokens': 7365186.0, 'completions/mean_length': 292.5, 'completions/min_length': 278.0, 'completions/max_length': 298.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 292.5, 'completions/min_terminated_length': 278.0, 'completions/max_terminated_length': 298.0, 'rewards/meter/mean': 0.9674590826034546, 'rewards/meter/std': 0.049329616129398346, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/repeat_penalty/mean': 0.9440359473228455, 'rewards/repeat_penalty/std': 0.042011942714452744, 'rewards/total_composite/mean': 0.6899652481079102, 'rewards/total_composite/std': 0.280244380235672, 'reward': 0.6899652481079102, 'reward_std': 0.280244380235672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03874696418642998, 'sampling/sampling_logp_difference/max': 2.491647720336914, 'sampling/importance_sampling_ratio/min': 0.08277346938848495, 'sampling/importance_sampling_ratio/mean': 1.0074106454849243, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3465934246778488, 'clip_ratio/low_mean': 0.002595155732706189, 'clip_ratio/low_min': 0.002595155732706189, 'clip_ratio/high_mean': 0.02566183707676828, 'clip_ratio/high_max': 0.02566183707676828, 'clip_ratio/region_mean': 0.028256992809474468, 'reward_total_mean': 0.6899652481079102, 'reward_meter_mean': 0.9674590826034546, 'reward_meter_std': 0.049329616129398346, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.9440359473228455, 'reward_repeat_penalty_std': 0.042011942714452744, 'reward_total_composite_mean': 0.6899652481079102, 'reward_total_composite_std': 0.280244380235672, 'epoch': 0.13}

 98%|█████████▊| 3236/3300 [4:59:25<05:58,  5.60s/it]INFO 04-12 03:36:16 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3237/3300 [4:59:29<05:30,  5.25s/it]2026-04-12 03:36:20,919 | INFO | train_grpo_train | metrics_logged mode=train step=3237

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.9393939393939395e-07, 'num_tokens': 7366962.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 5.0931172154378146e-05, 'sampling/sampling_logp_difference/max': 0.0009334206115454435, 'sampling/importance_sampling_ratio/min': 0.9994771480560303, 'sampling/importance_sampling_ratio/mean': 1.0000455379486084, 'sampling/importance_sampling_ratio/max': 1.0009338855743408, 'entropy': 0.0005204225053603295, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 98%|█████████▊| 3237/3300 [4:59:29<05:30,  5.25s/it]INFO 04-12 03:36:21 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3238/3300 [4:59:35<05:26,  5.27s/it]2026-04-12 03:36:26,233 | INFO | train_grpo_train | metrics_logged mode=train step=3238

                                                     
{'loss': -0.0084, 'grad_norm': 4.2172322273254395, 'learning_rate': 1.9090909090909094e-07, 'num_tokens': 7369424.0, 'completions/mean_length': 133.75, 'completions/min_length': 129.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.75, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.9989288449287415, 'rewards/meter/std': 0.0006227882695384324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9989288449287415, 'rewards/total_composite/std': 0.0006227882695384324, 'reward': 0.9989288449287415, 'reward_std': 0.0006227805861271918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035977549850940704, 'sampling/sampling_logp_difference/max': 5.998313903808594, 'sampling/importance_sampling_ratio/min': 0.0024829350877553225, 'sampling/importance_sampling_ratio/mean': 1.0042328834533691, 'sampling/importance_sampling_ratio/max': 1.8821479082107544, 'entropy': 0.22379975207149982, 'clip_ratio/low_mean': 0.007562352577224374, 'clip_ratio/low_min': 0.007562352577224374, 'clip_ratio/high_mean': 0.011160198540892452, 'clip_ratio/high_max': 0.011160198540892452, 'clip_ratio/region_mean': 0.018722551118116826, 'reward_total_mean': 0.9989288449287415, 'reward_meter_mean': 0.9989288449287415, 'reward_meter_std': 0.0006227882695384324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9989288449287415, 'reward_total_composite_std': 0.0006227882695384324, 'epoch': 0.13}

 98%|█████████▊| 3238/3300 [4:59:35<05:26,  5.27s/it]INFO 04-12 03:36:26 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3239/3300 [4:59:39<05:08,  5.06s/it]2026-04-12 03:36:30,812 | INFO | train_grpo_train | metrics_logged mode=train step=3239

                                                     
{'loss': -0.0012, 'grad_norm': 1.0152816772460938, 'learning_rate': 1.878787878787879e-07, 'num_tokens': 7371177.0, 'completions/mean_length': 68.125, 'completions/min_length': 68.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.999480128288269, 'rewards/meter/std': 7.362089672824368e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999480128288269, 'rewards/total_composite/std': 7.362089672824368e-05, 'reward': 0.999480128288269, 'reward_std': 7.363814802374691e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011260369792580605, 'sampling/sampling_logp_difference/max': 1.374150037765503, 'sampling/importance_sampling_ratio/min': 0.2629310190677643, 'sampling/importance_sampling_ratio/mean': 1.0008584260940552, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.04902258049696684, 'clip_ratio/low_mean': 0.0018382353009656072, 'clip_ratio/low_min': 0.0018382353009656072, 'clip_ratio/high_mean': 0.0018382353009656072, 'clip_ratio/high_max': 0.0018382353009656072, 'clip_ratio/region_mean': 0.0036764706019312143, 'reward_total_mean': 0.999480128288269, 'reward_meter_mean': 0.999480128288269, 'reward_meter_std': 7.362089672824368e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999480128288269, 'reward_total_composite_std': 7.362089672824368e-05, 'epoch': 0.13}

 98%|█████████▊| 3239/3300 [4:59:39<05:08,  5.06s/it]INFO 04-12 03:36:31 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3240/3300 [4:59:47<06:00,  6.01s/it]2026-04-12 03:36:39,055 | INFO | train_grpo_train | metrics_logged mode=train step=3240

                                                     
{'loss': -0.0318, 'grad_norm': 1.8787816762924194, 'learning_rate': 1.8484848484848486e-07, 'num_tokens': 7375719.0, 'completions/mean_length': 350.75, 'completions/min_length': 333.0, 'completions/max_length': 380.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 350.75, 'completions/min_terminated_length': 333.0, 'completions/max_terminated_length': 380.0, 'rewards/meter/mean': 0.9988770484924316, 'rewards/meter/std': 0.0009327766601927578, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.05345224589109421, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9926470518112183, 'rewards/repeat_penalty/std': 0.020797256380319595, 'rewards/total_composite/mean': 0.8923759460449219, 'rewards/total_composite/std': 0.056543249636888504, 'reward': 0.8923759460449219, 'reward_std': 0.056543245911598206, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05698416754603386, 'sampling/sampling_logp_difference/max': 4.723583698272705, 'sampling/importance_sampling_ratio/min': 0.008883286267518997, 'sampling/importance_sampling_ratio/mean': 1.0113011598587036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49749621003866196, 'clip_ratio/low_mean': 0.005920994793996215, 'clip_ratio/low_min': 0.005920994793996215, 'clip_ratio/high_mean': 0.024295849725604057, 'clip_ratio/high_max': 0.024295849725604057, 'clip_ratio/region_mean': 0.030216844519600272, 'reward_total_mean': 0.8923759460449219, 'reward_meter_mean': 0.9988770484924316, 'reward_meter_std': 0.0009327766601927578, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.05345224589109421, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9926470518112183, 'reward_repeat_penalty_std': 0.020797256380319595, 'reward_total_composite_mean': 0.8923759460449219, 'reward_total_composite_std': 0.056543249636888504, 'epoch': 0.13}

 98%|█████████▊| 3240/3300 [4:59:48<06:00,  6.01s/it]INFO 04-12 03:36:39 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3241/3300 [4:59:56<06:36,  6.72s/it]2026-04-12 03:36:47,429 | INFO | train_grpo_train | metrics_logged mode=train step=3241

                                                     
{'loss': -0.0301, 'grad_norm': 1.581771969795227, 'learning_rate': 1.8181818181818183e-07, 'num_tokens': 7380120.0, 'completions/mean_length': 353.125, 'completions/min_length': 342.0, 'completions/max_length': 392.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 353.125, 'completions/min_terminated_length': 342.0, 'completions/max_terminated_length': 392.0, 'rewards/meter/mean': 0.9989774823188782, 'rewards/meter/std': 0.0006195669411681592, 'rewards/count_adherence/mean': 0.7604166269302368, 'rewards/count_adherence/std': 0.029462775215506554, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9926470518112183, 'rewards/repeat_penalty/std': 0.020797256380319595, 'rewards/total_composite/mean': 0.754129946231842, 'rewards/total_composite/std': 0.03522041067481041, 'reward': 0.754129946231842, 'reward_std': 0.03522040322422981, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.048592861741781235, 'sampling/sampling_logp_difference/max': 1.755746841430664, 'sampling/importance_sampling_ratio/min': 0.1727781593799591, 'sampling/importance_sampling_ratio/mean': 1.0124188661575317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4680623523890972, 'clip_ratio/low_mean': 0.019117801217362285, 'clip_ratio/low_min': 0.019117801217362285, 'clip_ratio/high_mean': 0.004464285913854837, 'clip_ratio/high_max': 0.004464285913854837, 'clip_ratio/region_mean': 0.023582087131217122, 'reward_total_mean': 0.754129946231842, 'reward_meter_mean': 0.9989774823188782, 'reward_meter_std': 0.0006195669411681592, 'reward_count_adherence_mean': 0.7604166269302368, 'reward_count_adherence_std': 0.029462775215506554, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9926470518112183, 'reward_repeat_penalty_std': 0.020797256380319595, 'reward_total_composite_mean': 0.754129946231842, 'reward_total_composite_std': 0.03522041067481041, 'epoch': 0.13}

 98%|█████████▊| 3241/3300 [4:59:56<06:36,  6.72s/it]INFO 04-12 03:36:47 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3242/3300 [5:00:03<06:37,  6.85s/it]2026-04-12 03:36:54,565 | INFO | train_grpo_train | metrics_logged mode=train step=3242

                                                     
{'loss': 0.015, 'grad_norm': 2.48710560798645, 'learning_rate': 1.7878787878787882e-07, 'num_tokens': 7383827.0, 'completions/mean_length': 275.375, 'completions/min_length': 269.0, 'completions/max_length': 283.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 275.375, 'completions/min_terminated_length': 269.0, 'completions/max_terminated_length': 283.0, 'rewards/meter/mean': 0.9965729713439941, 'rewards/meter/std': 0.0008044191054068506, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9411764740943909, 'rewards/repeat_penalty/std': 0.03144249692559242, 'rewards/total_composite/mean': 0.8207110166549683, 'rewards/total_composite/std': 0.027550045400857925, 'reward': 0.8207110166549683, 'reward_std': 0.027550052851438522, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04147728160023689, 'sampling/sampling_logp_difference/max': 1.294790267944336, 'sampling/importance_sampling_ratio/min': 0.2739553153514862, 'sampling/importance_sampling_ratio/mean': 1.0088013410568237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39379945397377014, 'clip_ratio/low_mean': 0.012771393987350166, 'clip_ratio/low_min': 0.012771393987350166, 'clip_ratio/high_mean': 0.013120089075528085, 'clip_ratio/high_max': 0.013120089075528085, 'clip_ratio/region_mean': 0.02589148306287825, 'reward_total_mean': 0.8207110166549683, 'reward_meter_mean': 0.9965729713439941, 'reward_meter_std': 0.0008044191054068506, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9411764740943909, 'reward_repeat_penalty_std': 0.03144249692559242, 'reward_total_composite_mean': 0.8207110166549683, 'reward_total_composite_std': 0.027550045400857925, 'epoch': 0.13}

 98%|█████████▊| 3242/3300 [5:00:03<06:37,  6.85s/it]INFO 04-12 03:36:54 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3243/3300 [5:00:09<06:09,  6.48s/it]2026-04-12 03:37:00,175 | INFO | train_grpo_train | metrics_logged mode=train step=3243

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.7575757575757576e-07, 'num_tokens': 7386683.0, 'completions/mean_length': 158.0, 'completions/min_length': 158.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.0, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.6563846468925476, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.7272727489471436, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.47737064957618713, 'rewards/total_composite/std': 0.0, 'reward': 0.47737064957618713, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0011825277470052242, 'sampling/sampling_logp_difference/max': 0.05302273482084274, 'sampling/importance_sampling_ratio/min': 0.9712576866149902, 'sampling/importance_sampling_ratio/mean': 1.0010987520217896, 'sampling/importance_sampling_ratio/max': 1.0544536113739014, 'entropy': 0.009368695667944849, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.47737064957618713, 'reward_meter_mean': 0.6563846468925476, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.7272727489471436, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.47737064957618713, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 98%|█████████▊| 3243/3300 [5:00:09<06:09,  6.48s/it]INFO 04-12 03:37:00 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3244/3300 [5:00:14<05:50,  6.25s/it]2026-04-12 03:37:05,903 | INFO | train_grpo_train | metrics_logged mode=train step=3244

                                                     
{'loss': -0.0034, 'grad_norm': 1.8808863162994385, 'learning_rate': 1.7272727272727275e-07, 'num_tokens': 7389632.0, 'completions/mean_length': 155.625, 'completions/min_length': 153.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 155.625, 'completions/min_terminated_length': 153.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.998309850692749, 'rewards/meter/std': 0.0023174018133431673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998309850692749, 'rewards/total_composite/std': 0.0023174018133431673, 'reward': 0.998309850692749, 'reward_std': 0.002317406702786684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039463937282562256, 'sampling/sampling_logp_difference/max': 1.5400009155273438, 'sampling/importance_sampling_ratio/min': 0.2143809050321579, 'sampling/importance_sampling_ratio/mean': 1.0025235414505005, 'sampling/importance_sampling_ratio/max': 1.5186190605163574, 'entropy': 0.330673573538661, 'clip_ratio/low_mean': 0.003246753243729472, 'clip_ratio/low_min': 0.003246753243729472, 'clip_ratio/high_mean': 0.02562888152897358, 'clip_ratio/high_max': 0.02562888152897358, 'clip_ratio/region_mean': 0.02887563477270305, 'reward_total_mean': 0.998309850692749, 'reward_meter_mean': 0.998309850692749, 'reward_meter_std': 0.0023174018133431673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998309850692749, 'reward_total_composite_std': 0.0023174018133431673, 'epoch': 0.13}

 98%|█████████▊| 3244/3300 [5:00:14<05:50,  6.25s/it]INFO 04-12 03:37:06 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3245/3300 [5:00:20<05:31,  6.02s/it]2026-04-12 03:37:11,390 | INFO | train_grpo_train | metrics_logged mode=train step=3245

                                                     
{'loss': -0.0017, 'grad_norm': 2.231248378753662, 'learning_rate': 1.6969696969696974e-07, 'num_tokens': 7392187.0, 'completions/mean_length': 135.375, 'completions/min_length': 133.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 135.375, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9989526271820068, 'rewards/meter/std': 0.0006340565159916878, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9642857313156128, 'rewards/repeat_penalty/std': 0.06613000482320786, 'rewards/total_composite/mean': 0.9632641077041626, 'rewards/total_composite/std': 0.0658845379948616, 'reward': 0.9632641077041626, 'reward_std': 0.0658845454454422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02810845896601677, 'sampling/sampling_logp_difference/max': 0.8493685722351074, 'sampling/importance_sampling_ratio/min': 0.4276849031448364, 'sampling/importance_sampling_ratio/mean': 1.0051708221435547, 'sampling/importance_sampling_ratio/max': 1.8710554838180542, 'entropy': 0.206793999299407, 'clip_ratio/low_mean': 0.006537176435813308, 'clip_ratio/low_min': 0.006537176435813308, 'clip_ratio/high_mean': 0.01926107343751937, 'clip_ratio/high_max': 0.01926107343751937, 'clip_ratio/region_mean': 0.02579824987333268, 'reward_total_mean': 0.9632641077041626, 'reward_meter_mean': 0.9989526271820068, 'reward_meter_std': 0.0006340565159916878, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9642857313156128, 'reward_repeat_penalty_std': 0.06613000482320786, 'reward_total_composite_mean': 0.9632641077041626, 'reward_total_composite_std': 0.0658845379948616, 'epoch': 0.13}

