Hugging Face's logo

Shaer-AI-2
/
Shaer-adapters-grpo

Transformers
Safetensors
trl
grpo
arabic-poetry
classical-arabic
lora
Model card Files
xet
 Community
Shaer-adapters-grpo / train_stdout.log
AhmadAbbass's picture
AhmadAbbass
Training in progress, step 850
29768c9 verified
Raw
History Blame
491 kB
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