 98%|█████████▊| 3245/3300 [5:00:20<05:31,  6.02s/it]INFO 04-12 03:37:11 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3246/3300 [5:00:27<05:46,  6.42s/it]2026-04-12 03:37:18,742 | INFO | train_grpo_train | metrics_logged mode=train step=3246

                                                     
{'loss': 0.0198, 'grad_norm': 1.8668049573898315, 'learning_rate': 1.6666666666666668e-07, 'num_tokens': 7395817.0, 'completions/mean_length': 270.75, 'completions/min_length': 259.0, 'completions/max_length': 288.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 270.75, 'completions/min_terminated_length': 259.0, 'completions/max_terminated_length': 288.0, 'rewards/meter/mean': 0.9602240324020386, 'rewards/meter/std': 0.09070871770381927, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.915900707244873, 'rewards/repeat_penalty/std': 0.0559099018573761, 'rewards/total_composite/mean': 0.8224383592605591, 'rewards/total_composite/std': 0.09149650484323502, 'reward': 0.8224383592605591, 'reward_std': 0.09149649739265442, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03546025604009628, 'sampling/sampling_logp_difference/max': 1.9990062713623047, 'sampling/importance_sampling_ratio/min': 0.13546983897686005, 'sampling/importance_sampling_ratio/mean': 1.01006281375885, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.38703763484954834, 'clip_ratio/low_mean': 0.010340243112295866, 'clip_ratio/low_min': 0.010340243112295866, 'clip_ratio/high_mean': 0.01162644021678716, 'clip_ratio/high_max': 0.01162644021678716, 'clip_ratio/region_mean': 0.021966683329083025, 'reward_total_mean': 0.8224383592605591, 'reward_meter_mean': 0.9602240324020386, 'reward_meter_std': 0.09070871770381927, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.915900707244873, 'reward_repeat_penalty_std': 0.0559099018573761, 'reward_total_composite_mean': 0.8224383592605591, 'reward_total_composite_std': 0.09149650484323502, 'epoch': 0.13}

 98%|█████████▊| 3246/3300 [5:00:27<05:46,  6.42s/it]INFO 04-12 03:37:18 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3247/3300 [5:00:32<05:10,  5.85s/it]2026-04-12 03:37:23,260 | INFO | train_grpo_train | metrics_logged mode=train step=3247

                                                     
{'loss': 0.0001, 'grad_norm': 0.025763750076293945, 'learning_rate': 1.6363636363636367e-07, 'num_tokens': 7397609.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973357319831848, 'rewards/meter/std': 8.977292964118533e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973357319831848, 'rewards/total_composite/std': 8.977292964118533e-06, 'reward': 0.9973357319831848, 'reward_std': 8.97126938070869e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0030104033648967743, 'sampling/sampling_logp_difference/max': 0.4952678680419922, 'sampling/importance_sampling_ratio/min': 0.6253011226654053, 'sampling/importance_sampling_ratio/mean': 1.0012048482894897, 'sampling/importance_sampling_ratio/max': 1.6409376859664917, 'entropy': 0.012363885063678026, 'clip_ratio/low_mean': 0.004098360426723957, 'clip_ratio/low_min': 0.004098360426723957, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.004098360426723957, 'reward_total_mean': 0.9973357319831848, 'reward_meter_mean': 0.9973357319831848, 'reward_meter_std': 8.977292964118533e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973357319831848, 'reward_total_composite_std': 8.977292964118533e-06, 'epoch': 0.13}

 98%|█████████▊| 3247/3300 [5:00:32<05:10,  5.85s/it]INFO 04-12 03:37:23 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3248/3300 [5:00:37<05:04,  5.85s/it]2026-04-12 03:37:29,114 | INFO | train_grpo_train | metrics_logged mode=train step=3248

                                                     
{'loss': -0.0118, 'grad_norm': 4.575743675231934, 'learning_rate': 1.606060606060606e-07, 'num_tokens': 7400383.0, 'completions/mean_length': 166.75, 'completions/min_length': 159.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.75, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.9989683628082275, 'rewards/meter/std': 0.00019462761702015996, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9722222089767456, 'rewards/repeat_penalty/std': 0.05143444985151291, 'rewards/total_composite/mean': 0.9712226390838623, 'rewards/total_composite/std': 0.051451049745082855, 'reward': 0.9712226390838623, 'reward_std': 0.051451049745082855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.040822967886924744, 'sampling/sampling_logp_difference/max': 1.530099868774414, 'sampling/importance_sampling_ratio/min': 0.21651406586170197, 'sampling/importance_sampling_ratio/mean': 1.0083783864974976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3529275842010975, 'clip_ratio/low_mean': 0.0085039883852005, 'clip_ratio/low_min': 0.0085039883852005, 'clip_ratio/high_mean': 0.023729265900328755, 'clip_ratio/high_max': 0.023729265900328755, 'clip_ratio/region_mean': 0.032233254285529256, 'reward_total_mean': 0.9712226390838623, 'reward_meter_mean': 0.9989683628082275, 'reward_meter_std': 0.00019462761702015996, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9722222089767456, 'reward_repeat_penalty_std': 0.05143444985151291, 'reward_total_composite_mean': 0.9712226390838623, 'reward_total_composite_std': 0.051451049745082855, 'epoch': 0.13}

 98%|█████████▊| 3248/3300 [5:00:38<05:04,  5.85s/it]INFO 04-12 03:37:29 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3249/3300 [5:00:48<06:05,  7.16s/it]2026-04-12 03:37:39,335 | INFO | train_grpo_train | metrics_logged mode=train step=3249

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.575757575757576e-07, 'num_tokens': 7402303.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.998656690120697, 'rewards/meter/std': 0.0011350500863045454, 'rewards/count_adherence/mean': 0.7291666269302368, 'rewards/count_adherence/std': 0.01964186504483223, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9807692170143127, 'rewards/repeat_penalty/std': 0.02907419577240944, 'rewards/total_composite/mean': 0.7142915725708008, 'rewards/total_composite/std': 0.03164464980363846, 'reward': 0.7142915725708008, 'reward_std': 0.03164464607834816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.7142915725708008, 'reward_meter_mean': 0.998656690120697, 'reward_meter_std': 0.0011350500863045454, 'reward_count_adherence_mean': 0.7291666269302368, 'reward_count_adherence_std': 0.01964186504483223, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9807692170143127, 'reward_repeat_penalty_std': 0.02907419577240944, 'reward_total_composite_mean': 0.7142915725708008, 'reward_total_composite_std': 0.03164464980363846, 'epoch': 0.13}

 98%|█████████▊| 3249/3300 [5:00:48<06:05,  7.16s/it]INFO 04-12 03:37:39 [block_pool.py:378] Successfully reset prefix cache

 98%|█████████▊| 3250/3300 [5:00:53<05:32,  6.65s/it]2026-04-12 03:37:44,780 | INFO | train_grpo_train | metrics_logged mode=train step=3250

                                                     
{'loss': -0.0012, 'grad_norm': 1.1469806432724, 'learning_rate': 1.5454545454545456e-07, 'num_tokens': 7404902.0, 'completions/mean_length': 141.875, 'completions/min_length': 140.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.875, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9991378784179688, 'rewards/meter/std': 0.00011923787678824738, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9991378784179688, 'rewards/total_composite/std': 0.00011923787678824738, 'reward': 0.9991378784179688, 'reward_std': 0.00011922277917619795, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020554685965180397, 'sampling/sampling_logp_difference/max': 1.2523746490478516, 'sampling/importance_sampling_ratio/min': 0.285825252532959, 'sampling/importance_sampling_ratio/mean': 1.0037734508514404, 'sampling/importance_sampling_ratio/max': 1.4649658203125, 'entropy': 0.20848924107849598, 'clip_ratio/low_mean': 0.010569972451776266, 'clip_ratio/low_min': 0.010569972451776266, 'clip_ratio/high_mean': 0.00882849539630115, 'clip_ratio/high_max': 0.00882849539630115, 'clip_ratio/region_mean': 0.019398467848077416, 'reward_total_mean': 0.9991378784179688, 'reward_meter_mean': 0.9991378784179688, 'reward_meter_std': 0.00011923787678824738, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9991378784179688, 'reward_total_composite_std': 0.00011923787678824738, 'epoch': 0.13}

 98%|█████████▊| 3250/3300 [5:00:53<05:32,  6.65s/it]INFO 04-12 03:37:45 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:38,  3.53s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.42s/it]

 31%|███       | 4/13 [00:16<00:40,  4.47s/it]

 38%|███▊      | 5/13 [00:22<00:39,  4.90s/it]

 46%|████▌     | 6/13 [00:29<00:39,  5.58s/it]

 54%|█████▍    | 7/13 [00:34<00:33,  5.54s/it]

 62%|██████▏   | 8/13 [00:41<00:29,  5.99s/it]

 69%|██████▉   | 9/13 [00:48<00:25,  6.40s/it]

 77%|███████▋  | 10/13 [00:53<00:17,  5.97s/it]

 85%|████████▍ | 11/13 [00:59<00:11,  5.82s/it]

 92%|█████████▏| 12/13 [01:05<00:05,  5.96s/it]

100%|██████████| 13/13 [01:11<00:00,  5.81s/it]2026-04-12 03:39:02,963 | INFO | train_grpo_train | metrics_logged mode=eval step=3250

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 78.0663, 'eval_samples_per_second': 1.332, 'eval_steps_per_second': 0.167, 'eval_num_tokens': 7404902.0, 'eval_completions/mean_length': 214.33653846153845, 'eval_completions/min_length': 60.76923076923077, 'eval_completions/max_length': 421.38461538461536, 'eval_completions/clipped_ratio': 0.028846153846153848, 'eval_completions/mean_terminated_length': 204.00686880258414, 'eval_completions/min_terminated_length': 60.76923076923077, 'eval_completions/max_terminated_length': 398.7692307692308, 'eval_rewards/meter/mean': 0.7977132155345037, 'eval_rewards/meter/std': 0.3206370784542881, 'eval_rewards/count_adherence/mean': 0.95811671935595, 'eval_rewards/count_adherence/std': 0.06441009388520168, 'eval_rewards/arabic_clean/mean': 0.9807692307692307, 'eval_rewards/arabic_clean/std': 0.05439282839114849, 'eval_rewards/repeat_penalty/mean': 0.9612986537126395, 'eval_rewards/repeat_penalty/std': 0.06227556343835134, 'eval_rewards/total_composite/mean': 0.7313130910579975, 'eval_rewards/total_composite/std': 0.33714570047763676, 'eval_reward': 0.7313130910579975, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.034000108448358685, 'eval_sampling/sampling_logp_difference/max': 1.2143818781926081, 'eval_sampling/importance_sampling_ratio/min': 0.30588124233942765, 'eval_sampling/importance_sampling_ratio/mean': 1.0085856593572176, 'eval_sampling/importance_sampling_ratio/max': 1.4786042525218084, 'eval_entropy': 0.38327448299297917, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7313130910579975, 'eval_reward_meter_mean': 0.7977132155345037, 'eval_reward_meter_std': 0.3206370784542881, 'eval_reward_count_adherence_mean': 0.95811671935595, 'eval_reward_count_adherence_std': 0.06441009388520168, 'eval_reward_arabic_clean_mean': 0.9807692307692307, 'eval_reward_arabic_clean_std': 0.05439282839114849, 'eval_reward_repeat_penalty_mean': 0.9612986537126395, 'eval_reward_repeat_penalty_std': 0.06227556343835134, 'eval_reward_total_composite_mean': 0.7313130910579975, 'eval_reward_total_composite_std': 0.33714570047763676, 'epoch': 0.13}

 98%|█████████▊| 3250/3300 [5:02:11<05:32,  6.65s/it]

100%|██████████| 13/13 [01:11<00:00,  5.81s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 99%|█████████▊| 3251/3300 [5:02:18<24:36, 30.12s/it]2026-04-12 03:39:09,685 | INFO | train_grpo_train | metrics_logged mode=train step=3251

                                                     
{'loss': 0.003, 'grad_norm': 1.9768705368041992, 'learning_rate': 1.5151515151515152e-07, 'num_tokens': 7406604.0, 'completions/mean_length': 40.75, 'completions/min_length': 39.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.998726487159729, 'rewards/meter/std': 0.0001513757451903075, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.998726487159729, 'rewards/total_composite/std': 0.0001513757451903075, 'reward': 0.998726487159729, 'reward_std': 0.0001513654860900715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020130213350057602, 'sampling/sampling_logp_difference/max': 1.2893218994140625, 'sampling/importance_sampling_ratio/min': 0.275457501411438, 'sampling/importance_sampling_ratio/mean': 0.9942156076431274, 'sampling/importance_sampling_ratio/max': 1.5404958724975586, 'entropy': 0.12377766706049442, 'clip_ratio/low_mean': 0.006253908621147275, 'clip_ratio/low_min': 0.006253908621147275, 'clip_ratio/high_mean': 0.009146341122686863, 'clip_ratio/high_max': 0.009146341122686863, 'clip_ratio/region_mean': 0.015400249743834138, 'reward_total_mean': 0.998726487159729, 'reward_meter_mean': 0.998726487159729, 'reward_meter_std': 0.0001513757451903075, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.998726487159729, 'reward_total_composite_std': 0.0001513757451903075, 'epoch': 0.13}

 99%|█████████▊| 3251/3300 [5:02:18<24:36, 30.12s/it]INFO 04-12 03:39:09 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▊| 3252/3300 [5:02:24<18:23, 22.99s/it]2026-04-12 03:39:16,012 | INFO | train_grpo_train | metrics_logged mode=train step=3252

                                                     
{'loss': 0.0033, 'grad_norm': 3.5564746856689453, 'learning_rate': 1.484848484848485e-07, 'num_tokens': 7409050.0, 'completions/mean_length': 124.75, 'completions/min_length': 123.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.75, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.9973940253257751, 'rewards/meter/std': 0.0014456679346039891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973940253257751, 'rewards/total_composite/std': 0.0014456679346039891, 'reward': 0.9973940253257751, 'reward_std': 0.001445673406124115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.018116910010576248, 'sampling/sampling_logp_difference/max': 1.4764413833618164, 'sampling/importance_sampling_ratio/min': 0.22844921052455902, 'sampling/importance_sampling_ratio/mean': 1.0002362728118896, 'sampling/importance_sampling_ratio/max': 1.582746148109436, 'entropy': 0.12918496038764715, 'clip_ratio/low_mean': 0.0009920635493472219, 'clip_ratio/low_min': 0.0009920635493472219, 'clip_ratio/high_mean': 0.012048649485222995, 'clip_ratio/high_max': 0.012048649485222995, 'clip_ratio/region_mean': 0.013040713034570217, 'reward_total_mean': 0.9973940253257751, 'reward_meter_mean': 0.9973940253257751, 'reward_meter_std': 0.0014456679346039891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973940253257751, 'reward_total_composite_std': 0.0014456679346039891, 'epoch': 0.13}

 99%|█████████▊| 3252/3300 [5:02:25<18:23, 22.99s/it]INFO 04-12 03:39:16 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▊| 3253/3300 [5:02:30<13:58, 17.83s/it]2026-04-12 03:39:21,829 | INFO | train_grpo_train | metrics_logged mode=train step=3253

                                                     
{'loss': 0.008, 'grad_norm': 1.9089981317520142, 'learning_rate': 1.4545454545454548e-07, 'num_tokens': 7411813.0, 'completions/mean_length': 154.375, 'completions/min_length': 152.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.375, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9973270297050476, 'rewards/meter/std': 0.0007912821019999683, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9861111044883728, 'rewards/repeat_penalty/std': 0.03928370773792267, 'rewards/total_composite/mean': 0.983467161655426, 'rewards/total_composite/std': 0.038973402231931686, 'reward': 0.983467161655426, 'reward_std': 0.03897340968251228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029868151992559433, 'sampling/sampling_logp_difference/max': 1.405970573425293, 'sampling/importance_sampling_ratio/min': 0.24512901902198792, 'sampling/importance_sampling_ratio/mean': 1.0039182901382446, 'sampling/importance_sampling_ratio/max': 1.7957075834274292, 'entropy': 0.2324562668800354, 'clip_ratio/low_mean': 0.002419354859739542, 'clip_ratio/low_min': 0.002419354859739542, 'clip_ratio/high_mean': 0.017753356718458235, 'clip_ratio/high_max': 0.017753356718458235, 'clip_ratio/region_mean': 0.020172711578197777, 'reward_total_mean': 0.983467161655426, 'reward_meter_mean': 0.9973270297050476, 'reward_meter_std': 0.0007912821019999683, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9861111044883728, 'reward_repeat_penalty_std': 0.03928370773792267, 'reward_total_composite_mean': 0.983467161655426, 'reward_total_composite_std': 0.038973402231931686, 'epoch': 0.13}

 99%|█████████▊| 3253/3300 [5:02:30<13:58, 17.83s/it]INFO 04-12 03:39:22 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▊| 3254/3300 [5:02:37<11:08, 14.53s/it]2026-04-12 03:39:28,647 | INFO | train_grpo_train | metrics_logged mode=train step=3254

                                                     
{'loss': -0.0002, 'grad_norm': 2.1981167793273926, 'learning_rate': 1.4242424242424244e-07, 'num_tokens': 7415280.0, 'completions/mean_length': 232.375, 'completions/min_length': 227.0, 'completions/max_length': 241.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 232.375, 'completions/min_terminated_length': 227.0, 'completions/max_terminated_length': 241.0, 'rewards/meter/mean': 0.9978033304214478, 'rewards/meter/std': 0.0022864267230033875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9807692170143127, 'rewards/repeat_penalty/std': 0.03560846298933029, 'rewards/total_composite/mean': 0.978593111038208, 'rewards/total_composite/std': 0.03494200482964516, 'reward': 0.978593111038208, 'reward_std': 0.03494199737906456, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042390793561935425, 'sampling/sampling_logp_difference/max': 1.2398028373718262, 'sampling/importance_sampling_ratio/min': 0.2894412875175476, 'sampling/importance_sampling_ratio/mean': 1.0061860084533691, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34237322956323624, 'clip_ratio/low_mean': 0.007630055071786046, 'clip_ratio/low_min': 0.007630055071786046, 'clip_ratio/high_mean': 0.025140726240351796, 'clip_ratio/high_max': 0.025140726240351796, 'clip_ratio/region_mean': 0.03277078131213784, 'reward_total_mean': 0.978593111038208, 'reward_meter_mean': 0.9978033304214478, 'reward_meter_std': 0.0022864267230033875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9807692170143127, 'reward_repeat_penalty_std': 0.03560846298933029, 'reward_total_composite_mean': 0.978593111038208, 'reward_total_composite_std': 0.03494200482964516, 'epoch': 0.13}

 99%|█████████▊| 3254/3300 [5:02:37<11:08, 14.53s/it]INFO 04-12 03:39:28 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▊| 3255/3300 [5:02:42<08:39, 11.54s/it]2026-04-12 03:39:33,210 | INFO | train_grpo_train | metrics_logged mode=train step=3255

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.393939393939394e-07, 'num_tokens': 7417000.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 8.257458102889359e-05, 'sampling/sampling_logp_difference/max': 0.005363806616514921, 'sampling/importance_sampling_ratio/min': 0.9946505427360535, 'sampling/importance_sampling_ratio/mean': 1.0000584125518799, 'sampling/importance_sampling_ratio/max': 1.0015891790390015, 'entropy': 0.000789376012107823, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 99%|█████████▊| 3255/3300 [5:02:42<08:39, 11.54s/it]INFO 04-12 03:39:33 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▊| 3256/3300 [5:02:47<07:01,  9.58s/it]2026-04-12 03:39:38,220 | INFO | train_grpo_train | metrics_logged mode=train step=3256

                                                     
{'loss': 0.0005, 'grad_norm': 1.5352351665496826, 'learning_rate': 1.3636363636363637e-07, 'num_tokens': 7419231.0, 'completions/mean_length': 100.875, 'completions/min_length': 100.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.875, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9993561506271362, 'rewards/meter/std': 0.00013285702152643353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9743703603744507, 'rewards/total_composite/std': 0.07064089924097061, 'reward': 0.9743703603744507, 'reward_std': 0.07064089179039001, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01845725066959858, 'sampling/sampling_logp_difference/max': 1.2666641473770142, 'sampling/importance_sampling_ratio/min': 0.2817700207233429, 'sampling/importance_sampling_ratio/mean': 1.0004955530166626, 'sampling/importance_sampling_ratio/max': 1.6758545637130737, 'entropy': 0.13637553714215755, 'clip_ratio/low_mean': 0.0012376237427815795, 'clip_ratio/low_min': 0.0012376237427815795, 'clip_ratio/high_mean': 0.01858959416858852, 'clip_ratio/high_max': 0.01858959416858852, 'clip_ratio/region_mean': 0.0198272179113701, 'reward_total_mean': 0.9743703603744507, 'reward_meter_mean': 0.9993561506271362, 'reward_meter_std': 0.00013285702152643353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9743703603744507, 'reward_total_composite_std': 0.07064089924097061, 'epoch': 0.13}

 99%|█████████▊| 3256/3300 [5:02:47<07:01,  9.58s/it]INFO 04-12 03:39:38 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▊| 3257/3300 [5:02:51<05:47,  8.09s/it]2026-04-12 03:39:42,838 | INFO | train_grpo_train | metrics_logged mode=train step=3257

                                                     
{'loss': 0.0003, 'grad_norm': 1.7167513370513916, 'learning_rate': 1.3333333333333336e-07, 'num_tokens': 7421119.0, 'completions/mean_length': 71.0, 'completions/min_length': 70.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994316101074219, 'rewards/meter/std': 0.0001929309801198542, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994316101074219, 'rewards/total_composite/std': 0.0001929309801198542, 'reward': 0.9994316101074219, 'reward_std': 0.00019294493540655822, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.014849050901830196, 'sampling/sampling_logp_difference/max': 1.1428728103637695, 'sampling/importance_sampling_ratio/min': 0.31890156865119934, 'sampling/importance_sampling_ratio/mean': 0.9987840056419373, 'sampling/importance_sampling_ratio/max': 1.2882606983184814, 'entropy': 0.0983225479722023, 'clip_ratio/low_mean': 0.007067404338158667, 'clip_ratio/low_min': 0.007067404338158667, 'clip_ratio/high_mean': 0.008754611015319824, 'clip_ratio/high_max': 0.008754611015319824, 'clip_ratio/region_mean': 0.01582201535347849, 'reward_total_mean': 0.9994316101074219, 'reward_meter_mean': 0.9994316101074219, 'reward_meter_std': 0.0001929309801198542, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994316101074219, 'reward_total_composite_std': 0.0001929309801198542, 'epoch': 0.13}

 99%|█████████▊| 3257/3300 [5:02:51<05:47,  8.09s/it]INFO 04-12 03:39:43 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▊| 3258/3300 [5:02:56<04:55,  7.05s/it]2026-04-12 03:39:47,459 | INFO | train_grpo_train | metrics_logged mode=train step=3258

                                                     
{'loss': 0.0004, 'grad_norm': 0.16614317893981934, 'learning_rate': 1.3030303030303033e-07, 'num_tokens': 7422983.0, 'completions/mean_length': 71.0, 'completions/min_length': 71.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9994429349899292, 'rewards/meter/std': 1.4887036741129123e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994429349899292, 'rewards/total_composite/std': 1.4887036741129123e-05, 'reward': 0.9994429349899292, 'reward_std': 1.4891715181875043e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013734190724790096, 'sampling/sampling_logp_difference/max': 1.3592395782470703, 'sampling/importance_sampling_ratio/min': 0.2568560242652893, 'sampling/importance_sampling_ratio/mean': 0.9997973442077637, 'sampling/importance_sampling_ratio/max': 1.2386420965194702, 'entropy': 0.0794384153559804, 'clip_ratio/low_mean': 0.0017605633474886417, 'clip_ratio/low_min': 0.0017605633474886417, 'clip_ratio/high_mean': 0.01056338008493185, 'clip_ratio/high_max': 0.01056338008493185, 'clip_ratio/region_mean': 0.012323943432420492, 'reward_total_mean': 0.9994429349899292, 'reward_meter_mean': 0.9994429349899292, 'reward_meter_std': 1.4887036741129123e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994429349899292, 'reward_total_composite_std': 1.4887036741129123e-05, 'epoch': 0.13}

 99%|█████████▊| 3258/3300 [5:02:56<04:55,  7.05s/it]INFO 04-12 03:39:47 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3259/3300 [5:03:02<04:32,  6.65s/it]2026-04-12 03:39:53,186 | INFO | train_grpo_train | metrics_logged mode=train step=3259

                                                     
{'loss': -0.0021, 'grad_norm': 6.062583923339844, 'learning_rate': 1.272727272727273e-07, 'num_tokens': 7424765.0, 'completions/mean_length': 65.75, 'completions/min_length': 63.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9497079849243164, 'rewards/meter/std': 0.005244073923677206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9497079849243164, 'rewards/total_composite/std': 0.005244073923677206, 'reward': 0.9497079849243164, 'reward_std': 0.00524406973272562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030090492218732834, 'sampling/sampling_logp_difference/max': 1.032027244567871, 'sampling/importance_sampling_ratio/min': 0.3562839925289154, 'sampling/importance_sampling_ratio/mean': 1.0019686222076416, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1472517903894186, 'clip_ratio/low_mean': 0.01923532225191593, 'clip_ratio/low_min': 0.01923532225191593, 'clip_ratio/high_mean': 0.018950162804685533, 'clip_ratio/high_max': 0.018950162804685533, 'clip_ratio/region_mean': 0.038185485056601465, 'reward_total_mean': 0.9497079849243164, 'reward_meter_mean': 0.9497079849243164, 'reward_meter_std': 0.005244073923677206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9497079849243164, 'reward_total_composite_std': 0.005244073923677206, 'epoch': 0.13}

 99%|█████████▉| 3259/3300 [5:03:02<04:32,  6.65s/it]INFO 04-12 03:39:53 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3260/3300 [5:03:06<03:56,  5.92s/it]2026-04-12 03:39:57,379 | INFO | train_grpo_train | metrics_logged mode=train step=3260

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.2424242424242426e-07, 'num_tokens': 7426277.0, 'completions/mean_length': 32.0, 'completions/min_length': 32.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9992982149124146, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992982149124146, 'rewards/total_composite/std': 0.0, 'reward': 0.9992982149124146, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 5.620270167128183e-05, 'sampling/sampling_logp_difference/max': 0.001364716561511159, 'sampling/importance_sampling_ratio/min': 0.9986362457275391, 'sampling/importance_sampling_ratio/mean': 1.0000357627868652, 'sampling/importance_sampling_ratio/max': 1.0006037950515747, 'entropy': 0.0003866630977427121, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9992982149124146, 'reward_meter_mean': 0.9992982149124146, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992982149124146, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 99%|█████████▉| 3260/3300 [5:03:06<03:56,  5.92s/it]INFO 04-12 03:39:57 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3261/3300 [5:03:10<03:36,  5.56s/it]2026-04-12 03:40:02,120 | INFO | train_grpo_train | metrics_logged mode=train step=3261

                                                     
{'loss': 0.0044, 'grad_norm': 3.223155975341797, 'learning_rate': 1.2121212121212122e-07, 'num_tokens': 7428154.0, 'completions/mean_length': 69.625, 'completions/min_length': 66.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9926314949989319, 'rewards/meter/std': 0.006077465135604143, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9926314949989319, 'rewards/total_composite/std': 0.006077465135604143, 'reward': 0.9926314949989319, 'reward_std': 0.006077463272958994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029856102541089058, 'sampling/sampling_logp_difference/max': 1.1406583786010742, 'sampling/importance_sampling_ratio/min': 0.31960853934288025, 'sampling/importance_sampling_ratio/mean': 1.0040087699890137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1911148466169834, 'clip_ratio/low_mean': 0.007126289419829845, 'clip_ratio/low_min': 0.007126289419829845, 'clip_ratio/high_mean': 0.012933629681356251, 'clip_ratio/high_max': 0.012933629681356251, 'clip_ratio/region_mean': 0.020059919101186097, 'reward_total_mean': 0.9926314949989319, 'reward_meter_mean': 0.9926314949989319, 'reward_meter_std': 0.006077465135604143, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9926314949989319, 'reward_total_composite_std': 0.006077465135604143, 'epoch': 0.13}

 99%|█████████▉| 3261/3300 [5:03:11<03:36,  5.56s/it]INFO 04-12 03:40:02 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3262/3300 [5:03:15<03:21,  5.29s/it]2026-04-12 03:40:06,788 | INFO | train_grpo_train | metrics_logged mode=train step=3262

                                                     
{'loss': 0.0013, 'grad_norm': 0.49830862879753113, 'learning_rate': 1.1818181818181818e-07, 'num_tokens': 7429891.0, 'completions/mean_length': 71.125, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994339346885681, 'rewards/meter/std': 2.7899814085685648e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994339346885681, 'rewards/total_composite/std': 2.7899814085685648e-05, 'reward': 0.9994339346885681, 'reward_std': 2.791790757328272e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011753112077713013, 'sampling/sampling_logp_difference/max': 1.6477203369140625, 'sampling/importance_sampling_ratio/min': 0.19248820841312408, 'sampling/importance_sampling_ratio/mean': 1.0001577138900757, 'sampling/importance_sampling_ratio/max': 1.19877290725708, 'entropy': 0.07193375751376152, 'clip_ratio/low_mean': 0.0034966744715347886, 'clip_ratio/low_min': 0.0034966744715347886, 'clip_ratio/high_mean': 0.0017605633474886417, 'clip_ratio/high_max': 0.0017605633474886417, 'clip_ratio/region_mean': 0.00525723781902343, 'reward_total_mean': 0.9994339346885681, 'reward_meter_mean': 0.9994339346885681, 'reward_meter_std': 2.7899814085685648e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994339346885681, 'reward_total_composite_std': 2.7899814085685648e-05, 'epoch': 0.13}

 99%|█████████▉| 3262/3300 [5:03:15<03:21,  5.29s/it]INFO 04-12 03:40:07 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3263/3300 [5:03:20<03:13,  5.22s/it]2026-04-12 03:40:11,819 | INFO | train_grpo_train | metrics_logged mode=train step=3263

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.1515151515151516e-07, 'num_tokens': 7431611.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973388910293579, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973388910293579, 'rewards/total_composite/std': 0.0, 'reward': 0.9973388910293579, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0008966223103925586, 'sampling/sampling_logp_difference/max': 0.09105479717254639, 'sampling/importance_sampling_ratio/min': 0.9129676818847656, 'sampling/importance_sampling_ratio/mean': 1.0003302097320557, 'sampling/importance_sampling_ratio/max': 1.0381407737731934, 'entropy': 0.012014997191727161, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9973388910293579, 'reward_meter_mean': 0.9973388910293579, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973388910293579, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 99%|█████████▉| 3263/3300 [5:03:20<03:13,  5.22s/it]INFO 04-12 03:40:12 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3264/3300 [5:03:25<03:03,  5.11s/it]2026-04-12 03:40:16,667 | INFO | train_grpo_train | metrics_logged mode=train step=3264

                                                     
{'loss': 0.0001, 'grad_norm': 0.04755792394280434, 'learning_rate': 1.1212121212121213e-07, 'num_tokens': 7433682.0, 'completions/mean_length': 97.875, 'completions/min_length': 97.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.875, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.999409019947052, 'rewards/meter/std': 5.810459242638899e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999409019947052, 'rewards/total_composite/std': 5.810459242638899e-06, 'reward': 0.999409019947052, 'reward_std': 5.812751624034718e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.003859634278342128, 'sampling/sampling_logp_difference/max': 0.3599740266799927, 'sampling/importance_sampling_ratio/min': 0.7912850379943848, 'sampling/importance_sampling_ratio/mean': 1.0018383264541626, 'sampling/importance_sampling_ratio/max': 1.4332921504974365, 'entropy': 0.03584319236688316, 'clip_ratio/low_mean': 0.0038265305338427424, 'clip_ratio/low_min': 0.0038265305338427424, 'clip_ratio/high_mean': 0.003839680110104382, 'clip_ratio/high_max': 0.003839680110104382, 'clip_ratio/region_mean': 0.0076662106439471245, 'reward_total_mean': 0.999409019947052, 'reward_meter_mean': 0.999409019947052, 'reward_meter_std': 5.810459242638899e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999409019947052, 'reward_total_composite_std': 5.810459242638899e-06, 'epoch': 0.13}

 99%|█████████▉| 3264/3300 [5:03:25<03:03,  5.11s/it]INFO 04-12 03:40:16 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3265/3300 [5:03:31<03:06,  5.33s/it]2026-04-12 03:40:22,530 | INFO | train_grpo_train | metrics_logged mode=train step=3265

                                                     
{'loss': 0.0039, 'grad_norm': 2.038606643676758, 'learning_rate': 1.090909090909091e-07, 'num_tokens': 7436489.0, 'completions/mean_length': 166.875, 'completions/min_length': 162.0, 'completions/max_length': 169.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.875, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 169.0, 'rewards/meter/mean': 0.999045729637146, 'rewards/meter/std': 0.00020887328719254583, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999045729637146, 'rewards/total_composite/std': 0.00020887328719254583, 'reward': 0.999045729637146, 'reward_std': 0.00020887046412099153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03505001217126846, 'sampling/sampling_logp_difference/max': 1.5322790145874023, 'sampling/importance_sampling_ratio/min': 0.21604275703430176, 'sampling/importance_sampling_ratio/mean': 1.0036981105804443, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31825447641313076, 'clip_ratio/low_mean': 0.011210258584469557, 'clip_ratio/low_min': 0.011210258584469557, 'clip_ratio/high_mean': 0.015736370929516852, 'clip_ratio/high_max': 0.015736370929516852, 'clip_ratio/region_mean': 0.02694662951398641, 'reward_total_mean': 0.999045729637146, 'reward_meter_mean': 0.999045729637146, 'reward_meter_std': 0.00020887328719254583, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999045729637146, 'reward_total_composite_std': 0.00020887328719254583, 'epoch': 0.13}

 99%|█████████▉| 3265/3300 [5:03:31<03:06,  5.33s/it]INFO 04-12 03:40:22 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3266/3300 [5:03:37<03:07,  5.50s/it]2026-04-12 03:40:28,419 | INFO | train_grpo_train | metrics_logged mode=train step=3266

                                                     
{'loss': -0.003, 'grad_norm': 1.5206332206726074, 'learning_rate': 1.0606060606060608e-07, 'num_tokens': 7439156.0, 'completions/mean_length': 155.375, 'completions/min_length': 152.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 155.375, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.9990906119346619, 'rewards/meter/std': 0.0002633438853081316, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990906119346619, 'rewards/total_composite/std': 0.0002633438853081316, 'reward': 0.9990906119346619, 'reward_std': 0.0002633513940963894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.031446803361177444, 'sampling/sampling_logp_difference/max': 1.307551383972168, 'sampling/importance_sampling_ratio/min': 0.27048155665397644, 'sampling/importance_sampling_ratio/mean': 1.009647250175476, 'sampling/importance_sampling_ratio/max': 1.7514793872833252, 'entropy': 0.30485600233078003, 'clip_ratio/low_mean': 0.003257433301769197, 'clip_ratio/low_min': 0.003257433301769197, 'clip_ratio/high_mean': 0.018389825709164143, 'clip_ratio/high_max': 0.018389825709164143, 'clip_ratio/region_mean': 0.02164725901093334, 'reward_total_mean': 0.9990906119346619, 'reward_meter_mean': 0.9990906119346619, 'reward_meter_std': 0.0002633438853081316, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990906119346619, 'reward_total_composite_std': 0.0002633438853081316, 'epoch': 0.13}

 99%|█████████▉| 3266/3300 [5:03:37<03:07,  5.50s/it]INFO 04-12 03:40:28 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3267/3300 [5:03:42<03:03,  5.56s/it]2026-04-12 03:40:34,116 | INFO | train_grpo_train | metrics_logged mode=train step=3267

                                                     
{'loss': 0.0021, 'grad_norm': 1.932774305343628, 'learning_rate': 1.0303030303030304e-07, 'num_tokens': 7441984.0, 'completions/mean_length': 156.5, 'completions/min_length': 155.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 156.5, 'completions/min_terminated_length': 155.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.9990622997283936, 'rewards/meter/std': 0.00025987729895859957, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990622997283936, 'rewards/total_composite/std': 0.00025987729895859957, 'reward': 0.9990622997283936, 'reward_std': 0.0002598702849354595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.035560522228479385, 'sampling/sampling_logp_difference/max': 1.2120389938354492, 'sampling/importance_sampling_ratio/min': 0.29758986830711365, 'sampling/importance_sampling_ratio/mean': 1.0105011463165283, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3256886452436447, 'clip_ratio/low_mean': 0.004726932500489056, 'clip_ratio/low_min': 0.004726932500489056, 'clip_ratio/high_mean': 0.011238692968618125, 'clip_ratio/high_max': 0.011238692968618125, 'clip_ratio/region_mean': 0.01596562546910718, 'reward_total_mean': 0.9990622997283936, 'reward_meter_mean': 0.9990622997283936, 'reward_meter_std': 0.00025987729895859957, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990622997283936, 'reward_total_composite_std': 0.00025987729895859957, 'epoch': 0.13}

 99%|█████████▉| 3267/3300 [5:03:43<03:03,  5.56s/it]INFO 04-12 03:40:34 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3268/3300 [5:03:48<02:55,  5.48s/it]2026-04-12 03:40:39,416 | INFO | train_grpo_train | metrics_logged mode=train step=3268

                                                     
{'loss': -0.0011, 'grad_norm': 0.6131901741027832, 'learning_rate': 1.0000000000000001e-07, 'num_tokens': 7444602.0, 'completions/mean_length': 131.25, 'completions/min_length': 130.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.25, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.9992755651473999, 'rewards/meter/std': 0.00033690276904962957, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992755651473999, 'rewards/total_composite/std': 0.00033690276904962957, 'reward': 0.9992755651473999, 'reward_std': 0.0003369078040122986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011317918077111244, 'sampling/sampling_logp_difference/max': 2.1706619262695312, 'sampling/importance_sampling_ratio/min': 0.1141020655632019, 'sampling/importance_sampling_ratio/mean': 0.9999168515205383, 'sampling/importance_sampling_ratio/max': 1.4053065776824951, 'entropy': 0.08114278595894575, 'clip_ratio/low_mean': 0.0019083969527855515, 'clip_ratio/low_min': 0.0019083969527855515, 'clip_ratio/high_mean': 0.005740093358326703, 'clip_ratio/high_max': 0.005740093358326703, 'clip_ratio/region_mean': 0.007648490311112255, 'reward_total_mean': 0.9992755651473999, 'reward_meter_mean': 0.9992755651473999, 'reward_meter_std': 0.00033690276904962957, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992755651473999, 'reward_total_composite_std': 0.00033690276904962957, 'epoch': 0.13}

 99%|█████████▉| 3268/3300 [5:03:48<02:55,  5.48s/it]INFO 04-12 03:40:39 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3269/3300 [5:03:52<02:38,  5.12s/it]2026-04-12 03:40:43,691 | INFO | train_grpo_train | metrics_logged mode=train step=3269

                                                     
{'loss': 0.0089, 'grad_norm': 3.9373042583465576, 'learning_rate': 9.696969696969697e-08, 'num_tokens': 7446314.0, 'completions/mean_length': 46.0, 'completions/min_length': 46.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9440339803695679, 'rewards/meter/std': 0.0030637700110673904, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9440339803695679, 'rewards/total_composite/std': 0.0030637700110673904, 'reward': 0.9440339803695679, 'reward_std': 0.0030637700110673904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020654289051890373, 'sampling/sampling_logp_difference/max': 1.6016666889190674, 'sampling/importance_sampling_ratio/min': 0.20156030356884003, 'sampling/importance_sampling_ratio/mean': 1.0039082765579224, 'sampling/importance_sampling_ratio/max': 1.5834907293319702, 'entropy': 0.08023244515061378, 'clip_ratio/low_mean': 0.00815217406488955, 'clip_ratio/low_min': 0.00815217406488955, 'clip_ratio/high_mean': 0.00815217406488955, 'clip_ratio/high_max': 0.00815217406488955, 'clip_ratio/region_mean': 0.0163043481297791, 'reward_total_mean': 0.9440339803695679, 'reward_meter_mean': 0.9440339803695679, 'reward_meter_std': 0.0030637700110673904, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9440339803695679, 'reward_total_composite_std': 0.0030637700110673904, 'epoch': 0.13}

 99%|█████████▉| 3269/3300 [5:03:52<02:38,  5.12s/it]INFO 04-12 03:40:43 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3270/3300 [5:03:59<02:47,  5.59s/it]2026-04-12 03:40:50,402 | INFO | train_grpo_train | metrics_logged mode=train step=3270

                                                     
{'loss': 0.0097, 'grad_norm': 2.0776824951171875, 'learning_rate': 9.393939393939395e-08, 'num_tokens': 7449775.0, 'completions/mean_length': 232.625, 'completions/min_length': 226.0, 'completions/max_length': 238.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 232.625, 'completions/min_terminated_length': 226.0, 'completions/max_terminated_length': 238.0, 'rewards/meter/mean': 0.9988599419593811, 'rewards/meter/std': 0.0006400212296284735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9988599419593811, 'rewards/total_composite/std': 0.0006400212296284735, 'reward': 0.9988599419593811, 'reward_std': 0.0006400320562534034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04505442455410957, 'sampling/sampling_logp_difference/max': 1.3682456016540527, 'sampling/importance_sampling_ratio/min': 0.2545531392097473, 'sampling/importance_sampling_ratio/mean': 1.0114996433258057, 'sampling/importance_sampling_ratio/max': 1.781883716583252, 'entropy': 0.4155345596373081, 'clip_ratio/low_mean': 0.005398246110416949, 'clip_ratio/low_min': 0.005398246110416949, 'clip_ratio/high_mean': 0.016811800538562238, 'clip_ratio/high_max': 0.016811800538562238, 'clip_ratio/region_mean': 0.022210046648979187, 'reward_total_mean': 0.9988599419593811, 'reward_meter_mean': 0.9988599419593811, 'reward_meter_std': 0.0006400212296284735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9988599419593811, 'reward_total_composite_std': 0.0006400212296284735, 'epoch': 0.13}

 99%|█████████▉| 3270/3300 [5:03:59<02:47,  5.59s/it]INFO 04-12 03:40:50 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3271/3300 [5:04:04<02:38,  5.48s/it]2026-04-12 03:40:55,595 | INFO | train_grpo_train | metrics_logged mode=train step=3271

                                                     
{'loss': 0.0018, 'grad_norm': 0.9732382893562317, 'learning_rate': 9.090909090909091e-08, 'num_tokens': 7451918.0, 'completions/mean_length': 106.875, 'completions/min_length': 106.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.875, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9993022680282593, 'rewards/meter/std': 8.71433803695254e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993022680282593, 'rewards/total_composite/std': 8.71433803695254e-05, 'reward': 0.9993022680282593, 'reward_std': 8.713654096936807e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016620170325040817, 'sampling/sampling_logp_difference/max': 1.211838722229004, 'sampling/importance_sampling_ratio/min': 0.29764947295188904, 'sampling/importance_sampling_ratio/mean': 1.00331711769104, 'sampling/importance_sampling_ratio/max': 1.4888734817504883, 'entropy': 0.16768567822873592, 'clip_ratio/low_mean': 0.004673101240769029, 'clip_ratio/low_min': 0.004673101240769029, 'clip_ratio/high_mean': 0.0093678361736238, 'clip_ratio/high_max': 0.0093678361736238, 'clip_ratio/region_mean': 0.014040937414392829, 'reward_total_mean': 0.9993022680282593, 'reward_meter_mean': 0.9993022680282593, 'reward_meter_std': 8.71433803695254e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993022680282593, 'reward_total_composite_std': 8.71433803695254e-05, 'epoch': 0.13}

 99%|█████████▉| 3271/3300 [5:04:04<02:38,  5.48s/it]INFO 04-12 03:40:55 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3272/3300 [5:04:10<02:34,  5.52s/it]2026-04-12 03:41:01,240 | INFO | train_grpo_train | metrics_logged mode=train step=3272

                                                     
{'loss': -0.0028, 'grad_norm': 2.9168341159820557, 'learning_rate': 8.787878787878788e-08, 'num_tokens': 7454434.0, 'completions/mean_length': 136.5, 'completions/min_length': 132.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.5, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9926499128341675, 'rewards/meter/std': 0.0036937205586582422, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9926499128341675, 'rewards/total_composite/std': 0.0036937205586582422, 'reward': 0.9926499128341675, 'reward_std': 0.0036937135737389326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0317571647465229, 'sampling/sampling_logp_difference/max': 1.151773452758789, 'sampling/importance_sampling_ratio/min': 0.31607571244239807, 'sampling/importance_sampling_ratio/mean': 1.0109081268310547, 'sampling/importance_sampling_ratio/max': 1.8989595174789429, 'entropy': 0.310006458312273, 'clip_ratio/low_mean': 0.008297410036902875, 'clip_ratio/low_min': 0.008297410036902875, 'clip_ratio/high_mean': 0.017071759328246117, 'clip_ratio/high_max': 0.017071759328246117, 'clip_ratio/region_mean': 0.02536916936514899, 'reward_total_mean': 0.9926499128341675, 'reward_meter_mean': 0.9926499128341675, 'reward_meter_std': 0.0036937205586582422, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9926499128341675, 'reward_total_composite_std': 0.0036937205586582422, 'epoch': 0.13}

 99%|█████████▉| 3272/3300 [5:04:10<02:34,  5.52s/it]INFO 04-12 03:41:01 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3273/3300 [5:04:14<02:21,  5.26s/it]2026-04-12 03:41:05,871 | INFO | train_grpo_train | metrics_logged mode=train step=3273

                                                     
{'loss': 0.0013, 'grad_norm': 2.5891594886779785, 'learning_rate': 8.484848484848487e-08, 'num_tokens': 7456266.0, 'completions/mean_length': 67.0, 'completions/min_length': 67.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981439113616943, 'rewards/meter/std': 3.332228152430616e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981439113616943, 'rewards/total_composite/std': 3.332228152430616e-05, 'reward': 0.9981439113616943, 'reward_std': 3.332816413603723e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0074036987498402596, 'sampling/sampling_logp_difference/max': 0.8070247173309326, 'sampling/importance_sampling_ratio/min': 0.4461836516857147, 'sampling/importance_sampling_ratio/mean': 0.9988389611244202, 'sampling/importance_sampling_ratio/max': 1.3197448253631592, 'entropy': 0.04388477047905326, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.005597014795057476, 'clip_ratio/high_max': 0.005597014795057476, 'clip_ratio/region_mean': 0.005597014795057476, 'reward_total_mean': 0.9981439113616943, 'reward_meter_mean': 0.9981439113616943, 'reward_meter_std': 3.332228152430616e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981439113616943, 'reward_total_composite_std': 3.332228152430616e-05, 'epoch': 0.13}

 99%|█████████▉| 3273/3300 [5:04:14<02:21,  5.26s/it]INFO 04-12 03:41:06 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3274/3300 [5:04:19<02:11,  5.08s/it]2026-04-12 03:41:10,520 | INFO | train_grpo_train | metrics_logged mode=train step=3274

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.181818181818183e-08, 'num_tokens': 7458170.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00010550412116572261, 'sampling/sampling_logp_difference/max': 0.005598756484687328, 'sampling/importance_sampling_ratio/min': 0.9944168329238892, 'sampling/importance_sampling_ratio/mean': 1.0000689029693604, 'sampling/importance_sampling_ratio/max': 1.0018798112869263, 'entropy': 0.0009244889370165765, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 99%|█████████▉| 3274/3300 [5:04:19<02:11,  5.08s/it]INFO 04-12 03:41:10 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3275/3300 [5:04:23<02:02,  4.88s/it]2026-04-12 03:41:14,957 | INFO | train_grpo_train | metrics_logged mode=train step=3275

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.87878787878788e-08, 'num_tokens': 7459834.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00023801384668331593, 'sampling/sampling_logp_difference/max': 0.00208944920450449, 'sampling/importance_sampling_ratio/min': 0.9997556209564209, 'sampling/importance_sampling_ratio/mean': 1.0002355575561523, 'sampling/importance_sampling_ratio/max': 1.002091646194458, 'entropy': 0.0017418517236365005, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 99%|█████████▉| 3275/3300 [5:04:23<02:02,  4.88s/it]INFO 04-12 03:41:15 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3276/3300 [5:04:28<01:55,  4.82s/it]2026-04-12 03:41:19,623 | INFO | train_grpo_train | metrics_logged mode=train step=3276

                                                     
{'loss': -0.0024, 'grad_norm': 1.703597068786621, 'learning_rate': 7.575757575757576e-08, 'num_tokens': 7461675.0, 'completions/mean_length': 68.125, 'completions/min_length': 67.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.999450147151947, 'rewards/meter/std': 0.00012958116712979972, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.999450147151947, 'rewards/total_composite/std': 0.00012958116712979972, 'reward': 0.999450147151947, 'reward_std': 0.000129588384879753, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.011213197372853756, 'sampling/sampling_logp_difference/max': 0.7290372848510742, 'sampling/importance_sampling_ratio/min': 0.4823731780052185, 'sampling/importance_sampling_ratio/mean': 1.0002046823501587, 'sampling/importance_sampling_ratio/max': 1.684713363647461, 'entropy': 0.06052346946671605, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.999450147151947, 'reward_meter_mean': 0.999450147151947, 'reward_meter_std': 0.00012958116712979972, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.999450147151947, 'reward_total_composite_std': 0.00012958116712979972, 'epoch': 0.13}

 99%|█████████▉| 3276/3300 [5:04:28<01:55,  4.82s/it]INFO 04-12 03:41:19 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3277/3300 [5:04:36<02:09,  5.63s/it]2026-04-12 03:41:27,164 | INFO | train_grpo_train | metrics_logged mode=train step=3277

                                                     
{'loss': 0.0383, 'grad_norm': 3.1708502769470215, 'learning_rate': 7.272727272727274e-08, 'num_tokens': 7465634.0, 'completions/mean_length': 285.875, 'completions/min_length': 275.0, 'completions/max_length': 308.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 285.875, 'completions/min_terminated_length': 275.0, 'completions/max_terminated_length': 308.0, 'rewards/meter/mean': 0.9972749948501587, 'rewards/meter/std': 0.0005863041151314974, 'rewards/count_adherence/mean': 0.9722222089767456, 'rewards/count_adherence/std': 0.05143444985151291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9500774145126343, 'rewards/repeat_penalty/std': 0.03746138885617256, 'rewards/total_composite/mean': 0.9212331175804138, 'rewards/total_composite/std': 0.062395479530096054, 'reward': 0.9212331175804138, 'reward_std': 0.06239548325538635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039440274238586426, 'sampling/sampling_logp_difference/max': 1.6334123611450195, 'sampling/importance_sampling_ratio/min': 0.1952621191740036, 'sampling/importance_sampling_ratio/mean': 1.0104100704193115, 'sampling/importance_sampling_ratio/max': 1.906610369682312, 'entropy': 0.3647267259657383, 'clip_ratio/low_mean': 0.007274207135196775, 'clip_ratio/low_min': 0.007274207135196775, 'clip_ratio/high_mean': 0.018823198741301894, 'clip_ratio/high_max': 0.018823198741301894, 'clip_ratio/region_mean': 0.02609740587649867, 'reward_total_mean': 0.9212331175804138, 'reward_meter_mean': 0.9972749948501587, 'reward_meter_std': 0.0005863041151314974, 'reward_count_adherence_mean': 0.9722222089767456, 'reward_count_adherence_std': 0.05143444985151291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9500774145126343, 'reward_repeat_penalty_std': 0.03746138885617256, 'reward_total_composite_mean': 0.9212331175804138, 'reward_total_composite_std': 0.062395479530096054, 'epoch': 0.13}

 99%|█████████▉| 3277/3300 [5:04:36<02:09,  5.63s/it]INFO 04-12 03:41:27 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3278/3300 [5:04:43<02:18,  6.30s/it]2026-04-12 03:41:34,989 | INFO | train_grpo_train | metrics_logged mode=train step=3278

                                                     
{'loss': 0.0041, 'grad_norm': 2.639291286468506, 'learning_rate': 6.96969696969697e-08, 'num_tokens': 7469932.0, 'completions/mean_length': 333.25, 'completions/min_length': 331.0, 'completions/max_length': 337.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 333.25, 'completions/min_terminated_length': 331.0, 'completions/max_terminated_length': 337.0, 'rewards/meter/mean': 0.9980206489562988, 'rewards/meter/std': 0.0009543930646032095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9934210777282715, 'rewards/repeat_penalty/std': 0.01860806532204151, 'rewards/total_composite/mean': 0.9914476871490479, 'rewards/total_composite/std': 0.01817752607166767, 'reward': 0.9914476871490479, 'reward_std': 0.01817752793431282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0520353838801384, 'sampling/sampling_logp_difference/max': 1.9485530853271484, 'sampling/importance_sampling_ratio/min': 0.14248007535934448, 'sampling/importance_sampling_ratio/mean': 1.010366678237915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4740913324058056, 'clip_ratio/low_mean': 0.005647590383887291, 'clip_ratio/low_min': 0.005647590383887291, 'clip_ratio/high_mean': 0.03409638348966837, 'clip_ratio/high_max': 0.03409638348966837, 'clip_ratio/region_mean': 0.03974397387355566, 'reward_total_mean': 0.9914476871490479, 'reward_meter_mean': 0.9980206489562988, 'reward_meter_std': 0.0009543930646032095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9934210777282715, 'reward_repeat_penalty_std': 0.01860806532204151, 'reward_total_composite_mean': 0.9914476871490479, 'reward_total_composite_std': 0.01817752607166767, 'epoch': 0.13}

 99%|█████████▉| 3278/3300 [5:04:43<02:18,  6.30s/it]INFO 04-12 03:41:35 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3279/3300 [5:04:48<02:00,  5.75s/it]2026-04-12 03:41:39,452 | INFO | train_grpo_train | metrics_logged mode=train step=3279

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.666666666666668e-08, 'num_tokens': 7471564.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 8.241144678322598e-05, 'sampling/sampling_logp_difference/max': 0.0017932088812813163, 'sampling/importance_sampling_ratio/min': 0.9982084035873413, 'sampling/importance_sampling_ratio/mean': 1.0000587701797485, 'sampling/importance_sampling_ratio/max': 1.0012202262878418, 'entropy': 0.0007044602243695408, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 99%|█████████▉| 3279/3300 [5:04:48<02:00,  5.75s/it]INFO 04-12 03:41:39 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3280/3300 [5:04:53<01:49,  5.48s/it]2026-04-12 03:41:44,323 | INFO | train_grpo_train | metrics_logged mode=train step=3280

                                                     
{'loss': 0.0002, 'grad_norm': 0.038945455104112625, 'learning_rate': 6.363636363636365e-08, 'num_tokens': 7473684.0, 'completions/mean_length': 98.0, 'completions/min_length': 98.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9994074702262878, 'rewards/meter/std': 3.4306899578950834e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994074702262878, 'rewards/total_composite/std': 3.4306899578950834e-06, 'reward': 0.9994074702262878, 'reward_std': 3.428247964620823e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0032499232329428196, 'sampling/sampling_logp_difference/max': 0.3044092655181885, 'sampling/importance_sampling_ratio/min': 0.7375589609146118, 'sampling/importance_sampling_ratio/mean': 1.0016450881958008, 'sampling/importance_sampling_ratio/max': 1.345357894897461, 'entropy': 0.033778895856812596, 'clip_ratio/low_mean': 0.006377550889737904, 'clip_ratio/low_min': 0.006377550889737904, 'clip_ratio/high_mean': 0.0025510203558951616, 'clip_ratio/high_max': 0.0025510203558951616, 'clip_ratio/region_mean': 0.008928571245633066, 'reward_total_mean': 0.9994074702262878, 'reward_meter_mean': 0.9994074702262878, 'reward_meter_std': 3.4306899578950834e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994074702262878, 'reward_total_composite_std': 3.4306899578950834e-06, 'epoch': 0.13}

 99%|█████████▉| 3280/3300 [5:04:53<01:49,  5.48s/it]INFO 04-12 03:41:44 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3281/3300 [5:04:57<01:36,  5.09s/it]2026-04-12 03:41:48,498 | INFO | train_grpo_train | metrics_logged mode=train step=3281

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.060606060606061e-08, 'num_tokens': 7475108.0, 'completions/mean_length': 34.0, 'completions/min_length': 34.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9977458119392395, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9977458119392395, 'rewards/total_composite/std': 0.0, 'reward': 0.9977458119392395, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.002193058142438531, 'sampling/sampling_logp_difference/max': 0.06379712373018265, 'sampling/importance_sampling_ratio/min': 0.9427482485771179, 'sampling/importance_sampling_ratio/mean': 1.0015556812286377, 'sampling/importance_sampling_ratio/max': 1.0658761262893677, 'entropy': 0.024989124154672027, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9977458119392395, 'reward_meter_mean': 0.9977458119392395, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9977458119392395, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

 99%|█████████▉| 3281/3300 [5:04:57<01:36,  5.09s/it]INFO 04-12 03:41:48 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3282/3300 [5:05:02<01:30,  5.03s/it]2026-04-12 03:41:53,392 | INFO | train_grpo_train | metrics_logged mode=train step=3282

                                                     
{'loss': 0.0002, 'grad_norm': 0.13846832513809204, 'learning_rate': 5.757575757575758e-08, 'num_tokens': 7476965.0, 'completions/mean_length': 71.125, 'completions/min_length': 71.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9994271993637085, 'rewards/meter/std': 1.0313916391169187e-05, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9994271993637085, 'rewards/total_composite/std': 1.0313916391169187e-05, 'reward': 0.9994271993637085, 'reward_std': 1.0320004548702855e-05, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.008291705511510372, 'sampling/sampling_logp_difference/max': 0.5785994529724121, 'sampling/importance_sampling_ratio/min': 0.5606830716133118, 'sampling/importance_sampling_ratio/mean': 1.0017709732055664, 'sampling/importance_sampling_ratio/max': 1.1672364473342896, 'entropy': 0.0743112824857235, 'clip_ratio/low_mean': 0.0017361111240461469, 'clip_ratio/low_min': 0.0017361111240461469, 'clip_ratio/high_mean': 0.0017605633474886417, 'clip_ratio/high_max': 0.0017605633474886417, 'clip_ratio/region_mean': 0.0034966744715347886, 'reward_total_mean': 0.9994271993637085, 'reward_meter_mean': 0.9994271993637085, 'reward_meter_std': 1.0313916391169187e-05, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9994271993637085, 'reward_total_composite_std': 1.0313916391169187e-05, 'epoch': 0.13}

 99%|█████████▉| 3282/3300 [5:05:02<01:30,  5.03s/it]INFO 04-12 03:41:53 [block_pool.py:378] Successfully reset prefix cache

 99%|█████████▉| 3283/3300 [5:05:06<01:23,  4.94s/it]2026-04-12 03:41:58,113 | INFO | train_grpo_train | metrics_logged mode=train step=3283

                                                     
{'loss': -0.0006, 'grad_norm': 0.04218384996056557, 'learning_rate': 5.454545454545455e-08, 'num_tokens': 7478764.0, 'completions/mean_length': 66.875, 'completions/min_length': 66.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9981493353843689, 'rewards/meter/std': 8.558939043723512e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9981493353843689, 'rewards/total_composite/std': 8.558939043723512e-06, 'reward': 0.9981493353843689, 'reward_std': 8.561799404560588e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.005590750835835934, 'sampling/sampling_logp_difference/max': 0.4178617000579834, 'sampling/importance_sampling_ratio/min': 0.8036051392555237, 'sampling/importance_sampling_ratio/mean': 1.002824306488037, 'sampling/importance_sampling_ratio/max': 1.5187106132507324, 'entropy': 0.03944724937900901, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.005597014795057476, 'clip_ratio/high_max': 0.005597014795057476, 'clip_ratio/region_mean': 0.009384893695823848, 'reward_total_mean': 0.9981493353843689, 'reward_meter_mean': 0.9981493353843689, 'reward_meter_std': 8.558939043723512e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9981493353843689, 'reward_total_composite_std': 8.558939043723512e-06, 'epoch': 0.13}

 99%|█████████▉| 3283/3300 [5:05:07<01:23,  4.94s/it]INFO 04-12 03:41:58 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3284/3300 [5:05:12<01:20,  5.06s/it]2026-04-12 03:42:03,452 | INFO | train_grpo_train | metrics_logged mode=train step=3284

                                                     
{'loss': -0.0012, 'grad_norm': 1.1970688104629517, 'learning_rate': 5.151515151515152e-08, 'num_tokens': 7481306.0, 'completions/mean_length': 131.75, 'completions/min_length': 131.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.75, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.9992576837539673, 'rewards/meter/std': 0.00048146978951990604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9992576837539673, 'rewards/total_composite/std': 0.00048146978951990604, 'reward': 0.9992576837539673, 'reward_std': 0.0004814636486116797, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.009545176289975643, 'sampling/sampling_logp_difference/max': 0.7750411033630371, 'sampling/importance_sampling_ratio/min': 0.4606848359107971, 'sampling/importance_sampling_ratio/mean': 1.0037535429000854, 'sampling/importance_sampling_ratio/max': 1.4763288497924805, 'entropy': 0.08821968082338572, 'clip_ratio/low_mean': 0.0009469697251915932, 'clip_ratio/low_min': 0.0009469697251915932, 'clip_ratio/high_mean': 0.0019083969527855515, 'clip_ratio/high_max': 0.0019083969527855515, 'clip_ratio/region_mean': 0.0028553666779771447, 'reward_total_mean': 0.9992576837539673, 'reward_meter_mean': 0.9992576837539673, 'reward_meter_std': 0.00048146978951990604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9992576837539673, 'reward_total_composite_std': 0.00048146978951990604, 'epoch': 0.13}

100%|█████████▉| 3284/3300 [5:05:12<01:20,  5.06s/it]INFO 04-12 03:42:03 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3285/3300 [5:05:20<01:31,  6.08s/it]2026-04-12 03:42:11,948 | INFO | train_grpo_train | metrics_logged mode=train step=3285

                                                     
{'loss': -0.0227, 'grad_norm': 1.6411669254302979, 'learning_rate': 4.8484848484848486e-08, 'num_tokens': 7485883.0, 'completions/mean_length': 361.125, 'completions/min_length': 347.0, 'completions/max_length': 385.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 361.125, 'completions/min_terminated_length': 347.0, 'completions/max_terminated_length': 385.0, 'rewards/meter/mean': 0.9987712502479553, 'rewards/meter/std': 0.0013375915586948395, 'rewards/count_adherence/mean': 0.8409091234207153, 'rewards/count_adherence/std': 0.04208274558186531, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9852941036224365, 'rewards/repeat_penalty/std': 0.027230001986026764, 'rewards/total_composite/mean': 0.827822208404541, 'rewards/total_composite/std': 0.05304892361164093, 'reward': 0.827822208404541, 'reward_std': 0.053048938512802124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04496872425079346, 'sampling/sampling_logp_difference/max': 1.6645660400390625, 'sampling/importance_sampling_ratio/min': 0.18927277624607086, 'sampling/importance_sampling_ratio/mean': 1.01465904712677, 'sampling/importance_sampling_ratio/max': 1.6690725088119507, 'entropy': 0.4625861681997776, 'clip_ratio/low_mean': 0.01729804463684559, 'clip_ratio/low_min': 0.01729804463684559, 'clip_ratio/high_mean': 0.005871425149962306, 'clip_ratio/high_max': 0.005871425149962306, 'clip_ratio/region_mean': 0.023169469786807895, 'reward_total_mean': 0.827822208404541, 'reward_meter_mean': 0.9987712502479553, 'reward_meter_std': 0.0013375915586948395, 'reward_count_adherence_mean': 0.8409091234207153, 'reward_count_adherence_std': 0.04208274558186531, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9852941036224365, 'reward_repeat_penalty_std': 0.027230001986026764, 'reward_total_composite_mean': 0.827822208404541, 'reward_total_composite_std': 0.05304892361164093, 'epoch': 0.13}

100%|█████████▉| 3285/3300 [5:05:20<01:31,  6.08s/it]INFO 04-12 03:42:12 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3286/3300 [5:05:27<01:28,  6.34s/it]2026-04-12 03:42:18,876 | INFO | train_grpo_train | metrics_logged mode=train step=3286

                                                     
{'loss': 0.0016, 'grad_norm': 0.6697944402694702, 'learning_rate': 4.545454545454546e-08, 'num_tokens': 7488852.0, 'completions/mean_length': 178.125, 'completions/min_length': 177.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.125, 'completions/min_terminated_length': 177.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.9990102052688599, 'rewards/meter/std': 0.00010323245805921033, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9990102052688599, 'rewards/total_composite/std': 0.00010323245805921033, 'reward': 0.9990102052688599, 'reward_std': 0.0001032423970173113, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02063995786011219, 'sampling/sampling_logp_difference/max': 0.810206413269043, 'sampling/importance_sampling_ratio/min': 0.44476625323295593, 'sampling/importance_sampling_ratio/mean': 1.006458044052124, 'sampling/importance_sampling_ratio/max': 1.7193830013275146, 'entropy': 0.24275393225252628, 'clip_ratio/low_mean': 0.0021067415946163237, 'clip_ratio/low_min': 0.0021067415946163237, 'clip_ratio/high_mean': 0.008415504998993129, 'clip_ratio/high_max': 0.008415504998993129, 'clip_ratio/region_mean': 0.010522246593609452, 'reward_total_mean': 0.9990102052688599, 'reward_meter_mean': 0.9990102052688599, 'reward_meter_std': 0.00010323245805921033, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9990102052688599, 'reward_total_composite_std': 0.00010323245805921033, 'epoch': 0.13}

100%|█████████▉| 3286/3300 [5:05:27<01:28,  6.34s/it]INFO 04-12 03:42:19 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3287/3300 [5:05:32<01:18,  6.01s/it]2026-04-12 03:42:24,121 | INFO | train_grpo_train | metrics_logged mode=train step=3287

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 4.2424242424242435e-08, 'num_tokens': 7490484.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0002587019116617739, 'sampling/sampling_logp_difference/max': 0.0030488455668091774, 'sampling/importance_sampling_ratio/min': 0.9995039105415344, 'sampling/importance_sampling_ratio/mean': 1.0002548694610596, 'sampling/importance_sampling_ratio/max': 1.0030535459518433, 'entropy': 0.0020569458283716813, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

100%|█████████▉| 3287/3300 [5:05:33<01:18,  6.01s/it]INFO 04-12 03:42:24 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3288/3300 [5:05:38<01:08,  5.72s/it]2026-04-12 03:42:29,158 | INFO | train_grpo_train | metrics_logged mode=train step=3288

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 3.93939393939394e-08, 'num_tokens': 7492204.0, 'completions/mean_length': 64.0, 'completions/min_length': 64.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9993994235992432, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9993994235992432, 'rewards/total_composite/std': 0.0, 'reward': 0.9993994235992432, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 9.810684423428029e-05, 'sampling/sampling_logp_difference/max': 0.0028118479531258345, 'sampling/importance_sampling_ratio/min': 0.9984528422355652, 'sampling/importance_sampling_ratio/mean': 1.0000782012939453, 'sampling/importance_sampling_ratio/max': 1.002815842628479, 'entropy': 0.001106615709431935, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9993994235992432, 'reward_meter_mean': 0.9993994235992432, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9993994235992432, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

100%|█████████▉| 3288/3300 [5:05:38<01:08,  5.72s/it]INFO 04-12 03:42:29 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3289/3300 [5:05:43<01:00,  5.53s/it]2026-04-12 03:42:34,240 | INFO | train_grpo_train | metrics_logged mode=train step=3289

                                                     
{'loss': 0.0109, 'grad_norm': 3.059532642364502, 'learning_rate': 3.636363636363637e-08, 'num_tokens': 7494281.0, 'completions/mean_length': 103.625, 'completions/min_length': 101.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.625, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9926178455352783, 'rewards/meter/std': 0.0021126074716448784, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9750000238418579, 'rewards/repeat_penalty/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.9677305221557617, 'rewards/total_composite/std': 0.06925326585769653, 'reward': 0.9677305221557617, 'reward_std': 0.06925328075885773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03573079779744148, 'sampling/sampling_logp_difference/max': 1.4160699844360352, 'sampling/importance_sampling_ratio/min': 0.2426658272743225, 'sampling/importance_sampling_ratio/mean': 1.0063260793685913, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2910528890788555, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.025478466413915157, 'clip_ratio/high_max': 0.025478466413915157, 'clip_ratio/region_mean': 0.025478466413915157, 'reward_total_mean': 0.9677305221557617, 'reward_meter_mean': 0.9926178455352783, 'reward_meter_std': 0.0021126074716448784, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9750000238418579, 'reward_repeat_penalty_std': 0.0707106739282608, 'reward_total_composite_mean': 0.9677305221557617, 'reward_total_composite_std': 0.06925326585769653, 'epoch': 0.13}

100%|█████████▉| 3289/3300 [5:05:43<01:00,  5.53s/it]INFO 04-12 03:42:34 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3290/3300 [5:05:49<00:57,  5.76s/it]2026-04-12 03:42:40,514 | INFO | train_grpo_train | metrics_logged mode=train step=3290

                                                     
{'loss': 0.0003, 'grad_norm': 2.9757983684539795, 'learning_rate': 3.333333333333334e-08, 'num_tokens': 7497469.0, 'completions/mean_length': 203.5, 'completions/min_length': 201.0, 'completions/max_length': 209.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 203.5, 'completions/min_terminated_length': 201.0, 'completions/max_terminated_length': 209.0, 'rewards/meter/mean': 0.9778822660446167, 'rewards/meter/std': 0.03634575754404068, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9772727489471436, 'rewards/repeat_penalty/std': 0.04208271950483322, 'rewards/total_composite/mean': 0.9552866816520691, 'rewards/total_composite/std': 0.0472070537507534, 'reward': 0.9552866816520691, 'reward_std': 0.047207050025463104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.038450971245765686, 'sampling/sampling_logp_difference/max': 1.2187190055847168, 'sampling/importance_sampling_ratio/min': 0.29560860991477966, 'sampling/importance_sampling_ratio/mean': 1.0083707571029663, 'sampling/importance_sampling_ratio/max': 1.8156616687774658, 'entropy': 0.38689322397112846, 'clip_ratio/low_mean': 0.009898273507133126, 'clip_ratio/low_min': 0.009898273507133126, 'clip_ratio/high_mean': 0.025775008834898472, 'clip_ratio/high_max': 0.025775008834898472, 'clip_ratio/region_mean': 0.0356732823420316, 'reward_total_mean': 0.9552866816520691, 'reward_meter_mean': 0.9778822660446167, 'reward_meter_std': 0.03634575754404068, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9772727489471436, 'reward_repeat_penalty_std': 0.04208271950483322, 'reward_total_composite_mean': 0.9552866816520691, 'reward_total_composite_std': 0.0472070537507534, 'epoch': 0.13}

100%|█████████▉| 3290/3300 [5:05:49<00:57,  5.76s/it]INFO 04-12 03:42:40 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3291/3300 [5:05:54<00:49,  5.46s/it]2026-04-12 03:42:45,304 | INFO | train_grpo_train | metrics_logged mode=train step=3291

                                                     
{'loss': 0.0059, 'grad_norm': 2.9630963802337646, 'learning_rate': 3.0303030303030305e-08, 'num_tokens': 7499282.0, 'completions/mean_length': 65.625, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9937740564346313, 'rewards/meter/std': 0.0003091543912887573, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9937740564346313, 'rewards/total_composite/std': 0.0003091543912887573, 'reward': 0.9937740564346313, 'reward_std': 0.0003091563412453979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.01210538949817419, 'sampling/sampling_logp_difference/max': 0.9835647940635681, 'sampling/importance_sampling_ratio/min': 0.37397557497024536, 'sampling/importance_sampling_ratio/mean': 1.0025181770324707, 'sampling/importance_sampling_ratio/max': 1.6661887168884277, 'entropy': 0.06898282328620553, 'clip_ratio/low_mean': 0.005710955825634301, 'clip_ratio/low_min': 0.005710955825634301, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.005710955825634301, 'reward_total_mean': 0.9937740564346313, 'reward_meter_mean': 0.9937740564346313, 'reward_meter_std': 0.0003091543912887573, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9937740564346313, 'reward_total_composite_std': 0.0003091543912887573, 'epoch': 0.13}

100%|█████████▉| 3291/3300 [5:05:54<00:49,  5.46s/it]INFO 04-12 03:42:45 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3292/3300 [5:05:58<00:41,  5.16s/it]2026-04-12 03:42:49,749 | INFO | train_grpo_train | metrics_logged mode=train step=3292

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 2.7272727272727276e-08, 'num_tokens': 7501162.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.000192559469724074, 'sampling/sampling_logp_difference/max': 0.0017719701863825321, 'sampling/importance_sampling_ratio/min': 0.9987837076187134, 'sampling/importance_sampling_ratio/mean': 1.000186800956726, 'sampling/importance_sampling_ratio/max': 1.001773476600647, 'entropy': 0.0016719198902137578, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

100%|█████████▉| 3292/3300 [5:05:58<00:41,  5.16s/it]INFO 04-12 03:42:49 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3293/3300 [5:06:08<00:45,  6.46s/it]2026-04-12 03:42:59,245 | INFO | train_grpo_train | metrics_logged mode=train step=3293

                                                     
{'loss': -0.0169, 'grad_norm': 1.2802501916885376, 'learning_rate': 2.4242424242424243e-08, 'num_tokens': 7506578.0, 'completions/mean_length': 428.0, 'completions/min_length': 408.0, 'completions/max_length': 448.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 428.0, 'completions/min_terminated_length': 408.0, 'completions/max_terminated_length': 448.0, 'rewards/meter/mean': 0.9988534450531006, 'rewards/meter/std': 0.00014253715926315635, 'rewards/count_adherence/mean': 0.8392857313156128, 'rewards/count_adherence/std': 0.03306501731276512, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9562541246414185, 'rewards/repeat_penalty/std': 0.03288932517170906, 'rewards/total_composite/mean': 0.8016811609268188, 'rewards/total_composite/std': 0.04294924437999725, 'reward': 0.8016811609268188, 'reward_std': 0.042949263006448746, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03670765459537506, 'sampling/sampling_logp_difference/max': 1.6504631042480469, 'sampling/importance_sampling_ratio/min': 0.19196099042892456, 'sampling/importance_sampling_ratio/mean': 1.0091326236724854, 'sampling/importance_sampling_ratio/max': 1.804500699043274, 'entropy': 0.39450087025761604, 'clip_ratio/low_mean': 0.008654525852762163, 'clip_ratio/low_min': 0.008654525852762163, 'clip_ratio/high_mean': 0.014237138908356428, 'clip_ratio/high_max': 0.014237138908356428, 'clip_ratio/region_mean': 0.02289166476111859, 'reward_total_mean': 0.8016811609268188, 'reward_meter_mean': 0.9988534450531006, 'reward_meter_std': 0.00014253715926315635, 'reward_count_adherence_mean': 0.8392857313156128, 'reward_count_adherence_std': 0.03306501731276512, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9562541246414185, 'reward_repeat_penalty_std': 0.03288932517170906, 'reward_total_composite_mean': 0.8016811609268188, 'reward_total_composite_std': 0.04294924437999725, 'epoch': 0.13}

100%|█████████▉| 3293/3300 [5:06:08<00:45,  6.46s/it]INFO 04-12 03:42:59 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3294/3300 [5:06:12<00:35,  5.91s/it]2026-04-12 03:43:03,854 | INFO | train_grpo_train | metrics_logged mode=train step=3294

                                                     
{'loss': -0.0, 'grad_norm': 0.030057502910494804, 'learning_rate': 2.1212121212121217e-08, 'num_tokens': 7508354.0, 'completions/mean_length': 61.0, 'completions/min_length': 61.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9973385334014893, 'rewards/meter/std': 1.0115243185282452e-06, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9973385334014893, 'rewards/total_composite/std': 1.0115243185282452e-06, 'reward': 0.9973385334014893, 'reward_std': 1.0115243185282452e-06, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.002317759208381176, 'sampling/sampling_logp_difference/max': 0.680436372756958, 'sampling/importance_sampling_ratio/min': 0.5063959360122681, 'sampling/importance_sampling_ratio/mean': 0.9997027516365051, 'sampling/importance_sampling_ratio/max': 1.0437626838684082, 'entropy': 0.012461528764106333, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.9973385334014893, 'reward_meter_mean': 0.9973385334014893, 'reward_meter_std': 1.0115243185282452e-06, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9973385334014893, 'reward_total_composite_std': 1.0115243185282452e-06, 'epoch': 0.13}

100%|█████████▉| 3294/3300 [5:06:12<00:35,  5.91s/it]INFO 04-12 03:43:04 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3295/3300 [5:06:17<00:27,  5.45s/it]2026-04-12 03:43:08,258 | INFO | train_grpo_train | metrics_logged mode=train step=3295

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.8181818181818185e-08, 'num_tokens': 7510050.0, 'completions/mean_length': 54.0, 'completions/min_length': 54.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.787638783454895, 'rewards/meter/std': 0.0, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.787638783454895, 'rewards/total_composite/std': 0.0, 'reward': 0.787638783454895, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.00023673544637858868, 'sampling/sampling_logp_difference/max': 0.002077840268611908, 'sampling/importance_sampling_ratio/min': 0.9994556307792664, 'sampling/importance_sampling_ratio/mean': 1.000232458114624, 'sampling/importance_sampling_ratio/max': 1.002079963684082, 'entropy': 0.0021129547676537186, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.787638783454895, 'reward_meter_mean': 0.787638783454895, 'reward_meter_std': 0.0, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.787638783454895, 'reward_total_composite_std': 0.0, 'epoch': 0.13}

100%|█████████▉| 3295/3300 [5:06:17<00:27,  5.45s/it]INFO 04-12 03:43:08 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3296/3300 [5:06:23<00:22,  5.71s/it]2026-04-12 03:43:14,629 | INFO | train_grpo_train | metrics_logged mode=train step=3296

                                                     
{'loss': 0.0072, 'grad_norm': 1.9730912446975708, 'learning_rate': 1.5151515151515152e-08, 'num_tokens': 7513219.0, 'completions/mean_length': 197.125, 'completions/min_length': 192.0, 'completions/max_length': 200.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.125, 'completions/min_terminated_length': 192.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.9992784261703491, 'rewards/meter/std': 0.00010585635027382523, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9886363744735718, 'rewards/repeat_penalty/std': 0.03214120864868164, 'rewards/total_composite/mean': 0.9879227876663208, 'rewards/total_composite/std': 0.0321127250790596, 'reward': 0.9879227876663208, 'reward_std': 0.032112736254930496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021909546107053757, 'sampling/sampling_logp_difference/max': 0.8929929733276367, 'sampling/importance_sampling_ratio/min': 0.40942853689193726, 'sampling/importance_sampling_ratio/mean': 1.009697675704956, 'sampling/importance_sampling_ratio/max': 1.8454686403274536, 'entropy': 0.2336829099804163, 'clip_ratio/low_mean': 0.0025125627871602774, 'clip_ratio/low_min': 0.0025125627871602774, 'clip_ratio/high_mean': 0.012068168085534126, 'clip_ratio/high_max': 0.012068168085534126, 'clip_ratio/region_mean': 0.014580730872694403, 'reward_total_mean': 0.9879227876663208, 'reward_meter_mean': 0.9992784261703491, 'reward_meter_std': 0.00010585635027382523, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9886363744735718, 'reward_repeat_penalty_std': 0.03214120864868164, 'reward_total_composite_mean': 0.9879227876663208, 'reward_total_composite_std': 0.0321127250790596, 'epoch': 0.13}

100%|█████████▉| 3296/3300 [5:06:23<00:22,  5.71s/it]INFO 04-12 03:43:14 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3297/3300 [5:06:31<00:18,  6.32s/it]2026-04-12 03:43:22,327 | INFO | train_grpo_train | metrics_logged mode=train step=3297

                                                     
{'loss': -0.0017, 'grad_norm': 2.3195743560791016, 'learning_rate': 1.2121212121212122e-08, 'num_tokens': 7517109.0, 'completions/mean_length': 302.25, 'completions/min_length': 291.0, 'completions/max_length': 316.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 302.25, 'completions/min_terminated_length': 291.0, 'completions/max_terminated_length': 316.0, 'rewards/meter/mean': 0.9957056641578674, 'rewards/meter/std': 0.006368847563862801, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 0.9926470518112183, 'rewards/repeat_penalty/std': 0.020797256380319595, 'rewards/total_composite/mean': 0.9883631467819214, 'rewards/total_composite/std': 0.020583728328347206, 'reward': 0.9883631467819214, 'reward_std': 0.0205837395042181, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05743308737874031, 'sampling/sampling_logp_difference/max': 1.1259183883666992, 'sampling/importance_sampling_ratio/min': 0.3243544399738312, 'sampling/importance_sampling_ratio/mean': 1.014825701713562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5184807442128658, 'clip_ratio/low_mean': 0.010391832329332829, 'clip_ratio/low_min': 0.010391832329332829, 'clip_ratio/high_mean': 0.03054804727435112, 'clip_ratio/high_max': 0.03054804727435112, 'clip_ratio/region_mean': 0.04093987960368395, 'reward_total_mean': 0.9883631467819214, 'reward_meter_mean': 0.9957056641578674, 'reward_meter_std': 0.006368847563862801, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 0.9926470518112183, 'reward_repeat_penalty_std': 0.020797256380319595, 'reward_total_composite_mean': 0.9883631467819214, 'reward_total_composite_std': 0.020583728328347206, 'epoch': 0.13}

100%|█████████▉| 3297/3300 [5:06:31<00:18,  6.32s/it]INFO 04-12 03:43:22 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3298/3300 [5:06:35<00:11,  5.79s/it]2026-04-12 03:43:26,891 | INFO | train_grpo_train | metrics_logged mode=train step=3298

                                                     
{'loss': 0.0017, 'grad_norm': 2.183478593826294, 'learning_rate': 9.090909090909092e-09, 'num_tokens': 7518824.0, 'completions/mean_length': 65.375, 'completions/min_length': 65.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9922501444816589, 'rewards/meter/std': 0.004244488663971424, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9922501444816589, 'rewards/total_composite/std': 0.004244488663971424, 'reward': 0.9922501444816589, 'reward_std': 0.0042444937862455845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013568882830440998, 'sampling/sampling_logp_difference/max': 0.7026901245117188, 'sampling/importance_sampling_ratio/min': 0.49525120854377747, 'sampling/importance_sampling_ratio/mean': 1.0067819356918335, 'sampling/importance_sampling_ratio/max': 1.4506497383117676, 'entropy': 0.11512181628495455, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.005710955825634301, 'clip_ratio/high_max': 0.005710955825634301, 'clip_ratio/region_mean': 0.009498834726400673, 'reward_total_mean': 0.9922501444816589, 'reward_meter_mean': 0.9922501444816589, 'reward_meter_std': 0.004244488663971424, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9922501444816589, 'reward_total_composite_std': 0.004244488663971424, 'epoch': 0.13}

100%|█████████▉| 3298/3300 [5:06:35<00:11,  5.79s/it]INFO 04-12 03:43:27 [block_pool.py:378] Successfully reset prefix cache

100%|█████████▉| 3299/3300 [5:06:41<00:05,  5.77s/it]2026-04-12 03:43:32,611 | INFO | train_grpo_train | metrics_logged mode=train step=3299

                                                     
{'loss': 0.0057, 'grad_norm': 3.7170727252960205, 'learning_rate': 6.060606060606061e-09, 'num_tokens': 7521612.0, 'completions/mean_length': 166.5, 'completions/min_length': 159.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.5, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9987373352050781, 'rewards/meter/std': 0.000354176911059767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9987373352050781, 'rewards/total_composite/std': 0.000354176911059767, 'reward': 0.9987373352050781, 'reward_std': 0.00035417175968177617, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03916100785136223, 'sampling/sampling_logp_difference/max': 1.9973111152648926, 'sampling/importance_sampling_ratio/min': 0.13569968938827515, 'sampling/importance_sampling_ratio/mean': 1.001840591430664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3029062431305647, 'clip_ratio/low_mean': 0.01500677247531712, 'clip_ratio/low_min': 0.01500677247531712, 'clip_ratio/high_mean': 0.015017259865999222, 'clip_ratio/high_max': 0.015017259865999222, 'clip_ratio/region_mean': 0.030024032341316342, 'reward_total_mean': 0.9987373352050781, 'reward_meter_mean': 0.9987373352050781, 'reward_meter_std': 0.000354176911059767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9987373352050781, 'reward_total_composite_std': 0.000354176911059767, 'epoch': 0.13}

100%|█████████▉| 3299/3300 [5:06:41<00:05,  5.77s/it]INFO 04-12 03:43:32 [block_pool.py:378] Successfully reset prefix cache

100%|██████████| 3300/3300 [5:06:45<00:00,  5.30s/it]2026-04-12 03:43:36,813 | INFO | train_grpo_train | metrics_logged mode=train step=3300

                                                     
{'loss': 0.0153, 'grad_norm': 6.677674293518066, 'learning_rate': 3.0303030303030304e-09, 'num_tokens': 7523188.0, 'completions/mean_length': 46.0, 'completions/min_length': 46.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9442145824432373, 'rewards/meter/std': 0.003297017654404044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/total_composite/mean': 0.9442145824432373, 'rewards/total_composite/std': 0.003297017654404044, 'reward': 0.9442145824432373, 'reward_std': 0.0032970213796943426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.021698685362935066, 'sampling/sampling_logp_difference/max': 2.1988744735717773, 'sampling/importance_sampling_ratio/min': 0.11092793941497803, 'sampling/importance_sampling_ratio/mean': 1.0027215480804443, 'sampling/importance_sampling_ratio/max': 1.663463830947876, 'entropy': 0.09115941543132067, 'clip_ratio/low_mean': 0.00815217406488955, 'clip_ratio/low_min': 0.00815217406488955, 'clip_ratio/high_mean': 0.0027173913549631834, 'clip_ratio/high_max': 0.0027173913549631834, 'clip_ratio/region_mean': 0.010869565419852734, 'reward_total_mean': 0.9442145824432373, 'reward_meter_mean': 0.9442145824432373, 'reward_meter_std': 0.003297017654404044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_total_composite_mean': 0.9442145824432373, 'reward_total_composite_std': 0.003297017654404044, 'epoch': 0.13}

100%|██████████| 3300/3300 [5:06:45<00:00,  5.30s/it]INFO 04-12 03:43:37 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:38,  3.50s/it]

 23%|██▎       | 3/13 [00:10<00:34,  3.40s/it]

 31%|███       | 4/13 [00:16<00:40,  4.48s/it]

 38%|███▊      | 5/13 [00:22<00:40,  5.07s/it]

 46%|████▌     | 6/13 [00:29<00:39,  5.60s/it]

 54%|█████▍    | 7/13 [00:34<00:33,  5.59s/it]

 62%|██████▏   | 8/13 [00:41<00:30,  6.05s/it]

 69%|██████▉   | 9/13 [00:49<00:25,  6.45s/it]

 77%|███████▋  | 10/13 [00:54<00:18,  6.17s/it]

 85%|████████▍ | 11/13 [01:00<00:12,  6.04s/it]

 92%|█████████▏| 12/13 [01:08<00:06,  6.46s/it]

100%|██████████| 13/13 [01:13<00:00,  6.11s/it]2026-04-12 03:44:57,274 | INFO | train_grpo_train | metrics_logged mode=eval step=3300

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 80.3156, 'eval_samples_per_second': 1.295, 'eval_steps_per_second': 0.162, 'eval_num_tokens': 7523188.0, 'eval_completions/mean_length': 213.14423076923077, 'eval_completions/min_length': 61.23076923076923, 'eval_completions/max_length': 421.38461538461536, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 209.98489027756912, 'eval_completions/min_terminated_length': 61.23076923076923, 'eval_completions/max_terminated_length': 412.2307692307692, 'eval_rewards/meter/mean': 0.7804106657321637, 'eval_rewards/meter/std': 0.3429693900621854, 'eval_rewards/count_adherence/mean': 0.9621203220807589, 'eval_rewards/count_adherence/std': 0.06154714152216911, 'eval_rewards/arabic_clean/mean': 1.0, 'eval_rewards/arabic_clean/std': 0.0, 'eval_rewards/repeat_penalty/mean': 0.9574309633328364, 'eval_rewards/repeat_penalty/std': 0.07397974411455485, 'eval_rewards/total_composite/mean': 0.7239617155148432, 'eval_rewards/total_composite/std': 0.3431667788670613, 'eval_reward': 0.7239617155148432, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03249364231641476, 'eval_sampling/sampling_logp_difference/max': 1.2331663278432994, 'eval_sampling/importance_sampling_ratio/min': 0.2969068231490942, 'eval_sampling/importance_sampling_ratio/mean': 1.0095902956449068, 'eval_sampling/importance_sampling_ratio/max': 1.6007587084403405, 'eval_entropy': 0.38235602699793303, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.7239617155148432, 'eval_reward_meter_mean': 0.7804106657321637, 'eval_reward_meter_std': 0.3429693900621854, 'eval_reward_count_adherence_mean': 0.9621203220807589, 'eval_reward_count_adherence_std': 0.06154714152216911, 'eval_reward_arabic_clean_mean': 1.0, 'eval_reward_arabic_clean_std': 0.0, 'eval_reward_repeat_penalty_mean': 0.9574309633328364, 'eval_reward_repeat_penalty_std': 0.07397974411455485, 'eval_reward_total_composite_mean': 0.7239617155148432, 'eval_reward_total_composite_std': 0.3431667788670613, 'epoch': 0.13}

100%|██████████| 3300/3300 [5:08:06<00:00,  5.30s/it]

100%|██████████| 13/13 [01:13<00:00,  6.11s/it]

                                               2026-04-12 03:45:00,970 | INFO | train_grpo_train | metrics_logged mode=train step=3300

                                                     
{'train_runtime': 18489.879, 'train_samples_per_second': 1.428, 'train_steps_per_second': 0.178, 'train_loss': 0.0004697396774788627, 'epoch': 0.13}

100%|██████████| 3300/3300 [5:08:09<00:00,  5.30s/it]
100%|██████████| 3300/3300 [5:08:09<00:00,  5.60s/it]

Processing Files (0 / 0)      : |          |  0.00B /  0.00B            

New Data Upload               : |          |  0.00B /  0.00B            


  ..._223409/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            



  ...223409/plots/kl_chain.png: 100%|██████████| 33.4kB / 33.4kB            




  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            





  ...rd_panels_train_chain.png: 100%|██████████|  350kB /  350kB            






  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            







  ...adapter_model.safetensors:   7%|▋         | 47.9MB /  640MB            








  ..._arabic_clean_debug.jsonl:  45%|████▌     | 47.9MB /  106MB            









  ...unt_adherence_debug.jsonl:  45%|████▌     | 47.9MB /  106MB            










  ...409/all_generations.jsonl:   9%|▉         | 55.9MB /  614MB            











  ...tal_composite_debug.jsonl:  53%|█████▎    | 55.9MB /  106MB            


  ..._223409/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            



  ...223409/plots/kl_chain.png: 100%|██████████| 33.4kB / 33.4kB            




  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            





  ...rd_panels_train_chain.png: 100%|██████████|  350kB /  350kB            






  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            







  ...adapter_model.safetensors:   7%|▋         | 47.9MB /  640MB            








  ..._arabic_clean_debug.jsonl:  45%|████▌     | 47.9MB /  106MB            









  ...unt_adherence_debug.jsonl:  45%|████▌     | 47.9MB /  106MB            










  ...409/all_generations.jsonl:   9%|▉         | 55.9MB /  614MB            











  [+ 6 files]                 :  22%|██▏       |  208MB /  959MB            
Processing Files (8 / 16)     :  17%|█▋        |  410MB / 2.43GB,   ???B/s  


  ...adapter_model.safetensors:  15%|█▍        | 96.0MB /  640MB            



  ..._arabic_clean_debug.jsonl:  98%|█████████▊|  104MB /  106MB            




  ...unt_adherence_debug.jsonl:  90%|█████████ | 95.9MB /  106MB            





  ...409/all_generations.jsonl:  18%|█▊        |  112MB /  614MB            






  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            







  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            








  ...epeat_penalty_debug.jsonl:  98%|█████████▊|  104MB /  106MB            









  ...ard_panels_eval_chain.png: 100%|██████████|  319kB /  319kB            










  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            











  [+ 3 files]                 :  15%|█▌        | 97.3MB /  641MB            
Processing Files (13 / 19)    :  34%|███▍      |  824MB / 2.43GB, 2.07GB/s  


  ...adapter_model.safetensors:  23%|██▎       |  144MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  27%|██▋       |  168MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  24%|██▍       |  152MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  41%|████      |  999MB / 2.43GB, 1.47GB/s  


  ...adapter_model.safetensors:  31%|███▏      |  200MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  35%|███▌      |  216MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  34%|███▍      |  216MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  48%|████▊     | 1.17GB / 2.43GB, 1.26GB/s  


  ...adapter_model.safetensors:  41%|████▏     |  264MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  46%|████▌     |  280MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  41%|████▏     |  264MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  55%|█████▌    | 1.34GB / 2.43GB, 1.17GB/s  


  ...adapter_model.safetensors:  49%|████▉     |  312MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  53%|█████▎    |  328MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  50%|█████     |  320MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  62%|██████▏   | 1.49GB / 2.43GB, 1.08GB/s  


  ...adapter_model.safetensors:  58%|█████▊    |  368MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  64%|██████▍   |  392MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  59%|█████▉    |  376MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  69%|██████▉   | 1.67GB / 2.43GB, 1.05GB/s  


  ...adapter_model.safetensors:  66%|██████▋   |  424MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  73%|███████▎  |  448MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  66%|██████▋   |  424MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  75%|███████▌  | 1.83GB / 2.43GB, 1.02GB/s  


  ...adapter_model.safetensors:  76%|███████▋  |  488MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  82%|████████▏ |  504MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  75%|███████▌  |  480MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  83%|████████▎ | 2.01GB / 2.43GB,  998MB/s  


  ...adapter_model.safetensors:  85%|████████▌ |  544MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  90%|████████▉ |  552MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  84%|████████▍ |  536MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  89%|████████▉ | 2.17GB / 2.43GB,  976MB/s  


  ...adapter_model.safetensors:  94%|█████████▍|  600MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl:  99%|█████████▉|  608MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors:  93%|█████████▎|  592MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            
Processing Files (16 / 19)    :  96%|█████████▌| 2.33GB / 2.43GB,  963MB/s  


  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            



  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            





  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            






  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            








  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            









  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            










  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            











  [+ 3 files]                 :  84%|████████▍ | 1.68MB / 2.00MB            
Processing Files (19 / 22)    : 100%|█████████▉| 2.43GB / 2.43GB,  918MB/s  

New Data Upload               :  83%|████████▎ | 1.60MB / 1.92MB,  727kB/s  


  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            





  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            






  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            







  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            








  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            









  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            










  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            











  ...ts/meter_by_meter_run.png: 100%|██████████|  756kB /  756kB            
Processing Files (22 / 22)    : 100%|██████████| 2.43GB / 2.43GB,  842MB/s  

New Data Upload               : 100%|██████████| 1.92MB / 1.92MB,  799kB/s  


  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            





  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            






  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            







  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            








  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            









  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            










  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            











  ...ts/meter_by_meter_run.png: 100%|██████████|  756kB /  756kB            


  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            





  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            






  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            







  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            








  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            









  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            










  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            











  ...ts/meter_by_meter_run.png: 100%|██████████|  756kB /  756kB            


  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            





  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            






  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            







  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            








  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            









  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            










  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            











  ...ts/meter_by_meter_run.png: 100%|██████████|  756kB /  756kB            


  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            




  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            





  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            






  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            







  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            








  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            









  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            










  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            











  ...ts/meter_by_meter_run.png: 100%|██████████|  756kB /  756kB            
Processing Files (22 / 22)    : 100%|██████████| 2.43GB / 2.43GB,  673MB/s  

New Data Upload               : 100%|██████████| 1.92MB / 1.92MB,  639kB/s  

  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            

  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            

  ...eward_panels_eval_run.png: 100%|██████████|  309kB /  309kB            

  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            

  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            

  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            

  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            

  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            

  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            

  ...ts/meter_by_meter_run.png: 100%|██████████|  756kB /  756kB            
No files have been modified since last commit. Skipping to prevent empty commit.

Processing Files (0 / 0)      : |          |  0.00B /  0.00B            

New Data Upload               : |          |  0.00B /  0.00B            


  ..._223409/generations.jsonl:   8%|▊         | 47.5MB /  614MB            


  ..._223409/generations.jsonl:   8%|▊         | 47.5MB /  614MB            
Processing Files (0 / 1)      :   7%|▋         | 47.5MB /  668MB,   ???B/s  


  ..._223409/generations.jsonl:  15%|█▍        | 90.6MB /  614MB            
Processing Files (0 / 1)      :  14%|█▎        | 90.6MB /  668MB,  216MB/s  


  ..._223409/generations.jsonl:  22%|██▏       |  136MB /  614MB            
Processing Files (0 / 1)      :  20%|██        |  136MB /  668MB,  220MB/s  


  ..._223409/generations.jsonl:  29%|██▉       |  177MB /  614MB            
Processing Files (0 / 1)      :  26%|██▋       |  177MB /  668MB,  216MB/s  


  ..._223409/generations.jsonl:  46%|████▌     |  284MB /  614MB            
Processing Files (0 / 1)      :  42%|████▏     |  284MB /  668MB,  295MB/s  

New Data Upload               :  95%|█████████▍| 63.6MB / 67.0MB, 79.4MB/s  


  ..._223409/generations.jsonl:  52%|█████▏    |  321MB /  614MB            
Processing Files (0 / 1)      :  48%|████▊     |  321MB /  668MB,  273MB/s  

New Data Upload               : 100%|██████████| 67.0MB / 67.0MB, 67.0MB/s  


  ..._223409/generations.jsonl:  60%|█████▉    |  366MB /  614MB            
Processing Files (0 / 1)      :  55%|█████▍    |  366MB /  668MB,  265MB/s  


  ..._223409/generations.jsonl:  67%|██████▋   |  413MB /  614MB            
Processing Files (0 / 1)      :  62%|██████▏   |  413MB /  668MB,  261MB/s  


  ..._223409/generations.jsonl:  85%|████████▍ |  519MB /  614MB            
Processing Files (0 / 1)      :  78%|███████▊  |  519MB /  668MB,  295MB/s  

New Data Upload               :  99%|█████████▉|  133MB /  134MB, 83.1MB/s  


  ..._223409/generations.jsonl:  91%|█████████ |  560MB /  614MB            
Processing Files (0 / 1)      :  84%|████████▍ |  560MB /  668MB,  285MB/s  

New Data Upload               : 100%|██████████|  134MB /  134MB, 74.5MB/s  



  ...23409/generations.parquet: 100%|█████████▉| 53.5MB / 53.8MB            




  ...11_223409/metrics.parquet: 100%|█████████▉|  670kB /  673kB            





  ...409/reward_curves_run.png: 100%|█████████▉| 27.1kB / 27.2kB            


  ..._223409/generations.jsonl: 100%|█████████▉|  613MB /  614MB            



  ...23409/generations.parquet: 100%|█████████▉| 53.5MB / 53.8MB            




  ...11_223409/metrics.parquet: 100%|█████████▉|  670kB /  673kB            





  ...409/reward_curves_run.png: 100%|█████████▉| 27.1kB / 27.2kB            
Processing Files (0 / 4)      : 100%|█████████▉|  668MB /  668MB,  310MB/s  

New Data Upload               : 100%|█████████▉|  191MB /  192MB, 95.7MB/s  


  ..._223409/generations.jsonl: 100%|█████████▉|  613MB /  614MB            



  ...23409/generations.parquet: 100%|█████████▉| 53.5MB / 53.8MB            




  ...11_223409/metrics.parquet: 100%|█████████▉|  670kB /  673kB            





  ...409/reward_curves_run.png: 100%|█████████▉| 27.1kB / 27.2kB            


  ..._223409/generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...23409/generations.parquet: 100%|██████████| 53.8MB / 53.8MB            




  ...11_223409/metrics.parquet: 100%|██████████|  673kB /  673kB            





  ...409/reward_curves_run.png: 100%|██████████| 27.2kB / 27.2kB            
Processing Files (4 / 4)      : 100%|██████████|  668MB /  668MB,  259MB/s  

New Data Upload               : 100%|██████████|  192MB /  192MB, 79.8MB/s  


  ..._223409/generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...23409/generations.parquet: 100%|██████████| 53.8MB / 53.8MB            




  ...11_223409/metrics.parquet: 100%|██████████|  673kB /  673kB            





  ...409/reward_curves_run.png: 100%|██████████| 27.2kB / 27.2kB            


  ..._223409/generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...23409/generations.parquet: 100%|██████████| 53.8MB / 53.8MB            




  ...11_223409/metrics.parquet: 100%|██████████|  673kB /  673kB            





  ...409/reward_curves_run.png: 100%|██████████| 27.2kB / 27.2kB            


  ..._223409/generations.jsonl: 100%|██████████|  614MB /  614MB            



  ...23409/generations.parquet: 100%|██████████| 53.8MB / 53.8MB            




  ...11_223409/metrics.parquet: 100%|██████████|  673kB /  673kB            





  ...409/reward_curves_run.png: 100%|██████████| 27.2kB / 27.2kB            
Processing Files (4 / 4)      : 100%|██████████|  668MB /  668MB,  222MB/s  

New Data Upload               : 100%|██████████|  192MB /  192MB, 68.4MB/s  

  ..._223409/generations.jsonl: 100%|██████████|  614MB /  614MB            

  ...23409/generations.parquet: 100%|██████████| 53.8MB / 53.8MB            

  ...11_223409/metrics.parquet: 100%|██████████|  673kB /  673kB            

  ...409/reward_curves_run.png: 100%|██████████| 27.2kB / 27.2kB            

Processing Files (0 / 0)      : |          |  0.00B /  0.00B            

New Data Upload               : |          |  0.00B /  0.00B            


  ...107/chain_metrics.parquet: 100%|██████████|  982kB /  982kB            



  ...7/chain_reward_curves.png: 100%|██████████| 27.8kB / 27.8kB            


  ...107/chain_metrics.parquet: 100%|██████████|  982kB /  982kB            



  ...7/chain_reward_curves.png: 100%|██████████| 27.8kB / 27.8kB            
Processing Files (2 / 2)      : 100%|██████████| 1.01MB / 1.01MB,   ???B/s  

New Data Upload               : 100%|██████████|  985kB /  985kB,   ???B/s  


  ...107/chain_metrics.parquet: 100%|██████████|  982kB /  982kB            



  ...7/chain_reward_curves.png: 100%|██████████| 27.8kB / 27.8kB            


  ...107/chain_metrics.parquet: 100%|██████████|  982kB /  982kB            



  ...7/chain_reward_curves.png: 100%|██████████| 27.8kB / 27.8kB            
Processing Files (2 / 2)      : 100%|██████████| 1.01MB / 1.01MB,  0.00B/s  

New Data Upload               : 100%|██████████|  985kB /  985kB,  0.00B/s  

  ...107/chain_metrics.parquet: 100%|██████████|  982kB /  982kB            

  ...7/chain_reward_curves.png: 100%|██████████| 27.8kB / 27.8kB            
2026-04-12 03:46:19,803 | INFO | train_grpo_train | pushing final adapter to hub repo=Shaer-AI/Shaer-adapters-grpo

Processing Files (0 / 0)      : |          |  0.00B /  0.00B            

New Data Upload               : |          |  0.00B /  0.00B            


  ..._223409/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            



  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            




  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            





  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            






  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            







  ...1_223409/plots/kl_run.png: 100%|██████████| 24.6kB / 24.6kB            








  ...223409/plots/kl_chain.png: 100%|██████████| 33.4kB / 33.4kB            









  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            










  ...rd_panels_train_chain.png: 100%|██████████|  350kB /  350kB            











  ...adapter_model.safetensors:  10%|█         | 64.0MB /  640MB            


  ..._223409/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            



  ...plots/arabic_gate_run.png: 100%|██████████| 81.1kB / 81.1kB            




  ...ots/arabic_gate_chain.png: 100%|██████████| 98.8kB / 98.8kB            





  ...l_adapter/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            






  ...adapter/training_args.bin: 100%|██████████| 7.50kB / 7.50kB            







  ...1_223409/plots/kl_run.png: 100%|██████████| 24.6kB / 24.6kB            








  ...223409/plots/kl_chain.png: 100%|██████████| 33.4kB / 33.4kB            









  .../meter_by_meter_chain.png: 100%|██████████| 1.06MB / 1.06MB            










  ...rd_panels_train_chain.png: 100%|██████████|  350kB /  350kB            











  [+ 8 files]                 :  18%|█▊        |  440MB / 2.42GB            
Processing Files (10 / 18)    :  18%|█▊        |  443MB / 2.43GB,   ???B/s  


  ...rd_panels_train_chain.png: 100%|██████████|  350kB /  350kB            



  ...adapter_model.safetensors:  20%|██        |  128MB /  640MB            




  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            





  ...adapter_model.safetensors:  19%|█▉        |  120MB /  640MB            






  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...409/all_generations.jsonl:  20%|█▉        |  120MB /  614MB            








  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            









  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            











  [+ 1 files]                 : 100%|██████████|  107MB /  107MB            
Processing Files (16 / 19)    :  37%|███▋      |  903MB / 2.43GB, 2.30GB/s  


  ...adapter_model.safetensors:  28%|██▊       |  176MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  28%|██▊       |  176MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  29%|██▊       |  176MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  44%|████▍     | 1.06GB / 2.43GB, 1.55GB/s  


  ...adapter_model.safetensors:  38%|███▊      |  240MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  38%|███▊      |  240MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  38%|███▊      |  232MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  51%|█████▏    | 1.25GB / 2.43GB, 1.34GB/s  


  ...adapter_model.safetensors:  46%|████▋     |  296MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  46%|████▋     |  296MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  47%|████▋     |  288MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  58%|█████▊    | 1.42GB / 2.43GB, 1.21GB/s  


  ...adapter_model.safetensors:  55%|█████▌    |  352MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  55%|█████▌    |  352MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  56%|█████▌    |  344MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  65%|██████▌   | 1.58GB / 2.43GB, 1.14GB/s  


  ...adapter_model.safetensors:  64%|██████▍   |  408MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  64%|██████▍   |  408MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  65%|██████▌   |  400MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  72%|███████▏  | 1.75GB / 2.43GB, 1.09GB/s  


  ...adapter_model.safetensors:  73%|███████▎  |  464MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  74%|███████▍  |  472MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  74%|███████▍  |  456MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  79%|███████▉  | 1.93GB / 2.43GB, 1.06GB/s  


  ...adapter_model.safetensors:  83%|████████▎ |  528MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  83%|████████▎ |  528MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  85%|████████▍ |  520MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  87%|████████▋ | 2.11GB / 2.43GB, 1.04GB/s  


  ...adapter_model.safetensors:  91%|█████████▏|  584MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors:  91%|█████████▏|  584MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl:  94%|█████████▍|  576MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (16 / 19)    :  94%|█████████▍| 2.28GB / 2.43GB, 1.02GB/s  


  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            
Processing Files (19 / 19)    : 100%|█████████▉| 2.43GB / 2.43GB,  993MB/s  


  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            



  ...ward_panels_train_run.png: 100%|██████████|  342kB /  342kB            




  ...adapter_model.safetensors: 100%|██████████|  640MB /  640MB            





  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            







  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            








  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            









  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            










  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            











  [+ 3 files]                 : 100%|██████████| 2.61MB / 2.61MB            
Processing Files (22 / 22)    : 100%|██████████| 2.43GB / 2.43GB,  903MB/s  

New Data Upload               : 100%|██████████| 1.38MB / 1.38MB,  628kB/s  


  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            



  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            




  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            





  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            








  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            









  ...ts/meter_by_meter_run.png: 100%|██████████|  754kB /  754kB            










  ...ard_panels_eval_chain.png: 100%|██████████|  318kB /  318kB            











  ...eward_panels_eval_run.png: 100%|██████████|  310kB /  310kB            


  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            



  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            




  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            





  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            








  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            









  ...ts/meter_by_meter_run.png: 100%|██████████|  754kB /  754kB            










  ...ard_panels_eval_chain.png: 100%|██████████|  318kB /  318kB            











  ...eward_panels_eval_run.png: 100%|██████████|  310kB /  310kB            


  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            



  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            




  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            





  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            








  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            









  ...ts/meter_by_meter_run.png: 100%|██████████|  754kB /  754kB            










  ...ard_panels_eval_chain.png: 100%|██████████|  318kB /  318kB            











  ...eward_panels_eval_run.png: 100%|██████████|  310kB /  310kB            


  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            



  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            




  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            





  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            






  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            







  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            








  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            









  ...ts/meter_by_meter_run.png: 100%|██████████|  754kB /  754kB            










  ...ard_panels_eval_chain.png: 100%|██████████|  318kB /  318kB            











  ...eward_panels_eval_run.png: 100%|██████████|  310kB /  310kB            
Processing Files (22 / 22)    : 100%|██████████| 2.43GB / 2.43GB,  710MB/s  

New Data Upload               : 100%|██████████| 1.38MB / 1.38MB,  494kB/s  

  ...unt_adherence_debug.jsonl: 100%|██████████|  106MB /  106MB            

  ...409/all_generations.jsonl: 100%|██████████|  614MB /  614MB            

  ..._arabic_clean_debug.jsonl: 100%|██████████|  106MB /  106MB            

  .../reward_meter_debug.jsonl: 100%|██████████|  106MB /  106MB            

  ...tal_composite_debug.jsonl: 100%|██████████|  106MB /  106MB            

  ...epeat_penalty_debug.jsonl: 100%|██████████|  106MB /  106MB            

  ...11_223409/tokenizer.model: 100%|██████████| 1.23MB / 1.23MB            

  ...ts/meter_by_meter_run.png: 100%|██████████|  754kB /  754kB            

  ...ard_panels_eval_chain.png: 100%|██████████|  318kB /  318kB            

  ...eward_panels_eval_run.png: 100%|██████████|  310kB /  310kB